SCIENTIFIC

PROGRAMME
KEYNOTE
 ADVANCED PROCESS CONTROL
THE OPPORTUNITIES, BENEFITS AND BARRIERS

Hans H. Eder
ACT, Brussels office
B 3080, Tervuren,
Belgium
E-mail: [email protected]

Keywords: Advanced control, control performance, service DEFINITION

factor, economics,
To start, let us see what APC means. Unfortunately, already
the name “Advanced Control“ gives rise to many different
perceptions, most of them being too narrow as they relate just
ABSTRACT to the technological aspects. Therefore I would like to present
here a more complete definition:
Advanced Process Control (APC) has been around for more
than 35 years, yet is still very much underutilized: Only a few “Advanced Control is the intelligent, well managed use of
companies have recognized the tremendous potential and process control technology, systems and tools, based on
realized enormous credits. APC is still perceived often as too sound process knowledge, with the objective to enable and
difficult, too expensive and without really hard ‘incentives’. benefit from operations improvements in a most cost and
As a consequence, the industry is by far not exploiting the time effective way”.
possible benefits. To help improve this situation is the overall
intent of this paper. The first objective is to promote the use Let us briefly expand on two of the keywords:
of APC by raising the awareness of its many benefits and • ‘Intelligent’ means here that we are aware of the
opportunities, by describing the mechanism why and how currently available technologies, that we have a suitable
APC can deliver these credits and how to estimate the subset, a Technology Set, readily available on our control
economics. The second objective is to show the key factors systems and that we know when and how to use them.
for successful implementation and use. • ‘Well managed’ means here that all APC activities are
well planned, monitored and continuously improved, that
they are in line with the overall plant operations
objectives and based on sound standards and rules.
INTRODUCTION
Form this definition we can already derive that APC has in
Despite its many benefits, only a small percentage of the fact three main dimensions:
possible APC applications have been implemented in industry • Technology
to date. Also, APC is often used in a sporadic way rather than • Economics
as integral part of a complete operations and control strategy. • Organization
This calls for improvement. As far as a more widespread use
of the most powerful control technology is concerned, two
obstacles stand in the way: EXPECTATIONS FROM APC
• Technical staff often lacks the knowledge where APC
could bring extra credits and how to apply it in practice Process control is a service to plant operations, assisting them
• Management is typically not aware of the benefits and to reach the operating objectives. The goal of APC in
thus reluctant to support the required effort. particular is to maximize the contributions of process control,
its “credits”, over time. This has again two dimensions:
The second fact is only partly the fault of management, since
many technical experts are unable to quantify the credits. The a) The maximum achievable credits (the “gain”) when all
objective of the first part of this presentation is therefore to possible advanced controls are implemented and used.
raise the awareness of the benefits of APC and to give The economical gain is mainly influenced by the
guidance on when to use and how to justify it. technology used. Superior technology can deliver higher
performance, resulting in additional credits and thus will
The second objective of this paper is to help establish a solid raise this gain.
basis which enables that maximum benefits are achieved from b) The time needed to reach the maximum gain (the “time
the use of this powerful technology. constant”). The speed of the approach of the maximum
gain is mainly influenced by manpower, skills, planning
and the use of suitable productivity tools.
 WHERE CAN APC HELP THE MOST?

APC does not aim to replace every simple flow controller, but
should be used where the standard PID controllers is
overwhelmed and where operators need extra help. In general,
APC shall be used where we face

a) difficult process behavior

b) high performance and operability requirements.

Picture 1: “Credits over time” Typical examples for difficult process behavior are:

Long deadtime is known as a major difficulty for control.

This is a key reason why relatively few product qualities are
WHAT IS ‘CONTROL PERFORMANCE’? controlled in closed loop, even where stream analyzers exist,
since analysis time is often quite long.
To be able to improve operations, APC must deliver higher
performance. Lets us briefly look at what this means: High controllability ratio CR (the ratio deadtime to time
constant), which is another key indicator for the degree of
To be able to produce the products with the right properties, difficulty: Generally, up to a ratio of 2 the PID will deliver a
the process must follow certain targets as closely as possible. fine job, for values between 2 and 3, PID tuning becomes
These targets are seen as the optimal values, every deviation more difficult and time consuming and beyond 3 the PID is
from such target results therefore in some kind of economical not recommended any more. Using the CR to choose the right
loss (or safety hazard). If the process was always stable and controller, of course, requires at least some quantitative
without any influences then this would be quite easy: Once knowledge of the process dynamics.
the target conditions are reached no further interactions were
necessary. But the opposite is the case, the process is Inverse response and overshoot: For processes with inverse
permanently undergoing disturbing influences, dynamic response any reactive controller will apply initially the wrong
effects, which must be compensated. The challenge here is action on the process. This can be avoided with approaches
therefore to manage the process dynamics. And like Model Based Control (MBC). And for systems with
“performance” means in this context in general terms how overshoot much smoother yet tight control can be achieved.
well we are able to master these dynamics.
Strong interactions: Dynamic decoupling can be performed
The above implies of course that the targets are set correct or also with PID’s but is quite difficult and thus confined to
optimal. But they are typically calculated based on static small (2x2) systems. And there is the risk of instability when
calculations and not taking the process dynamics and the the decouplers are not active. Thus the controllers cannot be
variations into account: If the process has a highly non-linear tuned tightly. MBC e.g. is much easier to tune, delivers better
behavior and we must accept large variations in the variable performance and has practically no limitation in size.
under consideration then we need to adjust even the target.
Despite the common belief that APC mainly is justified in
To illustrate this let us look at the following example: For large oil refineries, successful applications have proven that it
Claus plants, which produce sulfur out of H2S containing has its value in the food industry as well: Applications exist in
“sour” gas, the maximum theoretical conversion is achieved the production of cookies, in batch freezing of ice cream, in
at a H2S / SO2 ratio of 2.8. But the conversion degrades with meat production etc. In dairy production, increased product
different slopes on both sides of the optimum. With a mean yield can be realized by targeting upper specification limit.
value of the H2S / SO2 ratio of exactly 2.8 but a certain Other benefits are increased productivity because more
variance, the actual conversion will, however, be much less product is produced and 'Right Time First' energy savings by
than the maximum value. And because of the different slopes not over-drying, reduction in waste/scrap, increased product
in the vicinity of the optimum, we can achieve higher quality. In beer production, temperature control of brew
conversion by reducing the variance and by moving the target kettle, fermenter, wort cooling, bottle washer can yield
ratio to a value higher than the theoretical one. significant energy savings and also water reduction in the
bottle washer. In the canning production, temperature control
Looking at the situation the other way round we can state: of blanching sections and of retorts have similar results. In
The more we can reduce the variations of that process sugar industry higher extraction rate, less energy and water
variable, the smaller are the overall deviations from the consumption and reduction of environmental issues are key
setpoint and therefore the economical losses and also the achievements.
closer we will operate to the true (static) optimal value.
WHAT ARE THE RESULTING BENEFITS? HOW CAN THE CREDITS BE DETERMINED?

APC aims at reducing the variance of the controlled variable The economics of the effects described above can be
and / or of the manipulated variable(s). Reducing the variance relatively easily quantified, given the process mechanisms are
of the controlled variable will yield the following effects: known and the economic data and some calculation tools are
available. Several steps are required:
1) If the controlled variable must be pushed against a limit,
then reducing the variance allows moving the average value Steps 1: Determine the current performance like the variance,
closer to this limit. In other words: We can squeeze the distance of the mean value from limiting constraints, etc., also
variance and shift the mean! This effect is very important, the overall consumption of the resource over time.
since about 60 to 70% of the total credits from APC come
from operation closer to limiting constraints. An example: A Step 2: Determine how much the variance can be reduced.
reactor shall be operated at maximum conversion. Conversion This is more difficult if no prior experience values exist.
increases with increasing inlet temperature, which may, Several methods exist which cannot be discussed here further.
however, be constrained by the maximum tube metal One easy way is to monitor the performance of the variable in
temperature of the preheat furnace. Therefore, tighter control question and to compare its average performance with the
allows to push the average furnace temperature closer to its best demonstrated performance. This gives at least a
constraint, achieves higher conversion and thus higher profits. conservative indication of the potential improvements. More
It should be noted that in this example (like in many other realistic figures are, of course, obtained by using suitable
cases) the final effect is not directly but indirectly related to simulation tools that allow to compare different techniques.
the variable we have to control. Tighter temperature control
itself does not make the money – but its effect on product Step 3: What is / are the effect(s) of the reduced variance?
yield, qualities etc.! Again, a variety of methods exist but one simple case shall be
described here to illustrate the procedure: Where several
2) Often the “money making” variable stands in a non-linear products are produced in a unit typically these products have
relationship with the controlled variable. If e.g. a temperature different prices and market outlets. Thus the product for
would have zero variance then this would result in a certain which a market exists and which yields the highest profit
product quality. However, the larger the temperature variance margin shall be maximized. We call it the marginal product.
the more the actual quality will differ form the ideal value.
Next, we need to know the mechanisms, the effects,
The average value of the dependent variable – here the quantitatively. Example: Increasing the reflux of a splitter
product quality – is a function of both the average value of tower will increase the fractionation and thus reduce the
the independent variable – here the temperature – and the impurities in both products. This allows in turn to change the
variance of the independent variable. cutpoint and to raise the impurity of the marginal product
back to the original value - which results in a higher yield of
This means that we are forced to run at a “non-ideal” the marginal product and thus higher overall profits.
temperature setpoint in order to stay on the target quality. The
typical consequence is a lower product yield. However, at the
end the product is on spec, there is no apparent problem and WHAT STANDS IN THE WAY OF APC?
therefore the loss of yield is often overlooked.
Despite its merits, APC is typically not aggressively applied
Having succeeded in minimizing the variance of the in industry, the contrary is the case, a main reason being the
controlled variable, we have reached our first performance lack of knowledge of the economical and operational benefits
goal. But we have also to look at the manipulated variable, of APC by both technical staff and management.
basically addressing the question: How have we achieved this
performance – with minimum, reasonable or excessive effort? Because of this deficiency, we hear many defensive and often
destructive arguments against APC: We will just take the two
Smooth action on the process is the second objective. This very common ones and see if they hold at all.
should be stressed especially since it is easily overlooked.
Good performance of the controlled variable can be often “My process is too simple.”
achieved with quite simple controllers, even on-off switches, This is a very destructive statement. The process may be
but at the price of drastic action on the process and thus over- simple but it still may carry a tremendous improvement
utilization (“consumption”) of the resource, the manipulated potential. By simply ignoring all simple processes by
variable. A typical example is again temperature control. principle, tremendous credits can be lost: A typical example is
Many controllers deliver good temperature behavior but at the liquid level control: A drum or vessel is a very simple piece
expense of excessive consumption of the heating or cooling of equipment and to control the level within a band is in
medium. Smooth action is especially important in integrated principle an extremely simple request. Yet, up to 80% of all
plants where any corrective action at one place means a level controllers perform poorly, many of them causing
disturbance for another variable. substantial problems downstream because of the permanent
swings in the manipulated flow. Using an error squared PID well as large scale control and optimization packages that
instead of the standard PID is a very simple yet effective require extra computing platforms.
‘advanced’ alternative, but is hardly known and used by
control professionals. The answer: The right attitude is to Technical selection criteria shall be: Performance, ability to
look at the cost of non-conformance – rather than at the handle a wide range of process types (integrating processes!),
complexity of the process. robustness against model errors, expandability (multivariable,
control, feedforward), built-in constraint handling, ease of use
“My process is too complex.” (implementation, tuning, maintenance and updating) and the
Even for simple processes we see typically quite some computing requirements.
differences in the performance from operator to operator and
from shift to shift. The objective should be clearly lo run the Important is also the vendors’ philosophy: Some supply not
plant permanently at least as good as the best operator or only control algos and applications but also transfer their
shift. One advantage of automatic control is to ensure this. know-how. This puts the user in full command and enables
And we see that with increasing complexity the spread in the him to maintain the application later himself. Others provide
performance between different operators and shifts is clearly only turn-key solutions, without any know how transfer at all,
increasing. Thus - in contrary to the statement - increasing which can bring the user in some dilemma at a later stage.
complexity demands even more intense use of APC.
A key consideration is to use the most simple approach
possible (Einstein: “All models shall be a s simple as possible
SUCCESSFUL APC DEVELOPMENT AND USE - but no simpler”). Most of the unsuccessful APC applications
for example failed because of their unnecessary complexity.
Once we have convinced ourselves and management to apply
higher levels of control, then the next step is to plan for its Success factor 3. Training
most effective use. Before we go on let us re-state: The main
objective of process control is to improve operations with Since the user has to choose the technology for a certain
controllers and applications that deliver top performance and application, he must be in a position to make a conscious
are operating whenever needed. It is important to keep in decision, he must have had adequate education before. Most
mind that the latter – a high service factor – does by no means have theoretical education and system specific knowledge.
come automatically together with top performance, we have Yet the know-how to turn theory into working and lasting
to take separate, special measures. control applications is rather scarce and so is the ability to
locate improvement opportunities and to quantify the credits.
But first of, to be able to fully exploit the APC capabilities a Therefore additional practice and profitability oriented
several changes are required. Of all the requirements for solid education of the control engineers comes first and is in fact
performance there is one factor that stands out: Process the second most important factor after process knowledge.
knowledge. Therefore we will address it up-front.
Success factor 4: Tools
Success factor 1: Process knowledge
Not only the maximum achievable credits are of concern but
This is by far the most important factor. It is often stated that also the speed by which they can be realized. This demands
one of its strongest advantages of the PID is that it can be for productivity tools, may it be for incentive calculation,
tuned without solid process knowledge. In my opinion, it one performance measurement, model development, comparison
of its strongest disadvantages. Even during studies to locate of different techniques - but also for training and refresher.
improvement opportunities the process knowledge was Measuring the performance plays a key role: Implementing a
always significantly increased and has lead in many cases to Performance Monitoring System is neither difficult nor work-
changes in targets and constraints. The consequence: Quite intensive yet it is a key instrument in application development
often the development effort has been already paid out before and follow-up. It helps to detect under-performing controls,
the first application was commissioned! For closed loop documents the base case situation and shows in quantitative,
control itself the most important success factors are known as indisputable terms the achieved improvement. To stress this
the 4 T’s: Technology, Training, Tools and Tactics. point I would like to use two citations, one from the US and
Success factor 2. Technology: one from The Netherlands:
“What I cannot measure I cannot manage” - “Meten is
Successful exploitation of advanced control demands the weten, gissen is missen” (measuring means knowing,
establishment of a Technology Set. This is the “arsenal” that guessing means missing)
allows the control engineer to find the right “weapon” for
every task. Many different approaches exist and it is very Success factor 5: Tactics
important to select the ones which are not only best suited for
the tasks but also matches the company’s situation with By “tactics” I mean all the organizational means and
respect to manpower, skills and available infrastructure. measures needed to steer process control to maximum
Small, yet powerful, DCS based approaches are available as success. This includes standards - from tag and display
naming conventions to standard control elements (e.g. allowed to spend using standard techniques like DCF or IRR.
feedforwards, analyzer controls,..), as well as careful design When the economics speak for the project then we can go
of human and also system interfaces: Higher level control ahead and select the technology and put this into our overall
also crosses sometimes the boundary between target setting process control business plan and start developing the
(planning/scheduling) and execution of the targets. applications according to this plan.

Giving the operator the right information is the key to high

application service factors. The operator decides over the use CONCLUSIONS
of the application, thus we must make sure he has all the
required information and that it is properly presented. The APC technology is mature, the computing platforms are
in most cases available, today there are no serious technical
Furthermore, there must be an overall “business plan” hurdles any more that prevent us from harvesting the benefits.
describing how and when the identified applications will be And once the incentives are determined, the remaining
developed, what skills, training, human and system resources barriers are typically breaking down. To achieve the
and tools are required and when. Only with such a sound plan maximum benefits, however, we must plan carefully and
we will be able to realize the full economical success. observe the key success factors – just as for any other project.
Its certainly means extra effort, but the rewards have proven
Success factor 6: Management involvement and support absolutely to be worth it - in many different industries, on
many different processes, in many different locations.
Last but by no means least, this is another crucial non-
technical success factor. Unfortunately, in most companies The key is any case is to undertake the first step and to start
management is hardly aware of the potential contributions measuring the current dynamic performance and to take a
and benefits of process control, seems often just concerned serious, unbiased look at the improvement potential in the
with average and not so much with optimal execution of the plant. To benefit the most on a long term basis, also all future
operating targets. The key to get management agreement and conventional and advanced controls should be defined before
support is to explain the reasons for applying APC, for hiring a new DCS system is purchased: Then the best suitable
and training engineers in the language of management, in platform is available and applications development can go on
economical, in business terms! But this, again, calls for a plan without running into capacity of functional limitations.
and it calls for knowledge of the economics.
Were a suitable DCS was already in place, APC has proven to
be the most cost and time effective way to improve plant
PLANNING FOR SUCCESS operations and profitability.

The steps towards a successful plan are not all that difficult:
The best start is to begin measuring the current performance AUTHOR BIOGRAPHY
of key variables with a suitable Performance Measurement
System. The next step is the conduct of an operations HANS HEINZ EDER was born in Vienna, Austria. He spent
improvement (“incentive”) study. The goal is to assess the over 20 years with EXXON in various European countries
current controls and to pinpoint areas for improvement and to and the USA. As APC engineer he has implemented MBC
locate and evaluate further opportunities - besides known applications already in 1979. Later he was APC manager and
operational problems. Untapped, overlooked opportunities are CIM Advisor. Today he heads ACT, a company providing
more difficult to locate and the exploitation strategies more consulting and know-how transfer in process control and esp.
difficult to “sell”. Therefore it is typically advisable to seek APC. A major recent achievement is the development of a
help from outside experts, provided that they have strong complete model based batch-reactor control scheme in just
experience in operations and esp. in economics. It is helpful one day. Hr. Eder is also a registered expert with European
to start always with the business drivers, with question like: Commission and helping projects in the field of eLearning,
Knowledge Management and Virtual Enterprises.
• What would it be worth to run 1 m3/hr more throughput?
• What would it be worth to increase product yield by 1 %?
• What would it be worth to decrease the giveaway in
product quality by 1%?
• What would it be worth to reduce the consumption of
energy or additives etc. by one unit?

Once the necessary improvements on the existing controls

and the future applications, their objectives, the current and
the targeted performance are defined then, just as for any
project, we have to determine the potential incentives. Once
they are known we can easily calculate how much we are
METHODOLOGY
 HYBRID MODELING FOR DEVELOPMENT AND OPTIMIZATION OF FOOD
PRODUCTION CHAINS INCLUDING COSTS AND FOOD QUALITY
Peter de Jong, Maykel Verschueren, Marc M.M. Vissers, Han Straatsma and Erik Smit
Department of Processing, Quality & Safety
NIZO food research
PO Box 20, 6710 BA Ede,
Netherlands
E-mail: [email protected]

KEYWORDS purposes the industry wants to compare both quality

Hybrid modeling, Model design, Agricultural, Global, and costs.
Industrial processes To shorten the time of model development more and more
black box modeling techniques are applied. However, the
ABSTRACT predictive capabilities of these models are not sufficient for
food product and process development. On the other hand,
The need for predictive models of food quality aspects is the industry has no time enough to wait until all process–
increasing. To shorten the time of model development product interactions are covered by mechanistic white box
more and more black box modeling techniques are models.
applied. However, the predictive capabilities of these New developments in modeling are both market-driven and
models are not sufficient for food product and process encouraged by advancing modelling techniques and
development. On the other hand, the industry has no increasing computer resources. Different modelling
time enough to wait until all process–product inter- techniques can be combined to obtain so-called hybrid
actions are covered by mechanistic white box models. models. For example, white box models representing
Thus, the solution is to use both model types: white box physical or chemical processes taking place can be
models for all the known and well mathematical combined with black box models (e.g. fuzzy logic, neural
described phenomena and black box models for the networks) representing classifications by a sensory panel.
complex and fuzzy phenomena. This approach is called Models describing food properties can also be combined
hybrid modeling and has been applied for some with advanced 3D calculations by Computational FLuid
applications in the dairy industry: (1) modeling of the Dynamics (CFD) representing complex processing events.
raw milk quality at the farm, (2) modeling of the cheese Hybrid modeling is a new approach in the food industry.
ripening process and (3) modeling of powder properties. Examples of hybrid modeling are found in majority in
The results show that by the hybrid modeling approach fermentation technology (Alves et al. 1998, Silva-
complex phenomena such as bacterial load in raw milk Henriques et al. 1999, Beluhan and Beluhan 2000). Energy
and the taste of Gouda cheese can be predicted. and mass balances as the most basic form of white box
models are combined with neural network modeling. More
INTRODUCTION recently Nantawan and Weibiao (2002) published a hybrid
neural network model to predict the electrical conductivity
The production of high quality, safe foods becomes more of recombined milk.
and more important. The question is how to predict and to In this paper the hybrid modeling approach is described for
guarantee the product properties based on specifications of prediction of food properties in production chains. Some
the raw material for different types of processes. examples of applications for the food industry are given.
Therefore, mathematical models to predict product
properties would largely benefit the food industry. PREDICTIVE PRODUCT MODELS
Especially, if it is possible to optimise the process operation
in relation to the desired product quality and safety. Mechanistic models (white box)
In general, three types of obstructions for full industrial use
of models can be recognised: Traditionally (food) science is focused on white box
1. The extensive effort needed for model development. modeling, mostly represented as a hypothesis. The most
Development of white box models is too time- basic form of white box models are mass and energy
consuming and black box models have very limited balances based on the conservation laws (e.g CFD). To
predictive capabilities. describe the transformation of food products during
2. Many models are focused on specific unit-operations, processing, good results are obtained with reaction kinetic
the impact on the production chain is ignored. modeling techniques. The model consists of a set of
3. Many models are either focused on certain food quality reaction rate equations based on a reaction scheme that
aspects or on production costs, while for marketing implies a mechanistic hypothesis. Depending on the
application this set might be very simple or very complex. approach for food production chains is presented. To model
Examples of more complex mechanistic models are the the effect of processing and composition on the food
polymerization model for prediction of heat-induced properties recognized by the consumer three levels of
protein denaturation and viscosity changes in milk (De Jong relevant knowledge are described.
and Van der Linden 1998) and reaction models of the 1. Phenomena that can be modeled by mechanistic white
Maillard process (Brands 2002). box models (energy balances, mass balances, reaction
Reaction kinetic modeling can also be used for product- kinetic models).
process interaction such as fouling. De Jong et al. (1992) 2. Phenomena that can be modeled by linguistic rules
developed a predictive fouling model based on the based on expert experience.
following reaction scheme: 3. Phenomena that have uncertain effects on the food
properties but appear to be relevant.
k1 k2
[native β - lg] → [unfolded β - lg] → [aggregatedβ - lg]
k3
[unfolded β - lg] → [aggregates of milk components ] Factors with known Set of white box
quantitative effect process
k4 for example:
[aggregates of milk components] →[deposits]
models
(1) • process design
• process conditions
• T/t profile production chain
• compositions
• (bio) fouling equipment

where β-lg stands for β-lactoglobulin, a reactive protein in

Set of white box
milk and k1 .. k4 are reaction rate constants depending on product
• enzyme activity
models
temperature. The protein denaturation affects the product Factors with known
• viscosity change
• colour
texture while the amount of deposits affects the heat qualitative effect
for example:
• ...

Consumer oriented
transfer and indirectly the process economics. This model • source of ingredients
• additions Set of black box
food properties
for example:
has been used for optimizing dairy production plants by Factors with
models
(neural, fuzzy)
• taste
related the amount of deposits to operating costs (De Jong unknown effect
for example:
• safety
• shelf-life

1996). • season
• climate
White box models are excellent for doing process and • location

product development. The model constants have a physical

meaning and are not depending on process design. The
Figure 1: The Hybrid Modeling Approach for Food
main disadvantage of white box models is the time of
Production
development but an increasing number of white box models
become available.
It is important to realize that factors which have a clear
effect (white box model) on food property x may have an
Neural network and fuzzy logic models (black box)
uncertain effect on food property y. The main advantage of
this approach is that all the knowledge available is
Black box models such as neural network and fuzzy logic
incorporated into the hybrid model.
models are data-driven. In principal, physical laws are
Summarizing, make first a long list of all the factors that
ignored. These black box models have no mechanistic basis
may affect the food properties. Then, collect all the
and are very specific for the process and product trained
available white box models that describe the relation
for. Training is performed by use of fitting the (numerous)
between these factors and the food properties. Fill the
model constants with a huge data set of measured process
‘white spots’ in the quantitative knowledge with black box
and product data. In the case of fuzzy logic models it is
models such as neural network and fuzzy logic models.
possible to translate qualitative relations (expert knowledge)
into linguistic rules. These rules avoid the model to produce
EXAMPLES
unrealistic numbers just outside the model training area.
For example: if temperature is high than bacterial product
Bacterial load of raw milk
load is high.
The main advantage of black box models is their simplicity
One of the major quality issues in the dairy industry is the
to develop them, even for complex phenomena such taste
contamination of raw milk arriving at the processing plant
development. Once there is enough data computer tools,
with micro-organisms, such as Bacillus cereus. This
with some expert help, can make a model within minutes.
bacteria is one of the major spoilage causing bacteria in
However, one small change in de process equipment or the
dairy products. The origin of B. cereus is at the farm and is
composition of the raw materials the model must be trained
transported from cow and milking equipment to the dairy
again. In most cases this makes black box models only
factory, see Figure 2. In order to control and reduce the
interesting for control of processes and bulk products that
levels of microorganisms, a predictive model has been
show no variation in requirements.
developed. The model, FarmQ, describes the contamination
and outgrowth of Bacillus cereus in raw milk at the farm
Hybrid modeling
environment. Special attention is given to the effects of
non-controlled factors such as the weather conditions and
Hybrid modeling combines the advantages of white and
hygiene at the farm (black box). Further more adherence of
black box modeling. In Figure 1 the hybrid modeling
vegetative cells and spores of Bacillus cereus to surfaces
and insufficient cleaning procedures of equipment are taken conditions). Because it is very complex to quantify these
into account (De Jong et al. 2002a). The model can be effects fuzzy logic is used to model the initial
extended to other micro-organisms, viruses, enzymes and contamination.
milk contaminants. To calculate the initial contamination the fuzzy logic model
As shown in Figure 3, the model distinguishes four “unit- follows three steps (see Figure 4 )
operations”: 1. Fuzzyfication: translation of numerical input into fuzzy
1. The cow values.
2. Milking equipment 2. Rule evaluation: fuzzy output values are computed
3. Milk bulk tank at the farm using the rule base.
4. Milk collection tanker 3. Defuzzification: fuzzy output values are translated into
Each step is influenced by different factors and therefore a numerical value.
various calculation modules are used. The modules can be The developed model for initial contamination can be
connected flexible in different ways to describe each applied for example to simulate temperature effects and to
possible situation. classify the farm’s hygiene. In Figure 4 an example is
given of a fuzzy logic modeling approach. The bacterial
load of the raw milk is modeled as a function of the outside
temperature and humidity (%RAH). A temperature of
23 °C is for 70% warm and 40% hot resulting in 70%
medium and 30% high bacterial load which is equivalent to
4.1 colony forming units of B. cereus per ml.
The model-units that describe the milking equipment,
collection tank and contamination during transport take
growth characteristics into account and have a white box
nature. To optimise milk handling at the farm and during
transport dimensions, the cooling capacity and interval
times can be changed. Cleaning efficiency is also a
variable.

Figure 2: The Hybrid FARMQ Model to Predict the

Occurence of Bacillus cereus at the farm

Figure 4: The Fuzzy Inference System (example)

Figure 3: The Combination of White (module 2-4) and
Black Box (module 1) Modeling in FARMQ
It can be concluded that the uncontrollable factors and
qualitative relations can be modeled using fuzzy logic. The
The initial contamination of raw milk is of major
hybrid FarmQ model allows a stepwise description of the
importance and is affected by various controllable (e.g.
milk proces and can be used for identification of critical
farm hygiene) and non-controllable factors, (e.g. weather
control points at the farm (HACCP approach). Simulations
can be performed to determine the effectiveness of Figure 5: Results of Hybrid Cheese Ripening Model: Age
measures to reduce the contamination level of raw milk.
Taste of cheese The hybrid model firstly relates the relevant processing
conditions to the relevant parameters (e.g. moisture, salt,
Research on cheese manufacturing has resulted in collection pH, enzyme activities) in the cheese at the end of the
of white box models focused on different processes and manufacturing process. After that the fermentation process
operations taking place during cheese manufacturing (e.g. during ripening is modelled by considering protein
pasteurisation, renneting, acidification, pressing, brining breakdown as a function of the storage conditions (e.g.
and storage). These models can be used to evaluate the temperature, relative humidity) and the relevant (time-
process of cheese manufacturing from cheese milk storage dependent) parameters (e.g. salt, moisture, enzyme
tanks to storage in warehouses and at retailers as a function activities). Finally, the sensory evaluations are related to
of various process parameters such as concentration of milk the relevant parameters in the cheese (e.g. concentration of
constituents, amount of rennet, heat load, cheese geometry, small proteins and peptides, pH).
storage conditions etc. Predicted values are cheese yield, The hybrid model accurately predicts the effect of
renneting time, the degree of denaturation of whey variations in the processing and storage conditions on the
proteins, inactivation of enzymes, moisture and salt relevant parameters in the cheese during ripening and the
content, pH of the cheese. sensory evaluations. The sensory evaluations are related to
Variations in processing conditions during the manu- the salt concentration, the moisture content, the pH and the
facturing of cheese often also have an effect on the taste of parameters representing protein breakdown. By using
the cheese. Taste parameters are often quantified by means white-box elements in the hybrid global model the
of sensory evaluations. White-box approaches can not be predictive capacities of the model increase. Moreover,
used to include the sensory data in the model. In this case hybrid modeling allows for an optimal use of the available
black-box approaches can be used (e.g. neural networks, data and knowledge since white-box models that were
fuzzy logic). Most of the white models are also relevant for developed in the past can be integrated.
the modelling of taste development. The effect of changes In Figure 5 and Figure 6 some preliminary results are given
in processing conditions on taste parameters can therefore of the hybrid modeling approach of cheese-ripening. In
be described most accurately by a hybrid model, consisting Figure 5 the model outcomes are shown with respect to the
of both white- and black-box elements. estimated age of the cheese by an expert panel. In
To be able to predict the taste development of cheese comparison with a linear black box model the used neural
during ripening as a function of processing and storage network model shows a better prediction of about 20%. In
conditions, a hybrid model was developed based on white- Figure 6 the hybrid model predicts the panel score of a
box model related to cheese-making. First, the key typical off-flavour in cheese. The results show that in
parameters for taste development and the relevant approximately 80 % of the cheeses the panel score could be
processing and storage conditions were identified. Based on predicted on forehand.
the results of this investigation the database on which the
white box models are based was extended with the relevant
experimental results (e.g. protein breakdown, enzyme 1

activities, sensory panel evaluations etc.). A large database

> 1 panel member tasted off-flavour
based on the evaluations of about 400 cheeses (Gouda) was
0.8
used.

70
modelvoorspelling

0.6
Predicted

MMLLRR
Linear model
60 vnoeNeural
ourrsp
a ael lnetwork
n e=twg e
d emr ke te n
0.4
v oPredict
o r sp e ld = g e m e te n
= measured

50
Predicted age (weeks)

0.2

40

0
30 0 0. 1 0. 2 0.3 0.4 0.5 0. 6 0. 7 0.8 0.9 1
≤ 1 panel member
tasted off-flavour
-0.2
20

stand.dev. = 2.3 wk ‘gemeten’

Panel fraction that tasted off-flavour

10
stand.dev. = 1.8 wk
Figure 6: Results of Hybrid Cheese Ripening Model:
0
0 10 20 30 40 50 60 70
Flavour attributes
geschatte ouderdom
Age estimated (wk)
by panel (weeks)
 product development activities of the industry. Besides, the
proposed combination of white and black models decrease
INDUSTRIAL USE the time for model development of predictive models. And,
finally, hybrid models have predictive capabilities with
NIZO Premia respect to black box models.

In order to combine hybrid models with other predictive REFERENCES

models or to construct a new hybrid model a software Alves, T.L.M.; A.C. da Costa; A.W. da Silva-Henriques and E.L.
system has been developed called Premia (Predictive Lima. 1998. “Adaptive optimal control of fed-batch alcoholic
Models ready for Industrial Application). Premia consists fermentation”. Applied Biochemistry and Biotechnology 70-72,
of a database of (white box) models covering a wide range 463-478.
of phenomena, which are often encountered in food Beluhan, D. and S. Beluhan. 2000. “Hybrid modeling approach to
processing, such as predictive models for reduction of on-line estimation of yeast biomass concentration in industrial
bioreactor”. Biotechnology Letters 22, 631-635.
fouling, reduction of contamination, loss of nutritional
Brands, C. 2002. “Kinetic modelling of the Maillard reaction
value, rheological properties, powder solubility and flavour between proteins and sugar”. Ph.D. thesis, Wageningen
development. Functionality is available to connect these University.
models with each other and with process models. For De Jong, P.; S. Bouman and H.J.L.J. van der Linden. 1992.
example, Figure 7 shows a screenshot of the cheese “Fouling of heat treatment equipment in relation to the
ripening model design (‘Chermo’) within NIZO Premia. denaturation of -lactoglobulin.” Journal of the Society of
Dairy Technology 45, 3-8.
De Jong, P. 1996. “Modelling and optimization of thermal
processes in the dairy industry”. ISBN 90-9009034-7,
Montfoort, the Netherlands.
dc
De Jong, P. and H.J.L.J. van der Linden. 1998. “Polymerisation
output
= kc + ...
input
dt
vocht / zout/ salt
moisture
model for prediction of heat-induced protein denaturation and
bprocess
ereidingsconditions
condities eiwitafbraak
proteolyses viscosity changes in milk”. Journal of Agricultural and Food
rijpings condities
ingredients
kaas geometrie
staste
maak
Chemistry 46, 2136-2142.
cheese geometry etc.
...
De Jong, P.; M.C. te Giffel and E.A. Kiezebrink. 2002a.
“Prediction of the adherence, growth and release of
microorganisms in production chains”. International Journal of
Food Microbiology 74, 13-25.
database / knowledge management
De Jong, P.; M.C. te Giffel; H. Straatsma and M.M.M. Vissers.
analyses data, operating data, panel
scores, ... 2002b. “Reduction of fouling and contamination by predictive
kinetic models”. International Dairy Journal 12, 285-292
Nantawan, T. 2002. “Hybrid neural modeling”. International
Figure 7: Cheese Ripening Model within NIZO Premia Journal of Food Properties 5, 49-61.
Silva-Henriques, A.W. da; A.C. da Costa; T.L.M. Alves and E.L.
Lima. 1999. “A hybrid neural model of ethanol production by
Cost minimization
Zymomonas mobilis”. Applied Biochemistry and Biotechnology
77-79, 277-291.
The operating costs of many food production chains
primarily depend on microbial and physical fouling of the ACKNOWLEDGEMENT
equipment. In general, process operating times at relatively
low temperatures (< 70 °C) are due to adherence and A part of this work (modeling cheese ripening) was
growth of bacteria. The operating time of equipment at supported by the J. Mesdag Foundation.
temperatures above 80 °C is determined largely by the
deposition of protein and minerals. The amount of BIOGRAPHY
(bio)fouling directly influences the costs for cleaning,
product losses, pollution and production efficiency. Studies PETER DE JONG was born in Utrecht, the Netherlands
show that fouling is responsible for 80 % of the operating and studied chemical engineering (B.Sc.) in The Hague and
costs in the dairy industry (De Jong 1996). obtained his Ph.D. degree at the University of Delft on
Within NIZO Premia it is relatively simple to connect cost modeling and optimization of thermal processes in the dairy
models to (hybrid) predictive models. These means that at a industry. In 1986 he went to the R&D department of
given set of desired product specifications the operating Campina, a major international dairy company. In 1987 he
costs can be minimized (De Jong et al. 2002b). started to work at NIZO food research as a process engineer.
In the nineties together with Han Straatsma he initiated a
CONCLUSIONS modeling group for applications in the food industry. At this
moment he is responsible for all the research activities on
Hybrid modeling offers great opportunities in the food processing, modeling and control within NIZO food research
industry. First of all, hybrid modeling makes it possible to and is the head of a group of 13 scientific workers.
predict complex food properties such as taste and even
consumer panel scores. These properties are essential for
 A PRELIMINARY QUANTITATIVE RISK ASSESSMENT OF E. COLI 0157:H7
IN IRISH MINCED BEEF
E. Cummins G. Duffy
P. Nally E. Carney
F. Butler J. Sheridan
Department of Agricultural and Food Engineering S. O’Brien
University College Dublin The National Food Centre
Earlsfort Terrace Teagasc
Dublin 2 Dunsinea
Ireland Castleknock
E-mail: [email protected] Dublin 15
Ireland

here. The module aims to assess the prevalence and counts

KEYWORDS of E. coli 0157:H7 in Irish beef trimmings destined for beef
Risk assessment, Simulation, E. coli 0157:H7, Minced beef. burger production.

ABSTRACT MATERIALS AND METHODS

This study looks at the development of a quantitative risk Model Development

assessment model for Escherichia coli (E. coli) 0157:H7 in The primary focus of this module was within the
beef products. A deterministic simulation was developed to slaughterhouse. The prevalence and numbers of E. coli
assess the prevalence and counts of E. coli 0157:H7 on beef 0157:H7 bacteria was modelled at various stages along the
trimmings. Beef trimmings are cuts of meat taken from an slaughter line. A flow diagram of the process is given in
animal carcass at the time of slaughter and are later Figure 1. The model was created in Excel with the add-on
processed into saleable products such as beef burgers. The package @Risk (Pallisade Corporation, New York, USA).
mean calculated prevalence of E. coli 0157:H7 on trimmings
Stunning
was 6% and the calculated mean number of counts on
contaminated trimmings was 2.26 log10 CFU/g. The model Shackling/Hoisting

developed here is one module of a quantitative risk

Ovals denote steps that Neck Hide
assessment investigating the risks of E. coli 0157:H7 may either increase or opening

contamination to humans and highlights the need for further decrease contamination.
quantitative risk assessments in the Irish food industry. Removal of hooves Sampling point 1

Rectangles denote steps

De-hiding
with little or no increase in
INTRODUCTION contamination Head removal

Verocytotoxigenic Escherichia coli, in particular serogroup Brisket sawing

0157, has emerged as a pathogen of major public concern.

High profile outbreaks have focused attention on outbreaks Evisceration

connected to food products, in particular minced beef and Carcass splitting

beef burgers (CDC 1993). The bacterium is present in faeces

and the intestines of healthy bovines (prevalence in Ireland Spinal Cord removal

of approx. 2-4%, McEvoy et al. 1998) and can contaminate Trimming

meat during the slaughter process (Chapman 2000). Over 25

Carcass grading, weighing
confirmed cases of E. coli 0157:H7 poisoning were & stamping

confirmed in Ireland during 2001. Consequences can vary Carcass washing

from severe illness to kidney failure, central nervous system Steam
Pasteurisation

damage and death. A quantitative risk assessment (QRA) is a Chilling

methodology that can help to identify those stages in the Boning out
manufacture, distribution, retail and consumption of Irish
minced beef products, which can contribute to food borne Trimmings

risks. A quantitative risk assessment currently in progress in Sampling point 2 Boxed Beef trimmings Boxed Beef trimmings Sampling point 3
Ireland aims to tackle this issue. This quantitative risk (90VL) (70VL)

assessment is broken into 3 main modules: 1. production of

boxed beef trimmings; 2. processing of trimmings and Figures 1: Flow diagram of beef trimmings production from
burger formation 3. retail and consumer behaviour. These the time of slaughter
modules are modelled separately with each proceeding
module acting as an input into the next. A preliminary risk
assessment model of E. coli O157:H7 in Irish minced beef
has been developed for the first module and is presented Model Inputs
The occurrence of E. coli 0157:H7 in live cattle in Ireland Reductions of E. coli 0157:H7 was found to be
was estimated from studies conducted by McEvoy et al. approximately 3.53 (± 0.49) log CFU/cm2. In this
(1998) in which 6 of 250 cattle tested were shedding preliminary analysis the model considers production in a
detectable levels of the pathogen in their feces. Bacon et al. plant which does not have steam pasteurization in place. The
(2000) evaluated microbial populations on animal hides and most likely effect from chilling is that there is no effect
changes in carcass microbial population at various stages in (Cassin et al. 1998). Growth or decline is assumed to occur
the slaughtering process in the USA. Sponge swab samples only on carcasses which are contaminated on entering the
yielded mean generic E. coli on the exterior hide of 5.5-7.5 chiller. The potential growth was modelled using a triangular
log cfu/100cm2 with corresponding concentration levels on distribution with minimum growth –2, a mode of 0 and a
the carcass (after hide removal but before any maximum of 5 generations. This factor and further inputs
decontamination intervention) of 2.6-5.3. log cfu/100cm2. and calculations used in the model are detailed in Table 1.
The concentration on the hide was obtained by fitting a
continuous empirical distribution to the data using a Second order modelling
cumulative distribution. The dehiding stage can contaminate There is increasing emphasis being placed on quantifying the
the carcass in many ways including direct contact between a impact of uncertainty and variability in risk assessments
carcass and a contaminated hide or cross contamination via (Kelly and Campbell 2000) Variability and uncertainty in the
worker handling. Elder et al. (2000) performed a survey in input parameters was incorporated in the construction of a
the USA to estimate the frequency of enterohemorrhagic second order model by means of probabilistic distributions.
Escherichia coli 0157:H7 in feces and on hides within As @Risk does not facilitate the separation of variability and
groups of cattle from single lots presented for slaughter at uncertainty a macro was developed in visual basic to deal
meat processing plants. The frequency of carcass with this task. 100,000 iterations were performed using Latin
contamination during processing from cattle was also Hypercube sampling with @Risk software. After 1000
surveyed. A cross contamination factor from hide to carcass iterations the macro was designed to change the uncertainty
can be created from data supplied by Elder et al. (2000). parameters, i.e. both uncertainty and variability were
From a sample of 355 hides 38 were E. coli positive. From a simulated with 1000 iterations performed for 100 uncertainty
sample of 341 carcasses, 148 carcasses were positive (pre- values (giving the required total of 100000 iterations). The
evisceration). The two beta distributions were combined to results are presented in Figures 2-4.
create a cross-contamination factor.
RESULTS AND DISCUSSION
Evisceration represents another opportunity for
contamination of the carcass to occur. If the gut of an animal A plot of the prevalence of E. coli 0157:H7 in Irish beef
is positive for E. coli 0157:H7 and the gut is inadvertently trimmings is given in Figure 2.
nicked during the evisceration process gross contamination
of the carcass may occur due to spilling of the gut contents. X <=0.03 X <=0.11
25
The prevalence of E. coli 0157:H7 in the gut of Irish cattle 5% 95%

was determined from McEvoy et al. (1998) where it was 20

Mean = 6.5E-2
discovered that two animals from a sample of 250 were
Probability Density

positive for the bacteria. The prevalence in the gut was thus 15
modeled using a beta distribution. The number of times
evisceration resulted in a nicked gut was determined to be 10
relatively small (between 1 in 100 to 1 in 1000, personal
communication, meat plant production manager). Thus the 5
probability of gut contents spilling and contaminating the
carcass during evisceration was modeled as a uniform 0
0.00 0.04 0.08 0.12 0.16
distribution with a minimum of 0.001 and a maximum 0.01.
Prevalence
After carcass splitting and spinal cord removal there is some
knife trimming to remove visible spots of fecal Figures 2: Prevalence of E. coli 0157:H7 in beef trimmings
contamination. The effectiveness of knife trimming is highly
variable (Prasai et al. 1995) in particular work by Sheridan et The mean calculated prevalence of contaminated trimmings
al. (1992) and Smeltzer et al. (1998) have shown that is 6% with 90 percentile range between 0.03 and 0.11%. The
equipment such as knives, gloves and aprons can act as calculated mean number of counts of E. coli 0157:H7 on
reservoirs for bacteria. contaminated trimmings was 2.26 log10 CFU/g (Figure 3).
Given the estimated small dose required to cause illness
Carcass washing can have the effect of reducing numbers of resulting from the ingestion of E.coli 0157:H7 (Cassin et al.
bacteria (Reagan et al. 1996). There is also evidence that 1998) these predictions may be a cause for concern. Results
some washing procedures may redistribute bacteria to other from this module will be validated by ongoing sampling
parts of the carcass (McEvoy et al. 1997). In this preliminary work currently being carried out at a major Irish beef
analysis a reduction of between 1 and 2.5 fold was assumed processing plant.
as suggested by Cassin et al. (1998) and is incorporated in
the model by means of a uniform distribution. Steam
pasteurization is used in a small number of meat plants in
Ireland. Its efficiency was assessed by Phebus et al. (1997).
 120 X <=2.26 X <=2.27
5% 95% REFERENCES
100 Mean = 2.26
Bacon, R.T.; K.E. Belk; J.N. Sofos; and G.C. Smith. 2000.
Probability Density

80
“Incidence of Escherichia Coli 0157:H7 on hide, carcass and
60
beef trimmings samples collected from United States packing
plants”. Presented at FSIS public meeting on E.coli 0157:H7,
40 February 29, 2000, Arlington, VA.
Cassin, MH.; A.M. Lammerding; ECD. Todd; W. Ross; RS.
20 McColl. 1998. “Quantitative risk assessment for Escherichia
coli O157 : H7 in ground beef hamburgers”. International
0
Journal of Food Microbiology. 41 (1): 21-44.
2.24 2.25 2.26 2.27 2.29
Chapman, P.A. 2000. “Sources of Escherichia coli 0157 and
Log10 CFU/g
experiences over the past 15 years in Sheffield, UK”. Journal
of Applied Microbiology Symposium Supplement. 88: 51S-60S.
Figures 3: Counts of E. coli 0157:H7 on contaminated beef
CDC, 1993. Centre for Disease Control. “Updata: multistate
trimmings outbreak of Escherichia coli 0157:H7 infections from
hamburgers- western United States, 1992-1993”. Morbid.
The sensitivity of the number of counts of E. coli 0157:H7 Mortal. Weekly Report 42:258-263.
on contaminated trimmings to input values was measured by Elder R.O.; J.E. Keen; G.R. Saragusa; G.A. Barkocy-Gallagher; M.
rank correlation. The results of the sensitivity analysis can be Koomaraie; and W.W. Laegreid. 2000. “Correlation of
seen in Figure 4. The input having greatest impact on model enterohemorrhagic E.coli 0157 prevalence in feces, hides and
results was the concentration of E. coli 0157:H7 on the carcasses of beef cattle during processing”. In Proceedings of
animal hide (correlation coefficient 0.86). This highlights the the National Academy of Science. 7: 2999-3003.
Kelly, E. and K. Campbell. 2000. “Seperating variability and
importance of having animals as clean as possible when
uncertainty- making choices”. In Human and Ecological Risk
presented for slaughter. assessment. 6, (1) 1-13.
McEvoy, J.M.; A.M. Doherty; J.J. Sheridan; F.M. Thomson-Carter;
Concentration on animal hide 0.86 P. Garvey; L. McGuire; I.S. Blair; and D.A. McDowell. 1998.
Growth 0.143 “The incidence and spread of Escherichia coli 0157:H7 at a
Reduction due to decontamination -0.147
commercial beef abattoir. Epidemiology of verocytotoxigenic
E.coli.” In Proceedings of Concerted action, CT98-3935,
Cross contamination factor 0.07
hosted by Teagasc, The National Food centre, Malihide,
Average surface area of Trimming 0.001
Dublin, Ireland 8-10th February 2001.
Animals with E.coli in gut 0.001
McEvoy, J., J.J. Sheridan; L. McGuire. 1997. “The effect of cold
Mass of trim per contaminated carcass -0.001 water washing on beef carcass contamination.” Unpublished
Number with nicked gut -0.001 data. The National Food Centre Dublin.
-1 -0.5 0 0.5 1 Phebus R.K.; A.L. Nutsch; D.E. Schafer; R.C. Wilson; M.J.
Correlation Coefficient Riemann; J.D. Leising; C.L. Kastner; J.R. Wolf; and R.K.
Prasai. 1997. “Comparison of steam pasteurization and other
Figures 4: Sensitivity of counts of E. coli 0157:H7 to model methods for reduction of pathogens on surfaces of freshly
inputs slaughtered beef.” Journal of Food Protection, 60: 476:484.
Prasai R.; R.K. Phebus; C.M. Garcia Zepeda; C.C. Kastner;
CONCLUSIONS A.E. Boyle; and D.Y.C. Fung. 1995. “Effectiveness of
trimming and/or washing on microbial quality of beef
The model described in this paper predicted the prevalence carcasses.” Journal of Food Protection, 58: 1114-1177
and counts of E. coli 0157:H7 in Irish beef trimmings. With Reagan J.O; GR. Acuff; D.R. Buege; M.J. Buyck; J.S.
the addition of further modules the model should encompass Dickson; C.L. Kastner; J.L. Marsden; J.B. Morgan; R.
available information about the processing, treatment and Nickelson; G.C. Smith; J.N. Sofos. 1996. “Trimming and
use of beef trimmings. The health risks posed to humans by washing of beef carcasses as a method of improving the
potential E. coli 0157:H7 contamination can therefore be microbiological quality of meat.” Journal of Food
quantitatively assessed. The model results indicate there may Protection. 59 (7): 751-756.
be cause for concern if E.coli counts are not reduced at a Sheridan, J.; B. Lynch; and D. Harrington. 1992. “The effect of
later stage in the processing process. The fact that E. coli boning plant cleaning on the contamination of beef cuts in
commercial boning hall.” Meat Science, 32: 155-164.
concentration on the hide is having the largest effect on E.
Smeltzer, T.; B. Peel; and G. Collins. 1998. “The role of equipment
coli counts highlights the importance clean animal policies that has direct contact with the carcass in the spread of
may be having on reducing E. coli numbers. The model can Salmonella in beef abattoir.” Australian Veterinary Journal,
provide an appropriate decision support tool aiding risk 55: 275-277.
mitigation strategies in an effort to protect human health.

ACKNOWLEDGEMENTS

The authors wish to acknowledge the Department of

Agriculture for their funding of this project under the Food
Institute Research Measure.
 Table 1: Model input details
Variable Description Distribution/model Units
S Slaughterings 2000000 animals/year
Ph Prevalence of E.coli on hide Beta(6+1,250-6+1)
Nh No. Animals with E.coli on hide Poisson(S*Ph) animals
Pg Prevalence of E.coli in gut Beta(2+1,250-6+1)
Ng No. Animals with E.coli in gut Poisson(S*Pg) animals
2
Ch Conc of E.coli on hide Cumulative log10 CFU/cm
Cm Mean 4.533
Cstd Standard deviation 1.1915
2
Cha Average Conc. of E.coli on all contaminated hides Normal(Cm, Cstd/SQRT(Nh)) log10 CFU/cm
Fcc Factor for cross-contamination Beta(148+1,341-148+1)/Beta(38+1,355-38+1)
Fccm Mean 4.5333
Fccstd Standard deviation 1.1915
Tcc Total contaminated carcasses Normal(Nh*Fccm, SQRT(Nh)*Fccstd) Carcasses
Pe Probability of evisceration Uniform(0.001, 0.01)
Ngc Number Animals cut at evisceration with gut infection Poisson(Pe*Ng) Animals
Fei Factor for evisceration increase 2
CCF Factor for increase/decrease from hide to carcass Cumulative
CCFm Mean 0.53586
CCFstd Standard deviation 0.2166
CCFav Average factor increase/decrease from hide to carcass Normal(CCFm, CCFstd/SQRT(Tcc))
Wr Washing reduction Uniform(1,2.5)
G Growth Triang(-2,0,5) Generations
2
Cctc Concentration on contaminated carcass Ch+CCFm-Wr+LOG(2^G) log10 CFU/cm
3
Cctcg Concentration on contam. carcass if gut nicked Ch+CCFm+Fei-Wr+LOG(2^G) log10 CFU/cm
2
St Surface area of trimmings RiskUniform(0.25,1) cm /g
Stm Mean 0.625
Ststd Standard deviation 0.21651
Stct Average surface area of contaminated trim Normal(Stm, Ststd/SQRT(Tcc))
Mt Mass of 90VL trimmings Normal(12000, 1000, Truncate(50, )) grams
Mtm Mean 12000
Mtstd Standard deviation 993.61
TMcc Total mass of trim from contaminated carcasses Normal(Tcc*Mtm, SQRT(Tcc)*Mtstd) grams
TMcg Total mass of trim from contam. carcasses when gut cut Normal(Ngc*Mtm, SQRT(Ngc)*Mtstd) grams
Nct Number of contaminated trimmings TMcc/Mtm
Nctg Number of contam. trimmings when gut nicked TMch/Mtm
Tt Total Number of trimmings produced S*2

Ptrim Prevalence of contaminated trimmings Nct+Nctg/Tt

Ccta Average concentration for each contaminated trimming Stct*Cctc CFU/g
Cctag Average concentration for each contam. trimming if gut nicked Stct*Cctcg CFU/g
 COMBINING LATENT CLASS ANALYSIS AND TREE-BASED MODELLING
TO SIMULATE THE DECISION PROCESS OF EXPERTS DURING
VISUAL QUALITY CLASSIFICATION

Schrevens Eddie and Coppenolle Hans

K.U.Leuven
Faculty of Agricultural and Biological Sciences
Research Unit for Quality Management
W. De Croylaan 42
B-3001 Heverlee, Belgium
Email: [email protected]

KEYWORDS descriptors. The grading process of chicory follows a dual

Image processing, chicory, quality, classification, grading mechanism. The dual system includes:
variability, simulation. • Classification based on intrinsic product
characteristics: shape, freshness, colour, openness of
the top, length of the kernel, …
ABSTRACT • Classification based on the size of the product: length
Study of uncertainty and enclosed subjectivity in expert and width
quality decision processes is a first step towards a structural Present-day classification rules for chicory include
comprehension of fuzzy classification concepts and the overlapping regions on the classification measurement scale
definition of optimal quality criteria. This paper relates to between quality classes. The fuzzy nature of the grading
visual quality classification of chicory by auction experts in system as well as the typical subjectivity in visual quality
a measurement space of quality descriptors length and classification contributes to uncertainty of the decision
width, in accordance with the traditional grading process. Differences in interpretation of classification rules,
mechanism. A uniform sample of products was optimally handling of different classification schemes and limitations
selected in the bivariate measurement space based on of the human perceptive system, result in inter- and intra-
digital imaging techniques. Image sequences of the selected individual variability of the quality decision. However,
products, were classified by five experts based on length optimal quality classification of horticultural products at the
and width in accordance with the traditional auction system. auction is necessary for an optimal management of the
The image sequences were visualized using a computer distribution channel.
based image system. Structural approaches that optimally recognize expert
Between-expert variability is analysed by latent class variability, are a pre-requisite for an objective
analysis. This statistical technique reveals latent clusters of comprehension of the probabilistic nature of human
homogeneous chicory products based on the observed decision processes. In this paper, a transparent statistical
mixture of expert classification scores. The contingency- approach to simulate the typical between-expert variability
based approach of latent class analysis simultaneously inherent in human quality classification is proposed. A
explains how the identified consensus structure relates to combination of latent class analysis and tree-based
the quality scoring of individual experts. The latent classes modelling is used to optimally detect uncertainty patterns in
of chicory products are thereafter related to the perceptual the quality classification of chicory by auction experts.
space of product quality descriptors using tree-based
modelling to simulate the variability construct inherent in MATERIAL AND METHODS
expert quality classification. Extraction of quality descriptors by quantitative
imaging techniques
INTRODUCTION
The study of visual quality evaluation of the fresh product
The definition of quality is often subjective and/or requires primarily an objective measurement method to
difficult to interpret. The uncertain nature of quality gives define quality. The visual character of quality evaluation
rise to the typical variability in quality decision between makes horticultural products very suitable for investigation
different experts. This work specifically concerns the by quantitative imaging techniques. Extraction of quality
quality classification of chicory (Cichorium intybus L. var. descriptors, as size, shape and colour, from digital video
Foliosum Helgi) by auction experts. At the auction, the images of a product, results in a quantitative, objective
quality of a perishable heterogeneous commodity of chicory description of external product features. Quality descriptors
products is controlled based on a set of visual quality of different varieties of chicory were objectively measured
during the season to compile a database of product features,
corresponding images and full-sized images sequences. The full-sized image sequences were shown in the same
This resulted in database of 1000 chicory crops. In this order to all experts. This limited positional effects by
experiment, only the quality parameters length and width controlling the influence of preceding image sequences on
are involved. The length of a chicory crop is defined as the quality judgement.
mean of the length of the head on four different images, the
product respectively rotated over 0, 90, 180 and 270
degrees. Width is calculated similarly.

Sampling strategy
In this experiment, only classification based on the size of
the product is studied. The dual grading mechanism for
chicory justifies the assumption of independence between
classification based on quality descriptors as shape, colour,
top openness, etc. and classification based on the size of the
product.
Chicory is selected in the 2-dimensional measurement
space of length and width by use of a 2-dimensional grid.
The distance between successive grid-points is determined Figure 2: Computer based image system
as 5 percent of the range of both quality descriptors.
Eliminating biologically impossible grid-points and Experts scored the products on a categorical five-point
selecting products as close as possible to the remaining scale, corresponding to the auction grading system. The
grid-points in the measurement space, results in a collection classes represent very short (quality class 1), short thin
of experimental subjects that are uniformly distributed over (quality class 2), short thick (quality class 3), thick (quality
the product space. The method resulted in 214 products class 4) and long chicory (quality class 5). The
selected from the constructed database of 1000 chicory classification scheme imposed on the quality experts is
crops (Figure 1). The sampling strategy assumes each part graphically represented in Figure 3. The auction experts
of the bivariate space of quality descriptors to be equally involved in the experiment were all familiar with the
important. In this way, the experimental design protocol imposed classification scheme. Figure 3 clearly shows the
enables an optimal exploration of the classification extensive overlap between the quality classes. Thus,
activities of auction experts with respect to the entire range variability between auction experts arises in this
of the length and the width of commercial chicory products. experimental context from the existing overlap between
quality classes as well as from subjectivity of auction
experts in quality classification. The defined five-point
categorical response scale exhibits a partially ordinal
character. The longer and thicker the product, the higher the
correspondence with a higher quality class. Quality
evaluation of selected chicory was not repeated within
experts.
8

Class 1
Class 2
Class 3
Class 4
7

Class 5 +
+ + + +
+ + + + ++ + +
+ + + +
+ + ++ + +
+ + + +
+ + + + + +
Width (cm)
6

+ + + ++
+ + + + ++ ++ + + +
+ + + ++ + ++ ++ + + + ++ ++ + + +
+ ++ + + ++ ++
+
Figure 1: Selection of products in a 2-dimensional space + +
++ + + ++ + + + + + + ++ ++ +
5

+ + ++ + + + + + + + + + +
+ ++ ++ + + ++ +
+ + + +
length-width (. = product, o = gridpoint, + = +
+ ++ + +
+ + + + ++ +
+ + + + +
+ + + + + ++ +
+
+
+ + + + +
+ + +
+ ++
+
+ + + + ++ + + +
selected product) + ++ + + + + ++ + +
4

+
+ ++ + ++ ++ + + + + + + + +
++ +
+ ++
+ + + +
3

Experiment by use of a computer based image system 9 10 11 12 13 14 15

Length (cm)
16 17 18 19 20

Full-sized image sequences of the 214 selected products,

uniformly distributed over length and width, were evaluated Figure 3: Auction grading criteria for chicory based on
by five experts (A1, B1, C1, D1, E1) using a computer length and width, superimposed on the selected
based image system (Figure 2). The use of a computer chicory crops (indicated as +)
visualisation system eliminated seasonal effects and
deterioration as influencing factors in the sampling strategy.
Latent class analysis can be derived to assign each product having response
Latent class analysis is proposed as a statistical technique pattern jklmn to that latent class i for which Pix| |jklmn
abcde
is
that is able to define a consensus structure over auction highest (McCutcheon 1987; McLachlan and Basford 1988).
experts. Quality classification typically results in a multi- Maximum likelihood estimates for the latent class model
way contingency table with the quality decision by different are obtained by the Expectation-Maximisation (EM)
auction experts as classification criteria. The psychometric algorithm (Goodman 1974; Goodman 1979; McLachlan
concept of latent class analysis relies on an additive and Thriyambakam 1997). The iterative procedure proceeds
decomposition of the observed frequencies for the expert by successively updating the initially random parameter
response patterns in a multi-way contingency table by an values using a maximum likelihood approach for the
unmeasured latent categorical variable x based on an assumed model concept.
optimality criterion (McCutcheon 1987). The The formulation of the latent class model agrees with a
decomposition by x is such that the classification by the conditional independence loglinear model where the quality
experts in each of the resulting equidimensional classification by the different experts is independent given
contingency table components, or latent classes, is the unobserved latent class i (Agresti 1990; Haberman
statistically independent. A latent class can be considered as 1979, Hagenaars 1993). In the specific context of quality
a group of homogeneous chicory crops based on the classification, theoretical independence between experts
mixture of expert classification scores. In this way, latent given a latent class corresponds to the intuitive assumption
class analysis results in the definition of a consensus model of homogeneity of scored products within a latent class.
over auction experts. The contingency-based approach of
latent class analysis simultaneously explains how the latent Tree-based modeling
classes relate to the quality scoring of individual experts.
The method of classification and regression trees (CART)
Let the five experts judge products on the five-point
categorical scale. The experiment implies five categorical is an approach to model the relationship between a
variables a, b, c, d, e for the five experts, with categorical- or numeric response variable with independent
variables possibly measured on different scales (Breiman et
corresponding quality class levels indexed by j, k, l, m, n
respectively. Categorical values j, k, l, m, n vary from 1 to al. 1984). In a statistical framework it can be considered as
5. The levels for the latent variable x, referred to as its a non-parametric discriminant analysis. Tree based
modelling generates decision trees as a succession of
latent classes, are indexed by i and assumed to vary from 1
to y. The associated latent class model is formalised as decision rules derived from an existing classified data set or
training set. For each binary split, an impurity function is
xabcde
Pijklmn = Pix P aj||ix P bk||ix P cl||ix P dm|x|i Pen||xi minimised over the set of all possible univariate decision
with rules. The rule resulting in minimal impurity is retained.
xabcde
Pijklmn The splitting procedure is stopped when a terminal node
receives less than 5 observations. From this point on, the
the expected joint probability of level i of the latent variable decision tree is pruned by minimising a cost-complexity
x and response pattern jklmn given by the experts. function, compromising between minimal cross-validated
Pix deviance and minimal number of terminal nodes.
is the latent class probability of falling into latent class i, A mixture model consensus structure over experts is
and obtained by means of latent class analysis. Once the latent
Paj||ix , P bk||ix , Plc||ix , Pdm|x|i , Pen||xi distribution known, probabilistic clustering assigns each
product to the latent class to which it has the highest
are conditional probabilities of falling into levels j, k, l, m estimated posterior probability of belonging. This step
and n respectively, given latent class i. The model enables to explain the relationship between the consensus
parameters are restricted as configuration represented by the latent classes with the
y z z z z z
∑ Pix = ∑ Paj||ix = ∑ Pbk||ix = ∑ Pcl||ix = ∑ Pdm|x|i = ∑ Pen||xi = 1 measurement space of product quality descriptors using
i j k l m n tree-based modelling.
xabcde
The expected probabilities Pijklmn are related to the
RESULTS
expected probabilities Pabcde
jklmn for a response pattern jklmn
In the context of the classification experiment with
in the original observed multi-way contingency table by chicory, a latent class can be interpreted as a group of
y homogeneous chicory crops. Based on the mixture of expert
jklmn = ∑ Pijklmn
Pabcde xabcde
i
classification data, we searched for optimal groups of
Posterior probabilities homogeneous chicory crops that reflected the best possible
xabcde consensus agreement among quality experts. Different
Pijklmn latent class models were run for an increasing sequence of
Pix| |jklmn
abcde
= y
xabcde latent classes. A latent class model with a higher number of
∑ Pijklmn
i latent classes implies a better goodness of fit for the
conditional independence structure or more homogeneous
chicory per latent class. Hence, an optimality criterion is The latent class model with six latent classes, showing the
necessary to select an optimal latent class model minimal AIC value of 2008.64 was selected. Conditional
compromising between model complexity and model bias. and latent class probabilities of the selected model are
Model selection was based on minimisation of Akaike’s represented in Table 2. From the conditional probabilities
information criterion (AIC). This parameter integrates follows that latent class 1 mainly agrees with chicory
information about model bias with a penalty for the number classified as quality class1 (very short), latent class 2 with
of estimated model parameters (Akaike 1974). chicory classified as quality class 2 (short thin) or 3 (short
A latent class analysis was performed based on the thick), latent class 3 with chicory classified as quality class
observed classification data of the five experts. The 3 (short thick) or 4 (thick), latent class 4 with chicory
marginal frequency distribution of the classification pattern classified as quality class 2 (short thin), latent class 5 with
for the five experts is represented in Table 1. Expert E chicory classified as quality class 5 (long) by expert E1,
classified many chicory crops in quality class 5 with respect latent class 6 with chicory classified as quality class 3
to other experts, expert B didn’t use quality class 5 at all (short thick). As a latent class includes homogeneous
when classifying chicory crops. chicory, uncertainty exists for particular chicory crops to
Table 1: Marginal probabilities for the five experts classify them as quality class 2 (short thin) or 3 (short
thick), for other crops to classify them as quality class 3
Expert Class 1 Class 2 Class 3 Class 4 Class 5 (short thick) or 4 (thick). This supposes uncertainty nearby
A1 29 70 55 55 5
B1 48 79 61 26 0
the boundaries of adjacent quality classes, which agrees
C1 40 63 95 15 1 with the existing overlapping regions between adjacent
D1 29 84 67 26 8 quality classes in the commercial grading system for
E1 25 54 69 41 25 chicory. Other latent classes (latent class 1, 4 and 6) reveal
evidence that some basic understanding of the appearance
As expert B did not use quality class five when of chicory typical for quality class 1, 2 and 3 exists. The
classifying chicory crops, the initial conditional latent class probabilities reported in Table 2 indicate the
probabilities
b|x
P5|i of expert B classifying chicory in quality relative magnitude of the different latent classes.
class 5 were restricted to 0 in the latent class initialisation Table 2: Latent and conditional probabilities of the selected
phase for all levels of latent variable x. Latent class models latent class model with six latent classes
for an assumed number of latent classes ranging from two Latent Latent Quality Expert Expert Expert Expert Expert
to eight were estimated. Each latent class model was class class class A1 B1 C1 D1 E1
estimated five times to circumvent the eventual occurrence probability
of local maxima in the estimation procedure. The values for 0.183 1 1 0.719 0.946 0.917 0.712 0.637
1 2 0.201 0.054 0.026 0.173 0.199
the log-likelihood and Akaike’s information criterion (AIC) 1 3 0.080 0.000 0.057 0.115 0.164
of the fitted models are represented in Figure 4. The figure 1 4 0.000 0.000 0.000 0.000 0.000
shows the log-likelihood as monotonously increasing with 1 5 0.000 0.000 0.000 0.000 0.000
an increasing number of latent classes. This essentially 0.111 2 1 0.000 0.000 0.039 0.000 0.000
agrees with the idea that a higher number of latent classes 2 2 0.501 0.576 0.186 0.738 0.064
implies either a better fit for the conditional independence 2 3 0.378 0.424 0.775 0.262 0.936
2 4 0.121 0.000 0.000 0.000 0.000
structure or the identification of groups of more 2 5 0.000 0.000 0.000 0.000 0.000
homogeneous chicory crops. 0.181 3 1 0.000 0.000 0.000 0.000 0.000
3 2 0.000 0.027 0.000 0.079 0.000
AIC 3 3 0.120 0.506 0.772 0.428 0.000
Log-likelihood
3 4 0.880 0.467 0.228 0.493 0.911
3 5 0.000 0.000 0.000 0.000 0.089
0.274 4 1 0.000 0.089 0.021 0.000 0.000
-900
2000 2100 2200 2300 2400 2500

4 2 0.856 0.911 0.897 0.965 0.761

4 3 0.000 0.000 0.082 0.017 0.068
-1000

4 4 0.144 0.000 0.000 0.000 0.017

Log-likelihood
AIC

4 5 0.000 0.000 0.000 0.018 0.154

-1100

0.065 5 1 0.000 0.000 0.000 0.000 0.000

5 2 0.000 0.428 0.357 0.000 0.000
5 3 0.000 0.000 0.217 0.000 0.000
-1200

5 4 0.639 0.572 0.354 0.499 0.093

5 5 0.361 0.000 0.072 0.501 0.907
2 3 4 5 6 7 8 0.186 6 1 0.020 0.141 0.046 0.026 0.000
Number of latent classes 6 2 0.000 0.070 0.000 0.000 0.000
6 3 0.961 0.789 0.922 0.974 0.914
6 4 0.019 0.000 0.032 0.000 0.086
Figure 4: Log-likelihood and Akaike’s information criterion 6 5 0.000 0.000 0.000 0.000 0.000
(AIC) of the fitted models
 Once the latent distribution is known, the defined latent CONCLUSIONS
classes are linked with the from images extracted quality
The beneficial use of combining latent class analysis and
descriptors length and width using tree-based modelling. tree-based modelling to define a consensus model based on
Based on the cost-complexity measure, an optimal pruned expert quality classification data has been demonstrated.
tree-based model with 6 terminal nodes and 21.5 %
Latent class analysis results in a consensus structure
misclassification error was selected. The pruned tree-based underlying the decision process while simultaneously
model which links the latent classes with the quality explaining its interaction with individual expert
measurement space is represented in Figure 5. The branches
information. The latent class analysis simply clusters
connecting the nodes, are labelled by the left and the right products with homogeneous uncertainty patterns in an
splitting rules. Ellipses denote interior nodes, while optimal way. In this way, the groups of homogeneous
rectangles represent terminal leaves. The predicted value of
chicory crops detected based on the observed classification
the quality classification response is centred in each node. scores, reflect the quality perception of an average expert.
The predicted value corresponds to the quality class to
Conditional probabilities resulting from latent class analysis
which the majority of the chicory crops arriving at a node relate the latent classes of homogeneous chicory products
are assigned. The number under each node denotes the with the classification of individual experts. The consensus
misclassification.
structure is finally linked with explanatory quality
descriptors using tree-based modelling.
By summarising all facets in the quality decision process
efficiently, the technique provides an informative construct
for redefinition of fuzzy classification systems. The
technique allows an elegant specification of between-expert
variability in model formulation. The integration of this
essential information in one single model defines a
statistical approach to the analysis of classification data
with a broad range of applicability.

REFERENCES
Agresti, A. 1990. Categorical data analysis. John Wiley & Sons,
New York.
Akaike, H. 1974. “A new look at the statistical model
identification.” IEEE Transactions On Automaic Control 19, 716-
723.
Breiman, L.; J. Friedman; R. Olshen; and C. Stone. 1984.
Classification and regression trees. Wadsworth International
Figure 5: Optimal, pruned tree-based model that links the Group, Belmont.
latent classes with the quality descriptors length Goodman, L.A. 1974. “Exploratory latent structure analysis using
and width both identifiable and unidentifiable models.” Biometrika 61, 215-
231.
Figure 6 shows the same tree-based model, represented as Goodman, L.A. 1979. “On the estimation of parameters in latent
predictor versus fitted values, superimposed on Figure 3. structure analysis.” Psychometrika 44, 123-128.
The location of the latent classes in the measurement space Hagenaars, J.A. 1993. Loglinear models with latent variables.
of length and width and their relation with the imposed Sage Publications, London.
grading rules can easily be read from the graphical display. McCutcheon, A.L. 1987. Latent class analysis. Sage Publications,
London.
McLachlan, G.J.; and K.E. Basford. 1988. Mixture models:
8

Class 1
Class 2 inference and applications to clustering. Marcel Dekker Inc., New
Class 3
Class 4 York.
7

Class 5
+ + +
+
+ McLachlan, G.J.; and K. Thriyambakam. 1997. The EM algorithm
+ + ++ + +
+ +
+ +
+ + +
+
++
+
+ +
+ +
+ and extensions. Wiley, New York.
+ + + + + +
Width (cm)
6

+ + + ++
+ + + + + + + + +
+ + + ++ + + + + + ++ + + + +
+ + + + + + +
+ ++ + + + ++ +
+ +
++ + + ++ + + + + + + ++ ++ +
5

+ + ++ + + + + + + + + + +
+ ++ ++ + + ++ +
+ + + ++ + + +
+ ++ + + + + + + + +
+ + + + + + + ++ +
+ + + + ++ + + + ++ + + + + + + ++
+ + + + ++ + ++ + + + + ++ ++ ++
4

+ ++ + ++ ++ + ++ + + +
+ ++ + + +
++ + +
3

9 10 11 12 13 14 15 16 17 18 19 20
Length (cm)

Figure 6: Latent class structure (black lines) in relation to

the auction classification scheme imposed on the
experts, chicory crops are indicated as +
 SYSTEMIC AND INTERPRETATION OF FOOD TEXTURE TESTS
Alain-Claude ROUDOT
ESMISAB, Université de Bretagne Occidentale
Laboratoire de Microbiologie et Sécurité Alimentaire
Technopole Brest-Iroise
F-29280 Plouzané
France
E-mail: [email protected]

KEYWORDS comparing the interpretation's possibilities of these

methods.
Food texture, tissue organisation, histological
modelling, systemic. COMPLEXITY THEORY APPLIED TO FOOD
STRUCTURE
ABSTRACT
A cellular tissue is assumed to be a complex system
Food texture analysis is a method of testing in which we can find different parts: the individual
agricultural product which is actually well known. cells, the physical organisation of these cells and the
However, interpretation of the results obtained intrinsic functions of the cells as parts of the tissue.
usually uses continuous theories which are not able These different elements exist in all complex systems
to give good mechanical interpretations. Different and need to be taken into account in a modelling
examples are given showing the limits of such process if further simulations have to be close to
theories. In order to solve this problem one can actual experimentations.
"mentally" interpret the results as global results of a Here, we will only consider two independent
particulate medium. This paper shows that it may domains: the domain of operation of the structural
possible to model food tissue and apply a components of the tissue (i.e. of the cells), and the
mechanical theory of cellular medium, which gives domain in which they will be considered as a whole
far more better results. That theory is based on (Maturana, 1987). The first one will be used during
complexity theories taking into account tissue creation of the model, and the second one during the
organisation as well as static and dynamic relations simulations using this created tissue. In this case,
between its components. what we can call the tissue organisation is the
configuration of both static and dynamic relations
INTRODUCTION between its components which specify its identity as
a whole. So, tissue organisation takes place partly in
A cellular food product such as vegetable or fruit, is the first domain and partly in the second one. In fact
constituted of various elements, each one having its it is the link between the two domains. During
own physical behaviour. A cellular tissue has a physical tests on the created tissue, interactions occur
behaviour in itself which is linked to its cells between external elements, the tissue as a whole and
properties and cells organisation. Classical methods individual cells. It follows that interactions can
of interpretation and/or testing are based on a modify both the first and second domains. Relations
continuous paradigm whose advantage is between external events and constitutive elements are
simplicity. However its drawback is that it is directed by classical physical laws, all relations with
impossible to link tests results to physical tissue as a whole being directed by structure
characteristics of the products. Another theoretical organisation, and hence its own internal laws.
approach only takes into account the individuals Finally, what appears, is that tissue behaviour is not
cells behaviour. Unfortunately, intercellular links determined by tissue structure or external medium,
are forgotten and then interpretation, when possible, but by interactions inside the tissue itself and between
is truncated. So, neither the macroscopic method, that tissue and this external medium. This is the
nor the microscopic one is sufficient to understand reason why theories which consider that tissue
cell tissue responses to external physical stresses. behaviour comes from cells' physical characteristics
Nowadays systemics and complexity theories (Le are not sufficient, as they forget the interactions
Moigne, 1991) seem to be the best approaches to inside the tissue itself and the high degree of
explain cell tissue behaviour and find the links recursivity in such a system.
between physical structure and texture of this kind
of food products (Roudot, 2002b). This can be
shown by studying different tests results and
Example of a compressed cellular tissue understand the test effects. So, a macroscopic load
creates a macroscopic strain, but the only way of
A cellular tissue is made of numerous cells placed explaining this last one is looking at the microscopic
together, each one in contact with some others. If a level and then generalise the results. In fact,
load is applied on this tissue (Fig. 1), one can see microscopic analysis of an isolated cell may be of
some global geometrical deformations. interest in order to better understand its modification
procedures in terms of stresses and strains. But one
can see that without a generalising process, this point
of view is only a reductionist one which is absolutely
insufficient in explaining the external effects on a
whole tissue.

SOME EFFECTS EXPLAINED BY SYSTEMIC

F ∆h σ
σ = ;ε = ;E = Tissue breakdown under mechanical stresses
S h ε
When submitted to a mechanical stress, a cellular
Figure 1: Continuous model of a tissue under load. tissue first accommodates, then fractures. This
One uses calculates global stress and strain. fracture is very different from the result obtained on a
continuous medium and is generally not explained:
But if the strain is supposed to be quasi-static, and with a continuous theory the explanation is
if one looks at a cellular level, the deformation is impossible, with a microscopic theory the
far more complex (Fig. 2). First the contact cell is explanation is insufficient. In this last case one
compressed and then is deformed because of its speaks of cell wall rupture, cell debonding and tissue
elastic wall. This cell strain creates strains and rupture, but gives no explanation on the localisation
stresses on the local other cells whose inner of these problems. Actually, this strain phenomenon
pressure grows and so on. Added to this first effect is only explainable at the tissue level using systemic
another one appears which is geometrical tools. The first tool is the percolation theory, which
modifications of the tissue, due to the void spaces explains that the external stress progresses chaotically
and the moving possibilities of the cells. because of the cell conglomerate (Cundall and Strack,
1979). Because of that structure, the stress may attain
very high values from place to place (Travers et al.,
1987) (Fig. 3).

Figure 2: Tissue analysis of the same test as on

figure 1. One can note structural (physical and
geometrical) modifications, and has to take into
account local stress and strain analysis.

Porosity of cells occurs too between their inner

part and void spaces. When the local stress has a
high value, destruction of cell(s) can occur, creating
large heterogeneities with their own consequences
such as variations of stress, strain and position of
other cells, possibility of a failure process and so
on. So a very simple test from a macroscopic point Figure 3: Lines of principal stresses for a granular
of view is clearly a complex one from a assembly under vertical load. Analysis of such effect
microscopic point of view. A good analysis of such uses percolation theory. (from Cundall and Strack,
a phenomenon needs a specific method using 1979)
complex theories and not a simple macroscopic
interpretation. A microscopic analysis (Gao and Another tool is the geometrical analysis of the
Pitt, 1991) is no more a good method because structure allowing detection of unstable areas where
dynamic relations between cells (strain and stress cells and voids organisation creates some local pre-
transfer), between cells and voids (mass transfer) or stressed areas (Lee, 1989). Another important point
between cells and tissue (geometrical changes) usable with complexity theories is the time constant
cannot be taken into account. Then this last which is different for one cell, a cell tissue or a
interpretation cannot be an actual one in order to continuous medium (Subhash, 1991). This time
constant is for instance the main factor to use in interpretation it is a cutting test with a knife
order to differentiate impact and compression tests. perpendicular to the sample fibres (Fig. 5). However,
It is also linked with tissue accommodation. So the classical interpretation considers that the sample
these different tools are necessary in the is undeformable and continuous. In fact this tests is
explanation of the strains obtained on cellular tissue an assembly of different deformations occuring
during these mechanical tests. altogether: shearing, extension, compression, and so
on (Roudot, 2002a).
Energy transfer

Some tests on texture analysis use energy transfer cutting

such as near infrared or mechanical vibration like
ultrasound. The problem is to catch a signal
resulting from its transmission inside a sample, or
its inner reflection on the different tissue interfaces.
In these cases the most important fact in order to Figure 5: Classical macroscopic analysis of the test
obtain good results is to have a constant acoustical
impedance over the sample. Variations in this An actual analysis of this test has to take into account
parameter attenuate the signal and can suppress it this fact (which is macroscopic) but also the actual
and one has to understand the reasons of this structure of the product which is fibrous. So the
attenuation or has to determine the depth attained deformations occurring at a global level also occurs
by the input signal before its reflection (Birth, at the fibre level, with some particularities such as
1978). Most of the impedance variations come from large deformations before failure, differences in
cell-voids interfaces (Fig. 4). elasticity between the fibres or local tissue collapses
(Fig. 6). These effects necessitate a local analysis
before a globalisation process, if a complete
interpretation of the failure process is needed.
Nevertheless, this test is rarely fully analysed, and
lack of knowledge on its interpretation does not seem
to be a problem for users.

crushing

shearing

Muscular Fibre
Large extension
Figure 4: Simulation of electromagnetic waves
transfer through a cellular tissue. The waves can go
Figure 6: Different phenomena occurring during the
through the cell walls, reflect or vanish.
test at the fibre level.
It seems clear that continuous theories cannot
Mechanical heterogeneity of plant food
explain the attenuation effects. In that case too,
microscopic theories seem an actual progress but
It is well known that plant tissues are heterogeneous,
they are quite limited because each element is
from a mechanical point of view. This is a direct
analysed separately from others. The only
consequence of the cellular structure, but also of the
theoretical solution is given by systemic. The tools
agronomic growing conditions. What is less studied
necessary in this explanation are the same as before
is the influence of the growing phenomenon itself:
except that the time constant problem is changed in
cell division, tissue organisation, intercellular
a wavelength problem.
stresses, and so on. In fact, this last point is
This kind of problem also occurs in food processing
classically forgotten because it cannot be studied with
with heat transfer. In that case, heat transfer is due
classical method, but only using systemic ones. The
to two independent facts: conduction along the cell
main consequence of this fact is that lots of things
wall (and then percolation theory appears), and
concerning tissue structure and relations between the
convection inside each cell (Gibson and Ashby,
different kinds of cells are unknown. Results from
1988).
such a science as histology cannot be verified
numerically using simulation because physical
Warner-Bratzler test
theories of such tissue are not developed enough in a
usable way. However, when used these new methods
This test is essentially used on meat and fish
show that the growing phenomenon is probably the
products. It is a shearing test made in order to
evaluate the sample tenderness. In its ideal
main parameter responsible for this mechanical Maturana, H.R. 1987. "The biological foundation of
heterogeneity (Roudot et al., 1994). self counsciousness and the physical domain of
existence". In: Physics of cognitive processes.
CONCLUSION Caianello E.R. (Ed.) World scientific, Singapore,
324-379
The various examples described permit to conclude
that systemic analyses are the best solutions in Niklas, K. 1989. "The cellular mechanics of plants".
order to understand strains and stresses in food American Scientist 77, 344-349
texture analysis tests. In fact, if classical methods
(microscopic and macroscopic) are currently used Ouwerkerk, C.E.D. 1991 "A micromechanical
in modelling these tests, their actual interpretations connection between the single-particle strength and
are very often empirically based on systemic. For the bulk strength of random packings of spherical
instance, the researcher uses classical mechanics in particles" Powder Technol. 65, 125-138
order to analyse the consequences of a loading
process such as the first described in this paper. Roudot, A-C., Studman, C.J.and F. Duprat. 1994.
Looking at the differences between the "Morphogénèse et fermeté de la pomme". Plant
experimental results and its continuous theory Biomechanics International Symposium, Montpellier,
(shearing bands for instance), he tries to explain 1994/09/05-09. Elsevier, Paris, 151-152
them by the cellular structure of the tissue.
However, though this empirical explanation is a Roudot, A-C. 2000. "Histological modelling, a new
good one, it does not give a clear answer to the paradigm for food texture analysis". 1st International
reasons why a cellular structure gives shear bands Conference on Simulation in Food and Bio Industries
under loading. The solution will only come from (FOODSIM’2000), Nantes, 2000/06/26-27 In: D.
the definition of a theory of cellular tissue. The next Thiel (Ed.), Society for Computer Simulation
point is the capacity of model cellular tissue and International, Delft, Pays-Bas, 200-204
simulate the loading test. As different theories
begin to appear ( Niklas, 1989; Ouwerkerk, 1991; Roudot, A-C., 2002a. "Rhéologie et analyse de
Roudot, 2002b)(all based on complexity theories), texture des aliments" Tec&Doc, Paris, 76-77
the main problem now encountered is to define a
systemic numerical model able to verify empirical Roudot, A-C. 2002b. "Theory of plant tissue
interpretation. For that, different methods can be mechanics". J. Theoret. Biol. (submitted for
used (Roudot, 2000) but none is still sufficiently publication)
universal and easy to use, to be largely used by
food technologists. Subhash, G., Nemat-Nasser, S., Mehrabadi, M.M.and
H.M. Shodja.1991. "Experimental investigation of
fabric-stress relations in granular materials". Mech.
REFERENCES Mat. 11, 87-106

Birth, G.S. 1978. "The light scattering properties of Travers, T., Ammi, M., Bideau, A., Gervois, A.,
food". J. Food Sci., 43, 916-925 Messager J.C. and J.P. Troadec. 1987. "Uniaxial
compression of 2d packings of cylinders. Effects of
Cundall, P.A. and O.D.L. Strack. 1979. "A discrete weak disorder". Europhys. Lett., 4 (3), 329-332
numerical model for granular assemblies".
Géotechnique, 29 (1), 47-65 BIOGRAPHY

Gao Q. and R.E. Pitt. 1991. "Mechanics of A-C Roudot received his PhD in mechanics in 1985.
parenchyma tissue based on cell orientation and Then he worked as a researcher at the National
microstructure" Trans. ASAE, 34 (1) : 232-238. Institute of Agronomic Research on the texture
analysis of vegetables and tissue structure modelling.
Gibson L.J. and M.F. Ashby. 1998. "Cellular solids, Since 1993, he has been senior lecturer at the School
structure and properties", Pergamon Press, Oxford, of Microbiology and Food Safety in Brest (France).
England, 212-240 His interests include image analysis in bacteriology
and toxicology and studies on biological cells
Le Moigne, J.L.1991."La modélisation des organisation.
systèmes complexes" Gauthier-Villars, Paris

Lee, Y.K. 1989. "Conditions for shear banding and

material instability in finite elastoplastic
deformation" Int. J. Plasticity, 5, 197-226
 SIMULATION-BASED ENVIRONMENTS FOR PRACTICING
DATA-COLLECTION SKILLS

Paul L. Darius Henk C.M. van der Knaap Kenneth M. Portier

Goele Massonnet
Eddie Schrevens
Katholieke Universiteit Leuven Unilever Research Vlaardingen University of Florida
Kasteelpark Arenberg 30 O. van Noortlaan 120 Gainesville, FL 32611-0339,
B3001 Leuven Belgium 3133 AT Vlaardingen, The USA
Netherlands

KEYWORDS VESTAC collection, previously developed for other

Data collection, design of experiments, statistical analysis, statistical purposes (Darius et al. 2000). The environments
web based training, VIRTEX applet library. are all realized in the form of JAVA applets, and are freely
available on the web at
ABSTRACT http://www.kuleuven.ac.be/ucs/virtex/ . They can be run on
any computer with a web browser that supports JAVA. Due
A set of web based simulation applets (called VIRTEX) is to the safety requirements of applets, they cannot read or
described. Each applet mimics a real situation of interest, write data from or to the user’s hard disk. Yet data can be
and invites the user to collect data to answer a research exchanged with other programs through cut and paste.
question. The data are noisy and generated by an underlying This paper describes some applets that are already available.
realistic model, invisible to the user. Once the data are Their level ranges from fairly simple to quite complicated.
collected, they can be transferred to a standard statistical The description given here is short and introductory. A
package. The user can train his/her data collection skills by detailed description can be found in each applet’s help
relating the quality of the statistical results obtained to the menu.
data collection strategy used.
SOME SAMPLING APPLETS
INTRODUCTION
The first applet introduces some relatively simple sampling
Engineers often have to make decisions based on noisy problems. There are three versions, which preferably should
data. Sometimes the data are already available (e.g. be used in sequence. The user sees a window, showing part
routinely measured production or quality data), sometimes of a large (50x100) array of candies with different colors.
they have to be collected first (e.g. problem solving in a The purpose is to get an idea of the percentage of candies
research lab, fine-tuning of a pilot plant). In the latter case, that are colored red (fig 1).
there is a vast range of situations about which data could be
collected, but only time and money to explore a few.
Efficient data collection (i.e. experimental design) is an
important skill, but there is typically little opportunity to get
experience. Textbooks introduce standard general-purpose
designs, and then proceed with the analysis of data already
collected.
Here we explore another approach: for some situations of
practical interest, software environments were developed
that mimic a real situation of interest. Data can be easily
collected, but this can be done in so many ways that before
doing so, many nontrivial decisions must be taken. Once
the data are collected, they can be transferred to a statistical
software package, and the user can relate the quality of the
analysis results to the data collection strategy used.

THE VIRTEX LIBRARY Fig 1. The candy applet, version 1.

When the user moves the mouse over a candy, its

Some simulation-based environments have been developed,
coordinates appear. With a left click, a candy is selected and
others are still in development. The collection is called
VIRTEX (VIRTual EXperiments), and is a companion to the
the counters for the number of candies selected and the (one “week” is about 3 minutes of real time) to complete the
number of red candies are incremented. experiment. Each experimental run takes a number of “days”
In the first version, the colors are (at first sight) evenly before the results are visible.
distributed over the window. In the second version however,
patches are clearly visible where one color occurs more
frequently.
Since there are so many candies, simply counting the
number of reds is not an option. The user will have to use
some form of sampling. Hence he/she will have to decide on
a sample size and a sampling scheme. In the first version, the
colors are distributed randomly and any form of selection
will do. In the second version the sampling must be done
much more carefully. The user can then choose for random
or stratified sampling, but will discover that each of those
requires careful planning before it becomes operational.
In the third version, the candies are wrapped in black and the
color of a candy is only revealed after it has been included in
the sample.

The second applet represents an environmental sampling

problem. The user sees an aerial photograph of a section of a
railway track (fig 2). It is explained that for many years a Fig 3. The factory applet.
diesel train uses this place to take a halt. Hence diesel
pollution of the site is suspected, and should be investigated. The production plant appears at the right hand side. The
current settings for the parameters temperature, time and
concentration are shown, as well as the current mean yield
per batch (averaged over a large number of batches). The left
hand side shows the pilot plant. This is a scaled down
version of the production process with (hopefully) similar
characteristics. The raw material for the pilot plant is stored
in a tank, which can contain enough material for 10 trials.
When the tank is empty (or upon request of the user) it will
be refilled, but the new raw material may have slightly
different characteristics.
The user sets up an experiment by clicking on the
“Experiment” button, which opens a new window. He/she
can fill in the parameter settings for any number of runs (as
well as instructions for handling the tank), or use an external
package to generate a design and paste it in. After hitting the
“Perform” button and waiting an appropriate amount of
“time”, the results become available in the “History”
Fig 2. The environmental sampling applet. window. This also shows the results of all previous
experiments. Data from this window can be copied and
The user is invited to take a set of soil samples (by clicking pasted into an external program for statistical analysis.
the mouse on the selected spot). Since there is a limited The user now has to decide whether sufficiently promising
budget and analysis is expensive, at most 20 samples can be new parameter settings have been found. If there are, the
taken. The results of the analyses are shown in a separate production plant can be halted and the new parameter
window, from where they can be cut and pasted into e.g. a settings installed. The resulting new average yield becomes
spreadsheet or a statistical package. available after 6 “weeks”.
Again the user is confronted with a sampling problem, but The “Profit” window gives an overview of the current
now simple random sampling will not work. The user will situation of costs and benefits. There are initial costs, a fixed
have to set up a sampling strategy that takes the context (the cost per experimental run with the pilot plant, and a fixed
layout of the site, the fact that part of it is covered by water cost each time the production plant is halted to change the
and hence cannot be sampled,..) is taken into account. settings. Benefits are computed from the change in yield in
the production plant, multiplied by the number of weeks
THE FACTORY APPLET these changed settings were operational. At the end of the 39
weeks, the balance should be positive and as large as
In the factory applet, the user has to experiment with a pilot possible.
plant in order to find optimized settings for the parameters of This applet allows the user to get some hands on experience
a production process. The experiment runs in “real” time: with several important concepts. All types of designs
when the start button is hit, time starts to run as shown in the discussed in Response Surface Methodology courses can be
top portion of the screen (fig 3). The user has 39 “weeks” tried out. The tank-to-tank variability has to be dealt with.
Moreover the user is confronted with two other problems, columns) to account for the heat gradient. The color of a
for which textbooks offer no immediate solution. The first circle shows which treatment was applied, the color of the
problem is: when are the pilot plant results sufficiently pie-part to which group (column) it belongs. A copy/paste
convincing to justify the cost for the production plant mechanism is available to save a current setup to disk and to
settings? The second problem deals with the overall strategy: restore a saved setup.
should one perform one large experiment and then decide,
do many very small experiments with a possibility to make
an adjustment each time, or do something in between ?

THE GREENHOUSE APPLET

The greenhouse applet requires the user to set up an

experiment with (tomato) plants in a greenhouse. To
succeed, the user has to deal with the problem of selecting
appropriate levels for a treatment variable, and with the
many problems caused by diversity in raw material and
experimental circumstances.
The purpose of the experiment is to find the optimal dose of
a new fertilizer. At the start, a set of 144 young tomato
plants is available (12 trays of 3x4 plants). These are the
green circles on the right hand side in fig 4. The young
plants are not all the same: the initial weight of each is
shown on the green circle. The user has to select some (or Fig 5. The greenhouse applet: a simple experimental setup
all) of the plants for the experiment and place them on the
greenhouse tablet in the middle. The tablet is bordered on When the plants have been properly placed, the user should
the left and the right with central heating devices (the thick select the time period for the plants to grow with the buttons
black vertical lines). The tablet is also lighted by four special at the bottom. When the “Grow” button is hit, the growth of
light bulbs. The resulting pattern can be seen on the screen. each plant is simulated on an hour-to-hour basis, and the
Consequently the positions on the tablet are not identical, final weight (along with all the other variables) is available
and the differences in light and heat can be expected to have in the “Output” window, from where it can be copied and
an effect on the growth. pasted into an external program for statistical analysis. The
growth simulation uses an adapted version of TOMGRO, a
well-known growth model for tomatoes (Jones et al. 1991),
as well as standard climatic data for the time period selected.
This applet allows the user to get comparative experience
with almost all classical designs (Box et al. 1978,
Montgomery 1997): completely randomized, complete or
incomplete block, Latin Square, etc. There is also ample
opportunity to invent and use new setups, made to
accommodate the specific features of the greenhouse
situation.

CONCLUSION

Since the applets were only recently developed, we have

little user experience to report on. However, the applets
show that current computer technology allows the creation
of special purpose, accessible and rich environments that
Fig 4. The greenhouse applet: initial view have, in the area of data collection and analysis, the potential
to give learning experiences well beyond those of traditional
The user also has to decide which plant gets which amount textbook exercises.
of fertilizer. Defining and selecting doses, and applying
doses to plants is done with simple mouse operations. REFERENCES
To account for the difference in locations and initial weights,
“grouping” factors can be defined. This is done on the left Box, G.E.P., W.G. Hunter, and J.S. Hunter. 1978. Statistics for
Experimenters. Wiley, New York.
side of the screen. Defining levels and assigning them to
Darius, P.L., J-P. Ottoy, A. Solomin, O. Thas, B. Raeymaekers, S.
plants is similar to the dose assignment mechanism. Michiels. “A Collection of Applets for Visualizing Statistical
Fig 5 shows a simple example of an experimental setup. 9 Concepts“ In Proceedings in Computational Statistics 2000,
plants were selected and placed on the tablet in 3 rows of 3. Bethlehem, J.G. and P.G.M. van der Heijden (eds). Physica
Here, only plants with initial weight 10 were selected and Verlag.
the light effect was ignored. Three doses (control, 4 and 8) Montgomery, D.C. 1997. Design and Analysis of Experiments.
were selected, and the plants grouped in 3 blocks (the Wiley, New York.
Jones, J., Dayan E., Allen L.H., Van Keulen, H. And Challa, H.
1991. A dynamic tomato growth and yield model (TOMGRO).
Trans ASAE. 34(2), 663-672
 MODELLING FOOD PRODUCTS VARIABILITY USING NON-LINEAR MIXED EFFECTS
MODELS

Julio C. Montañez Jeffrey K. Brecht

Fernanda A.R. Oliveira Department of Horticultural Sciences
Department of Process Engineering University of Florida, 1217 Fifield Hall
University College Cork, Ireland Gainesville FL 32611-0690, USA
E-mail: [email protected]; E-mail: [email protected]
[email protected]

Jesús M. Frías Rocio Luna

School of Food Science & Environmental Health Centro Tecnologico Nacional de la Conserva y Alimentacion
Dublin Institute of Technology, Cathal Brugha St. (C.T.C.), c/ Concordia s/n 30500 Molina de Segura
Dublin 1, Ireland Murcia, España
E-mail: [email protected] E-mail: [email protected]

Susana C. Fonseca Pablo Fernandez

Escola Superior de Biotecnologia Universidad Politecnica de Cartagena, E.T.S.I. Agronomica,
Universidade Católica Portuguesa Dpto. de Ingenieria de Alimentos y del Equipamiento Agricola,
Rua Dr. António Bernardino de Almeida Area de Tecnologia de Alimentos, Pº Alfonso XIII Nº 48
4200-072 Porto, Portugal 30203 Cartagena, España
E-mail: [email protected] E-mail: [email protected]

KEYWORDS The simplest case where a mixed effects model might be

C. sporogenes, modified atmosphere packaging, microbial considered is when there are two stages at sampling. In the first
thermal death kinetics, non-linear mixed effects, respiration stage units are selected at random from a population, and in the
rate, variability. second stage a number of measurements is made on each unit
sampled in the first stage (Pinheiro and Bates 2000).
ABSTRACT
Food processes are largely influenced by the natural Non linear mixed-effects models have recently received a
variability of the raw products. This variability constitutes a great deal of attention in the statistical literature because they
major constraint in process design, control and optimisation. are flexible and adequate for unbalanced repeated measurement
The development and building of mathematical models from data that occur in pharmacokinetics (e.g. drug intake), health
experimental data that describe not only the particular process sciences (study of human and animal medical experiments),
kinetics but also estimate different components of variance economics, biology, and other areas. In these models, some of
with know precision can help to establish more precise product the levels could be set as fixed effects (e.g. level of drug
and processing tolerances. Non-linear mixed effects models are concentration in the individual) but the individual variability is
a modelling technique widely used in the Agricultural and considered to be influential (Lindstron and Bates 1988).
Pharmaceutical research fields that can aid to estimate model
parameters and associated variances from experimental data. Model building techniques for mixed-effects models involve
The objective of this paper is to illustrate the potential of non- issues that do not have a parallel with conventional fixed
linear mixed effects models in food processing. Two cases are effects linear and non-linear models, such as: (i) use of
presented: analysis of the effect of batch-to-batch variation of structured random effects variance-covariance matrices (e.g.
respiration rate of Galega kale on the atmosphere reached diagonal matrices) to model variability associated to the
inside a modified atmosphere package, and assessment of parameters in the model; (ii) use of covariates to explain inter-
variability in thermal death kinetics of C. sporogenes. individual parameter variability; (iii) identification via
maximum likelihood tests of effects with an associated random
INTRODUCTION component. Using these tools mixed models can be built from
Non-linear mixed effects models have an intuitive appeal: the experimental data in a parsimonious way (Pinheiro and Bates
notion that individuals’ responses all follow a similar 1995). The modelling framework in pharmaceutical and
functional form, despite the inter-individual variability, with agricultural experiments, where mixed effect models have been
parameters that vary among individuals, seems to be successfully applied to estimate parameters and model
appropriate in many situations (Pinheiro and Bates 2000). Also, variances, resembles the case of kinetic modelling in food
this model framework may be viewed as an extension of the preservation and processing, where different random effects
standard non-linear modelling techniques, with the addition of may be relevant, often due to raw product variability.
inferential procedures to accommodate the different
components of the variance (within individuals and among One example is Modified Atmosphere Packaging (MAP) of
individuals) (Davidian and Giltinan 1996). fresh fruits and vegetables. This packaging technique consists
on the use of packaging materials with selective permeability
to O2 and CO2 that interplays with the natural product The R software (R 1.1.1, A Language and Environment
respiration rate to create a low O2/high CO2 atmosphere inside Copyright 2000) was used for statistical analysis. In particular,
the package, which extends the product shelf life. Engineering the nlme library was used for building the non-linear mixed
design of MAP is based on predictive models that include the effects models (Pinheiro and Bates 2000).
gas transfer trough the package and the O2 consumption/CO2
production due to respiration. Respiration rate often depends RESULTS AND DISCUSION
on batch and also varies between individual plants. For this
reason, it is important to estimate the variability of the product Case Study I - Design of MAP for Shredded Galega Kale
respiration rate from experimental data, and determine its
influence on the variability of gas composition inside the O2 consumption and CO2 production rates were modelled
package (Hertog 2002). using a Michaelis-Menten type equation, with uncompetitive
CO2 inhibition and model constants increasing exponentially
Another example is the determination of microbial death with temperature (Fonseca et al. 2002). The respiratory
rates, which provides the basis for the calculation of quotient (RQ) was considered to be constant in the range of
sterilisation values associated with a certain thermal process temperature and gas composition tested. To describe the
and/or for process design and optimisation. In this case, it variability among different product batches, the following
would be important to investigate if stochasticity due to batch- assumptions were made: (i) intra-batch variance is described by
to-batch variation can be significant, and if this variability can a power variance model; (ii) inter-batch variance is due to a
be associated with a particular sterility parameter (D or z- mixed additive effect with a normal distribution of the model
values) in order to design thermal processes that maximise constants on the respiration rate; (iii) the effect of season,
safety. harvest, etc. on variance is negligible. The first assumption was
confirmed by preliminary analysis of the data and agrees with
The main objective of this paper is to illustrate the potential of the hypothesis of proportional errors in the experimental
non-linear mixed effects models in food processing and measurements.
preservation, addressing in particular the two examples above
mentioned. The respiration rate of a particular sample from batch i taken
at time j, based on the above assumptions, is given by:

MATERIALS AND METHODS e ln(α i ) ⋅ O2ij

RO2 = + ε ij where ε ij ~ ℵ(0, σ 2 | RO2 |2δ ) (1)
ij
 CO2  ij

Case Study I - Design of MAP for Shredded Galega Kale e ln(φi ) + O2ij 1 + ln(γ i ij) 

 e 
Our study is based on data earlier reported in literature
(Fonseca et al. 2002). Shredded Galega kale is a traditional RCO2 = RQij ⋅ RO2 ij (2)
fresh-cut vegetable in Portugal. It is thinly shredded and
where:
consumed in a soup. O2 consumption and CO2 production rates
of shredded Galega kale were determined in the reported work.
The storage temperatures used were 1, 5, 10, 15 and 20°C. The ln(αi ) = ln(α 0 ) + α1T + bln(αi ) b ln( α i ) ~ ℵ ( 0 , σ ln( α ) )
atmospheres tested were all combinations of 1, 5 and 10% v/v ln(φi ) = ln(φ0 ) + φ1T + bln(φi )φ b ln( φ i ) ~ ℵ ( 0 , σ ln( φ ) )
O2 plus 0, 10 and 20% v/v CO2 with the balance being N2, as ln(γ i ) = ln(γ 0 ) + γ 1T + bln(γ i ) b ln( γ i ) ~ ℵ ( 0 , σ ln( γ ) )
well as ambient air.
RQi = RQ + bRQi b RQ i ~ ℵ( 0 , σ 2
RQ )
Case Study II – Thermal Degradation of C. Sporogenes
Spores The average population constants (RQ, ln (αo), ln (φo), ln (γo),
α1, φ1 and γ1) where estimated together with the inter-batch
Microorganisms were obtained from the National Culture variability parameters (σRQ, σln(α), σln(φ) and σln(γ)) and with he
type Collection (Spain). Sporulation, synchronisation, and
intra-batch variability parameters (σ and δ). In order to build a
collection of spores is detailed in Fernandez et al. (1996). 0.5
parsimonious model, the following procedure was followed: (i)
ml of the mother suspension were placed in a sterile tube, heat
a model with all the mixed effects σRQ, σα, σφ, σγ was built; (ii)
shocked at 83 °C during 13 min in a thermostatic bath (Polistat
nested models with all the combinations of three, two and one
CC2, Huber, Germany) and cooled in an ice bath. The
mixed effects were built; (iii) the likelihood ratio test was
suspension was then diluted to 350 ml using phosphate buffer
applied to all the models in order to select the nested model
(pH=7.0) and placed in 10 µl glass capillaries where they were with a minimum set of effects that would be statistically
subjected to a thermal treatment at the desired temperature different from the full model accounting for all mixed effects.
using a thermostatic bath (Stern and Proctor 1954). The This procedure led to the rejection of the random effect
temperatures tested where 110, 113, 115, 118 and 121ºC, with
associated to the ln(α) parameter (σln(α)), as shown in Table 1.
sampling times between 24 minutes and 30 seconds depending
The estimates of the constants of the best performing model are
on the temperature. Experiments were replicated with 6
presented in Tables 2 and 3. Figure 1 shows an example of the
different batches at least. Enumeration of the cfu/mL was
fit of the model to the experimental data.
performed using the media from Grischy et al. (1983).
 Table 1: Information Criteria Parameters for Selection of a 0 5 10 15 20
Non-Linear Mixed Effects Model to Predict Respiration Rate CO2 CO2
of Shredded Galega Kale
o
o o
o
Model df LogLik Test p-value
o o
o o
1 - No inter-batch o
o o
o
o o o o
o
o
9 -2676 o
oo
o
o
o
o
o
o
variability

Respiration rate (mL/kg h)

o
o o o o
o o
o
o o
o
o o
2 - Nested model o
o o
o
oo
o
o o
o
oo o
neglecting the 12 -2654 1vs.2 <0.001 o
o
o
oo
o
o
o
effect in ln(α) CO2 CO2 CO2
3 - Full model 13 -2654 2vs.3 0.97 o
o
o
o
o
o
o
o o o
o o
150 o
o o o
o o
o o o o
o o o o
Table 2: Estimates of the Michaelis-Menten Model Constants o o
o
o
o
o
o o o o
for the Non-Linear Mixed Effects Model with Three Random 100
o
o
o
o
o
o
o
o
o
o
o
o
o
Effects o
o o
o
o

o o o
o
50 o o o
o
o
o o
o o
o
o o o
Parameter Estimate ± 95% CI o

RQ 0.97 ± 0.05 0 5 10 15 20 0 5 10 15 20

ln (αo) 2.74 ± 0.09 Oxygen (%)

α1 0.132 ± 0.007 Figure 1: Fit of the Non-Linear Michaelis-Menten Mixed

ln (φo) -1.6 ± 0.5 Effects Model with Three Random Effects to the Experimental
φ1 0.16 ± 0.03 Data at 20 °C (different colours refer to different batches)
ln (γo) 2.7 ± 0.4
It can be observed that the distribution in terms of CO2
γ1 0.04 ± 0.02 composition is close to normal, with 5.5% of the packages
reaching a steady state gas composition outside the
Table 3: Estimates of the Parameters Accounting for recommended range. Regarding O2, the distribution curve is
Variability for the Non-linear Mixed Effects Model with Three skewed to the right but only 0.3% of the packages show a
Random Effects concentration above the maximum recommended value (2%).

Parameter Low Estimate High

95% CI 95% CI 20

σRQ 0.05 0.09 0.145

σln(γ) 0.27 0.42 0.67
15
P ercent of Total

σln(φ) 0.03 0.11 0.46 10

δ 0.62 0.69 0.75

5

σ 0.6 0.7 0.9

0

A perforation-mediated MA package was designed, using 0 .1 4 0 .1 6 0 .1 8 0 .2

the average population constants (Table 1) in the following set C a rb o n d io x id e [ % ]

of equations, which describe the gas mass balance at

equilibrium (Fonseca et al. 1999):
ROeq2
20

M (3)
y eq
O2 −y e
O2 =− ×
3.6 × 10 N × a × D b × Lc
7 15
P erc ent of Total

10
eq
RCO M RQ
eq
yCO − yCO
e
= 2
× × (4)
β
5
2 2
3.6 × 10 N × a × D × L
7 b c

The design target was to reach a steady state gas composition 0

0 .0 1 0 .0 1 2 0 .0 1 4 0 .0 1 6 0 .0 1 8 0 .0 2 0 .0 2 2

corresponding to the central point of the recommended gas

composition for extending the shelf life of Galega kale: C a r b o n d i o x id e [ % ]

5%<CO2< 20% and 1%<CO2< 2% (Fonseca et al. 1999). Then,

following the approach of Talasila and Cameron (1995) and Figure 2: Effect of Batch-to-Batch Respiration Rate Variability
using the estimated inter-variability of the constants of the on the Gas Concentration Inside a MAP Containing Shredded
respiration rate model (σRQ, σln(φ) and σln(γ) − Table 2), a Monte Galega Kale (the full vertical lines represent the limits of
Carlo simulation was performed to estimate the steady state gas recommended gas composition and the vertical broken lines
composition inside 3,000 packages. The results are shown in represent the target gas concentration)
Figure 2.
Case Study II – Thermal Degradation of C. Sporogenes fixed Exp
Spores 0 2 4 6 8 10 12

7 8 9
The thermal death kinetics of C. Sporogenes spores was 5
described by the Bigelow model, coupled to a mixed additive 4.5
effect with a normal distribution of the model constants, Dr and 4
z. For batch i at time j: 3.5
T − Tr
n ij t ij 3

log = − 10 zi
+ ε ij (5) 4 5 6
n oi D ri 5

Log(ufc/ml)
4.5
where:
4
D ri = D r + b D r bD r ~ ℵ(0,σ D r ) 3.5
i i
3
zi = z + bzi b z i ~ ℵ(0,σ z ) 1 2 3
5
The average population constants where estimated together
4.5
with the inter-batch variability parameters (σDr, σz). Different
4
models were tested, considering no random effects and random
3.5
effects in Dr and/or z. The information criteria parameters
3
(Table 4) show that mixed effects clearly improve the model,
0 2 4 6 8 10 12 0 2 4 6 8 10 12
particularly when the random effect in Dr is considered. Time (min)
Actually, when random effects in Dr and z are estimated
simultaneously, the latter is negligible (or unidentifiable with Figure 3: Fit of the Non-Linear Mixed Effects Bigelow Model
the experimental design used). Thus, the model that best fits with Random Effects in Dr to the Experimental Data at 100 °C
the data with the least number of parameters is the one that (the blue line represents the fit prediction using the average
accounts for random effects in Dr only. population parameters and the purple line represents best
individual predictions considering a random effect in Dr; the
Table 4: Information Criteria Parameters for Selection of a numbers in each plot refer to different experiments)
Non-Linear Mixed Effects Model to Predict Thermal
Degradation of C. Sporogenes Spores CONCLUSIONS

Model df AIC logLik Test L Ratio p-value The two case studies presented in this work show that non-
1 - No inter
batch 1 219 -108
linear mixed effects modelling is a statistical tool that can be
variability applied in food studies, to build stochastic models that account
2 - Random for product variability. These models may then be used to
4 212 -102 1vs2 13. 0.0041
effect in z assess the impact of the raw materials variability on the final
3 - Random product and to identify sources of variability, and eventually
4 162 -77.
effect in Dr
4 - Random the processes may be optimised to yield maximum product
effect in Dr 5 164 -77. 3vs4 0.003 0.96 quality whilst minimising variability.
and z
ACKNOWLEDGEMENTS
This finding may support the hypothesis often assumed in
literature that microbial thermal degradation rate (measured by Author J.C. Montañez is grateful to CONACyT (Mexico) for
the constant Dr) is quite dependent on endogenous and financial support.
exogenous conditions, whereas its sensitivity to temperature
changes (measured by the constant z) is characteristic of the NOMENCLATURE
particular specie, genus and strain of the bacteria. Thus, the z
parameter could therefore be considered as a “fingerprint” of a gas exchange model constant (6.42x10-6 m2.148 s-1)
the microorganism in question. The constants of this model are b additive mixed effect or gas exchange model
summarised in Table 5 and an example of the fit of the model constant (1.45)
to the experimental data is shown in Figure 3. c gas exchange model constant (-0.598)
D perforation diameter (m)
Table 5: Estimates of the Bigelow Model Constants for the Dr time required for a 90% reduction at Tr (min)
Non-Linear Mixed Effects Model with Random Effects in Dr L perforation length (m)
M product mass (kg)
Parameter Low 95% CI Estimate High 95% CI N number of perforations
Dr 0.85 0.94 1.03 n number of cell of C. sporogenes (log CFU ml-1)
no initial cell number of C. sporogenes (logCFU ml-1)
z 10.84 11.95 13.06
R CO 2 CO2 production rate (ml kg-1 h-1)
σDr 0.12 0.16 0.22
eq
σ 0.28 0.31 0.34 RCO 2
CO2 production rate at equilibrium (ml kg-1 h-1)
R O2 O2 consumption rate (ml kg-1 h-1)
AUTHORS BIOGRAPHY
ROeq2 O2 consumption rate at equilibrium (ml kg-1 h-1)
RQ respiratory quotient (dimensionless) JULIO CESAR MONTAÑEZ was born in Chihuahua, Mexico, and
T temperature (°C) obtained his undergraduate degree in Chemical Engineering at the
t time (s) University of Coahuila, Mexico. Currently he is a Ph.D. student at the
Department of Process Engineering, University College Cork, under
Tr reference temperature (121°C)
the supervision of Prof. Fernanda Oliveira and Dr. Jesús Frías. His
y eq
CO
CO2 concentration at equilibrium (% v/v) research work focuses on the engineering design of MAP systems for
2
fresh-cut produce.
y eCO CO2 external concentration (% v/v) http://www.ucc.ie/acad/departments/processeng/postgrad/montanezj.h
2 tml
y eq
O
O2 concentration at equilibrium (% v/v)
2 FERNANDA OLIVEIRA was born in Viana do Castelo, Portugal,
and obtained an undergraduate degree in Chemical Engineering in
y eO O2 external concentration (% v/v)
1985 at the Faculty of Engineering, University of Porto, Portugal, and
2

z slope of the Thermal Death Time graphic (oC) a PhD degree in Food Engineering in 1989 at University of Leeds,
ℵ Normal distribution function UK. She worked as Associate Professor and Associate Director at the
College of Biotechnology, Catholic University of Portugal, as
α, α1 Michaelis-Menten equation constant (ml kg-1 hr-1) Associate Professor and Vice-President of the Inter-University
α2, φ 2, γ2 Michaelis-Menten equation constant (°C-1) Institute of Macau, China, and as Senior Consultant and Director of
φ, φ1, γ, γ1 Michaelis-Menten equation constant (%v/v) the Portuguese Society for Innovation (Porto, Portugal) before moving
β permeability ratio (0.81) to Ireland, where she is currently Professor and Head of the
δ power index of intra-individual variance model Department of Process Engineering, at University College Cork.
http://www.ucc.ie/acad/departments/processeng/staff/oliveiraf.html
σ standard deviation
ε within-subject error term JESUS FRIAS was born in Tolosa (Spain) and obtained a degree in
Food Science and Technology in 1992 at Basque Country University
REFERENCES and a PhD. Degree in Biotecnhology at the Catholic University of
Portugal in 1998. He worked as Research Assistant at the ENSIA-
Davidian, M. and Giltinan, D. M. 1996. Nonlinear models for repeated INRA (France) and at the Department of Process Engineering at
measurement data. Chapman and Hall, London, UK. University College Cork before joining the Dublin Institute of
Fonseca, S.C.; Oliveira, F.A.R.; Brecht, J. K and Chau, K.V. 1999. Technology (Ireland), where he currently is Lecturer of Food
“Development of perforated-mediated modified atmosphere Chemical Analysis.
packaging for fresh-cut vegetables". In: Processing Foods: Quality
Optimisation and Process Assessment. F.A.R. Oliveira and J.C. SUSANA FONSECA was born in Porto, Portugal, and obtained an
Oliveira (eds.). CRC Press, United States. undergraduate degree (in 1993) and a PhD degree (in 2002) in Food
Fonseca, S.C.; Oliveira, F.A.R.; Frias, J.; Brecht, J.K. and Chau, K.V. Engineering at the College of Biotechnology, Catholic University of
2002. “Modelling respiration rate of shredded Galega kale for Portugal. She has worked as a teaching assistant at the College of
development of modified atmosphere packaging”. Journal of Food Biotechnology, Catholic University of Portugal, where she currently a
Engineering. In press. postdoctoral research fellow.
Hertog, M.L.A.T.M. 2002. “The impact of biological variation on
postharvest population dynamics”. Postharvest Biology and JEFFREY BRECHT obtained an undergraduate degree in Biology in
Technology. In press. 1979 at Whittier College, Californa, USA and a PhD degree in in
Lindstrom, J. J. and Bates, D. M. 1988. “Nonlinear mixed effects Plant Physiology in 1984 at University of California, Davis, USA. He
models for repeated measures data”. Biometrics 46, 673-687. worked as Research Assistant at Tri-Valley Growers, California, and
Pinheiro, J. C. and Bates, D. M. 1995. Model Building for Nonlinear as Graduate Research Assistant at the Department of Pomology,
Mixed-Effects Models Technical Report 91, Department of University of California, Davis, before moving to the Horticultural
Biostatistics, University of Wisconsin - Madison Sciences Department at the University of Florida where is currently
Pinheiro, J.J. and Bates, D.M. 2000. Mixed-Effects models in S and S- Professor of Postharvest Physiology and Horticulture.
PLUS. Statistics and Computing Series, Springer-Verlag, New http://www.hos.ufl.edu/jkbweb/
York, NY.
Stern, J.A. and Proctor, B.E. 1954. “A micro-method and apparatus ROCIO LUNA was born in Jerez de la Frontera (Spain) and obtained
for the multiple determination of rates of destruction of bacteria and an undergraduate degree in Veterinary in 1997 at the University of
bacterial spores subjected to heat”. Food Technology, 8: 139-143 Cordoba (Spain). She has been working since 1998 as analyst at the
Talasila, P.C. and Cameron, A.C. 1995. “Modelling frequency National Canning Technology Centre (Murcia, Spain) and started a
distribution of steady-state O2 partial pressures in modified- PhD. Programme in 1999, under the supervision of Prof. Pablo
atmosphere packages”. Journal of Food Process Engineering. Fernandez and Prof. Antonio Martinez (Univ. de Cartagena, Spain).
18:199.
Fernandez, P.S.; Ocio,M.J.; Rodrigo,F.; Rodrigo, M.; Martinez, A. PABLO FERNANDEZ is a lecturer at the Department of Food
1996. “Mathematical model for the combined effect of temperature Engineering and Agricultural Equipment of the Universidad
and pH on the thermal resistance of Bacillus stearothermophilus and Politecnica de Cartagena (Spain). He has published extensively in the
Clostridium sporogenes spores”. International-Journal-of-Food- area of microbial thermal destruction and mathematical modelling of
Microbiology, 32 (1/2) 225-233. spore inactivation.
Grischy, R.O.; Speck, R.V. and Adams, D.M. 1983. “New media for
enumeration and detection of Clostridium sporogenes (PA 3679)
spores”. Journal of Food Science, 48(5): 1466-1469.
 FOOD
ENGINEERING
CONTROL
 A NEW PROCESS CONTROL SYSTEM BASED ON PREDICTIVE KINETIC
MODELS FOR FOOD QUALITY AND OPERATING COSTS
Peter de Jong*, Erik Smit, Marc M.M. Vissers, Han Straatsma, Jan van de Wiel
Maykel Verschueren
Department of Processing, Quality & Safety Department of Industrial Automation & Control
NIZO food research Honeywell
PO Box 20, 6710 BA Ede, Laarderhoogtweg 18, 1101 EA Amsterdam,
Netherlands Netherlands
E-mail: [email protected]

KEYWORDS are implemented in the industrial R&D and production

Control systems, Agriculture, Global, Industrial Control, plants. A powerful method to realise this objective is
Computer Integrated Manufacturing. integration of these models into the process control
equipment. This ensures a continuous optimization of the
ABSTRACT production plant.
In recent studies some applications of model based
A new process control system for the food industry has applications are reported. Most of the models are of a black
been developed, called PREMIC (PREdictive Models box nature and increasingly based on neural networks. The
ready for Industrial Control). The system is focused on model-based control procedures are used for single unit-
predictable optimal product quality and minimal operations such as beer fermentation (Kurz et al. 2002,
operating costs. Majority of the models used are have a Corrieu et al. 2001), cane sugar evaporation (Benne et al.
white box nature and are based on reaction kinetics. The 2001) and extrusion (Haley and Mulvaney 2000). Trelea et
models predict the food properties (e.g. stability, shelf- al. (1998) used a more white box modeling approach and
life, nutritional value, texture, color) and the operation developed a physical model for optimal control of a batch
costs of production chains. An optimization algorithm refrigeration process. Six years earlier Straatsma et al.
converts this information together with actual process (1992) already developed a control system for spray drying
information (SCADA) to optimal set-points. The of dairy products based on physical models (mass and
estimated savings amount up to 10 – 50 % of the energy balances and sorption isotherms). However, non of
production costs. these control procedures applies a combination of physical
models and mechanistic (kinetic) models for prediction of
INTRODUCTION product properties such as protein denaturation degree,
enzyme activity and microbial contamination. In
A substantial amount of the knowledge and experience of contradiction with black box models, mechanistic kinetic
research institutes and universities never reach the food models have good predictive capabilities. In the case of
industry. The main bottleneck is the interpretation of the black box models such as neural networks every change in
scientific results. Thus the objective is to find a way to process configuration or product composition means a
benefit the food industry by application of these results in renewed necessary model training. Effects of food
practice. Experience with industrial projects shows that composition and small changes in process operation (e.g.
predictive kinetic models are an effective tool to apply changed piping, plate configuration heat exchangers) on
state-of-the-art knowledge in the factory (De Jong 1996, De the product properties are not included. Thus, for robust
Jong et al. 1999, De Jong et al. 2002a,b). Examples of control of food production plants it is essential to use a
these models are: mechanistic models.
1. Models to predict the concentration of micro-organisms In this paper a new process control system based on
in final products as a result of growth, adherence, predictive mechanistic (kinetic) models for food quality and
release and inactivation in process equipment; operating costs is presented. The system is called PREMIC
2. Kinetic models that predict the transformation of food (PREdictive Models for Industrial Control) and developed
components related to the food properties recognised in co-operation with Honeywell. The fundamentals of the
by the consumer (e.g. viscosity, colour, shelf-life); PREMIC system are described and some applications are
3. Predictive models to estimate the costs related to given.
process operation.
A number of these models (e.g. the kinetic models) are PREDICTIVE MODELS
mechanistic models based on reaction schemes describing
the product changes during processing. Contamination models
In order to increase the application of scientific results for
food production it is important that the models developed
The concentration of micro-organisms in food largely mL-1). However, the real contamination problems (108 cfu
depends on the adherence of micro-organisms to equipment mL-1) are due to adherence and growth on equipment walls.
walls (Bouman et al. 1982, De Jong et al. 2002a). In many
unit operations no contamination problems would occur if
bacteria would not adhere to equipment surfaces The
increase in the levels of bacteria in the product during
process operation is partly the result of growth in the
product, but the release of bacteria that have grown on the
walls of the equipment also plays a significant role. Growth
of micro-organisms in consumer food products leads to 10
8

spoilage and defects, e.g. excessive openness in cheese or 7

Production line (pasteuriser - buffertanks -
fermentor - centrifuge - cooler)
10
taste deviations of milk.

Bacteria count product (cfu/mL)

Raw milk (measured)
Processed whey (measured)
To model the contamination due to both adherence and 10
6 , , Prediction

growth of bacteria the following approach has been applied. 5

A production chain can be considered as a cascade of 10

no adherence
model reactors (process model). In each reactor, growth, 10
4

adherence and/or destruction of bacteria can occur

3

(transformation model). The basis of the proposed model is 10

formed by two mass balances: one balance for the wall on 10

2
no growth, no adherence

which bacteria may adhere and one for the liquid bulk
phase. The bacterial growth as a function of the operating
1
10
0 2 4 6 8 10
time t at position x at the wall in a tubular plug flow reactor Production time (h)
is defined by the transfer equation as:
d nw Figure 1: Concentration of Streptococcus thermophilus in
= µ T n w (1 - β ) + k a C (1) Product related to the Operating Time of Downstream
dt
Processing of Whey: Effect of the Local Adherence and
where nw is the wall coverage in cfu m-2, µT is the bacterial Growth in the Processing Equipment
growth rate at temperature T in s-1, β is the fraction of
generated bacteria which is released into the bulk and ka is
the adhesion constant in m s-1. The local bulk concentration
C in cfu m-3 at operating time t follows from the component Transformation of food components
(bacteria) balance:
2 During processing, food behaves like a complex reaction
dC πD
= (β µ T n w - k a C ) + πD (µ T − k d ,T ) C (2) system. A large number of chemical, physical and
dx φ 4φ biochemical reactions take place. Some of these
transformations are important because they change those
where φ is the milk flow in m3 s-1, kd the inactivation product characteristics that are easily recognized by a
(destruction) constant at temperature T in s-1 and D the consumer. Examples are: inactivation of enzymes,
hydraulic diameter of the reactor in m. In the case of a tank denaturation of proteins, loss of nutrients and formation of
reactor the concentration is independent of position x: new components. In general, most of these reactions can be
4φ described by nth order reactions (single or consecutive):
dC
= 2
(C in − C ) + 4 (β µ T n w - k a C ) + C (µ T − k d ,T ) dC1 dC1
d t πD L D = k n C1 n , = k n C1 n − k m C 2 m , etc. (4)
(3) dt dt
where L is the liquid level in the tank in m. Since the wall
temperature is a function of the position in the heat A large amount of kinetic data of several food components
exchanger (x) the differential equations 1, 2 and 3 have to has been collected. Models available (De Jong 1996) are
be solved numerically in parallel. the protein fouling model, models for bacterial spores
In Figure 1 the effect of adherence of micro-organisms on (quality related), models for vitamins and Maillard products
equipment walls on product contamination (industrial (flavour related), models for enzyme inactivation (shelf-life
conditions) is demonstrated. The lines represent the model related), models for protein denaturation and aggregation
calculations after determination of the process configuration (texture related) and models for protein breakdown.
and main process conditions in the factory and before
taking the samples. It is obvious that in the case that no Operating costs
bacteria would adhere to the equipment walls, the
contamination would be governed by growth in the liquid The operating costs of many food production chains
phase. The downstream processing of whey consists of primarily depend on microbial and physical fouling of the
equipment (e.g. vessels) with sufficient residence time in equipment. In general, process operating times at relatively
which the bacteria can grow in the liquid phase (>104 cfu low temperatures (< 70 °C) are due to adherence and
growth of bacteria. The operating time of equipment at
temperatures above 80 °C is determined largely by the
deposition of protein and minerals. Physical fouling due to composition of the raw materials the models can predict the
deposition of proteins and minerals can be described by an state of the process (e.g. amount of fouling, biofilm
adsorption reaction (De Jong et al. 1992): thickness, energy usage) and the state of product (degree of
J x,t = k ′′C x ,t 1.2 (5) contamination, stability, texture). This means that the
process can be controlled on product specifications instead
of process conditions. By adding cost related models the
where Jx,t is the local flux (kg m-2 s-1) of food components to system can continuously optimize the production process
the wall, k” the adsorption rate constant of proteins (m1.6 with respect to the product quality and the production costs.
kg-0.2 s-1) and Cx,t the local bulk concentration of the key The PREMIC system aims to provide this control approach.
component related to fouling (e.g. whey proteins). The In Figure 3 the PREMIC system is shown in general terms.
amount of fouling is obtained by integrating the flux over Based on process design (e.g. dimensions, apparatus
operating time and surface area of the equipment walls. configuration, in-line measured process conditions) the
This amount can be related to the costs due to cleaning, temperature time history of the product is calculated with
change-over (rinsing losses), depreciation, energy, the process model. Together with the given composition
operator, pollution and product losses (De Jong 1996). this information is used to predict the product properties
As an example, in Figure 2 the operating costs (energy, using the kinetic product models. Besides the operating
emission, product losses) of a sterilisation plant as a costs per ton product are estimated based on the fouling and
function of heating rate and heating temperature are shown. contamination models. The predicted product properties are
In the case of heating rates higher than 10 °C s-1 the compared with the given desired product properties. In the
operating costs increase with heating temperature, mainly optimization module the production process set points are
due to increased fouling rates. At extreme heating rates (> optimized to meet the desired product properties as close as
100 °C s-1) the operating costs become lower due to the fact possible with minimum operating costs.
that the residence time is very short. This results in a lower In the ideal situation the process is controlled based on the
concentration of the key component for fouling (e.g. desired product specifications and minimum operating
denatured proteins) and according to equation (5) in less costs. The system corrects itself automatically when:
fouling. • Fouling changes the temperature-time history the
product
• The product specifications change
• The composition of the raw materials change
• Disturbances occur (temperature changes, flow
instabilities)

7
raw materials
6
ingredients
5
)

120
Heating temperature (°C
ex

4 110
Cost ind

PROCESS PRODUCT
3 100
2 90
1 80
0 70 PREMIC sensors
-1
0
1 60
2
Log(hea 3
ting rate
in °C s -1)

Figure 2: Operating Costs (energy, emission, product process design

losses) of a Sterilisation Plant as a function of Heating desired product

Rate and Heating Temperature Process Product
properties
model models
composition

CONTROL APPROACH

General description of PREMIC

4 ,5

4 ,4
n
functio

Optimization
Objective

4 ,3

optimal 4 ,2

4 ,1

Predictive models as described turn out to be an effective setpoints

7
7 8 ,5 6
7 8 ,6 5
)
7 8 ,7 4 (s
3 e
T em 7 8 ,8
2 tim
p er c e
7 8 ,9
at
ur
1 en
e (° 7 9 ,00 id
C) es

tool to translate scientific knowledge to the food factories.

R

The most effective way to ensure the benefits of using

predictive models would be to integrate them into process
control systems. Based on actual process data and the
 Figure 3: Schematic Representation of the PREMIC
Control System and the production time per run:
trun = t if (CS .thermophilus > 0.0001 cfu mL-1) (10)

Optimization algorithm
10 °C 65 °C

An essential part of the PREMIC system is the optimization regenerative section (I) holder heater
(II) (III)
module. The main control factors for product and process 21 °C
regenerative section (V)
76 °C
optimization are the temperature-time relationship and the 5s
holder
configuration of the processing equipment. In order to cooler (IV)
determine the optimal values of the control variables, a (VI) 10 s
general objective function is used: 4 °C

F (u , x) = α cquality (u, x ) + β coperation (u ) (6)

Figure 4: Example Process
where u is a vector of process control variables (e.g.
temperature, flow) and x is a vector of desired product The weight factor αi is introduced to avoid trivial and
properties related to food quality and safety. The value of undesired solutions. The chosen values of the weight
cquality depends on the outcomes of the predictive models for factors are determined by the relative importance of the
contamination and transformation of food components. different product properties. However, since the
coperation is related to the operating costs. The optimal relationship between the weight factor values and the
configuration and operation of a production chain is optimization results are not clear on forehand, the
achieved by minimization of the objective function. To determination of the weight factor value is an iterative
avoid trivial and undesired solutions, the weight factors α process in consultation with industrial experts.
and β are introduced. These weight factors give the relative In this case the control variables (u) are limited to two, the
importance of each term of the objective function. For heating temperature and the residence time at this
example, too high a value of β may result in a very clean temperature in de second holder section. With two control
and cheap production process but an inferior product variables surface plots can present the results of the
quality. computer model simulations. Figure 5 shows the results of
the objective function evaluations. The optimal setpoints are
Example a heating temperature of 78.7 °C and a residence time of 3
seconds. Compared to the initial preliminary design (10 s,
To illustrate the application of the described procedure for 76 °C) the operating costs could be decreased by 14 %. At
optimizing food production chains the following case study an annual production of 4700 h this means an estimated
has been performed. A heating process with a capacity of cost saving of Euro 58,000. More details are given by De
40 tons skim milk per hour, consists of a regenerative Jong et al. (2002b). Detailed studies show that cost
section, a heating section and 2 holders sections and a reduction of more than 50 % are feasible (De Jong 1996).
cooler. In Figure 4 the scheme of the process is shown with
some preliminary temperatures and residence times. In
order to have a process model the equipment is transformed
to a cascade of model reactors. The objective function is
defined as: 4,4

2
3 x − x (u ) 
F (u , x) = ∑ α i  i , des i  + Fcos t (7)
 
ction

i =1  xi , des 
4,3
Objective fun

where 4,2

coperation ⋅ toperation + csolids ⋅ t production ⋅ ∫∫ J x ,t dt dx

x ,t
Fcos t =
4,1

t production ⋅ φ 78,5 6
7

78,6 5 )
(s
(8) 78,7 4
tim
e
Te 3 e
where αi is weight factor for the relative importance of mp
era
tur
78,8
2
si
de
nc
78,9 e
product property xi, xi,des is the desired product property, u e(
°C
) 79,00
1 R

is de set of control variables, c are costs, t is time and the

integral term is the total amount of deposits after 1 h of
production and φ is the capacity of the process in tons per Figure 5: Evaluation of the Objective Function
hour and:
toperation ⋅ trun
t production = (9)
trun + tcleaning
The PREMIC system using this optimization procedure for cleaning in food processing ’98, D.I. Wilson; P.J. Fryer and
the determination of the optimal set-points. Already with a A.P.M. Hasting (Eds.). EUR 18804, Cambridge, 39-46.
PC it is possible to repeat this procedure every 10 seconds. De Jong, P.; M.C. te Giffel and E.A. Kiezebrink. 2002a.
“Prediction of the adherence, growth and release of
microorganisms in production chains”. International Journal of
Food Microbiology 74, 13-25.
De Jong, P.; M.C. te Giffel; H. Straatsma and M.M.M. Vissers.
2002b. “Reduction of fouling and contamination by predictive
CONCLUSIONS kinetic models”. International Dairy Journal 12, 285-292.
Haley, T.A. and S.J. Mulvaney. 2000. “On-line system
The ultimate solution for the transfer of (scientific) research identification and control design of an extrusion cooking
process”. Food Control 11, 121-129.
results towards industrial practice is the application of Kurz, T.; M. Fellner; T. Becker and A. Delgado. 2002.
model-based control approaches. These models should be “Observation and control of the beer fermentation using
based on mechanistic (kinetic) modeling approaches as cognitive methods”. Journal of the Institute of Brewing 107,
much as possible and include a adequate process description 241-252.
(model). With the development of the PREMIC control Straatsma, J; G. van Houwelingen; A.P. Meulman; A.E.
system and the availability of a relatively large number of Steenbergen. 1992. “Dryspec2: a computer model of a two-
product models this advanced type of model-based control stage dryer”. Journal of the Society of Dairy Technology 44,
107-111.
will be realized in the food industry. As a consequence the Trelea, I.C.; G. Alvarez and G. Trystram. 1998. “Nonlinear
cleaning times will be shortened, the product quality will be predictive optimal control of a batch refrigeration process”.
more constant and process development time will be Journal of Food Process Engineering 21, 1-32.
minimized. Detailed studies show that the savings in
operating costs are estimated at more than 10-50 % .
ACKNOWLEDGEMENT
REFERENCES
Benne, M.; B. Grondin-Perez; J.P. Chabriat and P. Herve. 2001. This work has been supported by Senter, the Netherlands.
“Artificial neural networks for modelling and predictive control
of an industrial evaporation process”. Journal of Food BIOGRAPHY
Engineering 46, 227-234
Bouman, S.; D.B. Lund; F.M. Driessen and D.G. Schmidt. 1982. PETER DE JONG was born in Utrecht, the Netherlands
“Growth of thermoresistant streptococci and deposition of milk and studied chemical engineering (B.Sc.) in The Hague and
constituents on plates of heat exchangers during long operating obtained his Ph.D. degree at the University of Delft on
times”. Journal of Food Protection 45, 806-812.
modeling and optimization of thermal processes in the dairy
Corrieu, G.; I.C. Trelea and B. Perret. 2001. “On-line estimation
and prediction of density and ethanol evolution in the industry. In 1986 he went to R&D department of Campina, a
brewery”. Technical Quarterly of Masters Brewers’ major international dairy company. In 1987 he started to
Association of the Americas 37, 173-181. work at NIZO food research as a process engineer. In the
De Jong, P.; S. Bouman and H.J.L.J. van der Linden. 1992. nineties together with Han Straatsma he initiated a modeling
“Fouling of heat treatment equipment in relation to the group for applications in the food industry. At this moment
denaturation of -lactoglobulin.” Journal of the Society of he is responsible for all the research activities on processing,
Dairy Technology 45, 3-8. modeling and control within NIZO food research and is the
De Jong, P. 1996. “Modelling and optimization of thermal head of a group of 13 scientific workers.
processes in the dairy industry”. ISBN 90-9009034-7,
Montfoort, the Netherlands.
De Jong, P.; H.C. van der Horst and R. Waalewijn. 1999.
“Reduction of protein and mineral fouling”. In Fouling and
 TRACKING IN GENERALISED PREDICTIVE CONTROL USING THE
P TRANSFER FUNCTION
Ben Griffin* and Conor Downing
Department of Electronic Engineering
Cork Institute of Technology
Cork, Ireland
*
E-mail: [email protected]

KEYWORDS specified as a polynomial when used in model-following

GPC, Model-following control, servo response, tuning. control for second order processes. It will be shown that
there are advantages in selecting Pd as a polynomial, for
ABSTRACT example, reduced controller ringing and closer tracking of
the specified continuous time dynamics. However, the
One of the techniques cited in the literature on tracking polynomial form of Pd has an adverse effect on the
specifications in Generalised Predictive Control is the use sensitivity of the controller to measurement noise and the
of the so-called P polynomial, to shape the closed-loop disturbance rejection properties of the closed-loop. The
transient response. P(q-1) was originally defined as a sensitivity and disturbance rejection was worsened by the
transfer function written as P(q-1) = Pn(q-1)/Pd(q-1). inclusion of the T polynomial, which is normally associated
However, Pd is normally selected as a scalar equal to Pn(1). with improving the robustness of the closed loop.
The authors observed that there is no analysis in the The layout of the paper is as follows. A brief introduction
literature for any case other than scalar Pd. This paper to GPC is given, followed by a discussion on model-
analyses the case where P(q-1) is used to achieve model- following using P(q-1). Following this the equivalent linear
following control. It is shown that the use of a polynomial form of GPC is presented so that the closed loop can be
Pd, with roots in the left-half of the z-plane unit circle, analysed. An example is then given illustrating the
reduce the magnitude and ringing of the control signal in undesirable problems associated with model-following
the exact and detuned model-following tuning strategy. using P and the advantages and disadvantages of the
However, when the T polynomial, associated with inclusion of the polynomial form of Pd. Finally, an
robustness and disturbance rejection, was included in the alternative method for achieving model-following control,
design, the presence of the polynomial form of Pd(q-1) which does not suffer from the poor performance associated
severely degrades the disturbance rejection performance with using P, is briefly discussed.
and sensitivity to measurement noise. The poor
performance observed when using P to achieve model- GENERALISED PREDICTIVE CONTROL
following suggests that an alternative tuning strategy would
be more appropriate in a practical application.
The GPC algorithm belongs to the family of Model Based
Predictive Controllers (MBPC). A locally linearised
INTRODUCTION CARIMA (Controlled Autoregressive and Integrated
Moving Average) model of the process to be controlled is
An important feature of any control loop is the response to required:
a setpoint change, called the transient, tracking or servo
B(q−1) C(q−1)
response. For example, the plant engineer has to specify y(t) = −1
u(t −1) + ξ (t) (1)
how fast or slow it should take to fill a tank to a particular A(q ) A(q−1)∆
level or how fast a product is brought up to a particular where A, B and C are polynomials in the backward shift
temperature. This paper analyses some of the issues related operator q-1 of order na, nb and nc respectively. u(t) is the
to the specification of the desired transient response in process input (control signal) and y(t) is the process output.
Generalised Predictive Control (GPC) (Clarke et al. 1987, ξ(t) is zero mean white noise and ∆ is the differencing
Clarke and Mohtadi 1989). One of the techniques cited in operator (1 – q-1). The C polynomial in the noise term is
the literature to achieve a particular closed-loop transient rarely estimated in practice, but instead replaced by a
response is model-following control using the P design polynomial T(q-1). A j-step ahead predictor, ŷ(t + j),
polynomial, (McIntosh et al. 1991). P was originally is derived from the model in Equation (1) so that the future
defined as a unity gain transfer function as follows: process outputs can be predicted. In order to calculate the
P(q −1 ) = Pn (q −1 ) Pd (q −1 ) control signal a quadratic cost function is used to minimise
the sum of squared errors between the predicted and desired
However, Pd is normally selected as a scalar equal to Pn(1), outputs with an additional term weighting projected control
hence the reason why it is usually referred to as the P increments as follows:
polynomial. This paper examines the case when Pd is
 N2 NU sample intervals). Note that N2 = d + NU would yield exact
J= å[P(q
j = N1
−1
) yˆ (t + j ) − w(t + j )]2 + å λc [∆u(t + j −1)]2
j =1
(2) model-following.

subject to : ∆u(t + j –1) = 0; NU < j ≤ N2. If the desired closed loop model is first order than the
standard P polynomial can be easily used to represent the
In Equation (2), N1 is the initial prediction horizon, N2 is desired dynamics, since in this case Pd is a scalar, i.e.
the final prediction horizon and NU is the control horizon. P = 1/HM = (1 – aq-1)/(1– a). However, a first order closed
λc is the control weighting factor used to penalise control loop model is primarily applicable to plants that exhibit first
increments. P(q-1) is the P polynomial introduced to order characteristics only, because excessive control action
broaden the range of control objectives available in GPC, is required to make higher order processes to respond in a
for example, model-following control. w(t + j) is referred to first order manner.
as the reference trajectory and defines the desired trajectory
The z-transform of second and higher order continuous
that the output, ŷ(t + j), or auxiliary output, P(q-1)ŷ(t + j),
time models including zero order hold (ZOH) results in the
should follow. Future setpoint changes are generally not
discrete model having zero (numerator) as well as pole
considered in which case the reference trajectory is set
(denominator) dynamics. For example, the following
equal to the current setpoint value i.e. w(t + j) = r(t).
second order continuous time model has a 2% settling time
The tuning of the GPC controller involves setting the of 10s and a damping ratio ζ = 1:
various parameters (N1, N2, NU, λc , P, T). To make the
tuning process easier, various tuning strategies have been H M ( s ) = 0.3318 ( s 2 + 1.152s + 0.3318)
developed. In (McIntosh et al. 1991), McIntosh presents
Transforming this to the discrete domain including ZOH
three tuning strategies utilising these parameters, where at
and ignoring the unit delay, yields
the time of commissioning, a single tuning parameter
(either N2, λc or P) is used to vary the servo response while H M (q −1 ) = 0.114(1 + 0.68q −1 ) (1 − 1.124q −1 + 0.316q −2 ) (3)
the remaining parameters are set to default values. The T
polynomial may then be used to improve the robustness and In (McIntosh et al. 1991) only the poles of the reference
regulation without influencing the servo response (that is, if model in Equation (3) are used to specify the P polynomial
there is no plant-model mismatch). Furthermore, particular and the zero dynamics are ignored. McIntosh states that
settings of the GPC parameters lead to some standard ignoring the numerator dynamics makes little difference
control laws such as minimum variance, dead-beat and since the closed loop model is not followed precisely when
mean-level control. using detuned model-following. However, as will be seen
in the example, when P is specified using only the poles of
MODEL-FOLLOWING CONTROL the discrete reference model, the result is a very active
ringing control signal. However, a small value of λc can be
Model-following control involves specifying a reference or found, on a trial and error basis, which removes this ringing
closed loop model, HM(q-1), that describes the desired and has little effect on the servo response. Furthermore, the
closed loop response. A good way of selecting this step response of a second order discrete model without a
reference model is to consider an appropriate continuous zero in the left-half plane has a sharper initial response
time model and transform this to the discrete domain. The similar to a first order plant. Thus, larger control signal
continuous time model is usually derived from excursions will be required to track this faster response. In
specifications such as the rise time, settling time, light of this and the requirement of finding a suitable value
percentage overshoot, time constant etc. of λc to remove the ringing control, the author was
motivated to investigate the specification of Pd as a
GPC can be set up to follow the closed loop model HM polynomial rather than a scalar. The next section shows
exactly. In this case P is set equal the inverse of the desired how the GPC control law can be reformulated as an
closed loop model, i.e. P = 1/HM and the horizons are set to equivalent linear controller, see Figure 1, which will be
minimum variance (MV) settings. However, in this case the used to perform closed-loop analysis.
process zeros are cancelled, in which case if they are
outside the unit circle will lead to an unstable closed loop. EQUIVALENT LINEAR FORM OF GPC
In order to have a practical control scheme for plants with
arbitrary zeros the model-following capabilities must be
Minimising the cost function in Equation (2) with respect to
detuned. This can be achieved by either increasing the
output horizon relative to the control horizon or by adding the vector of future control increments, ∆u(t + j – 1), yields
the GPC control law, (Clarke et al. 1987). For the purpose
control weighting, i.e. λc > 0. The former method is used in
of carrying out closed loop analysis, the control law can be
the "Detuned Model-Following" tuning strategy (McIntosh
rearranged in the general linear form (Griffin 2002):
et al. 1991) and is achieved by commissioning GPC with:
NU = na + 1, N1 = 1, N2 > d + NU, λc = 0 R (q −1 )
S (q −1 )∆u (t ) = hT (q −1 )r (t ) − y (t )
and using the inverse closed-loop model, P = 1/HM(q-1), as Pd (q −1 )
the active tuning parameter (d is the process dead-time in where the scalar h and polynomials R(q-1) and S(q-1) are
functions of the GPC parameters and of the specified
process model. The influence of Pd(q-1) on the closed loop q −1 Pd P ∆A
is not only its appearance in the denominator of the y (t ) = r (t ) + d d y (t ) (7)
Pn b1hTPn
feedback term, but it also influences the R and S
polynomials (a step in the calculation of the control law and the input sensitivity function becomes:
involves the division of PnT by A∆Pd). Figure 1 illustrates
the block diagram for the general linear form of GPC, along u(t) −AR
Sudy(q−1) = = (8)
with the various sources of disturbances that can affect the d y (t) hTPn B
closed-loop. The derivation of equivalent three-term linear
control law obtained when Pd is a scalar and incorporated Equation 7 shows that exact model-following is achieved,
into Pn, such that P = Pn where P(1) = 1, can be found in i.e. the servo transfer function is q-1Pd/Pn = 1/P(q-1). This
(McIntosh et al. 1991). means that the output will track exactly the specified closed
loop model. Furthermore, T is present in the regulatory
(input disturbance) (output disturbance) transfer function but not in the servo transfer function.
Setpoint
du (t) dy(t) Therefore, T can be used to vary the regulatory response
r(t)
hT ( q −1 )
+ 1 + q −1 B ( q −1 ) + and sensitivity of the controller without affecting the servo
- ∆S (q − 1 ) + u(t) A(q − 1 ) + y(t)
response. Equation (8) shows that Sudy has poles at the roots
Plant of B, now if the roots of B are in the left half of the unit
R( q −1 ) + circle, this can lead to a ringing control signal during the
Pd (q −1 ) + m(t)
(measurement noise) rejection of an output step disturbance.

Figure 1: Block Diagram for the General Linear Form of To examine the control signal with respect to a setpoint
GPC. change, note that from Eq. (1) (ignoring the noise term):
A
From Figure 1, the transfer function between the output y(t) u (t ) = −1
y (t ) (9)
q B
and the setpoint r(t) (servo response) and the transfer
function between output and the output disturbance dy(t) Replacing y(t) by the servo response from Equation (7),
(output sensitivity function) are as follows: yields:
q −1hTBPd P ∆SA A Pd
y (t ) = r (t ) + d d y (t ) (4) u (t ) = . r (t ) (10)
Pclpd Pclpd B Pn

where Pclpd defines the closed-loop poles: Note, if the process numerator B contains roots in the left-
half of the unit circle, the presence of these roots in the
Pclpd = Pd (q −1 )∆S (q −1 ) A(q −1 ) + q −1B (q −1 ) R(q −1 ) denominator of Equation (10) will result in a ringing
control signal in response to a setpoint change. For control
The transfer function between the output disturbance dy(t) strategies that do not cancel the process zeros, for example,
(or measurement noise m(t)) and the control signal u(t) pole-placement, the servo response is (ignoring a prefilter)
(input sensitivity function) is given by: y(t) = q-1Br(t)/Pcl, where Pcl defines the desired closed loop
u(t) −AR servo poles. Replacing y(t) in Equation (9) with this servo
Sudy(q−1 ) = = (5) response, results in u(t) = Ar(t)/Pcl, a transfer function
d y (t) Pclpd
independent of B. However, examining Equation (10), note
This transfer function defines the influence of the output that if Pd(q-1) is selected so that its roots are in the left-half
disturbance and measurement noise on the control signal, plane, then Pd will reduce the ringing due to B. For the case
for example, how active the control signal will be in Pd = B, (assuming B is inverse stable) no ringing in the
response to high frequency measurement noise. control signal will take place during a setpoint change and
the servo response will be y(t) = q-1hBr(t)/Pn, similar to a
pole-placement strategy. This similarity however, is in
Closed Loop Analysis for Exact Model-Following
relation to the servo response only since the process zeros
are cancelled by the minimum variance control law,
For an exact model-following configuration, i.e. when whereas they are not cancelled in a pole-placement strategy.
minimum variance horizon settings are used (N1 = 1, Transforming a second order continuous time model to the
N2 = NU + d and λc = 0) and when there is no plant-model discrete domain results in a zero is in the left-half of the
mismatch, Pclpd becomes: unit circle, see Equation (3). Therefore, specifying this zero
for Pd will reduce the magnitude and ringing of the control
Pclpd = hPn (q −1 ) B (q −1 )T (q −1 ) (6)
signal.
Note that, with MV horizon settings, S = B(q-1)/b1, where b1
is the first element of the B polynomial (this applies for Closed Loop Analysis of Detuned Model following
plants with no dead-time). Substituting Equation (6) into
Equation (4) and (5) yields: The above analysis was for exact model-following strategy
using MV horizon settings. Similar results are established
here for the more practical detuned model-following
configuration. In detuned model-following the closed loop specified continuous time model. The effect of λc is to push
~ the left-half plane closed loop pole, introduced by Q,
poles are defined by the polynomial TPn Q . In this
~ towards the origin. This results in a much improved control
case Pn ≈ Pn and if B has roots in the left half of the unit signal during both the setpoint change and the disturbance.
circle then Q generally contains roots in the left-half plane Note also, the control signal is also less active in response
in order to minimise the effect of the process zeros on the to the measurement noise.
specified servo response, 1/P(q-1). The value of Q depends
mainly on B and P. The servo transfer function in this case 1.5
M e a s u re m e n t N o is e

is:

Output y(t)
1

1 / P (q - 1 ) O u t p u t S t e p D is t u rb a n c e
q −1hBP 0.5
y (t )
y (t ) = ~ d r (t )
PnQ 0
0 10 20 30 40 50

which is similar to the servo response in Equation (7) 5

except that B is not cancelled . An exception to this is when

Control u(t)
Pd = B, (assuming the roots of B are inside the unit circle), 0

in which case Q = B, and the servo transfer function

−1 ~ -5

becomes q hB Pn . The input sensitivity function will 0 10 20 30

Tim e (s )
40 50

be similar to that in Equation (8) but with PnB replaced by Figure 2: Detuned Model-Following with Pd a Scalar,
~
Pn Q , thus the possibility of a ringing control signal in λc = 0 and T = [1 – 0.8q-1].
response to an output disturbance. The control signal in 1.5
response to a setpoint change becomes: Output y(t)

A q −1hBP hAP 1 / P (q - 1 )
u (t ) = −1 . ~ d r (t ) = ~ d r (t ) (11) 0.5
y (t )
q B PnQ PnQ 0
0 10 20 30 40 50

As in the exact model-following case, specifying Pd as a 5

polynomial can be used to reduce the ringing during a
Control u(t)

setpoint change, caused by the roots of Q in Equation (11). 0

EXAMPLE -5
10 20 30 40 50
Tim e (s )

In this example the process is second order with a settling Figure 3: Detuned Model-Following with Pd a Scalar,
time of 23s given by λc = 0.01 and T = [1 –0.8q-1].
Gp(s) = 0.09/(s2 + 0.66s + 009)
The influence of specifying Pd as a polynomial will now be
With a sampling interval of 1s the discrete time model is examined. In this case P is specified to include the zeros as
well as the poles of the reference model as follows:
q −1 B 0.03628q −1 (1 + 0.8026q −1 )
G p (q −1 ) = =
A 1 − 1.451q −1 + 0.517 q − 2 1 P 1 − 1.124q −1 + 0.316q −2
P(q −1 ) = = n =
HM Pd 0.114(1 + 0.68q −1 )
Note the left half plane zero at z = –0.8026. To achieve
detuned model-following the horizons are set to N1 = 1, The closed loop response in this case is shown in Figures 4
N2 = 6 and NU = 3. The reference model with a settling and 5. The same horizon settings were used and λc = 0 was
time of 10s defined in Equation (3) will be used to specify used in both cases. Examining the servo response, the
P. Figures 2 and 3 shows the response when Pd is specified expected reduction in the initial control excursion and
as a scalar, in which case P is specified as follows: control signal ringing can be seen, in compared to Figure 1.
Furthermore, very accurate tracking of the specified model
P(q −1 ) = Pn Pn (1) = (1 − 1.124q −1 + 0.316q −2 ) 0.192 is achieved, which in this case is equivalent to the originally
In each case, T = [1 – 0.8q-1], was used to improve the specified continuous time reference model.
robustness and reduce the sensitivity of the controller to the However, the inclusion of Pd as a polynomial has an
measurement noise. Figure 2 shows the servo response unexpected adverse effect on the disturbance rejection and
when λc = 0. Note that the servo response is very close to the sensitivity of the controller to measurement noise.
the ideal, 1/P, however, at the expense of a very active Figure 4 shows the response when T = 1. The ringing
ringing control signal as expected from the previous closed- control in response to an output step disturbance and the
loop analysis. Note also the undesirable ringing control in very active control signal in response to measurement noise
response to the output step disturbance. Figure 3 shows the can be seen. In an attempt to improve this, Figure 5 shows
servo response when λc = 0.01. In this case the servo the response when T = [1 – 0.8q-1], which is the same as
response is not as close to 1/P, however, it is closer to the that used in Figures 2 and 3. However, this has made the
sensitivity and disturbance rejection of the controller even tuning strategies gives more predictable results than with
worse. Note, in Figure 5 the initial control signal, in model-following using the P polynomial.
response to the step disturbance, is in the same direction as
the disturbance, i.e. the Sudy transfer function has an inverse Model-Following Using the M Setpoint Prefilter
response characteristic. This has the effect of amplifying
the disturbance before removing it, and greatly The M setpoint prefilter (Griffin and Downing 2002,
amplification of the measurement noise. Specifying a Soeterboek 1992) offers superior model-following
slower T polynomial, while it may improve the gain, phase capabilities over using the P polynomial. In this case the
and delay margins, made this high amplification of the reference trajectory values, w(t + j), that appear in the cost
noise and inverse response of Sudy even worse. Furthermore, function of Eq. (2), are generated by filtering the setpoint
increasing the order of T also worsened the situation while r(t) with the reference model, M(q-1) = Mn(q-1)/Md(q-1), i.e.
not in general improving the robustness margins.
w(t + j ) = M (q −1 )r (t ); j = 1,..., N 2
This inverse response characteristic of Sudy was also seen to
occur for the case when Pd was a scalar, when much slower In this case the reference model can be either a FIR or IIR
T polynomials were specified, more so in the case when filter. Furthermore, the M setpoint prefilter has no influence
λc = 0. Interestingly, if this occurs and Pd is specified with on the closed loop dynamics and acts only on the setpoint,
a root in the right half of the unit circle (e.g. at z = 0.2) then therefore complete independent tracking and regulation is
this removed the inverse response of Sudy and greatly achieved.
reduced the high frequency gain of Sudy. However, a right
half plane zero will increase the ringing during a setpoint CONCLUSIONS
change.

1.5
Using the detuned model-following configuration to control
a process with zeros in the left-half of the unit circle results
Output y(t)

1
in a ringing control signal during a setpoint change when Pd
1 / P (q - 1 )
0.5 is a scalar. The ringing can be removed by including a
y (t )
0
'small' value of λc. The use of Pd as a polynomial with roots
0 10 20 30 40 50
in the left-half of the unit circle also reduced the magnitude
10 and ringing of the control signal during a setpoint change.
5 However, when the T polynomial was included in the
Control u(t)

0 design, it led to greatly compromised disturbance rejection,

-5 sensitivity and robustness properties. These problems
-1 0 associated with using detuned model-following, with either
0 10 20 30 40 50
Tim e (s ) Pd as a scalar or a polynomial more so, suggests that the
Figure 4: Detuned Model-Following with Pd as a Polynomial and other tuning strategies presented in the final section are
T = 1. more applicable to a practical application.
1.5
REFERENCES
Output y(t)

1 / P (q - 1 )
0.5 Clarke, D.W., C. Mohtadi, and P.S. Tuffs. 1987. "Generalised
y (t )
Predictive Control – Part I & Part II", Automatica, Vol. 23,
0
0 10 20 30 40 50 No. 2, 137-160.
10
Clarke, D.W. and C. Mohtadi. 1989. "Properties of Generalised
Predictive Control", Automatica, Vol. 25, No. 6, 859-875.
5
Control u(t)

Griffin, B. and C.J Downing. 2002. "Servo Response Specification

0
Using Setpoint Prefiltering in GPC", Proceedings of 2002
-5 Irish Signals and Systems Conference, Cork, Ireland.
-1 0
0 10 20 30 40 50
Griffin, B. 2002. "Servo Response and Modelling Issues in
Tim e (s ) Generalised Predictive Control". M. Eng. Thesis, Cork
Figure 5: Detuned Model-Following with Pd as Polynomial and Institute of Technology, Cork, Ireland.
T = [1 –0.8q-1]. McIntosh, A.R, S.L. Shah, D.G. Fisher. 1991. "Analysis and
Tuning of Adaptive Generalised Predictive Control", The
Canadian Journal of Chemical Engineering., Vol. 69 (Feb),
ALTERNATIVE TUNNING STRATEGIES 97-110.
Soeterboek, R. 1992. Predictive Control: A Unified Approach,
The two other tuning strategies presented in (McIntosh et Prentice Hall, 92-97
al. 1991), involving the output horizon N2 and the control
weighting factor λc, cannot in general be used to specify a Acknowledgements: The Authors wish to acknowledge the
specific response time, for example, a settling time of 20s. support of Enterprise Ireland under the Food Factory of the
However, they do allow the speed of the closed loop to be Future Program.
adjusted in a predictable manner over a wide range.
Furthermore, the inclusion of the T polynomial with these
 AN AUTOTUNED ROBUST GENERALISED PREDICTIVE CONTROLLER
Tom O’Mahony & Conor. J. Downing
Advanced Control Group,
Department of Electronics Engineering,
Cork Institute of Technology,
Cork, IRELAND.
E-mail: [email protected]

expended on the development of tuning guidelines to simplify

KEYWORDS the task of commissioning the GPC. Initially this research
focused on the GPC servo-parameters {N1, N2, NU, λ} and
Computer Aided Design, Genetic Algorithms, GPC, Robust
subsequently on the selection of the so-called T-polynomial.
Control, Optimisation, System Identification
Research into the latter is of considerable importance, since
this polynomial has a significant effect on the closed-loop
ABSTRACT
robustness and disturbance rejection properties. The difficulty
The contribution of this paper is the development of a in choosing this polynomial is partly associated with the
systematic tuning technique for the GPC such that an optimal robustness/performance trade-off and partly due to the fact
trade-off between robustness and performance is achieved. that its influence is not monotonic. Since the existing tuning
This trade-off is achieved by selecting the GPC parameters to rules do not adequately address this trade-off, many
minimise a mixed H2/H∞ design problem. A number of practitioners resort to the use of an iterative trial-and-error
simulation studies demonstrate the usefulness of this tool. design (Demirioglu & Karasu, 2000) or neglect to implement
this parameter altogether (Wu et al., 1998; Liu & Daley,
INTRODUCTION 1999). As a result sub-optimal performance is likely,
particularly in the latter case where it is necessary to detune
Model predictive control (MPC), also referred to as model the tracking performance in order to achieve sufficient
based predictive control (MBPC), is now regarded, by many, robustness.
as one of the most important developments in process control
The procedure advocated in this paper differs considerably
since the development of the PID controller over 50 years from previously published works, in that, while the latter have
ago. The technique now enjoys a large installed base in the invariably concentrated on the development of empirical
process industry in general (Qin & Badgwell, 1996) and the
tuning guidelines, this paper has concentrated on the
refining, petrochemical and chemical industries in particular. development of a generalised systematic tuning mechanism
In terms of the food industry, recent MPC applications which achieves a realistic trade-off between performance and
include riboflavin production (Kovarova-Kovar et al., 2000);
robustness. Thus to simplify the tuning process, an auto-
melter units and evaporators in the sugar industry (Zamarreno tuning tool was devised, which successfully negotiates a
& Vega, 1997, Lissane-Elhaq & Giri, 1998); crossflow grain practical trade-off between performance and robustness. This
dryers (Liu & Bakker-Arkema, 2001); drying towers (Clarke,
trade-off is achieved by selecting the GPC parameters to
1998) and greenhouse climate control (Gonzalez & Leyris,
minimise a mixed H2/H∞ design problem. Simulations
1996).
demonstrate the usefulness of this tool where the results are
The acclaim associated with MPC, both in industry and
compared with previously published tuning techniques. The
academia, is due to its unique combination of simplicity and
excellent results obtained support the belief that the advocated
control acumen. Unlike H∞ control (Kwakernaak 1993) the design is of practical significance and is the principal
tuning parameters associated with MBPC are based in the contribution of this paper.
time-domain and hence MPCs are considerably simpler to
tune and more intuitive to commission. More importantly,
GPC ALGORITHM
perhaps, MPCs have a unique predictive capacity and
constraint handling capability. The former endows the Since the literature abounds with extensive derivations of the
strategy with the ability to surmount the time-delay problem GPC algorithm, (Clarke et al, 1987; Soeterboek, 1992) such a
while the latter facility is considered essential in most derivation will be avoided here. Suffice it to say that in its
industrial implementations. Notwithstanding these formulation, the GPC algorithm utilises a CARIMA model of
advantages, most MPCs do not adequately address the design the process to predict the process output, y(t), over a finite
trade-off between robustness and performance. Hence the horizon, e.g. it generates the predictions { ŷ (t+N1|t),…,
focus in this paper has been to address this problem and ŷ (t+N2|t)}; N1, N2 ∈ Z+; N2 ≥ N1, based on the postulation
develop a design technique that encompasses the trade-off
alluded to above. that the future control sequence remains unchanged (i.e. open-
In particular, the generalised predictive control (GPC) law, loop prediction). Using these predictions a set of control
(Clarke et al., 1987) was considered for the control versatility increments, ∆ uˆ (t+j); 0 ≤ j ≤ N2 - N1; ∆ = (1–z-1), are
it allows and because of its combined academic and industrial calculated by minimising a quadratic cost comprising the
success. Since the late 1980's considerable effort has been predicted errors plus some weighting, λ, on the future control
signal increments. This philosophy yields a trajectory of Im

future control increments and therefore constitutes an open-

loop strategy. To close the loop, only the first element, ∆ uˆ (t)
is applied to the system and the optimisation recomputed at
1 + Lo
time instant t+1. This strategy is called the receding horizon
(-1,0)
principle and is one of the key features of model based L − Lo
Re

predictive control in general and GPC in particular. In the

absences of constraints on, for example, the process output,
input or states, the GPC algorithm may be re-arranged as a
general linear causal 2DoF control law of the form
HLL ĤLLo
R ( z −1 )∆u (t ) = T1 ( z −1 ) w(t ) + S ( z −1 ) y (t ) (1)
where u(t) is the process input, y(t) the process output and Figure 1: Nyquist plot of real and nominal loop transfer
w(t) the reference trajectory or set-point sequence. The functions
polynomials R(z-1), T1(z-1), and S(z-1) are polynomials in z-1
defined through the choice of model and GPC design then the perturbed closed-loop system is stable. In expression
parameters namely the prediction horizons and T(q-1) (4) T(jω) is the nominal complementary sensitivity function
polynomial (McIntosh et al., 1991). and the factor |L(jω)-Lo(jω)|/|Lo(jω)| is the relative size of the
perturbation of the loop gain L from its nominal value Lo and
ROBUST GPC denoted ∆M. Thus the relation (4) shows that the closed-loop
system is guaranteed to be stable as long as the loop
Regardless of the design technique used, controllers are perturbations satisfies
always designed based on (necessarily incomplete) (5)
∆ M ⋅ T ( jω ) < 1 ∀ω ∈ ℜ
information about the dynamic behaviour of the process. In
the context of linear time invariant models, the most Hence the complementary sensitivity function represents an
significant source of model uncertainty may be attributed to upper bound on the allowable magnitude of the proportional
the non-linear nature of real processes. Thus if the process loop perturbation - in other words, it defines the maximum
model is obtained via linearisation, then it is accurate only in amount of mismodelling that the closed-loop will tolerate
the neighbourhood of the reference chosen for linearisation prior to instability. The larger the magnitude of the
and different operating conditions generally lead to changes complementary sensitivity function, the smaller the allowable
in the parameters of the linear model. For example, increased uncertainty. Hence it is desirable that T be small, particularly
throughput and flowrates usually result in smaller dead-times at the frequencies where it is anticipated that the modelling
and time constants. Furthermore process models are usually errors may be severe.
simplified by neglecting fast dynamic phenomena, and More formally, the inequality (5) is written as W T ∞ < 1
additional parametric uncertainty arises as a result of system
where W represents an upper bound on the uncertainty i.e.
identification experiments. Neglecting the uncertainty
between the model and the process is not advisable as the W = ∆ M and the notation ||⋅||∞ represents the H∞ norm
resulting controller tends to be too "tight" and is likely to defined as H ∞
= sup H ( jω ) . The H∞ norm may be
become unstable in the real operating environment. ω∈ℜ

A basic mechanism for dealing with model uncertainty is interpreted as the maximum amplitude of the frequency
illustrated via the Nyquist plot, Figure 1. Denoting the true response of the transfer function. In the context of control
open-loop transfer function L(z-1) = C(z-1)G(z-1) where C(z-1) system design, the minimum requirement of robust stability is
represents the compensator and G(z-1) the plant. The nominal achieved by satisfying W T ∞ < 1 . An effective method of
loop transfer function is designated Lo(z-1) = C(z-1)Go(z-1) and seeking a compensator that satisfies this is to minimise
assume, for simplicity, that both L and Lo are stable. In order
to ensure the stability of the closed-loop system for an open- W T ∞ with respect to all compensators C that stabilise the
loop transfer function Lo which differs from the actual one L, system. This problem is one version of the general H∞-
the Nyquist plot of L should leave the critical point [-1, j0] on optimisation problem (Kwakernaak, 1993). In terms of the
the left when traversed in the sense of growing frequencies. GPC, robust stability may be achieved by minimising the
Thus a sufficient condition for stability is, that at each following H∞-norm with respect to the T-polynomial
frequency, the distance between L and Lo be less than the
min W T ∞ = min sup W (e − jω ) T (e − jω ) (6)
distance between the nominal open-loop transfer function and T T ω
the critical point. This may be expressed mathematically as The difficulty associated with this optimisation may readily
L( jω ) − Lo ( jω ) < 1 + Lo ( jω ) ∀ω ∈ ℜ (2) be appreciated by developing an expression for the
and is equivalent to complementary sensitivity function, T. From the equivalent
L( jω ) − Lo ( jω ) Lo ( jω ) linear control law of equation (1), this transfer function is: T =
⋅ <1 ∀ω ∈ ℜ (3) z-1SB/Pc where Pc is the closed-loop characteristic
Lo ( jω ) 1 + Lo ( jω ) polynomial. It can readily be shown that the T-polynomial is a
It follows that if factor of the closed-loop characteristic equation hence Pc may
L( jω ) − Lo ( jω ) be expressed as Pc = T⋅Pc’ and the robust stability
⋅ T ( jω ) < 1 ∀ω ∈ ℜ (4) 1 P 'T
Lo ( j ω ) requirement becomes: | W2 | < = c . It is clearly
|T | SB
evident that the T-polynomial may be manipulated to satisfy 4) Evaluate fitness: To generate a population of offspring, a
this. A difficulty arises due to the fact that the polynomial S set of individuals, taken from the current population, is
depends on T. Due to the complex non-linear relationship selected as parents. Individuals with higher fitness values
between T and S an analytical solution to the optimisation (i.e. those that are closer to the optimum solution) have
problem (6) is not feasible and numerical techniques must be an increased probability of survival. In this strategy the
employed. probability of an individual, xi, being selected, P(xi), is
To minimise the inequality (6) an optimisation technique defined by
known as genetic algorithms (GAs) was utilised. Inherently f ( xi )
GAs are very different from conventional optimisation P ( xi ) = (8)
∑
N
techniques. By dealing with several independent points, the f ( xi )
i =1
GA samples the solution space in parallel and hence is less where f(xi) is the fitness value of an individual and N is
susceptible to lingering in or converging to local optima. GAs the population size. The selected individuals are placed in
do not utilise gradient techniques and hence the search space a mating pool, to await the application of other GA
need not be differentiable or continuous. The optimisation operators e.g. crossover and mutation.
only requires objective information so that GAs can utilise 5) Crossover: The crossover operator is applied after
various kinds of objective function, even non-linear, multi- selection and produces a two new chromosomes or
objective or knowledge-based ones and are theoretically and offspring by combining information extracted from the
empirically proven to provide robust searches (Goldberg, two parent chromosomes. The parents are randomly
1989). Thus, with their robust features, simple structure and selected from the mating pool and single-point crossover,
the ability to attain global optimisation, GAs are now with a crossover probability, Pc = 0.95, is employed.
regarded as important mathematical tools for complex 6) Mutation: Mutation is the last genetic operator applied in
optimisation problems and are ideally suited for searching a simple genetic algorithm. It operates on each individual
discrete, noisy, multi-modal or complex spaces. by altering the value of a randomly selected bit position.
The salient characteristics of the GA utilised for the Again the success of the mutation operator depends on
optimisation of the T-polynomial may be described as follows the mutation probability, Pm, which is usually very low.
1) Parameter Encoding: Similar to their biological The combined use of the crossover and mutation
counterparts GAs utilise binary strings called operators ensures against total loss of any genes in the
chromosomes to represent the parameters of an population by their ability to introduce any gene which
optimisation problem. In this problem the optimisation may not have existed initially, or, may subsequently have
amounts to the determination of the optimum location of been lost.
the roots of the T-polynomial such that the criterion (6) is 7) Iteration: The GA runs iteratively and the evolution
minimised. Experience suggests that there are few strategy ensures the search converges to a highly fit
applications that require a high-order T-polynomial population representing optimal or near-optimal
therefore a fifth-order T-polynomial was selected i.e. solutions to the considered problem. The GA terminates
5
when some user-prescribed criterion is reached. This
T ( z −1 ) = ∏ (1 − t z
i =1
i
−1
) ; − 1 < ti < 1 (7) may be a predetermined number i.e. the GA is to stop
after 50 generations, or a prescribed degree of
The translation from real parameters to binary strings convergence.
may utilise any desired coding scheme e.g. standard
binary, Gray code, etc. AUTOTUNED GPC
2) Initial population: In a GA the population is the set of
possible solutions and is characterised by the binary In the previous section the GA tuning considered a single
chromosomes. The GA commences by randomly closed-loop design criterion; robust stability. In contrast a
generating an initial population consisting of N binary practical design requires a successful trade-off between
strings or chromosomes, each of which is of length m. In performance and stability. To achieve this it is necessary to
this application a population size of N = 30 was selected, extend the cost function utilised in the GA optimisation to
based on the principle that every possible point in the incorporate tracking and disturbance rejection performance
search space should be reachable from the initial metrics, in addition to the stability criterion previously
population (Reeves, 1995). discussed. In this section the new GA cost is defined to be
3) Evaluate fitness: Each individual of the initial and min {J }
{T , N1 , N 2 , λ } GA
subsequent generations must be evaluated to determine (9)
= min {η ⋅ J stability + J servo + J reg + J noise }
their fitness. This is accomplished by firstly decoding {T , N1 , N 2 ,λ }
each individual in the current population to obtain the where Jstability represents that component of the cost that
parameters {t1, t2, t3, t4, t5}. The T-polynomial is formed accounts for robust stability, Jservo specifications on the
from these roots and the controller polynomials S and R tracking performance and the latter two represent
calculated. Finally the complementary sensitivity specifications on the disturbance rejection performance. In
function is computed and the H∞-norm (6) determined. general good disturbance rejection implies that low-frequency
This norm is evaluated for the entire population, the or deterministic disturbances are rejected with suitably rapid
population is ranked by assigning a fitness value of one transients and also that high-frequency noise is appropriately
to the solution(s) that minimise (6), and a fitness value of attenuated. In the robust control literature the effect of high-
zero to the solution that maximises it. The intermediate frequency or measurement noise is frequently tackled by
solutions are appropriately scaled between zero and one. restricting the amplitude of the input sensitivity function i.e.
by minimising the norm
 J noise = U (10) being the model denominator polynomial. For example the
∞
most recent offering, Yoon & Clarke (1994), suggests setting:
The input sensitivity function, defined as U = C/(1+GC) (14)
T = A(1 − β z −1 ) n
relates measurement noise to the control signal and hence is a
suitable metric in that it will also tend to ensure a smooth with n = N1. The parameter β has a value ∈ (0, 1) and a rule-
control signal. The disturbance metric is defined as of-thumb is to place β in the neighbourhood of the dominant
τ root of the model denominator polynomial. To evaluate the
∫
J reg = | e(t ) | dt
0
(11) autotuning design advocated in this paper, a number of
simulation examples will be taken from the recent GPC
where τ represents the duration of the simulation and
literature. For each example a GPC is designed based on (i)
assuming a pure regulatory problem e(t) is defined as e(t) = -
the advocated autotuning procedure and (ii) equation 5.1.
y(t). Hence Jreg represents the integral of absolute error (IAE)
of the load disturbance response. In the problem definition (9)
the servo tracking criterion was simply defined using the Simulation Example 1
integral of squared error (ISE) criterion i.e. To evaluate the proposed tuning strategy the following
τ
= {r (t ) − y (t )} dt
∫ (12) example, taken from Normey-Rico & Camacho (1999), was
2
J
servo
0 utilised. This example, representing a typical industrial
though more complex definitions are possible. Finally Jstability process, is modelled by
represents the robust stability constraint, defined by the norm k
G p (s) = e −t d s (15)
J stability = W T ∞ (13) s+ p
while the term η represents a weighting factor that may be
where the nominal values of the parameters are td = 0.5, k=1
used to heavily penalise the robust stability criterion i.e.
and p=1. This process is subject to considerable parametric
increase the importance of this component. In this context, the
uncertainty where the uncertainties are defined as ±30% in
problem (9) may be interpreted as optimising the closed-loop
the delay and ±25% in the other two parameters. Sampling at
performance subject to the constraint of robust stability.
0.1second intervals yields the following discrete-time
Since both performance and stability criteria are now
nominal model:
included in the optimisation problem the parameter search
space is extended to incorporate all of the significant GPC 0.1z −1
Go ( z −1 ) = z −5 (16)
parameters. The exception is the parameter Nu as this 1 − 0.9 z −1
generally has a very restricted range - 1 ≤ Nu ≤ ~5, Nu ∈ Z+ - The uncertainty was defined as
and typically is set to the number of unstable or badly damped  G p ( z −1 ) 
open-loop poles. Thus the search is focused on the parameter ∆ M ( z −1 ) =  − 1 (17)
subset {T, N1, N2, λ}. Since the T-polynomial was chosen as  Go ( z −1 ) 
 
fifth-order the GA optimisation is accomplished over, where Gp(z-1) represents the discrete equivalent of one of the
effectively, eight parameters. plant extremes. Selecting various possible models for Gp and
Note that performance is automatically dictated by the examining the corresponding uncertainty model |∆M(jω)|
ISE criterion on the servo tracking error and the IAE on the revealed that the worst case scenario occurred for the process
regulator error. Hence the user is not required to either model Gp(s) = 1.25e-0.65s/(1+0.75s). This uncertainty model
specify performance or select the GPC tuning parameters. was used to characterise the upper bound for the uncertainty
What is required of the user is a rather significant modelling
W = ∆M .
effort to determine W = ∆ M = |G(jω)-Go(jω)|/|Go(jω)|. In a
The GPC servo parameters are defined as
typical industrial scenario the process is non-linear and Go
N1 = 1; N 2 = 70; N u = 1; λ = 0; P =1 (18)
therefore represents a low-order nominal model and G
selected as the model corresponding to the operating point and from the sampled plant transfer function (16) and with
furthest from the nominal. Thence W = ∆ M can be calculated. reference to (14) a reasonable parameterisation for the T-
polynomial is T = (1 –0.9z-1)(1-0.7z-1). Applying the
With W specified, equation (9) is optimised via the GA and autotuning procedure of the previous section yielded the
the optimal GPC parameters identified. In this context, parameter set
optimal implies that the joint IAE/ISE performance criterion
N1 = 6; N 2 = 45; λ = 0.89
is minimised subject to the robust stability constraint, eqn. (19)
(13). T = (1 − 0.85 z −1 )(1 − 0.51z −1 )(1 + 0.42 z −1 )
Interestingly, this choice of servo parameters closely parallels
EVALUATION VIA SIMULATION EXAMPLES those of equation (18); the effect of making N2 somewhat
Many researchers, McIntosh et al. (1991), Soeterboek (1992), smaller is counteracted by the increase in λ. Also the value of
Yoon & Clarke (1994) have examined the problem of N1 selected by the GA corresponds to the standard
designing a robust GPC control law in the presence of model recommendation, i.e. N1 = d +1; where d is the value of the
uncertainty. In each of these studies the selection of the T- delay in sampling periods. The resulting performance is
polynomial was the key element in enhancing robustness. illustrated in Figure 2. This comparison highlights a limitation
Many of these researchers suggested design guidelines to aid that is common to all the proposed recommendations for the
the selection of the T-polynomial. These proposals are quite T-polynomial. When developing these tuning rules, the
similar and the recommendation is to choose either T = (1 - respective authors have tended to focus on a single criterion -
µz-1)n, T = A, or a combination of these two expressions; A that of robust stability and the rules have been developed
from this perspective only. As noted previously however any A difficulty with any controller design is the selection of
practical design inherently involves a trade-off and focusing tuning parameters, intrinsic in which is the
1.5 performance/robustness trade-off. This paper reports on an
initial attempt to solve this problem by developing an
Process Output

1
autotuning tool, which automatically determines the tuning
0.5 parameters for the GPC, based on a nominal process model
and the estimated model uncertainty. The parameters are
0
0 2 4 6 8 10 12 14 16 18
selected by optimising the performance (ISE/IAE criterion)
Time (sec) subject to a robust stability constraint being satisfied. Future
1.5
Autotuned GPC work includes investigation of a multi-objective cost function,
alternative optimisation techniques and practical applications.
Control Signal

1 Yoon & Clarke

0.5
REFERENCES
0
Ansay, P. & V. Wertz, (1997), 'Model uncertainties in GPC: A
-0.5
0 2 4 6 8 10 12 14 16 18 systematic two-step design', European Control Conf., ECC'97,
Figure 2: Closed-loop GPC responses for worst case Bruxelles, Belgium
uncertainty Clarke, D.W., C. Mohtadi, and P.S. Tuffs, (1987), ‘Generalised
predictive control – Part 1. The basic algorithm, - Part II.
exclusively on one design criterion, such as robustness, is not Extensions and interpretations’, Automatica, 23, (2), pp 137-160
conducive to a balanced design process. Thus Yoon & Demirioglu, H. & E. Karasu, (2000), 'Generalised predictive
Clarke's recommendations result in a conservative control control’, IEEE Control Systems Mag., October 2000, pp 36-47
Goldberg, D.E., (1989), Genetic Algorithms in Search, Optimisation
law, for the given model uncertainty, and hence performance
and Machine Learning, Addison-Wesley, Reading, Mass.
is impinged upon. In contrast, the design based on the GA, Gonzalez, A. & J.P. Leyris, (1996), ‘Application of GPC in MIMO
which combines several criteria, results in a better trade-off systems to greenhouse climate regulation’, Int. J. System
between performance and robustness. Science, Vol. 27, No. 1, pp. 27
Kovarova-Kovar, K., S. Gehlen, A. Kunze & A.P. Van Loon,
Simulation Example 2 (2000), ‘application of model-predictive control based on
artificial neural networks to optimise the fed-batch process for
This example, also taken from Normey-Rico & Camacho riboflavin production’, J. Biotechnology, 79, (1), pp 39-52.
(1999), considers the following high-order plant process Kwakernaak, H., (1993), 'Robust control and H∞-optimisation -
(1 + 0.2)e − s Tutorial paper', Automatica, Vol. 29, No. 2, pp. 255-273
G p ( s) = (20) Lissane-Elhaq, S. & F. Giri, (1998), 'Modelling, identification and
(1 + 0.45s )(1 + 0.1s )(1 + 0.6 s )(1 + 0.125s ) control of sugar evaporation - Theoretical design and
This system can be approximated by a first-order model with experimental evaluation', IFAC Adapt. Sys. Ctrl. & Signal
time-delay where the model is defined by Process., Glasgow, Scotland.
Liu, G.P. & S. Daley, (1999), 'Design and implementation of an
e −1.2 s
Gm ( s ) = (21) adaptive predictive controller for combustor NOX emissions', J.
1 + 0.7 s of Process Control, Vol. 9, pp. 485-491
A sampling time of 0.12 seconds was used to obtain Gm(z-1). Liu, Q. & F.W. Bakker-Arkema, (2001), ‘A model-predictive
Based on this model, Normey-Rico & Camacho (1999), controller for grain drying’, J. Food Eng., Vol. 49, pp. 321-326
recommend setting the servo parameters as N = 15, λ =1, McIntosh, A.R., S.L. Shah & D.G. Fisher, (1991), 'Analysis and
tuning of adaptive generalised predictive control', Canadian J.
where the parameter N is the prediction horizon i.e. N = N2 -
Chem. Eng., Vol. 69, pp 97-110
N1+1. Choosing the initial prediction horizon equal to the Normey-Rico, J.E. & E.F. Camacho, (1999), 'Robustness effects of a
model time-delay yields following servo parameters – pre-filter in a Smith predictor-based generalised predictive
N1 = 10; N 2 = 25; N u = 1; λ = 1; P =1 (22) controller', IEE Proc., Vol. 146, No. 2, Pt. D, pp. 179-185
In this case, due to the large value of N1 it is unreasonable to Qin, S.J., & T.A. Badgwell, (1996), 'An overview of industrial
model predictive control technology', Chemical Process Control
utilise equation (14) to define the T-polynomial. Instead T
V, Tahoe City, California, 1996
was simply chosen as T = A = (1 -0.84z-1). Applying the GA Reeves, C.R., (1995), 'Using genetic algorithms with small
to autotune the GPC yielded – populations', Proc. 5th Int. Conf. on GAs, pp. 92-99
N1 = 3; N 2 = 19; λ = 0 Soeterboek, R., (1992), Predictive Control - A Unified Approach,
(23) Prentice Hall Int.
T = (1 − 0.79 z −1 )(1 − 0.57 z −1 )(1 + 0.34 z −1 ) Wu, G., N. Sepehri & K. Ziaei, (1998), 'Design of a hydraulic force
In this case identical performance resulted from both designs. control system using a generalised predictive control algorithm',
Thus the autotuner again negotiated a successful trade-off IEE Proc. Vol. 145, No. 5, Pt. D, pp. 428-436
between the various performance objectives, and considerably Yoon, T., & D.W. Clarke, (1994), 'Observer design in receding
simplifies the commissioning process for the user. Note that horizon predictive control', Int. J. Control, 61, (1), pp 171-191
considering the relatively large time delay (10 samples), Zamarreno, J.M. & P. Vega, (1997), ‘Identification and predictive
control of a melter unit used in the sugar industry’, Artificial
neglected high-order dynamics and mismodelling of the time- Intelligence Eng., Vol. 11, No. 4, pp 365-373
delay parameter, admirable performance was obtained from
the GPC. ACKNOWLEDGEMENT
CONCLUSION Funding for this research was provided the HEAPRTLI program,
2000-2003. The authors wish to gratefully acknowledge this support.
 A SURVEY OF PID CONTROLLER TUNING METHODS FOR PROCESSES
WITH TIME DELAY PROPOSED SINCE 1998

Aidan O’Dwyer
School of Control Systems and Electrical Engineering,
Dublin Institute of Technology, Kevin St., Dublin 8, Ireland
E-mail: [email protected]

KEYWORDS
Time delay, compensation, PID
Table 1: Control of delayed processes using PID
ABSTRACT controllers: publications by date (O’Dwyer, 2002)
Year Journal Total
A time delay may be defined as the time interval between articles publications
the start of an event at one point in a system and its 1942-1951 3 3
resulting action at another point in the system. Delays are 1952-1961 5 6
also known as transport lags or dead times; they arise in 1962-1971 11 14
physical, chemical, biological and economic systems, as 1972-1981 8 13
well as in the process of measurement and computation. In 1982-1991 48 87
the food processing industry, for example, time delays 1992-date 217 380
arise in steam heated dryers used to regulate moisture
content in a food product. Methods for the compensation 1992-1997 112 183
of time delayed processes may be broadly divided into 1998-date 105 197
proportional integral derivative (PID) based controllers, in
which the controller parameters are adapted to the
Due to space considerations, this paper will provide an
controller structure, and structurally optimised controllers,
overview of PID compensation techniques, proposed since
in which the controller structure and parameters are
1998, for processes with time delay.
adapted optimally to the structure and parameters of the
The PID controller may be implemented in
process model. The purpose of this paper is to extract the
continuous or discrete time, in a number of controller
essence of the developments in design, tuning and
structures. The ideal continuous time PID controller is
implementation of PID controllers for delayed processes
expressed in Laplace form as follows:
over the past few years, concentrating on journal
1
publications. The paper will provide a framework against G c (s) = K c (1 + + Td s) (1)
which the literature may be viewed. Tis
with K c = proportional gain, Ti = integral time constant
INTRODUCTION and Td = derivative time constant. If Ti = ∞ and Td = 0
(i.e. P control), then the closed loop measured value will
The use of the PID controller is ubiquitous in
always be less than the desired value for processes without
industry; it has been stated, for example, that in process
an integrator term, as a positive error is necessary to keep
control applications, more than 95% of the controllers are
the measured value constant, and less than the desired
of PID type (Astrom and Hagglund 1995). Despite the
value. The introduction of integral action facilitates the
development of a large number of alternative control
achievement of equality between the measured value and
algorithms over the past four decades, and the fact that PID
the desired value, as a constant error produces an
controllers have been used widely in industry for almost
increasing controller output. The introduction of derivative
sixty years, their popularity is growing; eighty-three
action means that changes in the desired value may be
publications on the control of delayed processes using PID
anticipated, and thus an appropriate correction may be
controllers have been recorded by the author in the year
added prior to the actual change. Thus, in simplified terms,
2000, for example (O’Dwyer 2002). However, Ender
the PID controller allows contributions from present, past
(1993) maintains that, in his testing of thousands of control
and future controller inputs.
loops in hundreds of plants, it has been found that more
In many cases, the design of PID controllers for
than 30% of installed controllers are operating in manual
delayed processes are based on methods that were
mode and 65% of loops operating in automatic mode
originally used for the controller design of delay-free
produce less variance in manual than in automatic (i.e. the
processes. However, there is general agreement that PID
automatic controllers are poorly tuned); this is rather
controllers are not well suited for the control of dominant
sobering, considering the wealth of information available
delay processes. It has been suggested that the PID
in the literature for determining controller parameters. The
implementation is recommended for the control of
following table in instructive in this regard:
processes of low to medium order, with small delays, when
controller parameter setting must be done using tuning
rules and when controller synthesis may be performed a stability of the closed loop control system occurs). The
number of times (Isermann 1989). first such tuning method was defined by Ziegler and
Nichols (1942) for the tuning of P, PI and PID controller
THE SPECIFICATION OF PI OR PID parameters of a process that may or may not include a
CONTROLLER PARAMETERS delay. The tuning rules implicitly build an adequate
frequency domain stability margin into the compensated
Iterative Methods system. Such tuning rules, to compensate delayed
processes by either minimising a performance criterion, or
The choice of appropriate compensator achieving a specified gain and/or phase margin, are
parameters may be achieved experimentally e.g. by manual discussed when the SISO process is modelled in IPD form
tuning. However, such an approach is time consuming and (Kookos et al. 1999), or stable or unstable SOSPD form
the process typically has to be driven to its stability limit. (Luyben 2000). Alternatively, ultimate cycle tuning rules,
Alternatively, a graphical or analytical approach to and modifications of the rules in which the proportional
controller tuning may be done in either the time or gain is set up to give a closed loop transient response
frequency domains. The time domain design is done using decay ratio of 0.25, or a phase lag of 1350 , may
root locus diagrams; it is, however, questionable that a compensate general, possibly delayed, stable or unstable
delayed process would be sufficiently well modelled by the processes (Hay 1998; Tan et al. 1999; Yu 1999; Prashanti
necessary second order model. The frequency domain and Chidambaram 2000; Tan et al. 2001; Robbins, 2002),
design is typically done using Bode plots to achieve a sometimes to achieve either a specified gain and/or phase
desired phase margin. Iterative methods for controller margin (Prashanti and Chidambaram 2000; Tan et al.
design provide a first approximation to desirable controller 2001) or a specified closed loop response (Vrancic et al.
parameters. 1999, 2001). The controller settings are easily calculated;
however, the system must generally be destabilised under
Tuning Rules proportional control, the empirical nature of the method
means that uniform performance is not achieved in general,
Process reaction curve tuning rules are based on several trials must typically be made to determine the
calculating the controller parameters, from the model ultimate gain, the resulting process upsets may be
parameters determined from the open loop process step detrimental to product quality, there is a danger of
response. This method was originally suggested by Ziegler misinterpreting a limit cycle as representing the stability
and Nichols (1942), who modelled the SISO process by a limit and the amplitude of the process variable signal may
first order lag plus delay (FOLPD) model, estimated the be so great that the experiment may not be carried out for
model parameters using a tangent and point method and cost or safety considerations.
defined tuning parameters for P, PI and PID controllers. Direct synthesis tuning rules result in a controller
Other process reaction curve tuning rules of this type are that facilitates a specified closed loop response. These
also described, sometimes in graphical form, to control methods include pole placement strategies and frequency
processes modelled by a FOLPD model (Shinskey 2001) domain techniques, such as gain margin and/or phase
or an integral plus delay (IPD) model (Hay 1998). The margin specification methods. Pole placement strategies
advantages of such tuning strategies are that only a single are described to compensate SISO processes modelled in
experimental test is necessary, a trial and error procedure is stable or unstable FOLPD form (Bi et al. 1999; Chien et al.
not required and the controller settings are easily 1999; Huang et al. 2000; Zhang and Xu 2000; Mann et al.
calculated; however, it is difficult to calculate an accurate 2001b), IPD form (Chen et al. 1999; Chien et al. 1999a) or
and parsimonious process model, load changes may occur SOSPD form (Wang et al. 1999a; Bi et al. 2000).
during the test which may distort the test results and a large Frequency domain based tuning rules are also described,
step input may be necessary to achieve a good signal to for processes modelled in stable or unstable FOLPD form
noise ratio. (Ho et al. 1998; Ho and Xu 1998; Chen et al. 1999a),
Performance (or optimisation) criteria, such as the stable or unstable SOSPD form (Ho and Xu 1998; Huang
minimisation of the integral of absolute error in a closed et al. 2000; Wang et al. 2001a), IPD form (Poulin and
loop environment, may be used to determine a unique set Pomerleau 1999; Cheng and Yu 2000), or more general
of controller parameter values. Tuning rules have been form (Yang and Shao 2000b).
described, sometimes in graphical form, to optimise either The presence of unmodelled process dynamics
the regulator response or the servo response, of a demands a robust design approach. The Internal Model
compensated SISO process, modelled in stable or unstable Control (IMC) design procedure, which allows uncertainty
FOLPD form (Wilton 1999; Majhi and Atherton 2000; on the process parameters to be specified, may be used to
Visioli 2001a), IPD form (Visioli 2001a), or stable or design appropriate PI and PID controllers for the
unstable second order system plus delay (SOSPD) form compensation of SISO processes modelled in stable or
(Wilton 1999; Kwak et al. 2000). Tuning rules to achieve unstable FOLPD form (Alvarez-Ramirez et al. 1998; Lee
specified servo and regulator responses simultaneously are et al. 1998; Chen et al. 1999b; Chun et al. 1999; Isaksson
also described (Tan et al. 1998; Yang and Shao 2000a). and Graebe 1999; Lee et al. 2000; Leva and Colombo
Ultimate cycle tuning rules are calculated from 2000; Rivera and Jun 2000; Marchetti and Scali 2000; Ho
the controller gain and oscillation period recorded at the et al. 2001), IPD form (Alvarez-Ramirez et al. 1998;
ultimate frequency (i.e. the frequency at which marginal Zhang et al. 1999; Rivera and Jun 2000) or stable and
unstable SOSPD form (Lee et al. 1998; Chen et al. 1999b; Control academics and practitioners remain
Rivera and Jun 2000). interested in the use of the PID controller to compensate
Tuning rules are easy to use, even in the absence processes with time delay. This paper provides a
of an accurate process model. These design methods are comprehensive summary of such compensation techniques
suitable for the achievement of a simple performance that have appeared in relevant journals since 1998. It is the
specification, for a compensated process with a non- hope of the author that the paper will provide a convenient
dominant delay. Comprehensive summaries of the tuning reference for application work. The work demonstrates that
rule formulae are available (O’Dwyer 2001). A summary new design techniques have been accumulating, each
of tuning rules published by year and medium is provided claiming that it is the best suited for the application. In
in Table 2; as this table shows, interest in the development general, there is a lack of comparative analysis with other
of tuning rules is growing. design techniques; associated with this is the lack of
benchmark examples for testing the different methods. The
Table 2: Tuning rules - publications by date (O’Dwyer, main priority for future research should be a critical
2001) analysis of available design techniques.
Year Journal Total
articles publications REFERENCES
1942-1951 3 3
1952-1961 5 6 Alvarez-Ramirez, J., A. Morales, and I. Cervantes. 1998.
1962-1971 5 8 “Robust proportional-integral control”, Industrial
Engineering Chemistry Research, 37, 4740-4747.
1972-1981 7 9 Astrom, K.J., H. Panagopoulos, and T. Hagglund. 1998. “Design
1982-1991 20 36 of PI controllers based on non-convex optimisation”,
1992-date 86 146 Automatica, 34, 1998, pp. 585-601.
Astrom, K.J. and T. Hagglund. 1995. PID controllers: theory,
1992-1997 46 73 design and tuning, Second Edition, Instrument Society of
1998-date 40 73 America.
Atherton, D.P. 1999. “PID controller tuning”, IEE Computing
and Control Engineering Journal, April, 44-50.
Analytical Techniques Bandyopadhyay, R. and D. Patranabis. 1998. “A fuzzy logic
based PI autotuner”, ISA Transactions, Vol. 37, 227-235.
Controller parameters may be determined using Bandyopadhyay, R. and D. Patranabis. 2001. “A new autotuning
analytical techniques. Some methods minimise an algorithm for PID controllers using dead-beat format”, ISA
appropriate performance index (Astrom et al. 1998; Ham Transactions, 40, 255-266.
and Kim 1998; Kookos et al. 1999; Liu and Daley 1999; Bi, Q., W.-J. Cai, E.-L. Lee, Q.-G. Wang, C.-C. Hang and Y.
Leva and Colombo 1999; He et al. 2000; Howell and Best Zhang. 1999. “Robust identification of first-order plus dead-
2000; Tan et al. 2000; Wang and Cluett 2000; Leva and time model from step response”, Control Engineering
Practice, 7, No. 1, 71-77.
Colombo 2001). Alternatively, a direct synthesis strategy Bi, Q., W.-J. Cai, Q.-G. Wang, C.-C. Hang, E.-L. Lee, Y. Sun,
may be used to determine the controller parameters. Such K.-D. Liu, Y. Zhang and B. Zou 2000. “Advanced controller
strategies may be defined in the time domain, possibly by auto-tuning and its application in HVAC systems”, Control
using pole placement (Atherton 1999; Daley and Liu 1999; Engineering Practice, 8, No. 6, 633-644.
Jung et al. 1999a, 1999b; Majhi and Atherton 1999) or in Blanchett, T.P., G.C. Kember and R. Dubay. 2000. “PID gain
the frequency domain, possibly by specifying a desired scheduling using fuzzy logic”, ISA Transactions, 39, No. 2,
gain and/or phase margin (Fung et al. 1998; Wang et al. 317-325.
1999b, 1999c; Grassi et al. 2001). Chen, C.-L., S.-H. Hsu and H.-P. Huang. 1999a. “Tuning PI/PD
controllers based on gain/phase margins and maximum closed
Robust methods may be used to design
loop magnitude”, Journal of the Chinese Institute of
analytically an appropriate PID controller (Huang and Chemical Engineers, 30, No. 1, 23-29.
Wang 2001; Wang et al. 2001b; Ge et al. 2002). Finally, Chen, C.-L., H.-P. Huang and C.-T. Hsieh. 1999b. “Tuning of
alternative design methods may be used to determine the PI/PID controllers based on specification of maximum
controller parameters, such as fuzzy logic (Bandyopadhyay closed-loop amplitude ratio”, Journal of Chemical
and Patranabis 1998, 2001; Blanchett et al. 2000; Xu et al. Engineering of Japan, 32, No. 6, 783-788.
1998, 2000; Li and Tso 1999, 2000; Mudi et al. 1999; Cheng, S.-L. and C. Hwang. 1998. “Designing PID controllers
Visioli 1999, 2001b; Wang et al. 1999d; Tao and Taur with a minimum IAE criterion by a differential evolution
2000; Mann et al. 2001a), genetic algorithms (Cheng and algorithm”, Chemical Engineering Communications, 170, 83-
115.
Hwang 1998; Mudi and Pal 1999) or neural networks
Cheng, Y.-C. and C.-C. Yu. 2000. “Nonlinear process control
(Huang et al. 1999; Sbarbaro et al. 2000; Shu and Pi using multiple models: relay feedback approach”, Industrial
2000). Engineering Chemistry Research, 39, No. 2, 420-431.
Analytical methods are suitable for the design of Chien, I.-L., H.-P. Huang and J.-C. Yang. 1999. “A simple
PI/PID controllers for non-dominant delay processes where multiloop tuning method for PID controllers with no
there are well-defined performance requirements to be proportional kick”, Industrial Engineering Chemistry
achieved. Research, 38, No. 4, 1456-1468.
Chun, D., J.Y. Choi and J. Lee. 1999. “Parallel compensation
CONCLUSIONS with a secondary measurement”, Industrial Engineering
Chemistry Research, 38, No. 4, 1575-1579.
Daley, S. and G.P. Liu. 1999. “Optimal PID tuning using direct
 search algorithms”, IEE Computing and Control Engineering Lee, Y., S. Park, M. Lee and C. Brosilow. 1998. “PID controller
Journal, April, 51-56. tuning to obtain desired closed-loop responses for SI/SO
Ender, D.B. 1993. “Process control performance: not as good as systems”, AIChE Journal, 44, 106-115.
you think”, Control Engineering, September, 180-190. Lee, Y., J. Lee and S. Park. 2000. “PID tuning for integrating and
Fung, H.-W., Q.-G. Wang and T.-H. Lee. 1998. “PI tuning in unstable processes with time delay”, Chemical Engineering
terms of gain and phase margins”, Automatica, 34, 1145- Science, 55, 3481-3493.
1149. Leva, A. and A.M. Colombo. 1999. “Method for optimising set-
Ge, M., M.-S. Chiu and Q.-G. Wang. 2002. “Robust PID point weights in ISA-PID autotuners”, IEE Proceedings –
controller design via LMI approach”, Journal of Process Control Theory and Applications, 146, No. 2, 137-146.
Control, 12, 3-13. Leva, A. and A.M. Colombo. 2000. “Estimating model mismatch
Grassi, E., K.S. Tsakalis, S. Dash, S.V. Gaikwad, W. MacArthur overbounds for the robust autotuning of industrial
and G. Stein. 2001. “Integrated system identification and PID regulators”, Automatica, 26, 1855-1861.
controller tuning by frequency loop-shaping”, IEEE Leva, A. and A.M. Colombo. 2001. “Implementation of a robust
Transactions on Control Systems Technology, 9, No. 2, 285- PID autotuner in a control design environment”,
293. Transactions of the Institute of Measurement and Control,
Ham, T.W. and Y.H. Kim. 1998. “Process identification using 23, No. 1, 1-20.
pulse response and proportional-integral-derivative controller Li, H.-X. and S.K. Tso. 1999. “Higher order fuzzy control
tuning with combined guidelines”, Industrial Engineering structure for higher order or time-delay systems”, IEEE
Chemistry Research, 37, 482-488. Transactions on Fuzzy Systems, 7, No. 5, 540-552.
Hay, J. 1998. Regeltechniek 1, Die Keure n.v., Brugge, Belgium, Li, H.-X. and S.K. Tso. 2000. “Quantitative design and analysis
197-198. of fuzzy proportional-integral-derivative control – a step
He, J.-B., Q.-G. Wang and T.-H. Lee. 2000. “PI/PID controller towards autotuning”, International Journal of Systems
tuning via LQR approach”, Chemical Engineering Science, Science, 31, No. 5, 545-553.
55, 2429-2439. Liu, G.P. and S. Daley. 1999. “Optimal tuning PID controller
Ho, W.K., K.W. Lim and W. Xu. 1998. “Optimal gain and phase design in the frequency domain with application to a rotary
margin tuning for PID controllers”, Automatica, 34, No. 8, hydraulic system”, Control Engineering Practice, 7, 821-830.
1009-1014. Luyben, W.L. 2000. “Tuning proportional-integral controllers for
Ho, W.K. and W. Xu. 1998. “PID tuning for unstable processes processes with both inverse response and deadtime”,
based on gain and phase-margin specifications”, IEE Industrial Engineering Chemistry Research, 39, No. 4, 973-
Proceedings - Control Theory and Applications, 145, No. 5, 976.
392-396. Majhi, S. and D.P. Atherton. 1999. “Autotuning and controller
Ho, W.K., T.H. Lee, H.P. Han and Y. Hong. 2001. “Self-tuning design for processes with small time delays”, IEE
IMC-PID control with interval gain and phase margins Proceedings – Control Theory and Applications, 146, No. 5,
assignment”, IEEE Transactions on Control Systems 415-425.
Technology, 9, No. 3, 535-541. Majhi, S. and D.P. Atherton. 2000. “Online tuning of controllers
Howell, M.N. and M.C. Best. 2000. “On-line PID tuning for for an unstable FOPDT process”, IEE Proceedings – Control
engine idle-speed control using continuous action Theory and Applications, 147, No. 4, 421-427.
reinforcement learning automata”, Control Engineering Mann, G.K.I, B.-G. Hu and R.G. Gosine. 2001a. “Two-level
Practice, 8, 147-154. tuning of fuzzy PID controllers”, IEEE Transactions on
Huang, C.-T., M.-Y. Lin and M.-C. Huang. 1999. “Tuning PID Systems, Man and Cybernetics – Part B: Cybernetics, 31, No.
controllers for processes with inverse response using artificial 2, 263-269.
neural networks”, Journal of the Chinese Institute of Mann, G.K.I., B.-G. Hu and R.G. Gosine. 2001b. “Time-domain
Chemical Engineers, 30, No. 3, 223-232. based design and analysis of new PID tuning rules”, IEE
Huang, H.-P., M.-W. Lee and C.-L. Chen. 2000. “Inverse-based Proceedings – Control Theory and Applications, 148, No. 3,
design for a modified PID controller”, Journal of the Chinese 251-261.
Institute of Chemical Engineers, 31, No. 3, 225-236. Marchetti, G. and C. Scali. 2000. “Use of modified relay
Huang, Y.J. and Y.-J. Wang. 2001. “Robust PID controller techniques for the design of model-based controllers for
design for non-minimum phase time delay systems”, ISA chemical processes”, Industrial Engineering Chemistry
Transactions, 40, 31-39. Research, 39, 3325-3334.
Isaksson, A.J. and S.F. Graebe. 1999. “Analytical PID parameter Mudi, R.K. and N.R. Pal. 1999. “A robust self-tuning scheme for
expressions for higher order systems”, Automatica, 35, 1121- PI- and PD-type fuzzy controllers”, IEEE Transactions on
1130. Fuzzy Systems, 7, No. 1, 2-16.
Isermann, R., Digital control systems Volume 1. Fundamentals, O’Dwyer, A. 2001. “PI and PID controller tuning rules for time
deterministic control, 2nd Revised Edition, Springer-Verlag, delay processes: a summary”, Technical Report AOD-00-01,
1989. Version 2, School of Control Systems, Dublin Institute of
Jung, C.L., H.K. Song and J.C. Hyun. 1999a. “A new direct- Technology, Ireland (Dec).
synthesis tuning method for PI-controllers”, The Canadian O’Dwyer, A. 2002. “PID compensation of time delayed
Journal of Chemical Engineering, 77, 180-185. processes: a full survey”, Technical Report AOD-00-05,
Jung, C.L., H.K. Song and J.C. Hyun. 1999b. “A direct synthesis Version 3, School of Control Systems, Dublin Institute of
tuning method of unstable first-order-plus-time-delay Technology, Ireland (Feb).
processes”, Journal of Process Control, 9, No. 3, 265-269. Poulin, E. and A. Pomerleau. 1999. “PI settings for integrating
Kookos, I.K., A.I. Lygeros and K.G. Arvanitis. 1999. “On-line PI processes based on ultimate cycle information”, IEEE
controller tuning for integrator/dead time processes”, Transactions on Control Systems Technology, 7, No. 4, 509-
European Journal of Control, 5, 19-31. 511.
Kwak, H.J., S.W. Sung and I.-B. Lee. 2000. “Stabilizability Prashanti, G. and M. Chidambaram. 2000. “Set-point weighted
conditions and controller design for unstable processes”, PID controllers for unstable systems”, Journal of the
Transactions of the Institute of Chemical Engineers, 78, Part Franklin Institute, 337, No. 2-3, 201-215.
A, 549-556. Rivera, D.E. and K.S. Jun. 2000. “An integrated identification
 and control design methodology for multivariable process Wilton, S.R. 1999. “Controller tuning”, ISA Transactions, 38,
system applications”, IEEE Control Systems Magazine, June, 157-170.
25-37. Xu, J.-X., Y.-M. Pok, C. Liu and C.-C. Hang. 1998. “Tuning and
Robbins, L.A. 2002. “Tune control loops for minimum analysis of a fuzzy PI controller based on gain and phase
variability”, Chemical Engineering Progress, January, 68-70. margins”, IEEE Transactions on Systems, Man and
Sbarbaro, D., J.P. Segovia, S. Alcozer and J. Gonzales. 2000. Cybernetics - Part A: Systems and Humans, 28, 685-691.
“Applications of radial basis network technology to process Xu, J.-X., C.-C. Hang and C. Liu. 2000. “Parallel structure and
control”, IEEE Transactions on Control Systems Technology, tuning of a fuzzy PID controller”, Automatica, 36, 673-684.
8, No. 1, 14-22. Yang, Y.G. and Shao, H.-H. 2000a. “PID auto-tuner based on
Shinskey, F.G. 2001. “PID-deadtime control of distributed sensitivity specification”, Transactions of the Institute of
processes”, Control Engineering Practice, 9, 1177-1183. Chemical Engineers, 75, Part A, 312-316.
Shu, H. and Y. Pi. 2000. “PID neural networks for time-delay Yang, Y.-G. and Shao, H.-H. 2000b. “Optimal tuning for PI
systems”, Computers and Chemical Engineering, Vol. 24, controller”, Automatica, 36, 147-152.
859-862. Yu, C.-C. 1999. Autotuning of PID controllers, Advances in
Tan, K.K., Q.G. Wang, C.C. Hang and T.J. Hagglund. 1999. Industrial Control Series, Springer-Verlag London Ltd.
Advances in PID control, Advances in Industrial Control Zhang, W., X. Xu and Y. Sun. 1999. “Quantitative performance
Series, Springer-Verlag London. design for integrating processes with time delay”,
Tan, K.K., T.H. Lee and X. Jiang. 2000. “Robust on-line relay Automatica, 35, 719-723.
automatic tuning of PID control systems”, ISA Transactions, Zhang, W. and X. Xu. 2000. “On the model reference method for
39, No. 2, 219-232. unstable time delay systems”. Computers and Chemical
Tan, K.K., T.H. Lee and X. Jiang. 2001. “On-line relay Engineering, 24, 2765-2766.
identification, assessment and tuning of PID controller”, Ziegler, J.G. and N.B. Nichols. 1942. “Optimum settings for
Journal of Process Control, 11, pp. 483-496. automatic controllers”, Transactions of the ASME, 64, 759-
Tan, W., J. Liu and P.K.S. Tam. 1998. “PID tuning based on 768.
loop-shaping H ∞ control”, IEE Proceedings - Control
Theory and Applications, 145, No. 6, 485-490. AUTHOR BIOGRAPHY
Tao, C.W. and J.S. Taur. 2000. “Flexible complexity reduced
PID-like fuzzy controllers”, IEEE Transactions on Systems, AIDAN O’DWYER was born in Portumna, Co. Galway,
Man and Cybernetics – Part B: Cybernetics, 30, No. 4, 510- Ireland and was educated at Portumna College, University
516. College Galway, and Dublin City University, from which
Visioli, A. 1999. “Fuzzy logic based set-point weight tuning of he received B.E., M.Eng. and Ph.D. degrees in 1984, 1992
PID controllers”, IEEE Transactions on Systems, Man and
and 1996, respectively. Since 1990, he has been a lecturer
Cybernetics – Part A: Systems and Humans, 29, No. 6, 587-
592. in electrical engineering at Dublin Institute of Technology.
Visioli, A. 2001a. “Optimal tuning of PID controllers for integral His research interests are the identification and control of
and unstable processes”, IEE Proceedings – Control Theory time delay systems. He is the author or co-author of thirty-
and Applications, 148, No. 2, 180-184. four published papers on these topics.
Visioli, A. 2001b. “Tuning of PID controllers with fuzzy logic”,
IEE Proceedings – Control Theory and Applications, 148,
No. 1, 1-8.
Vrancic, D., Y. Peng and S. Strmcnik. 1999. “A new PID
controller tuning method based on multiple integrations”,
Control Engineering Practice, 7, No. 5, 623-633.
Vrancic, D., S. Strmcnik and D. Juricic. 2001. “A magnitude
optimum multiple integration tuning method for filtered PID
controller”, Automatica, 37, 1473-1479.
Wang, L. and W.R. Cluett. 2000. From plant data to process
control, Taylor and Francis Ltd.
Wang, Q.-G., T.-H. Lee, H.-W. Fung, Q. Bi and Y. Zhang.
1999a. “PID tuning for improved performance”, IEEE
Transactions on Control Systems Technology, 7, No. 4, 457-
465.
Wang, Q-G., H.-W. Fung and Y. Zhang. 1999b. “PID tuning
with exact gain and phase margins”, ISA Transactions, 38,
243-249.
Wang, Q.-G., C.-C. Hang, S.-A. Zhu and Q. Bi. 1999c.
“Implementation and testing of an advanced relay auto-
tuner”, Journal of Process Control, 9, No. 4, 291-300.
Wang, L., M.L. Desarmo and W.R. Cluett. 1999d. “Real-time
estimation of process frequency response and step response
from relay feedback experiments”, Automatica, 35, 1427-
1436.
Wang, Q.-G., Y. Zhang and X. Guo. 2001a. “Robust closed-loop
identification with application to auto-tuning”, Journal of
Process Control, 11, 519-530.
Wang, Q.-G., C.C. Hang and X.-P. Yang. 2001b. “Single-loop
controller design via IMC principles”, Automatica, 37, 2041-
2048.
 A COMPARISON BETWEEN A SELF-TUNING FLC AND A FLC WITH GAIN
SCHEDULER FOR VARIABLE GAIN
Denis Herlihy & Dr. Barry O’Connor
AMT Ireland
University College Cork
[email protected]
[email protected]

ABSTRACT In this paper, the problem of variable deadtime in a process

is addressed. Two FLC’s are designed to minimise this
Ebrahim Mamdani constructed the first Fuzzy Logic problem followed by a performance comparison between
Controller (FLC) in 1973 and since then FLC’s are both techniques. Consider a liquid level control application
becoming increasingly popular in process industries. This shown in block diagram form in Figure 1.
paper proposes the design of two different FLC’s to control
the liquid level in a tank that possesses variable deadtime.
An attractive feature of this closed loop system would be to
minimise the deadtime. It will be shown that this can be
achieved in FLC’s by scaling the control signal of the
controller, by means of a controller gain. Two designs are
proposed in this paper to attain minimisation of variable
Figure 1
deadtime, namely a Self-tuning FLC and a FLC with Gain
A suitable sensor is used to measure the actual liquid level
Scheduler. The Self-tuning FLC consists of two FLC’s, the
in the process (tank) and it is then compared to the setpoint.
first FLC outputs the control signal for the process and the
The error is evaluated and fed to the controller, which
second FLC outputs a value (α), which further scales the calculates the appropriate control signal to the processes
control signal. This value of α is determined from a rule- pump. This forces it to increase or decrease its flow rate
base defined on both error (e) and gradient of liquid level depending on the error magnitude, maintaining liquid level
(∆e) of the controlled variable. The FLC can be tuned by at its desired value.
dynamically adjusting the controller gain in each sampling
instance by a scaling factor α. For the second FLC (with It can be easily shown that a simple PD-type FLC will
Gain Scheduler) design, it will be shown that the effect of sufficiently control this process. However, the presence of
deadtime is minimised using an automatic procedure. deadtime causes serious performance degradation. These
Which estimates the deadtime on start-up and an include longer rise and settling times, larger overshoot and
appropriate gain is then selected from the gain scheduler to even the possibility of unwanted oscillation Varying
minimise the effect of this deadtime. This paper presents deadtime causes a greater problem still, which occurs from
the design of both FLC’s followed by a performance manufacturing discrepancies, new product development and
comparison between them. Several performance criteria physical constraints placed on sensors at different locations.
such as, rise time, settling time, disturbance rejection and Several control design methods for systems with varying
Integral Absolute Error will be employed. deadtime have appeared in recent literature including forms
of self-tuning and self-organising FLC’s (HE 1993; Maeda
INTRODUCTION and Murakami 1992; Procyk and Mamdani 1979; Shao
1988). Further research employed neural networks and
Fuzzy control denotes the field in control engineering in genetic algorithms to design such controllers, to achieve
which fuzzy set theory and fuzzy inference are used to more improved performance and increased robustness
derive control laws. In 1973, Ebrahim Mamdani (Chen and Chen 1994; Homaifar and McCormick 1995;
constructed the first Fuzzy Controller to control a steam Karr and Gentry 1993; Li 1997). Also MathWorks have a
engine and since then FLC’s have been rapidly gaining function in their Fuzzy Logic Toolbox called ANFIS, which
popularity among practicing engineers. This increased tunes the membership functions of the controller off-line
popularity can be attributed to the fact that fuzzy logic using input/output training data.
provides a powerful vehicle that allows engineers to
incorporate human reasoning in the control algorithm. This One method of minimising deadtime is to adjust the
characteristic of FLC’s is further utilized in process magnitude of the control signal, by varying the controller
industries where it is sometimes difficult to obtain an gain. In practice a skilled human operator manipulates this
accurate plant model. controller gain in an ad hoc fashion based on the current
process state to achieve optimal control (Mudi and Pal
2000). However, this paper proposes two automated
 Figure 2: PD-Type FLC Block Diagram

procedures to tune the controller gain on-line, listed as (e), difference between desired and actual liquid height and
RATE (∆y) rate of change of process output. Furthermore,
1. A FLC with Gain Scheduler (FLCGS) one output referred to as VOLTAGE (u(k)), corresponds to
2. A Self-tuning PD-type FLC (STFLC) the input voltage to the pump. The linguistic terms decided
for the input and output variables were:
The FLCGS system consists of a PD-type FLC and an
automatic procedure. That measures the deadtime on start- Error = {large_neg, med_neg, okay, med_pos,
up followed by a gain scheduler that selects the appropriate large_pos}
gain for the measured deadtime. On the other hand, the
STFLC system is slightly more complex. The controller Rate ={neg_fast, negative, okay, positive,
gain is adjusted on each sampling instant using a scaling pos_fast}
factor (α), the value of which is determined by a second
FLC. Voltage = {very_small, small, no_change, high,
very_high}
Both controllers are simulated using a mathematical model
of the CE105 Coupled Tank Apparatus from TecQuipment Fuzzy sets must now be associated with each of these
Inc., which mimics liquid level control. linguistic terms, whereupon the designer selects the size
and shape of these fuzzy sets based on apriori or heuristic
This paper is organised as follows: Section 2 shows the knowledge of the plant. Figures 4, 5 and 6 show the fuzzy
controller design procedure and performance evaluation of sets for the inputs and output. The above linguistic terms
the FLCGS with different values of deadtime. Design and then use a set of If-Then rules (rule-base) that are specified
testing of the STFLC is seen in Section 3, again with to emulate the designer’s knowledge of the process and its
different values of deadtime. Following this, Section 4 appropriate control. In this case, nine rules are required to
compares the performance of both controllers. Finally, efficiently track the setpoint, which state the output actions
conclusions are presented in Section 5. to be performed depending on the status of the input
variables. These rules can be specified as follows:

FLCGS DESIGN If error is large_pos Then voltage is very_high

If error is med_pos Then voltage is very_fast
The FLCGS is designed using the Fuzzy Logic Toolbox in If error is okay Then voltage is no_change
MATLAB and is simulated using Simulink, a software If error is med_neg Then voltage is very_small
package for modelling, simulating, and analysing If error is large_neg Then voltage very_small
dynamical systems. However a mathematical model of the If error is okay and rate is positive Then voltage is small
CE105 Coupled Tank Apparatus is obligatory when If error is okay and rate is negative Then voltage is high
simulating the designed FLCGS. Using classical system If error is okay and rate is pos_fast Then voltage is small
identification techniques outlined in (Ljung 1987) the If error is okay and rate is neg_fast Then voltage is high
mathematical model for the process (ignoring deadtime)
was found to be, With these rules designed, the PD-type FLC is now
implemented and tested using the mathematical model of
G(s) = 4.245 the process. The resultant process response of the closed
140s + 1 loop system is depicted in Figure 7. It can be seen that good
setpoint tracking is achieved after a rise time of
In order to design the FLCGS, a PD-type FLC must be approximately 60 seconds. However, such a satisfactory
designed initially to control the process without deadtime, performance with the PD-type FLC requires a good
Figure 2 shows the block diagram of the PD-type FLC. In a understanding of the physics of the underlying system.
fuzzy control system, both input and output data are
represented by linguistic variables, which are intuitively
expressed by natural language words (Ross 1995; Jamshidi
et al. 1993). In this system the FLC has two inputs, ERROR
 Figure 3: FLCGS Block Diagram

In processes where the deadtime is known apriori a

controller gain (scaling factor) can placed after the
controller and can be tuned off-line to minimise this
deadtime. However, this gives poor performance in the case
where the deadtime varies.

Here, an automatic procedure will be used to measure the

variable deadtime on start-up and is then requested to
minimise the measured deadtime using a gain scheduler.
Figure 4: Input variable Error
A simple If-Then procedure is used to measure the
deadtime. The process output is monitored on
commencement of the process and the deadtime is
estimated. The If-Then procedure to perform such
estimation is as follows,

Count = 0
If (process output = 0) then
Count = Count + 1
Endif
Figure 5: Input variable Rate
Deadtime = Count

This value of deadtime is then fed to the gain scheduler,

which switches to the appropriate gain for that specific
deadtime. Table 1 contains the appropriate gains for their
respective deadtime. The values of these gains are found
through trial and error.

Table 1
Figure 6: Output variable Voltage
Deadtime Gain
0 1
1 0.8
2 0.6
3 0.29
4 0.22
5 0.19
6 0.17
7 0.165

Figure 7: Response of PD-type FLC

Simulated output responses for the process with deadtimes
of 1 and 3 seconds can be seen in Figures 8 and 9
respectively.

Figure 10: Input variable Error

Figure 8: Response of FLCGS, deadtime = 1s

Figure 11: Input variable Rate

Figure 12: Output variable α

As can be seen form Figure 13, there are three gains

Figure 9: Response of FLCGS, deadtime = 3s (scaling factors) in the STFLC system, two input and one
output. Determination of suitable values for these gains is
made based on designer’s knowledge of the workings of the
STFLC DESIGN process and sometimes through trial and error. The
relationship between the gains and the input and output
Similar to the pervious FLC technique, the STFLC system variables of the STFLC are as follows:
is designed using MATLAB and Simulink. Two FLC’s are
required for the STFLC system, a PD-type FLC as the main En = Ge . e
control strategy and a FLC (α-FLC) to compute the value Yn = Gy . ∆y
of the scaling factor α. The PD-type FLC used is identical Un = (φUn . Gu) . α
to the one designed in Section 2, same input and output
membership functions and rules apply. The controller output for the STFLC is derived using the
actual scaling factor (α . Gu). The value of α is determined
Designing the α FLC is similar to that of the PD-type FLC, from a rule-base, whose inputs are ERROR and RATE and
in that it has the same two inputs, e and ∆y but a different its output is α as described in the previous section.
output α. The shape and domain of the e and ∆y
membership functions are the same as the PD-type FLC but When designing this rule-base, it is important to consider
their size differs slightly, as can be seen from Figures 10 the classical closed loop performance trade-off. For
and 11. The membership functions for the output α are example, setting the controller gain at a small value when
symmetric triangles with equal base and 50% overlap with the error is very big achieves a reduced overshoot and
neighbouring membership functions, with a domain of [0, settling time, but must be coupled with an increase in rise
1] as shown in Figure 12. The linguistic terms associated time. In this situation the fuzzy rule will be as follows,
with the input variables are the same as that of the PD-type
FLC but the linguistic terms for the output variable α are: If error is large_pos and rate is negative then α is VS

α= {ze, very_small, small, small_big,

med_big, big, very_big}
 Figure 13: STFLC Block Diagram

Tuning the Controller

The tuning of the STFLC requires a number of design steps

given by,

1. The scaling factors of the STFLC must be tuned first.

This is achieved by disconnecting the self-tuning FLC
and adjusting the three scaling factors to yield a
satisfactory output response.

2. The self-tuning FLC is then reconnected and the value

of Gu scaling factor is increased to three times its value
in step 1, keeping the values of the other two scaling Figure 14: Response of STFLC, deadtime = 0s
factors the same. This value of Gu may be slightly
adjusted to improve the performance of the controller
at this point. At the end of this step a STFLC has been
designed and tuned.

3. Depending on the response required, changes could be

made to the rule base to achieve the output response
needed. For example, if reduced overshoot is required,
the value of α should be kept very small up to the
medium values of α.

The fully designed STFLC was then tested using the

mathematical model of the Tank Apparatus described in Figure 15: Response of STFLC, deadtime =0.5s
section 1. Figures 14, 15 and 16 show the output responses
of the process with deadtimes of 0s, 0.5s and 1.5s
respectively. These figures illustrate the effectiveness of the
STFLC to efficiently track the setpoint. It was decided for
these simulations to design the rules to keep the percentage
overshoot as minimum as possible. However, this had the
adverse effect of increasing the rise times.

Figure 16: Response of STFLC, deadtime = 1.5s

PERFORMANCE COMPARISON The controllers were tested with three values of deadtime,
i.e. 1s, 2s & 3s, a constant input and a load step disturbance
The performance of the FLCGS is compared with the of magnitude 0.3 at 150s. In Figures 17, 18 & 19 the solid
STFLC for the tank apparatus process. Several performance curves represent the responses due to the STFLC while the
metrics such as, rise time, settling time, disturbance dashed curves depict the responses due to the FPIC. Table 2
rejection and integral absolute error (IAE) are used to illustrates the resulting performance metrics of both
illustrate the difference between both controllers. The controllers with each value of deadtime. The performance
results of each metric will be tabulated for both controllers of the STFLC exceeds that of the FLCGS for all operating
simulated with different values of deadtime. The conditions. One of the main differences to be observed is
performance of both controllers will then further compared that the rise times of the STFLC did not vary with
with load disturbances. deadtime. However, the rise times of the FLCGS degraded
with increasing deadtimes.

Table 2: Performance Analysis of the STFLC & FLCGS

FLC Deadtime (s) Rise Time (s) Settling Time (s) Time to Reject IAE
Disturbances (s)
FLCGS 1 41 54 33 31.41
STFLC 1 30 43 29 23.4
FLCGS 2 41 70 35 32.58
STFLC 2 30 45 30 24.78
FLCGS 3 83 125 66 57.75
STFLC 3 30 55 41 26.64

Figure 17: Response of the STFLC & FLCGS, deadtime = 1s

Figure 18: Response of the STFLC & FLCGS, deadtime = 2s

Figure 19: Response of the STFLC & FLCGS, deadtime = 3s

CONCLUSIONS He, S.Z. Tan, S. Xu, F.L. Wang, P.Z. 1993. “Fuzzy self-
tuning of PID controller.” Fuzzy Sets and
This paper proposed the design of two different FLC’s, the Systems, 56, 37-46.
FLCGS and the STFLC, to minimise the effect of variable Homaifar, A. and McCormick, E. 1995. “Simultaneous
deadtime on a process. Minimisation of deadtime can be design of membership functions and rule sets for
achieved in FLC’s by scaling the control signal of the fuzzy controllers using genetic algorithms.” IEEE
controller, by means of a controller gain. The two designs Trans. Fuzzy Systems 3, (2), 129-139.
proposed use this technique, but the attractiveness of both Jamshidi, M. et al. 1993. “Fuzzy Logic And Control.” Prentice-
designs is that they tune the controller gain on-line. The Hall, Inc., NJ.
FLCGS measures the deadtime on process start-up and Karr, C.L. and Gentry, E.J. 1993. “Fuzzy control of PH using
selects the appropriate gain to yield minimisation of the genetic algorithm.” IEEE Trans. Fuzzy Systems 1, (1),
46-53.
measured deadtime. On the other hand, the STFLC uses Li, W. 1997. “A method for design of a hybrid neuro-fuzzy
two FLC’s, the first produces the control signal, while the control system based on behaviour modelling.” IEEE
other outputs a value of α, which scales the control signal Trans. Fuzzy Systems 5, (1), 791-797.
on each sampling instant. Both the controllers successively Ljung, L. 1987. “System Identification.” Prentice-Hall, Inc., NJ.
controlled the process for different values of deadtime. Maeda, M and Murakami, S. 1992. “A self-tuning fuzzy
Results from a performance comparison between both controller.” Fuzzy Sets and Systems 51, 29-40.
controllers indicated that the STFLC exhibited better Mudi, R.K and Pal, N.R. 2000. “A self-tuning PI controller.”
quality performance than that of the FLCGS. Fuzzy Sets and Systems, 115, 327-338.
Procyk, T.J. and Mamdani, E.H. 1979. “A linguistic self-
organising process controller.” Automatica 15, (1), 53-
REFERENCES 65.
Ross, T. J. 1995. “Fuzzy Logic with Engineering Applications.”
Chen, C.L. and Chen, W.C. 1994. “Fuzzy controller design Mc Graw Hill, E. U.
by using neural network techniques.” IEEE Trans. Shao, S. 1988. “Fuzzy self-organising control and its application
Fuzzy Systems 2, (3), 235-244. for dynamical systems.” Fuzzy Sets and Systems, 26,
151-164.
MODELLING
 MODELLING THE HYDRATION OF FOODSTUFFS

A.H. Weerts1, D.R. Martin1, G. Lian1*, J.R. Melrose1

1
Unilever R&D Colworth
Sharnbrook, Bedford
MK44 1LQ Bedfordshire
United Kingdom
*
Corresponding author: guoping [email protected]

KEYWORDS INTRODUCTION
rehydration, porous media, mass transfer,
NMR measurements, biomaterials Dried foodstuffs often need to be rehydrated
before they are consumed. It is desirable that
ABSTRACT these foodstuffs hydrate as fast as possible and
The rehydration kinetics of dried foodstuffs is show adequate structural and chemical
of critical importance to their sensory characteristics. Information about water
properties and delivery of flavour and absorption as a function of temperature of
functional molecules. Based on the dynamics those food materials is critically important to
of capillary flow in partially saturated porous their shelf life and product usage. Rehydration
media, a finite element model is developed to of food materials also has an important Impact
predict the infiltration of water into dried food on their nutritional and sensorial properties.
products taking into account temperature A number of studies have been reported to
effects. The finite element model is based on model the hydration kinetics of foodstuffs and
the mixed form of the mass conservation different types of models have been used. Two
equation. The constitutive relationships of main approaches can be identified. One
water retention and hydraulic conductivity are approach uses the empirical and semi-
adopted from the fields of hydrology and soil empirical models like for instance the Peleg
science. The transfer properties of water in the and the Weibull equation (Machado et al.,
porous medium depend on the moisture 1999; Machado et al., 1998; Lim et al., 1995;
content and the microstructure. This is contrast Sopade et al., 1992; Sopade and Obekpa,
to the constant transfer properties often used in 1990; Peleg, 1988). The other approach
the heat and mass transfer models developed employs diffusive models based on Fick’s
for foods. Rehydration of green tea as a second law of diffusion (Sanjuan et al., 1999;
function of Temperature has been simulated Sanjuan et al., 2001; Thorvaldsson and
and results are compared with NMR Janestad, 1999; Simal et al., 2000; Hsu, 1983).
measurements. There is good agreement. Despite numerous studies using Fick’s law to
model liquid water transport in porous
foodstuffs, the liquid water movement cannot mass conservation equation (Celia et al. 1991;
be simply defined as a diffusion process. Weerts et al., 2002ab)
Hydration occurs by capillary flow, driven by ∂θ
= ∇ ⋅ (K∇h ) − ∇ ⋅ ( K sin ϕ ) (1)
an energy potential gradient, rather then by ∂t
diffusion. Some studies using capillary flow where K (m/s)
approach to model hydration and/or drying of kk r ρ w g
K= (2)
foodstuffs have been reported recently (Ni et µw
al., 1999; Feng et al., 2001; Irudayaraj and
where k is the intrinsic permeability (m2), kr is
Wu, 1999; Turner and Jolly, 1991; Lian et al.,
the relative water permeability, where ρw is the
1997). However, the capillary flow approach is
density of water (1000 kg/m3), g is gravity (-
still not widely used.
9.81 m2/s) and µw (Pa.s) is the viscosity of
The objective of this work is to show the
water, h is the pressure head (m), or pressure
feasibility of modelling the rehydration
in equivalent water column and ϕ is the angle
process of foodstuffs using the capillary flow
away from the horizontal.
approach applying constitutive relationships
often used in soil science and hydrology. In
Constitutive Relationships
particular, the effect of temperature is
modelled. We first demonstrate that the effect
To solve the equation of water transfer in
of temperature on hydration can be directly
partially-saturated porous media, constitutive
taken into account in the constitutive
relationships describing the relative water
properties. A finite element model is then
permeability and capillary potential are
proposed. As a model system of porous
needed. Heat and mass transfer models
foodstuff, we use dried green tea leaf material
developed for foods often assume simple
(Sencha). Hydration behavior of green tea at
Fick’s diffusion theory applies. As a result,
various temperatures is investigated using time
empirically-fitted weak constitutive
domain NMR. Consequently we compare our
relationships are often used. Here, the
model predictions with the experimental data
constitutive relationships adopted are
of green leaf tea hydration. For a range of
physically-based and characteristic to porous
temperatures studied, we find good agreement
media. We first introduce the water retention
between the model prediction and the
curves most often used in soil physics and
experimental data. More details of this work
hydrology. The water retention curve describes
can be found in Weerts et al. (2002ab).
capillary potential of a porous medium as a
function of moisture content and temperature.
THEORY
One of the commonly used equations in soil
science is given by van Genuchten et al.
Capillary flow model
(1980)
θ 1
We use the following mixed form of the = (3)
θ s [1 + αh n ] m
unsaturated flow equations derived from the
where h is the pressure head (m), θ is the
volumetric water content (m3/m3), θs (m3/m3) Tea leaf model
is the saturated water content assumed to be
equal to the porosity φ, α (m-1) and n (>1) are The hydration of green tea was measured at

two parameters defining the shape of the water different temperatures (288, 298, 308, 318,

activity curve, and m=1-1/n. The water 328,338,348 K). Computer simulations of

retention curve can be related to the water hydration have been also performed at these

activity curve (Campbell et al., 1993; Gee et temperatures. The tea leaf is modelled as a two

al., 1992) by using the Kelvin Equation: dimensional block of 0.0002 m by 0.0025 m,

RT representing the average thickness and size of

h= ln( Aw ) (4)
tea leaves.
ρ w gM w
where Aw is the water activity, R is the gas
Water activity model parameters
constant and Mw is the molecular mass of
water. The value of R/Mw is 0.461 MPa/K.
Water activity curves of tea were taken from
According to Mualem (1976), the relative
Panchariya et al. (2001). Using the least square
permeability may be modelled by the
method, the data were fitted to equation (3) to
following relationship as a function of water
obtain the values of α and n. Extrapolation and
content and porosity
2
interpolation between these water retention
0.5   1/ m

m

θ    θ    parameters as a function of temperature
k r =   1 − 1 −  θ   (5)
θs     s    yielded the water retention curves at the
   temperatures of the hydration data.
The viscosity of water depends on
temperature. The relationship may be Hydraulic conductivity
approximately described as follows (Thomas
and King, 1992) From our study on the hydration of black tea
µ w = 661.2 (T − 229) −1.562
× 10 −3
(6) as a function of tea leaf size (Weerts et al.,
2002b), we know that the hydraulic
where T is the temperature in Kelvin.
conductivity in the horizontal direction (from
the cut edges of the leaf) is much greater than
MATERIAL AND METHODS
that in the vertical direction (top-bottom),
about 100-250 times. Here we assumed that
Leaf tea hydration data
the hydraulic conductivity in the horizontal
directions is 125 times greater than that in the
Isothermal hydration experiments have been
vertical direction. The hydraulic conductivity
performed on green tea using Time Domain
at saturation is calibrated on the hydration data
NMR. which is a well established technique
obtained at 25 0C, the only temperature at
for the measurement of rehydration kinetics of
which the water activity curve was also
porous systems such as dried foods (Hills,
measured. Hydraulic conductivity at other
1998).
 temperatures was directly predicted by
3

substituting the viscosity equation (6) into (2). T=298 K

)
f
ae
l
ae
t
y 2
r
d.
gr
re/
7
10
measurements t
model a
w.
10
6
gr( 1
n
oi
) t
ar
m ( 105 d
d y measurements
ae h
model
h
er 0
0 200 400 600 800 1000
us 104
es time (s)
pr
h
, Figure 2. Comparison of the modelled (dots)
3
10 and measured (circles) rehydration data on
green tea at T=298 K as a function of time
10
2
Temperature is indicated in the graph.
0 0.1 0.2 0.3
3 3
θ , volumetric water content (m /m )
RESULTS AND DISCUSSION experimental measurements, suggesting that,
by using the physically-based constitutive
Figure 1. Measured water activity data relationships, the effect of temperature can be
(Panchariya et al., 2001) and fitted water
retention function as a function of water directly predicted with good accuracy. The
content at T=298 K. only model parameters that varied with
temperature are the viscosity and activity (i.e.
Figure 1 shows the measured water activity
surface tension) of water. Their variations with
data compared with the fitted water retention
temperature are well documented and are
equation proposed by van Genuchten et al.
given by simple equations.
(1980) at T=298K. The model fits the data
The experimental data of leaf tea hydration
well. Results at the other temperatures are
showed moderate variations in the equilibrium
similar. Figure 2 shows a comparison of the
moisture content. There is no clear trend in
modelled and measured rehydration data at
terms of temperature effect. In the model, we
T=298K. The data at this temperature was
assumed that the equilibrium moisture content
used to calibrate the values of the hydraulic
was constant for all temperatures. This has
conductivity. The modelled hydration curve
resulted in some under and over predictions in
compares well with the measured data. The
the equilibrium moisture uptake at some
calibrated value of the hydraulic conductivity
temperatures. However, the differences are
at saturation is 10-11 m/s. This value of
very small (<10%). Experimental studies on
hydraulic conductivity is comparable to the
other foods stuffs showed that the equilibrium
value obtained by Weerts et al. (2002b) for
moisture content varied (decreased) as a
black tea.
function of temperature (Abu-Ghannam, 1998;
Figure 3 shows the predicted hydration of
Lim et al., 1995; Sanjuan et al., 1999). From
green tea at T=288 and T=308, 318, 328 , 338,
the current hydration data of leaf tea, the effect
and 348 K. The measured data are also given
of temperature on the equilibrium water uptake
in the figure for comparison. The predicted
is not clear.
hydration curves agree well with the
 current model can easily be applied to model
3
T=288 K T=308 K the hydration and dehydration of foodstuffs.
2

) 1 Moreover, this approach can easily be

aef
ae
l 0 extended to include the coupling with heat and
t
y 3
r T=318 K T=328 K
d
. solute transport
gr 2
/
ert 1
a
w
. 0
gr(
n
LITERATURE CITED
oi 4
t T=338 K T=348 K
ar
d
y
3 Abu-Ghannam,N., Modelling textural changes in
h 2
measurements
the hydration process of red beans, Journal of
1 model Food Engineering, 38, 341-352 (1998).
0
0 200 400 600 800 1000 0 200 400 600 800 1000 Campbell,G.S., J.D. Jungbauer, S. Shiozawa and
time (s) time (s)
R.D. Hungerford, A one-parameter equation for
water sorption isotherms of soils, Soil Sci., 156,
Figure 3 Comparison of the model predicted 302-305 (1993).
(dots) and measured (circles) rehydration data Celia,M.A. and P.Binning, A mass conservative
on green tea at T=288 and T=308-348 K as a numerical-solution for 2-phase flow in porous-
media with application to unsaturated flow,
function of time.
Water Resour.Res., 28, 2819-2828 (1992).
Celia,M.A., E.T.Bouloutas and R.L.Zarba, A
CONCLUSIONS general mass-conservative numerical-solution
for the unsaturated flow equation, Water
Resour.Res., 26, 1483-1496 (1990).
Chen,P.S. and D.C.T.Pei, A mathematical-model
Hydration of foods and biomaterials is greatly of drying processes, Int.J.Heat Mass Transfer,
influenced by temperature. Presented in this 32, 297-310 (1989).
Feng,H., J.Tang, R.P.Cavalieri and O.A.Plumb,
paper is a mixed form of the unsaturated flow Heat and mass transport in microwave drying
of porous materials in a spouted bed, AIChE J.,
equation based on the capillary flow in porous 47, 1499-1512 (2001).
media. Using this approach, the hydration of Gee,G.W., M.D.Campbell, G.S.Campbell and
J.H.Campbell, Rapid measurement of low soil-
green tea at a range of temperatures has been water potentials using a water activity meter,
Soil Sci.Soc.Am.J., 56, 1068-1070 (1992).
successfully modelled. The predicted Hills, B. Magnetic resonance in food science, J.
hydration curves agree well with the Wiley and Sons, New York (1998).
Hsu,K.H., A diffusion-model with a concentration-
experimental data of NMR. The hydraulic dependent diffusion- coefficient for describing
water-movement in legumes during soaking,
conductivity of green tea was estimated to be J.Food Sci., 48, 618-&(1983).
10-11 m/s at 298 K in the horizontal direction Irudayaraj,J. and Y.Wu, Finite-element analysis of
coupled heat, mass, and pressure transfer in
(from the cut edges of the leaves). The porous biomaterials, Numerical Heat Transfer
Part A-Applications, 26, 337-350 (1994).
hydraulic conductivity in the vertical direction Irudayaraj,J. and Y.Wu, Heat and mass transfer
is 8.0x10-14 m/s. coefficients in drying of starch based food
systems, J.Food Sci., 64, 323-327 (1999).
The mixed form capillary flow model Lian,G., C.S.Harris, R.Evans and M.Warboys,
Coupled heat and moisture transfer during
presented in this paper directly predicts the microwave vacuum drying, Journal of
effect of temperature on hydration. Increase in Microwave Power and Electromagnetic Energy,
32, 34-44 (1997).
the hydration rate with temperature is due to Lim,L.T., J.M.Tang and J.S.He, Moisture sorption
characteristics of freeze-dried blueberries,
mainly viscosity and surface tension of water J.Food Sci., 60, 810-814 (1995).
contained in the physically-based constitutive Lin,T.M., T.D.Durance and C.D.Scaman,
Characterization of vacuum microwave, air and
relationships. The current model is apparently freeze dried carrot slices, Food Res.Int., 31,
111-117 (1998).
advantageous over the models based on so- Machado,M.D., F.A.R.Oliveira, V.Gekas and
called Fick’s second law of diffusion. The R.P.Singh, Kinetics of moisture uptake and
soluble-solids loss by puffed breakfast cereals
 immersed in water, International Journal of Van Genuchten,M.T., A closed-form equation for
Food Science and Technology, 33, 225-237 predicting the hydraulic conductivity of
(1998). unsaturated soils, Soil Sci.Soc.Am.J., 44, 892-
Machado,M.F., F.A.R.Oliveira and L.M.Cunha, 898 (1980).
Effect of milk fat and total solids concentration Weerts, A.H., G. Lian, D. Martin, Modelling of
on the kinetics of moisture uptake by ready-to- hydration of foodstuffs: Temperature effects,
eat breakfast cereal, International Journal of Submitted to AIChE. (2002a)
Food Science and Technology, 34, 47-57 Weerts, A.H., G. Lian, D. Martin, Modelling of
(1999). rehydration of porous biomaterials: Anisotropy
Mualem,Y., New model for predicting hydraulic effects, Submitted to J. of Food Sci. (2002b)
conductivity of unsaturated porous-media,
Water Resour.Res., 12, 513-522 (1976).
Ni,H., A.K.Datta and K.E.Torrance, Moisture
transport in intensive microwave heating of
biomaterials: a multiphase porous media model,
Int.J.Heat Mass Transfer, 42, 1501-1512
(1999).
Panchariya,P.C., D.Popovic and A.L.Sharma,
Modeling of desorption isotherm of black tea,
Drying Technology, 19, 1177-1188 (2001).
Peleg,M., An empirical-model for the description
of moisture sorption curves, J.Food Sci., 53,
1216-&(1988).
Sanjuan,N., J.A.Carcel, G.Clemente and A.Mulet,
Modelling of the rehydration process of brocolli
florets, European Food Research and
Technology, 212, 449-453 (2001).
Sanjuan,N., S.Simal and A.Mulet, Modelling of
brocolli stems rehydration process, J.Food Eng.,
42, 27-31 (1999).
Simal,S., A.Femenia, P.Llull and C.Rossello,
Dehydration of aloe vera: simulation of drying
curves and evaluation of functional properties,
Journal of Food Engineering, 43, 109-114
(2000).
Simunek, J., T. Vogel and M.Th. van Genuchten,
The SWMS_2D code for simulating water flow
and solute transport in two-dimensional variably
saturated media, USSL, ARS, US Dept. of
Agric., Riverside, USA, Research Report No
132, (1994).
Sopade,P.A., E.S.Ajisegiri and M.H.Badau, The
use of pelegs equation to model water-
absorption in some cereal-grains during
soaking, Journal of Food Engineering, 15, 269-
283 (1992).
Sopade,P.A. and J.A.Obekpa, Modeling water-
absorption in soybean, cowpea and peanuts at 3
temperatures using peleg equation, J.Food Sci.,
55, 1084-1087 (1990).
Thomas,H.R. and S.D.King, Coupled heat and
mass-transfer in unsaturated soil - a potential-
based solution, International Journal for
Numerical and Analytical Methods in
Geomechanics, 16, 757-773 (1992).
Thorvaldsson,K. and H.Janestad, A model for
simultaneous heat, water and vapour diffusion,
Journal of Food Engineering, 40, 167-172
(1999).
Tiktak,A. and W.Bouten, Modeling soil-water
dynamics in a forested ecosystem .3. model
description and evaluation of discretization,
Hydrological Processes, 6, 455-465 (1992).
Turner,I.W. and P.G.Jolly, Combined microwave
and convective drying of a porous material,
Drying Technology, 9, 1209-1269 (1991).
 MODELLING THE KINETICS OF TEXTURAL CHANGES IN HAZELNUTS
DURING ROASTING
Ali Doğan Demir
Kevin Cronin
Department of Process Engineering
University College, Cork
Ireland
E-mail: [email protected]

KEYWORDS industry and the catering business it is important to predict

the roasting behaviour and the quality of hazelnuts.
Hazelnut roasting, Texture kinetics, Exponential Integral
function. During hazelnut roasting, the time-temperature relationship
affects the resulting texture strongly, as well as other quality
ABSTRACT
factors. In industry, hazelnuts are generally roasted at
The texture changes in hazelnuts that were dry roasted at temperatures between 100 and 180 °C from 5 to 60 minutes
four temperatures (120, 140, 160 and 180 °C), for times (Özdemir and Devres 2000; Perren and Escher 1997). Low
ranging from 5 up to 60 minutes were studied using temperature roasting for long times can be considered as an
instrumental analysis. Textural changes were determined iso-thermal process because for the majority of the process,
from the force deformation curves obtained from the nut temperature is asymptotically equal to the air
compression tests. The first break point of the maximum temperature after the short initial period. For this schedule,
force (Newtons) was used as a measure of texture. The most of the quality changes that occur during roasting will
magnitudes of the parameters for a corresponding texture take place during this ‘iso-thermal’ period. In high
change model were determined using linear regression. temperature roasting, as the required process time is much
Statistical analysis of the data showed that the change in shorter and air temperature considerably higher, hazelnut
texture with roasting time was represented by a first-order temperature exponentially rises throughout the process and
reaction. The reaction rate constants were assumed to have quality changes occur over the entire temperature range.
an Arrhenius-type dependence on temperature. The Thus, a complementary model is required for systems in
activation energy, Ea and the frequency (pre-exponential) which the quality change occurs during both exponentially
factor, k∞ were determined as 22.5 kJ mol-1 and 0.338 s-1 increasing temperature period and iso-thermal temperature
respectively. Theoretical equations, based on the period.
Exponential Integral function (E1), to predict texture changes Force deformation curves have been widely used to
when product temperature is varying exponentially were measure textural changes in food products. Initial slope,
developed. These equations were validated by comparison maximum force, energy, fracture points or other curve-
with the experimental data. The results of this study showed related parameters have been analysed and correlated with
that both time and temperature effects were significant (p < these textural parameters. In literature, the fracture points on
0.001) on the textural changes in hazelnuts. The Exponential force deformation curves have been named as maximum
Integral prediction of texture change when product force, peak force or first peak of break (Saklar et al. 1999).
temperature varies during roasting was superior to the First-order reaction kinetics and Arrhenius temperature
isothermal prediction throughout the process. dependence have been commonly used in evaluating the
INTRODUCTION textural quality changes of thermally processed foods (Abu-
Ghannam 1998; Verlinden and Baerdemaeker 1997).
The most common form of nuts used in the food- Califano et al. (1997) evaluated the hardness of cooked beef
processing sector are shelled nuts which undergo processes with a first-order kinetics equation to describe the textural
including blanching, dicing, roasting and grinding to fit changes during the cooking process. The texture changes of
product formulation needs or add consumer appeal. Roasting dry peas in long time cooking followed first-order reaction
is one of the most important processes giving necessary kinetics (Xie et al. 1998). Lau et al. (2000) showed that the
alterations to the product. Currently, 17.5 % of commercially thermal softening of green asparagus followed a first-order
produced nuts throughout the world is hazelnut (Corylus kinetic reaction. The texture change of potatoes during
avellana L.) production (Somogyi et al. 1996). One of the cooking was described using a first-order model (Verlinden
several objectives to hazelnut roasting is to alter and et al. 1995).
significantly enhance the flavour, texture, colour and
appearance of the product. Roasting gives hazelnut a variety In this study the textural changes occurring in hazelnuts
of texture and colours increasing the crispness and during roasting were instrumentally measured using
crunchiness (Özdemir and Devres 1999; Özdemir et al. compression tests. The objective of this study is to determine
2001; Saklar et al. 1999). Roasted hazelnuts are widely the reaction kinetics to characterise textural changes defined
consumed as an appetizer and as raw materials in sweets, by force deformation curves and to apply a theoretical
confectionary, chocolate and biscuits. For the processing formula using Exponential Integral function that can predict
the texture changes at both high temperature-short time  E 
(where hazelnut temperature exponentially increases) and k R = k ∞ exp − a  (7)
low temperature-long time (where the temperature is iso-  RT∞ 
thermal) processes. where k∞ is the frequency factor, Ea is the activation energy,
THEORY R is the universal gas constant and T∞ is the absolute oven
temperature. For first-order reactions (n = 1), Equation (6)
Temperature Model can be solved introducing Equation (7):
The thermal geometry of hazelnuts is regarded as a solid
sphere placed at the centre of a fan oven. Heat transfer is by   E 
F = F0 exp − k ∞ exp − a t  (8)
convection from the air and by radiation from the oven wall   RT∞  
to the hazelnut. Mass transfer and shrinkage effects have
been shown to be negligible on sun-dried hazelnuts. The where F0 is the initial texture. The first-order model in
temperature model consists of first term of the infinite series Equation (8) can be linearized into the following form:
solution to the Fourier field equation for a sphere
undergoing unsteady state heat conduction. Average  F 
ln  = −k R t
 (9)
hazelnut texture is the parameter of interest in the analysis of  F0 
product quality and hence average temperature through the
nut is required. The average temperature, of the hazelnut can Non-isothermal Texture Change Model
thus be defined as (Ilicali 1989): Equation (8) predicts the texture change versus time once
− rc t the product temperature is constant. For a non-isothermal
T = (T0 − T∞ )A1e + T∞ (1) process, replacing T∞ with Equation (1) in Equation (7), and
solving Equation (6) for a first-order process, the texture
where T0 is the initial hazelnut temperature and T∞ is the air
change can be predicted when the temperature is
temperature in the oven. The thermal rate constant, rc can be
exponentially varying by Equation (10) (Demir et al. 2001):
defined as:
 k   Ea   − E a   Ea E  
exp ∞ E1   − exp  E1  − a  
λ1 2 k  rc   RT 
 
 RT∞   RT RT∞

 
rc = (2) F = Fo
ρ C p r0 2  k
exp ∞
 
E1 
Ea   − Ea  
 − exp  
Ea Ea  

 RT  E1  R[(T − T )A + T ] − RT  
  R[(T0 − T∞ )A1 + T∞ ] 

 rc  ∞  0 ∞ 1 ∞ ∞  
The equation constant A1 for the average temperature is: (10)

12(Sin λ1 − λ1 Cos λ1 )2 The solutions of Equation (10) is valid for times where
A1 = (3) Fourier number is greater than 0.2. The function E1 is the
λ1 (2λ1 − Sin 2λ1 )
3
Exponential Integral, which can be defined by:
where λ1 is the function of Biot number: ∞
e −u
E1 (x ) = ∫ du (11)
1 − λ1Cot λ1 = Bi (4) u
x
and the Biot Number for the spherical geometry is: where x is a positive number. The Exponential Integral can
hr be directly evaluated in packages such as MATLAB or
Bi = 0 (5) MATHEMATICA and a large number of different
k
approximations, based on simple polynomial and
The heat transfer to the hazelnuts in the oven is a exponential functions, are available for different ranges of
combination of both convective and radiative heating and the argument (Abramowitz and Stegun 1970).
thus the overall heat transfer coefficient, h is the sum of the MATERIALS AND METHODS
convective and radiative heat transfer coefficients.
Restricting the solution to the first term of the infinite series Experimental Roasting Procedure
expansion of the Fourier equation places a lower limit of Sun-dried raw hazelnuts with skin were supplied. The
about 90 seconds on solution validity for these conditions skins of the hazelnuts were removed by applying hot
(Çengel 1997). glycerine-sodium carbonate solution before roasting. A
Texture Kinetics digital vernier and a digital analytical balance (Bosch SAE
200) were used to measure hazelnuts radius and mass
The texture change of hazelnuts during roasting can be respectively. The average density value for hazelnuts was
described using an nth order model: found by mass to volume ratio. The thermal properties of the
dF hazelnuts were experimentally measured by a thermal
= −k R F n (6) properties meter (Thermolink Sensor Model TL-1) before
dt
the roasting process. Hazelnuts were placed in an incubator
where F is the texture parameter (maximum force at the first to guarantee an initial temperature of 20 °C. A Memmert
fracture point) and kR is the reaction rate constant. The effect ULP 600 conventional fan oven was used to carry out the
of temperature on the reaction rate constant can be expressed roasting trials. Airflow in the oven was kept constant at 0.3
using an Arrhenius type relationship: m s-1 at the centre of the oven by a fan operating throughout
the experiments. The aim of the experiments was to obtain term, k∞ two-step linear regression analysis was performed
experimental hazelnut texture versus time curves at four using the package MATLAB. The activation energy and the
different oven temperatures of 120, 140, 160 and 180 °C. frequency factor were obtained from the slope and intercept
These oven temperatures cover the commonly used defined by the Arrhenius equation.
temperatures in the hazelnut industry for roasting. From an
initial batch of hazelnuts, sets consisting of five randomly RESULTS AND DISCUSSIONS
selected hazelnuts were assembled. Each set was placed on a Physical And Thermal Properties
small table located at the centre of the oven. Each of the five
hazelnuts was placed apart with no contact between each The average hazelnut radius, r0 was measured as 7.07 mm.
other to prevent any shielding effect on the nuts during The mean density, ρ value for hazelnuts was found as 875
roasting. A 1.5 mm diameter hole was drilled through one kg m-3. Thermal conductivity, k of hazelnuts was measured
side of one of the five hazelnuts and a thermocouple inserted as 0.204 W m-1 K-1. Specific heat capacity, Cp was found to
and sealed in there to measure the centre temperature and be 1994 J kg-1 K-1 (giving a thermal diffusivity, α of
data-log by a computer. The reason for forming sets with 1.17*10-7 m2 s-1). The overall heat transfer coefficient for
five hazelnuts was to be able to eliminate any unexpected hazelnuts in the oven was experimentally evaluated as 33.3
texture formation that could result. The roasting times were W m-2 K-1. Using this data, the Biot Number for heat transfer
ranging from 5 min up to 60 min. As it was not possible to to the hazelnut was evaluated as 1.15 meaning that the
measure texture change on-line, the roasting trials were constants A1 and λ1 were 0.981 and 1.66 respectively. The
repeated with 5 min increments for each set. Each set was thermal rate constant, rc was calculated as 6.52*10-3 s-1.
roasted at four different temperatures and same time Texture Data Analysis And Reaction Order
combination to enable the texture versus time curves to be
assembled. At the end of each experiment, the roasted Table (1) gives the results obtained from two-way analysis
hazelnuts were removed from the oven and placed in a of variance (ANOVA). As can be seen from the values of
freezer at 0 °C to prevent any further roasting and thus any temperature and time effects, the effect of temperature on the
textural changes after removing from the oven. They were texture response force, F is greater than the time effect. The
cooled to room temperature and kept in an incubator at 4 °C analysis showed that both time and temperature effects are
until the texture analysis was undertaken off-line. significant (p < 0.001) on the textural changes in hazelnuts.
The texture response F demonstrated the highest sensitivity
Texture Measurements to temperature change.
The texture of the raw and roasted hazelnuts were Table 1: F-Statistics of the Time and Temperature Effects on
measured using a computer interfaced JJ Lloyd Tensile the Texture Response, F from two-way ANOVA.
Testing Machine (model T5K). A load cell of 500 N was
mounted on the moving head of the tester. Each hazelnut Texture Response Force (F)
kernel was placed on a small table in the longitudinal F-Statistics of temperature effect 48.97
position. The same position was applied to all hazelnuts. F- Statistics of time effect 4.66
Hazelnuts, which were not able to stand in that position, df (temperature - time) 3 - 11
were slightly flattened on the bottom surface with sandpaper. Standard deviation, σ 13.35
The samples were compressed individually at a constant Range 77.97
deformation speed of 10 mm min-1. The force deformation Coefficient of variation, cv 0.45
curves were recorded by a computer. Recorded data
represents 260 hazelnuts. The force deformation curves were The goodness-of-fit of first-order model with the
analysed for the first fracture point in Newtons and experimental data was assessed from the coefficient of
deformation in millimetres. The maximum force (F) at the determination, R2 values at each temperature. The average R2
first peak of the break was used as the measure of texture. value was found to be 0.6912 indicating that the overall fit
of the first-order reaction to F response was satisfactory at
Statistical Analysis the temperatures applied. Hence the first-order reaction was
The force (F) data obtained from the experiments were selected to be the model representing the F response. The
analysed by analysis of variance technique (ANOVA) to main reason for low R2 values was found to be the high
find out the effect of time and temperature on those variation between nuts at the roasting times and relatively
responses. From the two-way ANOVA technique, the small textural changes that occurs at the temperatures
sensitivity of F response was determined to time and applied. The high variations in the experimental texture
temperature change (Devore and Farnum 1999). Hazelnut responses were due to the size, shape and ripeness of
force versus time curves were available at four different hazelnuts. There were other factors, which affect the
temperatures. The fit of data to first-order model given by variation in hazelnut texture. First of all, hazelnuts were not
Equation (8) was assessed at each temperature by perfectly spherical and they did not have exactly the same
determining the coefficient of determination values using shape and size. There was an internal cavity at the core of
one-step regression in MATLAB. The slopes obtained from each nut, which is independent of the size and the shape of
Equation (9) yield the reaction rate constants, kR for each the hazelnut. The size of this cavity also varied between
case. Having decided the order of reaction to represent the nuts. The first break point was dependent on this cavity, as a
texture changes satisfactorily, the reaction rate constants larger internal gap leads to a lower force whereas a smaller
were checked for the Arrhenius temperature dependence. In gap requires a greater force to break the hazelnut.
order to calculate the activation energy, Ea and the frequency
 Arrhenius Temperature Dependence than zero, showed that the mean values of experimental data
were in good agreement with the predicted values.
The Arrhenius temperature dependence of the reaction rate
constants was checked with the plot of ln (kR) versus 1/T. Table 2: One-way ANOVA Results of the Experimental and
Figure (1) shows that the reaction rate constants obtained Predicted Values of the Texture Response, F.
from the first-order model follows the Arrhenius equation
well giving a coefficient of determination value of 0.9925. Texture Force (F)
From the slope and intercept of the straight line described by Temperature (°C) 120 140 160 180
the Arrhenius equation, the activation energy, Ea and the p-value 0.859 0.552 0.777 0.381
frequency term, k∞ were determined as 22.5 kJ mol-1 and MSE 98.38 66.19 46.13 48.62
0.338 s-1 respectively. SS (interaction) 3475 4244 5940 5081
F-Statistics 3.21 5.83 11.71 9.5
df (temp. - time) 1-11 1-11 1-11 1-11
1/T (K-1 )

-3
CONCLUSIONS
0.002175 0.002275 0.002375 0.002475 0.002575 The results of this study showed that the effect of both
roasting air temperature and roasting times were significant
-3.5
on the textural changes in hazelnuts. The maximum force at
the first fracture point was used to monitor the textural
ln (k R)

-4 changes in hazelnuts during roasting. The texture response,

F had higher sensitivity for temperature than the time
y = -2760.2x + 2.725 change. Textural change represented by the maximum
-4.5
R2 = 0.9925 fracture force was a first-order reaction where the fit of the
first-order model was found to be satisfactory. The reaction
-5 kinetic sensitivity to temperature was described well by the
Arrhenius equation. The theoretical formula to predict the
Figure 1: Arrhenius Temperature Dependence of Reaction textural changes during the non-isothermal part of the
Rate Constants. roasting resulted in good agreement with the experimental
data. The validation of Equation (10) using the Exponential
Validation Of Non-Isothermal Texture Change Model
Integral function showed that the theoretically predicted
Figure (2) shows the non-isothermal predictions of texture texture changes during the heating up (non-isothermal) part
change versus time by Equation (10) with the experimentally of the process was superior to the isothermal prediction.
measured (F) values. The theoretical profile is given for
Further research is required on the high temperature short
times greater than 90 seconds corresponding to the lower
time roasting of hazelnuts in order to obtain more
limit on solution validity. The initial value of F was
information on the texture changes occurring in the non-
measured as 66.35 N. As can be seen from the figure, there
isothermal part of the process that can be modelled by the
is a good agreement between the experimental results and
application of the Exponential Integral function. The
theoretical predictions of Equation (10).
development of a validated texture model over a large
Expe rime ntal Expe rime ntal
roasting temperature range allows optimisation studies to be
90 120 °C 90 140 °C
75 The ore tical 75
The ore tical performed to determine the best roasting schedule in terms
60 60 of delivering required product texture.
F (N )
F (N )

45 45

30 30 NOMENCLATURE
15 15

0 0 A1 Equation constant
0 10 20 30 40 50 60
Bi Biot number
0 10 20 30 40 50 60
Ti m e (m i n ) Time (m in)

90 Expe rime ntal 90

cv Coefficient of variation
Expe rime ntal
75
160 °C
The ore tical 75
180 °C
The ore tical
Cp Specific heat capacity (J kg-1 K-1)
60 60
Ea Activation energy (kJ mol-1)
F (N)

F (N)

45 45
30 30
E1 Exponential Integral function
15 15 F Fracture force (Newton)
0
0 10 20 30 40 50 60
0
0 10 20 30 40 50 60
h Overall heat transfer coefficient (W m-2 K-1)
Time (min) Time (min)
k∞ Frequency factor (s-1)
Figure 2: Non-isothermal Prediction of Texture Changes (F) k Thermal conductivity (W m-1 K-1)
together with the Experimental Data. kR Reaction rate constant (s-1)
n Reaction order
Table (2) gives the one-way ANOVA results performed r0 Radius (m)
between the experimental data and theoretical predictions of rc Thermal rate constant (s-1)
Equation (10) at the times and temperatures studied. As it is R Universal gas constant (8.314*10-3 kJ mol-1 K-1)
seen in the table, the low mean square error values indicated R2 Coefficient of determination
that the theoretical prediction of the textural changes in t time (s, min)
hazelnuts was satisfactory. The p-values, which are different T Temperature (°C, K)
 Perren, R. and F.E. Escher. 1997. “Investigations on the Hot Air
T Average Temperature (°C, K) Roasting of Nuts.” The Manufacturing Confectioner, June, 123-
u Exponential Integral function parameter 127.
x Exponential Integral function parameter Saklar, S.; S. Ungan and S. Katnas. 1999. “Instrumental Crispness
and Crunchiness of Roasted Hazelnuts and Correlations with
Greek Symbols
Sensory Assessment.” Journal of Food Science, 64, (6), 1015-
ρ Density (kg m-3) 1019.
λ1 Thermal model constant Somogyi, L.P.; D.M. Barrett and Y.H. Hui. 1996. Major Processed
Products. Technomic Publishing, Pennsylvania.
σ Standard deviation
Verlinden, B.E.; B.M. Nicolai and J.D. Baerdemaeker. 1995. “The
Subscripts Starch Gelatinisation in Potatoes during Cooking in Relation to
the Modelling of Texture Kinetics.” Journal of Food
∞ Asymptotic value Engineering, 24, 165-179.
0 Initial value Verlinden, B.E. and J.D. Baerdemaeker. 1997. “Modelling Low
Temperature Blanched Carrot Based on Heat Induced Processes
REFERENCES and Enzyme Activity.” Journal of Food Science, 62, (2), 213-
Abramowitz, M. and I. Stegun. 1970. Handbook of Mathematical 229.
Functions. Dover Publications, New York. Xie, G.; R. Xiong and I. Church. 1998. “Modelling of Texture
Abu-Ghannam, N. 1998. “Modelling Textural Changes During the Changes of Dry Peas in Long Time Cooking.” Food Control, 9,
Hydration Process of Red Beans.” Journal of Food (4), 233-241.
Engineering, 38, 341-352.
AUTHOR BIOGRAPHY
Çengel, Y.A. 1997. Introduction to Thermodynamics and Heat
Transfer. McGraw-Hill, New York. ALI DOĞAN DEMIR was born in 1974 in Turkey. He
Califano, A.N.; N.C. Bertola; A.E. Bevilacqua and N.E. Zaritzky. went to Middle East Technical University, Ankara, where he
1997. “Effect of Processing Conditions on the Hardness of studied and obtained his B.E. degree in Food Engineering
Cooked Beef.” Journal of Food Engineering, 34, 41-54.
between 1993 and 1998. He has studied for a year in the
Demir, A.D.; J.M. Frías; K. Abodayeh; K. Cronin and F.A.R.
Oliveira. 2001. “Modelling of the Thermal Kinetics of Colour
Department of Environmental Science and Engineering,
Change in Hazelnuts During Roasting.” In Proceedings of the Technical University of Denmark. He started his Ph.D.
International Symposium on Applications of Modelling as an degree in 1999 in the Department of Food Engineering,
Innovative Technology in the Agri-Food-Chain (Palmerston University College Cork, Ireland. Currently he is doing his
North, New Zealand, Dec. 9-13). Model-IT, Acta Horticulturae, PhD and working in the Department of Process Engineering,
566, 317-322. University College, Cork. Further information can be
Devore, J.L. and N.R. Farnum. 1999. Applied Statistics for obtained from his e-mail address, [email protected] or
Engineers and Scientists. Duxbury Press ITP, Pacific Grove, from the web address, which is http://www.ucc.ie/acad/
USA. departments/foodeng/postgrad/dogandemir.html.
Ilicali, C. 1989. “A Simplified Analytical Model for Thawing Time
Calculation in Foods.” Journal of Food Science, 54 (4), 1031- KEVIN CRONIN was born in 1963 in Ireland. He went
1039. to University College Dublin where he studied and obtained
Lau, M.H.; J. Tang and B.G. Swanson. 2000. “Kinetics of Textural his B.E. (Mech.) degree between 1980 and 1984. He got his
and Colour Changes in Green Asparagus during Thermal
Master of Engineering Science (M.Eng.Sc) degree in
Treatments.” Journal of Food Engineering, 45, 231-236.
Özdemir, M. and O. Devres. 1999. “The Thin Layer Drying
University College Dublin in 1984 and 1986. He obtained
Characteristics of Hazelnuts during Roasting.” Journal of Food his Ph.D. (Mechanical Engineering) degree in 1988 and1991
Engineering, 42, 225-233. in Brunel University, UK. He worked in SKF, Engineering
Özdemir, M. and O. Devres. 2000. “Kinetics of Colour Changes of and Research Centre in Holland from 1986 to 1987.
Hazelnuts during Roasting.” Journal of Food Engineering, 44, Between 1991 and 1994, he worked as a research officer in
31-38. University of Ulster. Currently he is working as a lecturer in
Özdemir, M.; F.G. Seyhan; A.K. Bakan; G.Ö. İlter and O. Devres. the Department of Process Engineering, University College,
2001. “Analysis of Internal Browning of Roasted Hazelnuts.” Cork. Further information can be obtained from his e-mail
Food Chemistry, 73, 191-196. address, [email protected] or from the web address in
http://www.ucc.ie/acad/departments/foodeng/staff/cronink.ht
ml.
 MODELLING AND OPTIMISATION APPROACHES IN INDUSTRIAL
FERMENTATION PROCESSES
T. Kurz
S. Arnold
M. Fellner
E. Murnleitner
M. Mitzscherling
T. Becker
A. Delgado
TU München / Weihenstephan
Chair for Fluid Mechanics and Process Automation
85350 Freising / Germany
Phone: + 49 (0)8161 / 71-3247
Facsimile: + 49 (0)8161 / 71-4510
Mail: [email protected]

KEYWORDS Usually, biotechnology - and in particular food technology -

Metabolic modelling, fuzzy logic, artificial neural network, is characterised by a high complexity of the matrix as well
software sensor as dynamic and non-linear processes. Therefore, most of
the processes and the matrix are not describable with
ABSTRACT mathematical methods in an effective way. However a lot
of technological knowledge or large quantities of data exist,
In this paper different approaches for process modelling which can be used for cognitive approaches e.g. based on
and control of industrial fermentations are presented. A fuzzy-logic or artificial neural networks.
method for the online determination of diacetyl is depicted.
The concentration of diacetyl, a key component during the THE COGNITIVE ESTIMATOR FOR DIACETYL
fermentation of beer, is estimated using an artificial neural
network model (ANN). A metabolic kinetic modelling Diacetyl, an indirect by-product of yeast metabolism, is a
approach was used to describe the yeast propagation of key component in beer. It causes an off-flavour if its
brewing yeasts under industrial conditions. A similar concentration exceeds a threshold of about 0.1 mg/L.
approach was used for the description of the process Diacetyl and its precursors are present in beer during
behaviour of an industrial anaerobic wastewater treatment fermentation and maturation. The concentration of diacetyl
plant. Here, it served to evaluate the state detection and is considered to be the most important criterion in the
process control based on an adaptive fuzzy logic system. shortening of the maturation process, making its online
State detection and adaptive process control based on fuzzy determination of special economic interest. The current
logic was also applied in the beer fermentation. In another state of the art in diacetyl measurement technology is the
application fuzzy logic was used to adapt microorganisms off-line laboratory analysis using specially adopted
to extreme conditions in the industrial vinegar fermentation headspace gas chromatography. The major disadvantages of
to improve the acid yield. this method are the high cost of the laboratory equipment,
the required manual sample handling and pre-treatment and
INTRODUCTION the delay of several hours before results are available.
These disadvantages make this method unsuited for an on-
In order to reach a precise, economic, and homogenous line fermentation control. To date there are no alternative
production in food and biotechnological industry an active methods available.
process control is obligatory. Prerequisites of an active There are, however, readily measurable process parameters
process control and process optimisation are the which provide indirect information about the diacetyl
observability of the process and a comprehensive concentration in the fermenting beer. The main physical
knowledge about the influence of manipulated variables and chemical process parameters used in our studies are
upon the process behaviour or process variables. temperature, specific gravity, pressure, turbidity and pH. In
Observability here is defined as the possibility to receive order to facilitate the estimation of the diacetyl
relevant information from the process in order to describe concentration the cognitive estimator needs a model of the
the process behaviour comprehensively at every time. dynamic properties of the process and their dependence on
Particularly of interest in food and biotechnology is the the changes in diacetyl. Dynamic artificial neural networks
online determination of biochemical substances, which were used in this project to deduce these properties from
normally are only measurable after a sample preparation the available measurements. Since it was not sufficient to
using gas- or liquid-chromatographical methods. only use the current set of measurements, a measurement
history was used in order to internally reconstruct parts of The system was adapted to online determination of diac etyl
the unknown process dynamics. for an industrial beer fermentation. Based on the results
In the present problem a multilayer feed forward network obtained, it can be said that the accuracy and reliability of
was applied. A special neuron with a FIR filter was inserted the software sensor is sufficient for practical applications.
at each connection (tapped delay line). A supervised
method using the sum of the squared errors as the cost METABOLIC MODELLING OF BREWING YEAST
function was used to train the neural network. Gradient PROPAGATION
information was obtained using the IC-1 algorithm
(instantaneous cost, instantaneous gradient), weight updates For industrial yeast fermentations, e.g. brewing or baker’s
were calculated with different learning algorithms such as yeast fermentations, it is essential to provide the inoculum
Rprop, SuperSAB and Quickprop. Reliable results, in terms at the right time in a desired quantity with a high
of training time and convergence stability have been fermentation activity, in order to guarantee an efficient
obtained in the prediction phase using these configurations. fermentation process and high quality ferme ntation
In Fig. 1 and 2 results of the “Cognitive Estimator” are products. Therefore, in this work a comprehensive
presented. Fig. 1 shows the online estimation of diacetyl modelling approach of the fed-batch-wise yeast propagation
during the maturation examplarily for two fermentations is introduced. Also physiological data of a cell cycle
(line). Additionally the reference values are shown (points). analysis are correlated to experimental data. Both
It is obvious, that the ANN represents the reference values approaches represent the basis for a control strategy aiming
very well. Fig. 2 shows a parity plot of online and offline at the provision of an optimal inoculum at the starting time
diacetyl determination. The “Cognitive Estimator” works of subsequent fermentations. In the presented metabolic
not only during the maturation phase, but, as can be seen, modelling approach (formulated with the software tool
over the whole range of possible diacetyl contents. Almost ‘Aquasim’) respiratory metabolism on sugars and ethanol
all points are within the marked tolerance range of 0.05 and fermentative metabolism respectively, are included.
ppm, which is required from the online estimation to be Occurring limitations, due to specific nutritional data of the
used as an input for a closed loop control. beer-wort, were taken into account for sugars, nitrogen,
ethanol and oxygen. Correspondingly, inhibitions of the
metabolism by ethanol and high sugar concentrations were
formulated. Especially the Crabtree-effect is represented by
the model. For model identification, literature data were
used and selected experiments out of the relevant range of
manipulated variables (temperature, dissolved oxygen)
were carried out. Model based simulations matched these
data with an accuracy of 2-8% with standard deviations of
5% after an adaptation of three variable parameters, which
were extracted in a sensitivity analysis before. A relation
between the temperature and the variable parameters could
be formulated. The influence of oxygen, in particular at low
concentrations (oxygen limitation) could be quantified too.
So, predictive simulations of the yeast propagation process
using non isothermic trajectories were possible. Figure 3
Figure 1: Online estimation of diacetyl presents a case scenario for a brewery, where the yeast
propagation process has to be delayed for several hours.

600 300
Concentration [mmol/L]

400 294
Temperature [K]

200 288

0 282

Figure 3: Application of a temperature scenario for the

yeast propagation process.

A temperature profile (20-10-20°C) was applied (dotted

Figure 2: Parity plot of online and offline diacetyl line). Reference values for the progressions of biomass (¢)
determination and substrate (♦) and the model based predictive
simulations (lines) of the concentrations of biomass and Control 1
substrate (as glucose equivalents) for the given temperature ∆ setcurve ∆ temperature
temperature
profile are presented. The change of temperature caused a
deceleration and acceleration respectively of the
propagation process, which can be found in the Control 2
progressions of the reference values as well as in the ∆ pressure
gravity pressure
belonging predictive simulation runs. The accuracy of the
simulation was about 8% with standard deviations of 7%. pH-value
Also in other scenarios applying temperature and dissolved turbidity
oxygen profiles it could be shown, that with a precise
adjustment of trajectories the crop time of the inoculum temperature
fermentation - fermentation phase
could be varied within a period of 7 to 50 hours while diacetyl phase detection
maintaining a high fermentation activity for the subsequent
anaerobic fermentation. In the same way the actual Figure 4: Scheme of the fuzzy system
physiological state could be comprehended by evaluation of
cell cycle analysis and correlation to experimental data. So, The fuzzy logic fermentation phase detection worked very
a basis for a reliable process optimisation could be robust and reliable. Even trials with simulated failures of
established, which takes into account biomass growth as measurement values produced sufficient results. By using
well as information on the physiological state of the this fuzzy system it was possible to reduce the complete
biomass. fermentation time up to 25%, by an automatic change
between the fermentation phases in the sequential control.
THE FUZZY-LOGIC FERMENTATION CONTROL The superimposed closed loop control resulted in an
increased process homogeneity. The process spread of
The main topic of interest was to realise an active process gravity could be reduced to 50%.
control for the fermentation of beer. As a lot of
technological knowledge exits, the use of fuzzy logic for a ANAEROBIC WASTEWATER TREATMENT WITH
process control suggested itself. The implemented system FUZZY LOGIC
combined a process state (fermentation phase) detection
module and a process control module based on fuzzy logic. For an efficient anaerobic wastewater treatment,
In addition to the online measurement of physical or supervision by highly qualified staff or a proper process
chemical values it was necessary to record information control is required, particularly if the volumetric loading or
about biochemical values by integrating the described the substrate composition is variable or if toxic inhibition
“Cognitive Estimator”. can occur.
The gained values of online measurement and estimation However, anaerobic treatment is difficult to control,
are used as input for the fuzzy logic fermentation phase because a system overload can halt the process, whereby a
detection. The result is the actual fermentation phase like system restart can take up to weeks. Fuzzy logic was used
“main-fermentation” or “maturation”. The well known to model and control the process, because qualitative or
coherences between input data and the fermentation phase semi-quantitative expert knowledge is available. Moreover,
are described in rules. states like “hydrogen is normal” and “hydrogen is high”,
A second fuzzy logic system for the control of the which cannot be distinguished sharply because the
fermentation uses the absolute deviation of the measured transition from “high” to “normal” is not well known (it is
data from given set curves of defined partial processes as fuzzy), are handled with the Fuzzy logic’s concept of
the input. Dependent on the input data and the detected partial truth. Hydrogen was chosen as the key parameter for
fermentation phase the manipulated variables (temperature, the detection of an overload, as its concentration increases
pressure) can be adjusted. sharply after an overload. Not only the hydrogen
In Fig. 4 the scheme of the fuzzy system is presented. Data concentration, but also pH, the first derivative of pH, the
from online measurement is used to determine the concentration of methane, the gas production rate, and – if
fermentation phase. A superimposed control system uses necessary - also the concentration of organic compounds
the detected phase and the deviation of the online values (TOC, COD, or BOD) are used as input variables.
from given set curves. The systems tries to steer the process With these variables, the state of the reactors, like
towards the set curves by manipulating temperature and/or “shortage”, “normal”, “overload”,” toxic” is estimated.
pressure. Using these biological states together with the current
reactor conditions (e. g. acidification buffer tank is full), the
conditions are altered or the feed rate is changed. The
conditions are influenced by changing temperature, pH,
velocity of circulation and the recirculation rate.
In Fig. 5 data from a two stage anaerobic treatment of
potato processing waste water are shown. The first graph
shows the feed rate into the acidification buffer tank
(divided by the methane reactor volume) and its fuel level.
The latter is the result of the feed and the control action. the consideration of available know-how for situational set
Gas concentrations in the methane reactor are depicted in point management. The fuzzy module as the cognitive core
the second graph (% methane, ppm hydrogen, and the gas of the computational system was adapted to the acetic acid
flow normalised by the volume of the methane reactor). fermentation by setting up a specific structure of input
The third graph shows truth values (membership value) of variables, rule bases and output variables. This structure
the sets from the state of the methane reactor ( shortage, (Fig. 6) bases on expert knowledge collected in reports and
normal, overload). As shown in the third graph, the interviews, combined with experimental data.
states could be predicted well, so proper control actions
could be taken, too. Additionally a kinetic modelling
approach was developed to allow simulation runs and to
evaluate the process state detection and control.

16 60
% fuel level of acidification buffer tank
14
50
12
Figure 6: Structure of the fuzzy system
40
10
The final total concentration (FTC) was considered as the
8 30 manipulated variable. This value can be regarded as a
6 measure for the maximum stress conditions or selection
20
4 feed rate acidification buffer tank acid concentration. As input information for the fuzzy
10 system physical measurement data of two bygone
2
fermentation cycles were logged and prepared (FTC, cycle
0 0 time, degradation rate, turbidity development). This input
0 2 4 6 8 10 12 14
information is processed in a multilevel rulebase, where
1000 70 intermediate variables are calculated. The intermediate
900 variables enable a “fuzzy” grading of the two bygone cycles
60
% methane by respecting the conversion rate, the growth rate and the
800

700 50 not directly measurable value “acidity tolerance” of the

600
bacteria. Finally from these intermediate variables it is
40

500
decided, whether “selective conditions” or “recovering
400
30 conditions” should be chosen for the culture in the
following cycle. This is done by setting the FTC value for
300 20
the following cycle. Preliminary fermentation tests and the
200
ppm hydrogen 10
results of expert questionnaires showed clear limits. It is not
100
gasrate possible to choose too much acidity stress, especially after
0
0 2 4 6 8 10 12 14
0
several cycles, without losing average productivity.
After refining the parameters of the fuzzy system in respect
1 to further fermentation experiments, both focused factors -
shortage normal overload
0.9 a better conversion rate and a higher product concentration
shortage
0.8 normal - could be improved.
0.7 Conventional process strategies are able to reach about
0.6 18,5% product concentration in a two stage configuration.
0.5 As illustrated in Fig. 7a, immediately after a first selection
0.4 cycle by forcing up the FTC, the final concentration in the
0.3 second stage could be risen to 19,2%. Instancing the fuzzy
0.2 managed fermentation serial given in Fig. 7b, two selection
0.1 cycles with a higher FTC were determined. In this case the
0 fermentation temperature was not kept on a constant level.
0 2 4 6 8 10 12 14
It was lowered with the increasing inhibition of the
microorganisms, respectively with decreasing conversion
Figure 5: Data from a two stage anaerobic treatment of rate. After three cycles with situational inhibition of the
potato processing waste water biological material, an increase in the forth cycle from
19,6% to nearly 20% final acetic acid concentration
FUZZY-LOGIC CONTROL STRATEGIES FOR THE indicated an improvement by selection effects. In the 5th
“VIRTUAL VINEGAR BREWER” cycle, the final acidity could lifted to more than 20,3%,
combined with a shortened cycle time. These improvements
In this fuzzy logic approach process data from a two -stage could be reproduced in all cases after selection cycles with
vinegar fermentation, prepared or directly measured, a calculated FTC in the first stage.
provide the input values for an expert knowledge based According to practical experience, the productivity of a
fuzzy control system. This "virtual vinegar brewer" allows vinegar plant can decrease seriously, if the FTC is forced up
merely 0,2% and is kept for several cycles. With respect to development of hybrid approaches. Especially the
this, the performed increase of 2% in the maximum product integration of mathematical or technological knowledge in
concentration with the same given apparatuses is the learning process of ANN is suggested. This allows to
remarkable. improve the accuracy of process predictions. On the other
hand fuzzy logic systems should be developed or adapted
a) which use self generating methods for rules or sets. With
these methods it would be possible to obtain technological
knowledge about processes, which was not accessible so
far, since most of the considered processes could only be
modelled by black box approaches. In particular, the
description of the actual physiological state based on
knowledge based expert systems is a main topic in current
research. Hence, it can be stated that there is still a high
potential of implementing modelling and optimisation
algorithms in various problems of food technology.

REFERENCES

Arnold, S.; R. Schmidt and T.Becker. 2001. "Online determination

b) of Acetic acid during an aerobic acetobacter cultivation."
Anal. Chim. Acta, submitted.
Murnleitner, E.; T. Becker and A. Delgado. 2001. "State Detection
and Control of Overloads in the Anaerobic Wastewater
Treatment using Fuzzy Logic." Water Research. accepted.
Kurz, T.; T. Becker and A. Delgado. 2002. "Model based control
of the metabolism of saccharomyces cerevisiae under
growth limiting conditions." Journal of Biotechnology.
submitted.
Fellner, M.; A. Delgado and T. Becker. 2001. "Functional neurons
in dynamic neural networks for bioprocess modelling."
Bioprocess and Biosystems Engineering. submitted.
Murnleitner, E.; A. Delgado and T. Becker. 2001. "Simulation and
optimisation of a two stage anaerobic digestion using
knowledge based methods." Journal of Hydroinformatics.
submitted.
(?) final acidity in the second stage Arnold, S.; T. Becker; A. Delgado; F. Emde and A. Enenkel.
(O) selection acid concentration in the first stage, scaled to 2001. "Optimizing high strength acetic acid bioprocess by
their maximum level cognitive methods in an unsteady state cultivation."
Journal of Biotechnology. submitted.
Figure 7: Effect of fuzzy moderated selection acidity in the Murnleitner, E.; A. Delgado; T. Becker: 2001. “Optimisation and
first stage on finally reached acid concentration in the validation of a fuzzy-based closed-loop control strategy
second stage using a dynamic simulation model.” Biotechnology and
Bioengineering. submitted.
Kurz, T.; M. Fellner; T. Becker; A. Delgado. 2001. “Observation
CONCLUSION and control of the beer fermentation using cognitive
methods” Journal of the Institute of Brewing 107, No. 4,
The paper shows that mathematical and cognitive 241-252.
approaches can be used in many fields of biotechnology
and food-technology. Main topic in the future will be the
 MODELING OF THE COOKING LOSSES DURING THE COOKING
OF BEEF PATTIES USING FAR INFRARED RADIATION
Nicholas Shilton1, P. Mallikarjunan2, P. Sheridan3
1
Department of Agricultural and Food Engineering, University College Dublin, Earlsfort Terrace, Dublin 2,
Ireland. 2 Department of Biological Systems Engineering, Virginia Polytechnic Institute and State University.,
Blacksburg, VA, 24061. 3 Waterford Institute of Technology, Waterford, Ireland

KEYWORDS
Food, Biology, Differential equation solvers, Greek Letters
Model testing. α Thermal diffusivity (m2.s-1)
ρ Density (kg.m-3)
ABSTRACT σ Stefan Boltzmann constant (W.m-2.K-4)

During the cooking of beef patties, material is lost Subscripts

from the patty by two mechanisms. The first is f Final
evaporation of water, which can easily be described s Surface
by standard drying theory, while the second
mechanism is the expression of juices from the INTRODUCTION
patty during cooking. These ‘juices’ will contain a During the cooking of beef patties, material is lost
significant amount of fat and moisture in addition from the patty by two mechanisms. The first is by
to proteins, and can be characterized as ‘cook evaporation of water, which can easily be described
losses’. The objective of this study is to investigate by standard drying theory. The second mechanism
the moisture and fat losses from beef patties during is the expression of juices from the patty during
cooking by far infrared radiation and to develop a cooking. These ‘juices’ will contain a significant
model to describe these cooking losses. Fresh amount of fat and moisture in addition to proteins,
minced beef patties were prepared at 2 different fat and can be characterized as ‘cook losses’. The total
levels, (15% & 22 % fat). These were cooked using amount of product remaining can be viewed as the
a far infrared radiation heat source for 20 minutes. ‘cooking yield’. Some researchers will refer to the
The resulting data was modelled using a coupled ‘cook loss’ as being a combination of vaporized
finite difference model to predict heat transfer, losses and juices.
moisture losses and fat losses. Results showed heat
and mass transfer profiles similar to those found in Previous papers, (Sheridan and Shilton, 2002a;
similar studies. Good agreement was found in the Shilton et al., 2002) have described moisture losses
prediction of heat transfer, moisture losses and fat due to evaporation and the simulation of these
losses. The model also showed that internal losses, in this paper the cook losses in terms of total
movement of the fat through the beef patty moisture loss and the fat loss will be described and
influences heat transfer through the beef patty. This simulated. Sheridan (2000) found that as fat content
paper describes a simple model that is suitable for of beef patties increased, so did the amount of fat in
the prediction of heat and mass transfer during the the cook losses. A number of different researchers
cooking of beef patties. have examined the affect of increasing fat content
on the fat losses from beef patties during cooking
by a variety of methods including grilling (Troutt et
NOMENCLATURE
al., 1992; Sheard et al., 1998; Kirchner et al.,
C Moisture content (kg moisture / kg 2000). These researchers were mainly examining
solid) the affects of the cooking process on the chemical
Deff Mass diffusivity (m2.s-1) and sensory properties of the beef patties, no
Dx, DT, DC Constants in Equation 12 modelling of the cooking process was included.
h Heat transfer coefficient (W.m-2.K-1) The fat contents examined ranged from 0% to 50%
heff1, heff2 Constants in Equation 8 added fat. The affect of using far infrared radiation
k Thermal conductivity (W.m-1.K-1) was not examined by any of the above researchers.
k1, k2, k3 Constants in Equation 6 Andersson et al., (2000) report that loss of fat
Km Mass transfer coefficient (m.s-1) during the frying of beef patties increases linearly
r1, r2, r3 Constants in Equation 7 with fat content.
t Time (s)
T Temperature (K) Due to the risk of food poisoning by E. coli
T∞ Infinite temperature (K) O157:H7 from undercooked beef patties it is
x Thickness of beef patty (m) important to heat such products so that they are
cooked thoroughly. There have been a number of Industrial Group, Independence, Kansas) mounted
models proposed for the cooking of beef patties in the roof of the oven. Samples were placed 75
that account for heat transfer (Pan et al., 2000; mm below the heating element for the duration of
Shilton et al., 2002). The Food Safety Authority of the cooking time.
Ireland recommends that the internal temperature
exceed 70 °C for 2 minutes (Food Safety Authority Mathematical Model
of Ireland, 1999) to ensure adequate cooking and
The beef patty was assumed to be homogeneous
thermal destruction of micro-organisms. It has also
and isotropic as a result of the mixing during
been postulated that heat transfer in beef patties is
preparation. As heating was from one side only, the
influenced by the movement of fat during cooking
beef patty was assumed to be an infinite slab, due
(Shilton et al., 2002), thus the model developed in
to the sample geometry, thus heat and mass transfer
this paper will account for this.
could be considered in one dimension only. It was
assumed that the change in moisture content can be
The objective of this study is to investigate the
used to account for both evaporative and drip losses
moisture and fat losses from beef patties during
of moisture from the beef patty, and that no fat was
cooking by far infra red radiation and to develop a
lost as a result of vaporisation, but mainly as drip.
model to describe these cooking losses, while also
The model described here was developed in Shilton
accounting for heat transfer.
et al., (2002) for prediction of heat transfer and
moisture vapour losses. It has now been modified
MATERIALS AND METHODS
to account for total moisture lost as well as fat
losses by drip.
Minced meat (fat content 15% & 22%) was
obtained from a local supermarket. Beef patties
were prepared by pressing the minced beef into
HEAT TRANSFER
round stainless steel sample holders, (50 mm
internal diameter, 20 mm depth). The sample Unsteady state one dimensional heat transfer in a
holder was placed on a wide mesh (5 mm), and was slab can be described by:
allowed to sit on a drip tray to collect cooking
losses. Holes at 8 and 12 mm were drilled in the ∂T ∂ 2T
side of the sample holder, to allow for the radial =α 2 (1)
∂t ∂x
insertion of thermocouples. The temperature was
recorded using type T thermocouples, with data where T is the temperature of the beef patty, α is
being collected by a datalogger (Campbell the thermal diffusivity and x is the thickness of the
Scientific 21X; Campbell Scientific Inc., Logan , beef patty. The initial conditions are T=Ti for all
Utah) every 30 s. locations and the boundary conditions are given by:
dT
Samples were cooked for 5, 10, 15 and 20 minutes k = h ⋅ (T∞ − Ts ) at x = X, 0 < t ≤ tf,
using a far infrared heater. Once the cooking cycle dx
was complete the cook losses were measured. (2)

The cooked samples were sliced into 3 samples ∂T

along the longitudinal axis on cooling, representing
= 0 at x = 0 (3)
∂x
the top, middle and bottom of the beef patty. The
top slice represents the portion of the beef patty where k (W.m-1.K-1) is the thermal conductivity of
closest to the far infrared heat source, whereas the the beef patty and h is the heat transfer coefficient
bottom slice represents the portion of the beef patty (W.m-1.K-1). The heat transfer coefficient was
furthest away from the heat source during cooking. calculated by:
Proximate analysis was carried out in order to
h = σ (T∞ − Ts )
3
determine the moisture and fat profiles in the (4)
different slices during cooking. The moisture and
fat content of the sliced beef patties was determined where σ is the Stefan-Boltzman constant (5.669 x
using AOAC methods (AOAC, 1990). 10-8 m2.K4). This allows for the absorption of the
far infra-red radiation by the beef patty, and enables
Far infrared heater the temperature of the infra red source to be used as
an initial boundary condition. Physical property
Samples were cooked in a far infra oven, with data for the beef patties at the different fat contents
internal dimensions; height 330 mm, width 330 mm will vary with increasing temperature. Data was
and depth 340 mm. The heating element consisted calculated using the expressions of Choi & Okos
of a far infrared gas catalytic heater (Bruest (1985) for thermal conductivity and density, as
Flameless Catalytic Heater S12-12FM; Catalytic follows;
k = k1 + k 2 ⋅ T − k 3 ⋅ T 2 (5) Moisture loss during cooking can be considered
analogous to moisture loss during drying, thus use
of the equation here is justified.
ρ = r1 − r2 ⋅ T − r3 ⋅ T 2 (6)

where k1, k2, k3, r1, r2 and r3 are constants. Results

for the heat transfer in higher content beef patties The equations were solved in a discretised finite
indicated that heat transfer was not occurring by difference domain, using the differential equation
convection alone, but that movement of moisture solver, DESIRE. The beef patty was divided into
and fat contributed to heat transfer by internal 10 nodes of equal thickness.
convection. Thus equation 6 was modified to
include a convection term, designated an effective RESULTS AND DISCUSSION
heat transfer coefficient, as follows; HEAT TRANSFER

The results for the temperature penetration into the

(
k = k1 + k 2 ⋅ T − k 3 ⋅ T 2
) + (heff 1 + heff 2 ⋅ T ) beef patties during cooking can be seen in Figures 2
(7) and 3 for the 15 and 22 % fat beef patties
respectively. The time temperature profiles were
where heff1 and heff2 are constants. found to be very similar to those found previously
by Sheridan & Shilton (2002b). This data was
modelled using the model developed in Shilton et
MASS TRANSFER al., (2002), the results are shown in Figures 1 and
2. There is an excellent agreement between the
In a previous paper (Shilton et al., 2002) the
experimental and predicted data, the mean square
evaporative moisture loss from beef patties was
predicted. In this paper the total moisture loss is deviation is in the range 1.91 to 3.51 °C for the two
being considered by the following equations. The fat contents. From Figures 1 and 2 it can be seen
same equation set could be used in the prediction of that after 600 s (10 minutes) the temperature in the
fat loss from the beef patties. When predicting centre of the beef patty has risen above 80 °C for
moisture loss, C represents moisture content both the 15 and 22% fat beef patties. If this were a
(d.w.b.), when predicting fat loss, C represents fat thinner patty, then it could be considered cooked at
content (d.w.b.). Unsteady state one dimensional this point, as the temperature has been exceeded 70
mass transfer in a slab can be described by: °C for 2 minutes (Food Safety Authority, 1999).

∂C ∂ 2C 150.00
=α 2 (8)
∂t ∂x 100.00

where C is the concentration of the moisture in the 50.00

beef patty (kg moisture / kg solid). The initial
0.00
conditions are C=Ci for all locations and the
0.00 500.00 1000.00 1500.00
boundary conditions are given by:
t7 Sheridan 8mm Sheridan 12mm
dT
Deff = K m ⋅ (C ∞ − C s ) at x = X,
dx Figure 1: A comparison between experimentally derived
data and the finite difference model of the heat transfer in
0 < t ≤ tf, (9)
beef patties containing 15% fat during cooking by far-
∂C infra red radiation.
= 0 at x = 0 (10)
∂x
120.00
where Deff is the mass diffusion coefficient (m2.s-1) 100.00

and Km is the mass transfer 80.00

coefficient (m.s-1). The mass diffusion coefficient 60.00

40.00
for both moisture and fat loss predictions was 20.00
calculated using the following expression (Maroulis 0.00
et al., 1998); 0.00 200.00 400.00 600.00 800.00 1000.0 1200.0 1400.0
0 0 0

= D x ⋅ exp T  ⋅ exp c  (11)

D D
Deff t7 Sher idan 8mm Sher idan 12mm

 T   C
Figure 2: A comparison between experimentally
where Dx, DT and Dc are constants. Maroulis et al. derived data and the finite difference model of the heat
originally derived this expression for calculating transfer in beef patties containing 22% fat during cooking
mass diffusivities during air drying processes. by far-infra red radiation.
MASS TRANSFER Change in fat content
Change in moisture content
Figures 5 & 6 show the change in fat composition
From Figures 3 & 4 it can be seen that during the of the three layers in the beef patties for the two
cooking process the middle portion of the beef different fat compositions. In both cases the bottom
patty of both fat contents maintains a higher layer undergoes the largest decrease in fat
moisture content than the top or bottom portions. composition during the initial stages of the cooking
The top portion initially is the source of most of the process. However by the end of the cooking
moisture loss, by evaporation from the surface. process this bottom layer can be seen to have the
Once heat has penetrated sufficiently into the beef highest fat content of the three layers, this is
patty, the bottom portion loses moisture more pronounced in the lower (15%) fat content beef
rapidly, and moisture loss is occurring from both patty. Thus fat is moving from the top and middle
surfaces. This was confirmed by the presence of layers into the bottom layers as a result of
water in the drip fraction, however it is postulated diffusion. This process is probably accelerated as
that evaporation will also be occurring during the the temperature of the beef patty increases. The hot
latter stages of the cooking process. Formation of a fat is being driven downwards through the beef
surface crust on top of the beef patty may well patty. It can be seen that the change in fat content is
provide a barrier for mass transfer. much greater in the higher (22% fat) fat samples
than with the lower (15%) samples, this is in
agreement with the results of Andersson et al.
Moisture Content (kg/kg)

2.6

2.2 (2000).
1.8
0.4
Fat Content (Dry weight basis)

1.4
0.38
1
0 200 400 600 800 1000 1200 1400 0.36
Time (s)
0.34

Top Middle Bottom Predicted

0.32
0 200 400 600 800 1000 1200 1400
Figure 3: Change in the moisture content of slices of Time (s)

15% beef patties during cooking by far infra red radiation Top Middle Bottom Predicted
compared against an average moisture content predicted
using the model. Figure 5: Change in the fat content of slices of 15% beef
patties during cooking by far infra red radiation
compared against an average fat content predicted using
1.90 the model.
Moisture Content (kg/kg)

1.65
0.60
Fat Content (Dry weight basis)

1.40
0.55

1.15
0.50

0.90
0.45
0 200 400 600 800 1000 1200 1400
Time (s) 0.40
0 200 400 600 800 1000 1200 1400
Top Middle Bottom Predicted Time (s)

Top Middle Bottom Predicted

Figure 4: Change in the moisture content of slices of Figure 6: Change in the fat content of slices of 22% beef
22% beef patties during cooking by far infra red radiation patties during cooking by far infra red radiation
compared against an average moisture content predicted compared against an average fat content predicted using
using the model. the model.

All of the layers tend towards the same final fat

The average predicted moisture loss is also shown content towards the end of the cooking process,
in Figures 3 & 4. It can be seen that in both cases regardless of the initial fat content. This would
there is very good prediction of the loss of total imply that there is an equilibrium fat content for
moisture from the beef patties during drying. The products such as beef patties. In this case it would
mean square deviations are 0.08 and 0.08 kg/kg for appear to be around 0.35 (d.w.b.). Goihl et al.
the 15% and 22 % fat contents respectively. These (1992) monitored the fat loss from beef steaks
moisture loss profiles have a similar profile to that during contact grilling. Most fat is lost from the
observed by Pan et al., (2000) for 24 % fat content bottom of the beef steak, in this case in contact with
beef patties cooked using a contact heating process. the heating surface. The researchers also found that
untrimmed beef steaks experienced a high degree Transport Phenomena, (M. Le Maguer & P. Jelen
of fat penetration from the external fat during Eds.), Volume 1, p.p. 93-101, Elsevier Science,
cooking, with the fat content inside the meat rising London.
to over 100% of the original fat content. Olsson & Food Safety Authority of Ireland, (1999), E. coli O157:
Preventing the Spread of Infection in Catering and
Tornberg (1991) found that the protein matrix Retailing, Food Safety Authority of Ireland, Dublin.
played an important role in regulating fat losses Goihl DM, KB Harris, JW Savell & HR Cross, (1992),
from beef patties during cooking. They also found Location of fat migration during cooking of beef
that fat migration from the inner to the outer loin steaks with external fat still attached, Journal of
portions of beef patty are significant in influencing Food Composition and Analysis, 5, 246-251.
fat losses during cooking. Kirchner JM, LC Beasley, KB Harris, JW Savell, 2000,
Evaluating the cooking and chemical characteristics
Figure 5 shows the prediction of the change in fat of low fat ground beef patties, Journal of Food
content from the lower (15%) fat beef patty. The Composition and Analysis, 13, 253-264
Maroulis ZB, Kiranoudis CT & D. Marinos-Kouris,
predicted fat content is an average value for the 1998, Heat and mass transfer modeling in air drying
entire beef patty. It can be seen that there is an foods, Journal of Food Engineering, 26, 113-130.
excellent agreement between the predicted data and Olsson A & E Tornberg, 1991, Fat-holding in
that observed for the top and middle portions, but hamburgers as influenced by the different
the observed results for the bottom portion do not constituents of beef adipose tissue, Food Structure,
allow for prediction by the model. Analysis of the 10, 333-344.
predicted results for the top and middle portions, Pan Z., R.P. Singh & T.R. Rumsey, (2000), Predictive
gives a similar profile with the bottom portion still modeling of contact heating process for cooking of a
giving an anomalous result. It is postulated that hamburger patty, Journal of Food Engineering, 46,
9-19.
surface tension on the bottom of the beef patty did Sheard, PR, GR Nute & AG Chappell, 1998, The effect
not allow the fat to fall as drip, thus this extra fat of cookingh on the chemical composition of meat
was retained in the portion during analysis. This is products with special reference to fat loss, Meat
despite the fact that the patty was placed on a paper Science, 49 (2), 175-191
towel during recording of the cooked weight. Sheridan P., (2000), Heat and mass transfer aspects of
cooking beefburger patties in a range of fat contents,
Figure 6 shows the prediction of the change in the with mid and far infra red, ceramic and catalytic gas
average fat content from the higher (22%) fat beef radiant sources, Ph.D., National University of
patty as compared to the observed fat contents of Ireland, Dublin.
Sheridan P.S. & Shilton N.C., (2002a), Analysis of yield
the 3 layers. In this case it can be seen that there is while cooking beefburger patties using far infrared
an excellent agreement between the model and radiation, Journal of Food Engineering, 51, (1), 3-
experimental data, with a mean square deviation of 11.
0.01 kg/kg between the predicted data and the Sheridan P.S. & N.C. Shilton, (2002b) Determination of
observed mean fat content. the thermal diffusivity of ground beef patties under
infrared radiation oven-shelf cooking, Journal of
CONCLUSIONS Food Engineering, In press.
Shilton N., P. Mallikarjunan & P. Sheridan, (2002),
A model was developed that successfully described Modeling of heat transfer and evaporative mass
losses during the cooking of beef patties using far
heat transfer, moisture losses and fat losses during infrared radiation, Journal of Food Engineering, In
cooking by far infrared radiation. Moisture is lost review.
from the top surface as a result of evaporation, and Troutt ES, MC Hunt, DE Johnson, JR Claus, CL Kastner,
from the bottom of the beef patty as a result of drip. DH Kropf & S Stroda, 1992, Chemical, physical
The results showed that fat is lost primarily from and sensory characterization of ground beef
the bottom of the patty as a result of drip, and that containing 5 to 30 % fat, Journal of Food Science,
fat moves from the top to the bottom of the beef 25 (1), 25-29
patty during the cooking process.

REFERENCES

Andersson A, K. Andersson & E. Tornberg, (2000), A

comparison of fat-holding between beefburgers and
emulsion sausages, Journal of the Science of Food
and Agriculture, 80, 555-560.
AOAC, (1990), Official Methods of Analysis, 15th ed.,
Association of Official Analytical Chemists,
Washington, DC.
Choi Y & Okos M, (1985), Effects of temperature and
composition on the thermal properties of foods, In:
Food Engineering and Process Applications,
 FUZZY
CONTROL
 A FUZZY SYMBOLIC APPROACH TO ESTIMATE THE SENSORY
STATE OF A FOOD PRODUCT : APPLICATION TO CHEESE
RIPENING
Nathalie Perrot, Laure Agioux, Irina Ioannou, Vincent Gilles Mauris
Chevallereau, Gilles Trystram
UMR Génie Industriel Alimentaire – Equipe Automatique et Qualité LAMII Université de Savoie 74000 Annecy
Alimentaire ( Cemagref), 24, Avenue des Landais, BP 50085, 63172 [email protected]
Aubière cedex, tel: (33) 473-44-06-69 – fax: (33) 473-44-06-97 email
: [email protected]

Georges Corrieu
UMR Génie et microbiologie des procédés alimentaires INRA,
INA-PG,78850 Thiverval Grignon, 330 130 815 488
[email protected]

INTRODUCTION An interpolation is applied by the operators between these

classes in a control context.
In the food industry, end-products must conform to precise
specifications. Among the latter, sensory properties are Different ways can be developed to solve this problem,
essential because they influence the choice and preferences including:
of consumers. These properties depend on the choice and Development of new sensors.
combination of “raw materials” and on an appropriate feed- Estimation of the sensory measurements of the operators
back control of the process. The capacity to manage these using an adapted mathematical framework.
properties right from the fabrication stage in an automatic
framework constitutes a key issue for companies. Little We focus our study on the second point. If lot of
research has been done in this area thus leaving the field mathematical methods can be used to estimate numerical
open to study. data (Valet and al. 2000) from linear regressions to more
elaborated one like statistical methods (Bayesian for
Managing sensory properties with the aim of controlling example), theory of evidence, neural networks, theory of
them is no easy task. This is due to different factors: fuzzy sets and theory of possibilities, few are adapted to treat
Firstly, from the point of view of automation, human sensory the data at the linguistic level manipulated by the experts.
perception of, for example, the texture of a biscuit or the The fuzzy set theory (Zadeh 1971; Davidson 1996; Wide
aspect of a sausage slice, cannot simply be manipulated on 1999) is one interesting method adapted to treat numeric and
line using existing sensors (Perrot and al. 2000). On the one linguistic format of data at a symbolic level which can be the
hand, few suitable sensors are available to carry out such linguistic level manipulated by the operators.
measurement (Trystram and Courtois 1996; Wide 1999); on
the other hand, it is difficult to develop a classical automated PROBLEMATIC
approach due in part to the many dimensions that must be
taken into account in parallel and in part to the non linearity The aim of our study is to present an application of this
and coupling between the variables involved in the system. approach to estimate the symbolic state of a cheese during
ripening, in view to help the operator in his task. This study
Secondly, on another level these sensory properties are in was developed in accordance with a research program
practice evaluated on line by the operators who use their concerning the application of the EStimation to the SENSory
evaluation in combination with existing sensors to adjust the evaluation with the application on an operation of cheese
process (Perrot and al. 2000). Product properties that ripening (ESSENS supported by MENRT and in
contribute to quality and process productivity depend mainly collaboration between the UMR GENIAL, the LGMPA of
on the accuracy of their evaluation and reaction. The INA-PG and a factory H. Mons).
integration of such knowledge in a control frame is a Cheese ripening still remind a process where the operators
relevant direction(Davidson 1994; Trystram and al. 1998). evaluation and reasoning keep a major role (Lemoine 2001).
Nevertheless, this type of measurement is specific and the Some studies have developed (i) methods to follow in line
conventional methods of descriptive sensory analysis that some sensory properties of the product like texture using
have been developed are ill-adapted (Curt and al. à paraître) sensors (O'Callahan and al. 2000) or (ii) methods to
for acquiring these evaluations on line and a specific format characterize in laboratory the sensory properties of a cheese
for variables must be used. Last but not least, the operator’s (Lesage and al. 1992; Martin and al. 1999). Few studies are
measurements are not on the same scale as sensor data. They focus on the estimation versus time of the sensory trajectory
are expressed on a discontinuous graduated scale , such as and standard deviation trajectory of the product during
classes of color defined to characterize the degree of cooking processing. Nevertheless, it is a pertinent way to help the
of a biscuit during the baking process (Perrot and al. 1996). operator in his task, and as a consequence improve the
productivity. The example treated is focus on the ripening “late” and a quantification of this state is calculate. For
of a soft mould cheese so called “rogeret” (figure 1). example at t = 10 days of ripening, a cheese is supposed to
be in a state B called “little ripened” or “state 3.5 weeks”.
Suppose a cheese evaluate to have a ripening delay of two
days : the answer of the system would be B-2days.

Our approach implements two levels of the experts'

knowledge: (i) their knowledge of the standard sensory
behavior of the cheese, and (ii) their knowledge of the
different possible trajectories around this standard behavior
stemming from experience. By hypothesis, a standard
behavior is associated to the mean trajectory of the cheese in
standard conditions of ripening.

Figure 1 : a cheese called “rogeret”

knowledge of the symbolic standard sensory behavior
.
Low
Peu humide
middle
Moyennement high
Très humide
Low
Peu humide
middle
Moyennement high
Très humide
humide humide
1 1

THE SYMBOLIC MODEL TO ESTIMATE THE

SENSORY STATE OF THE CHEESE DURING 0 1 2 3 4 5 6 0 1 2 3 4 5 6

RIPENING Humidité
Humidity Humidity
Humidité

few little middle coat few little middle coat

Très peu Couvert Peu couvert Moyentcouvert Très couvert Très peu Couvert Peu couvert Moyent couvert Très couvert
1 1

The rogeret is ripened in 1 month. The evolution of the

0 1 2 3 4 5 6 0 1 2 3 4 5 6

product which is strongly linked to complex microbiological Geotrichum

Couverture
candidumcoat
Candidum
Geotrichum Geotrichum Candidum
Geotrichum
Couverture
candidum coat
reactions (Eck and Gillis 1997) is evaluated on line by the few little middle
Pas proteolyséPeu protéolyséMoyennement high
Protéolysé

operators through 7 different sensory characteristics that

protéolysé
1

have been formalized upon the concept of sensory indicators

(Curt and al. 2001). Sensory measurements are made on a 0 1 2 3 4 5 6

CheeseProtéolyse
consistency
structured ratio scale between [0,6]. Tests of reproducibility
of their measurements have been achieved (A level of at
mean 92% on the point tested). 4 3.5
Symbolic state
Figure 3 : example of the standard symbolic behavior of the
Fuzzy symbolic approach State of
cheese at stage 4 and 3.5
Time the
Computing
cheese –
fusion of
delay to Experts have memorized the standard sensory behavior of
symbols
the
standard
their specific cheese in terms of symbols like : at a stage
Sensory
Sensory measurements measurements called “3.5 weeks” a cheese in a standard behavior should be
in symbolic
On a structured ratio scale “middle” in term of humidity, “little” in term of Geotrichum
4 indicators Candidum coat” and “few” in term of cheese consistency.
Expert Knowledge of the
standard sensory trajectory of
We express this symbolic knowledge using the theory of
the cheese fuzzy sets. Logical AND and OR functions are respectively
associated to product and bounded sum. Fuzzy membership
Figure 2 : the fuzzy symbolic approach built to estimate the functions of each sensory measurements ensure the link
sensory state of the cheese between symbols manipulate by the experts and the scale of
the sensory measurements. Figure 3 presents an example of
implemented knowledge for the two state in the beginning of
The fuzzy symbolic approach was built to inform an the cheese ripening figure 3. It involves at this stage three
operator on the potential derive of the sensory trajectory of indicators on four : humidity, Geotrichum candidum coat
the cheese at t time as to help him to control the process. The and cheese consistency.
configuration of the symbolic approach in this case is
presented figure 2. Global symbolic sensory behavior of the cheese

Around the standard behavior of the cheese, other

On the basis of 4 of the 7 indicators of the cheese quality characteristics trajectories are also memorized and expressed
(humidity, Geotrichum candidum coat, cheese consistency, in the symbolic model implemented with an example for one
color) and the expert knowledge, an information on the state of this corresponding period of ripening (state “B” or “3.5
of the product is extract : cheese “in advance”, “standard” or weeks”) presented table 1. Fusion between symbols are
calculated upon the fuzzy formalism presented below. This level of compatibility is defined upon the sensibility of
Equations (1) and (2) express the calculus applied to table 1.

Ripening state at time t=RS(t) 4,5

November 2001
4
standard
RS(t)=standard(RS)(t)-RFR(t) (1) 3,5
standard trajectory of the cheese
Expert
i =1

cheese state (weeks)

model

RFR(t ) = ∑ µRi * (delay ( Ri ))

3

2,5

n (2) 2

1,5

µRi = activation degree of each rule Ri 1

0,5

standard(RS)(t) = standard ripening state of the 0

corresponding period t : for example 3.5 weeks for 4< t<6 0 5 10 15 20 25 30 35

time(days)
days
the operator measurement.
Table 1 : example of knowledge implemented at Stage B –
« 3.5 weeks »
Figure 4 : Comparison between the evaluation of the
operator and the model on estimation of the state of the
Gc coat few little middle coat cheese during an experiment at 9°C and 92% of relative
Humidity humidity of the air
low late delay Normal In advance
=–0.5 with delay = +0.5
with drying for a cheese
drying default consistency
1,2
default middle
middle late delay Standard In advance
deviation to the standard trajectory of the cheese

1
deviation model
=–0.2 Delay=0 delay = +0.5 0,8
for a cheese deviation expert

consistency 0,6

middle 0,4

high Late Late delay

delay = – = –0.5 0,2

0.5 0
0 2 4 6 8 10 12 14 16

-0,2

-0,4

METHOD OF VALIDATION
-0,6

time (days)

A basis of 54 data is acquired in a pilot and processed to

validate our approach. The kinetics of cheese ripening are Figure 5 : estimation of the derive to the standard symbolic
measured in the pilot unit at four constant relative air sensory kinetic by the model and the operator during an
moistures and air temperatures representative of different experiment lead at at 15°C and 98% of relative humidity
trajectories of the cheese ([15°C, 98%], [9°C, 92%], [9°C,
96%], [12°C, 94%]) in the experimental domain of [9°C,
15°C] for temperature and [92%, 98%] for relative humidity.
The result is treated in the form of an evaluation of the
trajectory of the considered cheese by comparison to it’s
traditional standard behavior registered by the experts from
RESULTS experience and during experiments. An example of such an
answer of the model is presented figure 5 for a kinetic at
15°C and 98% of relative humidity. We can see that at 10
The results obtained show the ability of the approach to days, the symbolic model point out an important and
follow the sensory trajectory of a cheese during ripening and constant derive of the trajectory of the cheese which is
help the operator to diagnosis a derive of it. On an example validated by the expert and is linked to the experimental
of a kinetic at 9°C and 92 % of relative humidity of the air conditions of ripening. Such an approach, combined with a
(figure 4), we can see the estimation of the symbolic model of kinetics of some indicators could be used to predict
approach by comparison to the operator evaluation. 96% on some derive of the quality of the cheese before it appears
54 of the tested points were compatible at a grade of 0.5. and thus help the operator in it’s task.
 Perrot, N., G. Trystram, et al. 2000. “Feed-back quality
control in the baking industry using fuzzy sets.” Journal
of Food Process Engineering 23: 249-279.
Perrot, N., G. Trystram, et al. 1996. “Sensor fusion for real
CONCLUSIONS time quality evaluation of biscuit during baking.
Comparison between Bayesian and Fuzzy approaches.”
A fuzzy symbolic approach has been implemented and Journal of Food Engineering 29: 301-315.
validated on 54 data to estimate the sensory state of a cheese Trystram, G. and F. Courtois 1996. Food process modeling
during ripening. and simulation. Computerized Control Systems in the
Food Industry. G. Mittal. New York, Basel, Hong Kong,
The results are relevant with regard to the objective of Marcel Dekker Inc.: 55-85.
controlling the sensory properties of the products at the Trystram, G., J. Hossenlopp, et al. 1998. "Experiences about
fabrication stage. These first examples are encouraging for man/technology cooperation through decision systems
different reasons : for the control of food unit operations". Biodecision'98,
This approach is a simple way to integrate operator Montpellier (France).
reasoning whenever the operator is involved in Valet, L., G. Mauris, et al. 2000. "A statistical overview of
measurement diagnosis or control of the process. recent literature in information fusion". ISIF, Paris.
It allows treating the measurements from the Wide, P. 1999. “The human decision making in the dough
sensors and operators at the same level. mixing process estimated in an artificial sensor system.”
It introduces interpolation between symbols Journal of food engineering 39: 39-46.
manipulated by the operators that provides Zadeh, L. 1971. “Quantitative fuzzy semantics.” Inf. Sci. 3:
robustness but which should be quantified more 159-176.
precisely and linked to processing uncertainty.

REFERENCES
Curt, C., L. Agioux, et al. 2001."Formalisation of at line
human evaluations". 4th symposium Rose Marie
Pangborn, 22-26 July, Dijon: 97.
Curt, C., G. Trystram, et al. (à paraître). “Formalisation of at
line human evaluations. Integration of human decision in
the dry sausage ripening process.” Science des aliments.
Davidson, V. 1994. “Expert systems in process control.”
Food Research International 27: 121-128.
Davidson, V. 1996. Fuzzy control of food processes.
Computerized Control Systems in the Food Industry. G.
Mittal. New York, Basel, Hong Kong, Marcel Dekker
Inc.: 179-205.
Eck, A. and J. Gillis 1997. "Le Fromage", Lavoisier Tec et
Doc.
Lemoine, R. 2001. “L'affinage, une étape décisive pour la
qualité du fromage.” Revue Laitière Française(614): 16-
19.
Lesage, L., F. Sauvageot, et al. 1992. “Influence de la teneur
en NaCl et de la durée d'affinage sur les caractéristiques
sensorielles d'un fromage type camembert enrichi en
magnésium.” Lait (72): 73-85.
Martin, N., S. Savonitto, et al. 1999. “Flavor generation in
cheese curd by coculturing with selected yeast, mold and
bacteria.” Journal of Dairy Sciences (82): 1072-1080.
O'Callahan, D., C. O'Donnell, et al. 2000. “On-line sensing
techniques for coagulum setting in renneted milks.”
Journal of food engineering 43: 155-165.
Perrot, N., G. Mauris, et al. 2000. "Modeling of the expert
sensory evaluation of the sausage crusting using fuzzy
symbolic approach". 1st International Conference on
Simulation in Food and Bio Industries - FoodSim, Nantes,
26-27 June, France: 36-40.
 ADA PTIVE FU ZZY CON TROL AS AD VANCED CON TROL TECHNIQU E IN
M ILK PA ST EUR ISA TION .

C. Riv ero l and B. O’ Connor

Adv anced Pro cess Techno log y Group, Food and Scien ce Bu ild ing, Un iv ersity
Co lleg e Cork, Co rk, I rela nd.
email: c.rivero l@u cc.ie, b.o connor@u cc.ie
Con tro l in to n ew ar eas of app lica tion
a s d e scr ib ed h er e.

K EYWORDS
M A TH EMA T ICA L M O D EL L IN G O F
P roc es s contro l, milk, ad ap tive , TH E PLATE H EAT EXCHANGER
r o b u s tn e s s .
P l a t e h e a t ex ch an g e r ( P H Es ) exh ib i t
e x c e l len t h e a t tr an sfe r c h a r a c te r ist i c s
AB STRACT w h i ch a l low mo r e co mp a c t d e sig n s
th an are a ch ievab le w ith conv en tion a l
Th is sho r t co mmu n ic a tion ou tline s th e shell- and- tub e h eat ex ch anger s. A long
u se if an a dap tiv e fu zz y con tro l w ith th e ir ea s y ma in tena nc e and o th er
s y s t e m a s al t e r n a t iv e me t h o d f o r mi l k a dvan tag es, th is fa c tor h as ma d e PHE s
p asteur isation. A comp ar ison w ith th e pr ime c ho ice for ma n y eng in e er ing
r esp e c t to P I D c l as s ic sys t e m i s ma d e. app lications, invo lv ing liqu id -liqu id
h ea t t r an sfe r d u t i es. Th e y ar e mo s t
co mmo n ly used in th e dair y, b everag e,
INTR ODUCTION g ener a l food processing and
p h ar ma c eu t ic a l in d u str ie s d u e to th e ir
e a se o f c le an in g an d t h e i r th erma l
Th is app eal is primarily mo tiv ated b y c o n tr o l c h a r ac t er i s t ic s. The
th e d e s ire to ma in ta in a h igh d egre e of fund amen ta l en erg y conserv a tion law s
produ ct qu ality wh ile op erating at th e le ad to equ a tions that de scr ib e the
h igh est po ssib le produ ction rates h eat tr ansf er betw een an y ch annel and
( Edan , 1992). Adv an ced pro cess its neighbour ing p lates. Th ese ar e
c on tro l te chn iqu es h ave b een typ ically o f th e fo llowing fo rm:
successfu lly used to imp rov e
produ ctiv ity in o th er pro cess
industr ie s, such as the p e tro l- chemical Chann e l j:
industr ie s. Th is f act sugg ests th at th e
food industr y can also realise similar
 ∂T j ∂T j 
b enef its fro m advan ced con tro l ρC p Ax  − vj ( ) (
 = U j A j T pj −1 − T j + U j A j T pj − T j )
te chno log y. Th er e are , howev er, a
 ∂t ∂z 
nu mb er of ob stacles, wh ich mitig ate
again s t th e transfer o f th is techno log y
in to th e food /b io
P l a t e j:

 ∂T pj ∂T j 
ρ pC pp Axp  − vj (
 = U j A j T pj −1 − T pj )
 ∂t ∂z 
s e c to r . Th e se inc l u d e i ss u e s
concern ing lack /unav a ilab ility of (
+ U j A j T j +1 − T pj )
produ ct and pro cess d a ta and
know ledge, econo mic f acto rs and th e
wh ere t deno te time and z ax ia l
trad itional conserv a tism o f th e sector.
po sition w ith a chann e l, T j (z , t) is th e
Adv anc es in ha rdware sp eed a nd
te mp e ra tur e of th e mil k in channe l j
c ap a c it y , co m mu n i ca t io n s an d
a n d T p j (t) t h e t e mp era tu r e o f t h e me t a l
suppor ting so ftwar e r esu lting in
i n p l a te j. The qu an tity n j takes a
in creased pr esen ce of f acto r y ‘ shop-
v a lue of +1 o r –1, d ep end ing on th e
f loor’ informa tio n s ys te ms h av e ma de
d ir ec tion of the f low in ch ann e l j. Th is
po ss ib le the e x ten s ion of Adaptiv e
g ener a l mod e l can b e u sed to s imu la te
th e tr ansien t b ehav iour o f several R ESU LTS AND C ONC LUSIONS
typ es of PH Es, in clud ing counter-
c urr en t, co-c urr en t, mu lti- chann e l etc .
Th e ad ap t iv e c o n tr o l sys t e m a c h i ev es
pro cess r egu lation b y specif ying th e
ADA PTIV E FU ZZY SYS TEM d es ir ed p lan t ou tpu t (s e t po in t 72.3
CON TROL D ESIGN ºC). Th e system g e ts th e settling time
shor tly w ithou t ov ershoo t as d ep icted
th e Figur e 2.
Th e ad ap tive con tro l used in th is paper
i s d ep ic t e d i n t h e F ig u r e 1 . I n t h i s Tmilk(x,t)

a pproa ch the “ad ap ta tion me ch an is m”

ob serv es the sign als fro m th e contro l 73

s y s t e m a n d a d ap t th e p ar a me t e r s o f t h e 72
con tro ller to ma in tain p erforman ce 71
e ven if th ere ch ange s in the p lan t. Th e 70
d es ir ed p erfor ma n c e is ch ar ac teris ed 69
w i th a “re f e r en c e mo d e l ” as t h e
68
d es cr ib ed abov e, and th e con tro lle r 20
15 0.2
th en s eek s to ma ke th e c lo sed-loop 10 0.15
0.1
s y s t e m b eh a v e a s th e r ef er en c e mod e l 5
0
0.05
0
wou ld even if th e p lant c hang es . Th is Time t Distance x

is called “mo d e l reference ad aptiv e

con tro l “ (MRA C) ( San chez, 1997).
Figur e 2 . Milk temp er atur e as function
of time and ax ia l position (z)

Adaptive
Adaptivemechanism
mechanism
Th e te mp era t u r e o f the h o t w a t er w as
SP
SP
simu lated as well as d ep icted th e
controller
controller plant
plant Figur e 3 wh ere th e inpu t te mp er atur e
w as 73.8 ºC (v er y closed to the
s e tpo in t) for avo id ov er p as teur isa tion.
Model
Model

T hotwater(x,t)

F igur e 1 . MRA C mod e l

74

Th e learn ing mechanism tu n es th e 72

r u l e- b as e o f d ir e c t f u z zy c o n tr o l le r s o
t h a t the c lo sed- lo o p sys t e m b eh av es 70

l i k e the r ef e r en c e mo d e l . Th e se r u l e-
b as e mo d if ica tion s a re ma d e b y 68
20
ob serv ing d a ta fro m th e con tro lled 15
10 0.15
0.2

p r o ce ss , r ef er en c e mod e l and f u z z y 5 0.05

0.1

0
c o n tr o l le r . Th e l ea r n in g me c h a n is m Time t 0
Distance x

c o n s i s ts in t w o p ar t s : “ f u zz y in v e r s e
mo d e l” and a “ know ledge-b ase F ig u r e 3 . H o tw a ter te m p era tu r e a s
mo d if ier ”. Th e fu zz y inver s e mod e l fun c tion of th e time and ax ial position
p erfor ms th e fun c tion of ma pp ing
(r epr es en ting the de v ia tion fro m
d es ir ed b eha v iour), to c hang e in th e Th e mo de l w as u s ed w i th a P ID
p r o ce ss in p u ts th a t are n ec e ss ar y t o con tro ller. Th e loop turn ing was ma d e
for ce to zero. Th e know ledge-b ase u sing the trad itional Zieg ler-Nicho ls
mo d if ier perfor ms the fun c tion of Method . The resu lt is d ep icted in the
mo d if ying th e fu z z y con tro ller ´s ru le- Figur e 4 .
b as e to affe c t th e n eed ed ch ang es in
th e pro cess inpu ts .
 80

70

60

50

40

30

20

10

0
0 2 4 6 8 10
Time

Figur e 4 . Milk temp er atur e u s ing PID

c o n tr o l le r .

Th e system d id no t show overshoo t

with a qu ic k settling ti me b u t th e
off se t is 2.4 ºC in con sequ enc e th e
qu ality of th e produ ct wou ld b e
s er io u s l y aff ec t ed.

A g l anc e f o r thr e e f ig u r es a l low s t o

conclud e that th e MRAC me thod used
in th is work gu aran te e th e quality of
th e milk avo id overheating and quick
c hang es in th eprof ile of te mp er a ture .

R EF ER ENC ES

Ed an Y , P. Gr insp en a nd H. K ahn
(1992) . “Fuzzy log ic app lications in
d a ir y indu str y”, In ternational W in ter
Me e t ing of A SAE, N ashv il l e ,
T enn es se e, (D ec ), 1-15 .

San chez R. and J. Rub io (1997 )

“Con tro l adap tivo y Robusto ”,
Un iv ersid ad d e Sev illa, Sp ain, 55-66.
NUTRITION
 FREEZING IN POROUS MEDIA
Jean-Yves Monteau
Nasser Hamdami
Alain Le Bail
GEPEA – ENITIAA
Laboratoire de Génie des Procédés Alimentaires
Rue de la Géraudière, BP82225
44322 Nantes Cedex 3, France
E-mail: [email protected]

Korin 1997; Miyata 1998; Bronfenbrener and Korin 1999),

KEYWORDS others on the frost damage of concrete by threeze-thaw
Freezing, porous media, heat and mass transfer, modeling. cycles are presented by (Chatterji 1999a; Chatterji 1999b).
Besides, freezing of porous food material are not
numerous. Immersion freezing is presented and modeled
ABSTRACT by (Lucas et al. 1999; Lucas et al. 2000). These works are
related to heat and mass transfer in a non-porous biological
Heat and mass transfers are important phenomena matrix. Most of these articles are general and consider the
involved in the freezing of porous media; this is especially problem as a theoretical one with minimal concern on the
valid for product with strong porosity. However, few application (Weaver and Viskanta 1986; Sasaki 2000;
studies are available in the case of foodstuffs. In this paper Song and Viskanta 2001).
a model is proposed for the freezing of bread. The
dominating physical phenomena are conduction for the According to the problem which is considered, equations
heat transfer inside the product and diffusion for the mass describing the phenomena are more or less complex. For
transfer. However, evaporation-condensation and example, very often for foodstuffs, the mass transfer has a
capillarity play an important role in both heat and mass minimal influence on the heat transfer, especially for low
transfer. On the external surface heat transfer is driven by porosity media. By the way, mass transfer is often
convection, radiation and evaporation or sublimation. neglected or modeled in a simple approach. In the case of
Mass transfer is driven by evaporation or sublimation. The high porosity products such as bakery products, mass
model presented in this work formalizes these exchanges transfer can’t be neglected. Mass transfer, which is mainly
by introducing an effective thermal conductivity which related to water diffusion in the case of porous biological
accommodates evaporation-condensation phenomena. substances, is proportional to the gradient of partial
Mass transfer is modeled by the Fick’s law. Thermal pressure of water in the product. Water has a very high
conductivity is modeled by using a Maxwell type model latent heat and a quite moderate specific heat. Water
with four components: ice, liquid water, dry matter and air. diffusion by evaporation-condensation can be taken into
The resolution is achieved by a hybrid method which uses account thanks to an effective thermal conductivity which
an enthalpic formulation and an apparent specific heat. accommodates conductivity of the matrix and water
Temperature changes and mass loss are compared with diffusion. This concept can be used to model coupled heat
experimental results obtained with a sponge which was and mass transfer (Pham 2001).
used as a model product. Experimental and numerical
results are satisfactory, but show that the thermal In some simple cases, an analytical solution can be
conductivity model has to be improved. obtained (Mikhailov 1976; Marcus and Tarzia 2000).
Generally a computational solution is necessary
(Campañone et al. 1998). Specific software can be
INTRODUCTION developed such as “Baktix” for the simulation of cooling
and freezing of bakery products (van der Sluis 1993).
Heat and mass transfers are important phenomena This work is presenting a model developed for bread
involved in the freezing of porous media. freezing. The industrial application concerns part baked
frozen bread. Transport phenomena involved in the case of
Porosity permits to quantify the void fraction of a porous bread freezing are first presented and the model used is
medium. Freezing of porous media has been studied quite presented. Results obtained with this model are then
extensively for low porosity media (i.e. rock, sand, soil…) discussed.
whereas few studies are available for high porosity media
(i.e. bread, …). For example, studies on phase front
propagation in soils are proposed by (Bronfenbrener and
TRANSPORT PHENOMENA DURING THE Governing Equations
FREEZING
The equation describing the temperature change can either
Transfer takes place on the surface by convection, use the temperature as primary variable, or use the
radiation, and evaporation. Evaporation results in water enthalpy. In the first case a variable called apparent
loss, which can interact with the final quality of the specific heat is defined. The apparent specific heat
product such as crust flaking. includes the heat capacity and the enthalpy of fusion of
A temperature gradient is established between surface and ice. The equation is
center of the bread. The water vapor pressure is higher for ∂T ∂ 2T
c(W, T) = λ(W, T) 2 (1)
higher temperature. Thus, water diffuses through the ∂t ∂x
crumb from the warmer zones towards the cooler zones, The drawback of this method is that the apparent specific
i.e. from center towards the periphery. Condensation might heat traced versus the temperature presents a peak, which
occur if the temperature of the crumb reaches the dew can lead to numerical non-stability.
point of the humid air contained in the bread. This An enthalpy formulation such as the equation (2) permits
evaporation-condensation mechanism plays a significant to avoid these problems and was used for the model.
part in the heat transfer inside the product (de Vries 1958).
∂H  ∂ 2T 
Heat and mass transfer also occurs by capillarity inside the = λ(W, T) 2  (2)
wall of the porous matrix (van der Sluis 1993). ∂t  ∂x 
The thermal conductivity used in this model was an
Radford et al. (Radford et al. 1976), while working on the effective thermal conductivity that includes evaporation-
meat chilling, indicated that at the beginning of chilling, condensation and capillarity phenomena. It is modeled
evaporation is higher than the diffusion of water in the with a Maxwell type model, which considers a dispersed
product, and thus the surface dries out rapidly. Then a phase in a continuous one. The effective thermal
balance settles between the diffusion and evaporation. In a conductivity of Maxwell model is written
third phase, the difference in water vapor pressure between  λ + 2λ c − 2ε d (λ c − λ d ) 
the product surface and the ambient air decreases. The λ = λc  d 
diffusion becomes the governing phenomena compared to  λ d + 2λ c − 2ε d (λ c − λ d ) 
evaporation, and the surface re-wets. where λ c stands for the continuous phase, λ d for the
dispersed, and ε d for the occupation ratio (by volume) of
Freezing starts when the surface temperature is sufficiently
low. A phase front travels through the product while the dispersed phase.
frozenable water turns to ice. The freezing of water does Four constituents were considered in our case: ice, liquid
not prevent a transfer of water in vapor form (Campañone, water, dry matter and air. In a first stage ice is the
Salvadori et al. 1998). At the same time, the frozen zone dispersed phase, and water the continuous one. In a second
dries out by sublimation of the ice, from a front receding stage this system constitutes the dispersed phase, and the
into the product much more slowly than the phase front of continuous one is dry matter. In a third and last stage the
freezing (Pham and Willix 1984). The thermal transfer is continuous phase is formed by the previous part, and the
done mainly by conduction in the dehydrated zone. dispersed phase is air. Thermal conductivity expressions
for ice and water are from (Singh 1992), and for air from
(Özişik 1985). Thermal conductivity of dry matter has
been previously measured with a transient hot source
MODELING
method. Air conductivity is modified to take into account
evaporation-condensation (Sakiyama et al. 1999).
The model was developed for a slab (1D) of thickness 2l D Pair dP
and of infinite length and width. λ ∗air = λ air + wa L a w sat
RT Pair − a w Psat (T) dT
The mass transfer is modeled by the Fick’s law amended
with a source term to express the ice formation.
∂W  ∂ 2 W  ∂X i
x = D(Ta ) 2  −
∂t  ∂x  ∂t
-l 0 l where
xi
X i = (W + 1)
1− x i
The mass diffusivity D was modeled by the Arrhénius’s
law.
Figure 1: Slab of Infinite Length The change of mass fraction of ice was modeled by the
conventional equation below where x b is calculated
according to the Pham’s method (Pham 1987).
  T − 273,15  for example. The enthalpy change / temperature change
x i = (x tw − x b )1 − f  if T < Tf ratio gives an estimate of voluminal heat at time step m+1,
 T − 273,15 
c m +1 . The use of the three time-steps Lees’s scheme (Lees
xi = 0 if T ≥ Tf 1966) enables to solve the temperature heat equation (1) at
The method of resolution used requires the knowledge of a the time step m+2. To not jump the peak of phase change,
table T-H. This table is calculated using the apparent the temperature is corrected according to the following
specific heat method: enthalpy at time-step m+2 is calculated for every
dx space-step with
Cp app = Cp s (T)x s + Cp w (T)x w + Cp i (T)x i ( x, T) − L i
dt H m + 2 = H(T m ) + c m +1 (T m + 2 − T m )
with L, latent heat of fusion. H(T m ) is calculated by interpolation in table H-T. Then
Polynomial expressions for the Cp functions are drawn
from (Singh 1992). T m + 2 is calculated in the same way by T(H m + 2 ) .
H(T) is then calculated by integration of Cp since –40 °C.
H or T is calculated by interpolating from the values of EXPERIMENTS
this table.
A vegetable sponge Cello G, reference F44110 (Euréponge
Boundary Conditions – France) was chosen as a model product. The size of the
sample was 10 cm×10 cm×4 cm. After washing and
Forced convection, radiation and evaporation or rinsing, the sponge was humidified to obtain a water
sublimation were considered for the heat balance at the content of 0.6 kg water/kg product. Four K-thermocouples
external surface of the product. were placed on the surface, at 1 cm, 1.7 cm and 2.6 cm
∂T below the surface respectively. The sample was installed
− λ(W, T) (l, t) = h m (Ts − Ta ) + εσ(Ts4 − Ta4 ) + k g L(Ps − Pa )
∂x in a polystyrene box to insulate the side walls and the
lower face. The system (sponge in the polystyrene box)
L is the latent heat of water vaporization if T ≥ Tf and of was placed in a freezing tunnel at –40 °C. The relative
water sublimation if T < Tf . humidity of air was logged throughout the experiment. Air
flowed parallel to the product surface at a velocity of 1
h m is given in experiments. k g is calculated by (Özişik m/s. Sample mass was measured before and after the
1985): experiment. Thermocouples temperatures were logged on
hm M 2 a data acquisition system every 4 s.
= a Patm Cp a (Le) 3
kg Mw
freezing tunnel
Ps and Pa are calculated by
Ps = a w Psat (Ts )
and
Pa = H a Psat (Ta ) air flow
where a w is water activity on surface calculated according •
• thermocouples
to the sorption curves. •
•
On surface, the mass transfer is driven by evaporation or
sublimation.
∂W
− D(Ta )ρ s (l, t) = k g (Ps − Pa )
∂x
A null heat flux and mass transfer conditions were
assumed at center of the slab.

Initial Conditions polystyrene

po lystyrene sponge

Temperature and moisture were supposed constant and

Figure 2: Experimental Assembly
uniform in the product at time zero.
RESULTS AND DISCUSSION
SIMULATOR
Experimental and computed temperature profiles are
presented in figures (3) to (6). For the curves at 1 mm
The method of resolution used is a hybrid temperature-
below the surface, some oscillations of the computed
enthalpy method suggested by Pham in (Pham 1985). This
temperature profile were observed until 500 s. After this
method is unconditionally stable. It is based on a finite
limit, the agreement was acceptable, the computed
differences scheme. The equation (2) enables the enthalpy
temperature being 2 K lower than the measured one. This
change estimation on two time-steps, from H m to H m + 2 difference vanishes with time. For the 1.1 cm position, the
computed temperature starts above the measured one T (K)
because the real initial temperature was not strictly 295
uniform in the sample. Then, up to 1700 s, the two curves
290
were superposed, although the computed one exhibited
some oscillations during phase change (time 1000 s). Then 285
the two curves diverge after the phase change plateau. At
280
1.8 cm the curves evolve in the same way, but phase
change was delayed. The two curves for the 2.6 cm 275
position are the same except at the beginning, but are
270
distinct at the phase change time, at 3200 s.
The experimental mass loss during freezing was 1.23 % of 265
the total mass whereas the model indicated only 0.5 %.
260
This difference can be explained by the fact that the 0 500 1000 1500 2000 2500 3000 3500 4000
computed surface temperature decreased much faster than t (s)
the real temperature. This yielded in a lower vapor
pressure difference between the surface and the ambient Figure 5: Temperature Profile at 1.8 cm below the Surface
air in the simulation in comparison with experiments. experimental
computed
T (K) T (K)
300 295
290
290
280
270 285
260
280
250
240 275

230 270
0 500 1000 1500 2000 2500 3000 3500 4000
0 500 1000 1500 2000 2500 3000 3500 4000
t (s)
t (s)
Figure 3: Temperature Profile at 0.1 cm below the Surface
experimental Figure 6: Temperature Profile at 2.6 cm below the Surface
computed experimental
computed
T (K)
295
290 CONCLUSION
285
280 The simulator developed in this study gave satisfactory
results. A good agreement was observed between
275
experimental and computed temperature profiles whereas a
270
larger difference was obtained for the water loss. The
265 differences observed can be explained by an under
260 estimation of the evaporation-condensation phenomena in
255 the thermal conductivity. The continuation of this work
250 will consist in developing more realistic models of thermal
0 500 1000 1500 2000 2500 3000 3500 4000 conductivity.
t (s)
Figure 4: Temperature Profile at 1.1 cm below the Surface NOMENCLATURE
experimental
computed aw water activity
c voluminal heat, J/(m3.K)
Cp specific heat, J/(kg.K)
D mass diffusivity, m2/s
Dwa diffusivity of water vapor in air, m2/s
hm thermal convection coefficient, W/(m2.K)
H voluminal enthalpy, J/m3 frozen porous materials.” Cement and Concrete Research,
Ha relative humidity 29, No. 5, 781-784.
de Vries, D. A. 1958. “Simultaneous transfer of heat and
kg mass convection coefficient, s/m moisture in porous media.” Transactions, American
L latent heat or fusion or sublimation, J/kg Geophysical Union, 39, No. 5, 909-916.
Le Lewis number Lees, M. 1966. “A linear three-level difference scheme for
M molecular mass, kg quasilinear parabolic equation.” Mathematical and
Computer Modelling, 20, 516-522.
P pressure, Pa Lucas, T., D. Flick, J. M. Chourot and A. L. Raoult-Wack. 2000.
R molar gas constant, 8.314510 J/(mol.K) “Modeling and control of thawing phenomena in solute-
t time, s impregnated frozen foods.” Journal of Food Engineering,
T temperature, K 45, No. 4, 209-218.
W humidity, kg liquid water/kg dry matter Lucas, T., D. Flick and A. L. Raoult-Wack. 1999. “Mass and
x spatial coordinate, m thermal behaviour of the food surface during immersion
x mass fraction, kg/kg product freezing.” Journal of Food Engineering, 41, 23-32.
Xi local ice content, kg ice/kg dry matter Marcus, E. A. S. and D. A. Tarzia. 2000. “Explicit solution for
freezing of humid porous half-space with a heat flux
condition.” International Journal of Engineering Science,
Greek letters 38, No. 15, 1651-1665.
Mikhailov, M. D. 1976. “Exact solution for freezing of humid
ε product emissivity porous half-space.” International Journal of Heat and Mass
εd voluminal ratio of the dispersed phase Transfer, 19, 651-655.
Miyata, Y. 1998. “A thermodynamic study of liquid
λ thermal conductivity, W/(m.K) transportation in freezing porous media.” JSME
ρ density, kg/m3 International Journal - Series B (Fluids and thermal
σ Stefan-Boltzmann constant, 5.67051 10-8 engineering), 41, No. 3, 601-609.
Özişik, M. N. 1985. Heat transfer - A basic approach.
W/(m2.K4)
McGRAW-HILL INTERNATIONAL EDITIONS,
Singapore.
Subscripts Pham, Q. T. 1985. “A Fast, Unconditionally Stable Finite-
Difference Scheme for Heat Conduction with Phase
a ambient Change.” International Journal of Heat and Mass Transfer,
air air 28, No. 11, 2079-2084.
app apparent Pham, Q. T. 1987. “Calculation of bound water in frozen food.”
atm atmospheric Journal of Food Science, 52, No. 1, 210-212.
b bound water Pham, Q. T. 2001. Modelling thermal processes: cooling and
freezing. In Food Process Modelling, L. M. M. Tijskens, M.
c continue
L. A. T. M. Hertog and B. M. Nicolaï (Eds.), CRC Press.
d dispersed England, Cambridge, 312-339.
f freezing point Pham, Q. T. and J. Willix. 1984. “A model for food dessication
i ice in frozen storage.” Journal of Food Science, 49, 1275-1294.
s solid Radford, R. D., L. S. Herbert and D. A. Lovett. 1976. “Chilling
s surface of meat - A mathematical model for heat and mass transfer.”
sat saturation International Institute of Refrigeration. Bulletin Annexe,
tw total water 1976, No. 1, 323-330.
w liquid water Sakiyama, T., M. Akutsu, O. Miyawaki and T. Yano. 1999.
“Effective thermal diffusivity of food gels impregnated with
air bubbles.” Journal of Food Engineering, 39, 323-328.
Sasaki, A. 2000. “Transient heat characteristics of water-
REFERENCES saturated porous media with freezing.” Heat Transfer-Asian
Research, 29, No. 4, 300-316.
Bronfenbrener, L. and E. Korin. 1997. “Kinetic model for Singh, R. P. 1992. Heating and cooling processes for foods. In
crystallization in porous media.” International Journal of Handbook of Food Engineering, D. R. Heldman and D. B.
Heat and Mass Transfer, 40, No. 5, 1053-1059. Lund (Eds.), Marcel Dekker, INC. New York, 247-276.
Bronfenbrener, L. and E. Korin. 1999. “Two-phases zone Song, M. and R. Viskanta. 2001. “Lateral freezing of an
formation conditions under freezing of porous media.” anisotropic porous medium saturated with an aqueous salt
Journal of Crystal Growth, 198/199, 89-95. solution.” International Journal of Heat and Mass Transfer,
Campañone, L. A., V. O. Salvadori and R. H. Mascheroni. 1998. 44, No. 4, 733-751.
"Modelling and simulation of heat and mass transfer during van der Sluis, S. M. 1993. "Cooling and freezing simulation of
freezing and storage of unpacked foods". In IIF-IIR - bakery products". In IIR Meeting Comm. B1, B2, D1, D2/3
Commissions B1, C1 and C2 (Orsay), 180-197 (Palmerston North, November)
Chatterji, S. 1999a. “Aspects of the freezing process in a porous Weaver, J. A. and R. Viskanta. 1986. “Freezing of liquid-
material-water system - Part 1. Freezing and the properties saturated porous media.” Transactions of the ASME, 108,
of water and ice.” Cement and Concrete Research, 29, No. No. August, 654-659.
4, 627-630.
Chatterji, S. 1999b. “Aspects of the freezing process in a porous
material-water system - Part 2. Freezing and properties of
 A STRUCTURED MULTI-FACTOR APPROACH TO INVESTIGATE THE
EFFECTS OF FAT REDUCTION IN THE CREAM OF A BISCUIT

Eddie Schrevens Theo Neutelings

Gerry De Rijck
Hans Coppenolle
Faculty of Agricultural and Applied Biological Sciences
Research Unit for Quality Management
K. Universiteit Leuven LU-Belgium, DANONE
Kasteelpark Arenberg 30, 3001 Heverlee, Belgium De Beukelaer-Pareinlaan 1, 2200 Herentals,
Belgium
Email: [email protected]

KEYWORDS (texture) are considered as response variables. In the

Mixture experiments with process variables, optimization, original experiment also organoleptic properties are
measured, these are omitted due to lack of space.
fat reduction, aroma, emulsifier, food sciences.
ABSTRACT MATERIAL AND METHODS

This research investigates the possibility to reduce the fat Experimental design
content in the cream of a cracker without a loss in quality The major components of the cream: cacao, sugar and fat
(texture). The cream consists of 3 ingredients: cacao, sugar can be considered as a 3 component mixture system (Figure
and fat. The effect of both an emulsifier and an aroma in 1). The operating composition 1 (12 % cacao, 56 % sugar,
compensating the fat reduction was investigated. 32 % fat) is used in the cream production. The aim of the
Using a unifactorial approach, reducing only the amount of experiment is to reduce the fat proportion and add aromas
fat in the cream, also changes the total amount of the cream and/or emulsifier to compensate for the quality loss.
and the proportions cacao, sugar and fat. In this approach, The experimental design points 1, 2 and 3 form a {3,1}
the experimental factors are mutually confounded. To simplex lattice screening design (Figure 1, Table 1). This
avoid this correlation between the experimental factors, a experimental design is D-optimal to fit a first degree
multifactorial approach is used based on “mixture theory”. canonical polynomial. This polynomial can be used to
The properties of mixture systems are determined by the represent the response surface over the experimental region,
proportions of their components, rather than by their situated between the three experimental design points.
quantitative amounts. It is not possible to experiment with
the composition of mixture systems in a classical Table 1: Proportions of the experimental design points in
orthogonal or in a unifactorial way. Mixture systems the mixture components
demand a specific design and analysis of the experiments. Mixture Cacao Sugar Fat
The cream was considered as a 3 component mixture
1 0.12 0.56 0.32
system consisting of the components cacao, sugar and fat.
2 0.17 0.56 0.27
The emulsifier and the aroma were considered as process
3 0.12 0.61 0.27
variables. Process variables are variables that are not
mixture variables by may effect the properties of the To prevent high collinearity between the independent
mixture systems. variables due to the small range of the emulsifier (0 – 1 %)
A {3,1} simplex lattice design in the mixture variables was and the aroma (0 – 0.2 %), these variables are considered as
combined with a 22 factorial design in the process variables, “process variables”. Process variables are variables that are
resulting in 12 treatments. The textural properties of the not mixture components, but may impact the blending
cream were measured with a penetrometer and analysed properties of the mixtures. Considering the emulsifier and
using specific mixture models. the aroma as process variables makes it possible to generate
INTRODUCTION for these variables an orthogonal design (22 factorial
design).
The aim of this research is to investigate if it is possible In a 22 factorial design 2 factors (variables) are investigated
to reduce the fat content in the cream of a biscuit from 32 to in two doses (levels). Each level of the first factor is
27 % without quality loss. Reduction of the fat proportion combined with each level of the second factor, resulting in
involves an increase in the sugar and the cacao proportion. a total of 4 treatments (Table 2).
To keep the same high quality of the cream two The combination of both experimental designs (Table 1 and
components are added in very low proportions: aroma and Table 2) results in a total of 12 (3 x 4) treatments and
emulsifier. Measurements of intrusion forces in time
investigates a four dimensional experimental region, The 12 creams were prepared according to general practice
determined by 5 independent variables (Figure 2). and both textural and organoleptical measurements were
To investigate if it is possible to compensate for the quality carried out. The latter are not discussed in this paper.
loss due to the reduction of fat by adding aroma and/or The texture of the cream was measured 1 hour, 1 day and
emulsifier within the 4 dimensional experimental region, 1 week after preparation of the cream at a temperature of 22
the 12 creams need to be prepared and all the response °C. In this way, time can be considered as a process
variables (texture, taste) need to be measured. variable, extending the 4 dimensional experimental region
to 5 dimensions. “Force 1” (penetrometer intrusion force)
Table 2: Proportions of the experimental design points in
is considered as an objective measurement for the textural
the process variables
properties of the cream.
Exp Unit Emulsifier Aroma LU-DANONE prepared 7 times a cream with the same
1 0 0 composition. The texture (force 1) of these creams was
2 0 0.002 measured after 1 hour, 1 day and 1 week. These
3 0.01 0 measurements are used as an error estimate for the
4 0.01 0.002 preparation of the cream and the measurement of force 1
and are used to test if there is a significant difference in
texture within the 5 dimensional experimental region.
cacao=1 RESULTS
Analysis of variance
Both the mixture components (Mix: fat, cacao and sugar),
fat=0.27 the emulsifier and time significantly effect the texture of the
cream (Table 3). The investigated aroma doses did not
significantly affect the texture of the cream.
There exist a significant interaction between the mixture
components and emulsifier, the mixture components and
aroma, emulsifier and aroma and emulsifier and time for
the texture of the cream.
2 Table 3: Analysis of variance for the texture of the cream
4
Source DF Mean square Pr > F
1 3
Mix 2 23175 0.00002
fat=1 sugar=1 Emulsifier 1 1763584 0.00000
Aroma 1 1547 0.25529
Figure 1: Experimental design in the cream components Time 2 1101169 0.00000
cacao, sugar and fat Mix*emulsifier 2 58303 0.00000
Mix*aroma 2 8648 0.00380
Mix*time 4 1291 0.36261
emulsifier Emulsifier*aroma 1 6188 0.03038
emulsifier Emulsifier*time 2 516782 0.00000
cacao=0.17
aroma 0.01
Aroma*time 2 713 0.54043
Error 18 1120

0 0.002
Response surface regression
aroma
0
The texture of the creams is measured on 3 different times.
The variable time can therefore be considered as a process
fat=0.32 sugar=0.61 variable, expanding the experimental region from 4 to 5
dimensions. To model the response of the combined
a b experiment and the time, a combination is made of the
interaction model in the three process variables emulsifier
(z1), aroma (z2) and time (z3):

Figure 2: Combined experimental design of a 22 factorial f(z) = α0 + α1z1 + α 2z2 + α 3z3 + α 12z1z2 + α 13z1z3 + α 23z2z3
design in emulsifier and aroma and a {3,1} simplex lattice + α 123z1z2z3
design in cacao, sugar and fat. Factorial design nested in
the mixture design (a). Mixture nested in factorial (b). and the linear mixture model in the mixture components
cacao (x1), sugar (x2) and fat (x3):
f(x) = β1x1 + β2x2 + β 3x3 (2)
Table 4: Regression statistics for the complete combined
model
This results in the following combined model:
Variable Parameter estimate Prob > T V.I.F.
f(x,z) = N000100x1 + N000010x2 + N000001x3 + Ca 536.69 0.0215 7.03
(N100100x1+ N100010x2 + N100001x3) z1 + Su 563.57 0.0168 7.03
(N010100x1 + N010010x2 + N010001x3) z2 + Fa 299.57 0.1661 7.03
(N001100x1 + N001010x2 + N001001x3) z3 + ca*em -34.56 0.2522 7.03
su*em -25.85 0.3858 7.03
(N110100x1 + N110010x2 + N110001x3) z1z2 +
fa*em -1.66 0.9549 7.03
(N101100x1 + N101010x2 + N101001x3) z1z3 + ca*ar -90.32 0.7821 7.03
(N011100x1 + N011010x2 + N011001x3) z2z3 + su*ar -123.43 0.7058 7.03
(N111100x1 + N111010x2 + N111001x3) z1z2z3 fa*ar 121.96 0.7091
(3) 7.03
The complete combined model in pseudocomponent ca*ti 4.97 0.0336 7.03
proportions was fitted to the data. This transformation su*ti 4.76 0.0404 7.03
reduces the negative effects of the multicollinearity fa*ti 4.63 0.0452 7.03
between mixture components on the variance properties of ca*em*ar 17.52 0.7047 7.03
the estimates. The following L-pseudocomponents su*em*ar 16.06 0.7282 7.03
transformations transform the weights of the mixture fa*em*ar -11.25 0.8075 7.03
components in pseudocomponent proportions: ca*em*ti -0.43 0.1700 7.03
su*em*ti -0.41 0.1832 7.03
cacao' − 238 fa*em*ti -0.38 0.2201
(4) 7.03
cacao =
100 ca*ar*ti 0.87 0.7930 7.03
su*ar*ti 0.07 0.9831 7.03
sugar' -1000 -0.20 (5)
0.9527 7.03
sugar = fa*ar*ti
100 ca*em*ar*ti -0.10 0.8267 7.03
fat' - 500 su*em*ar*ti -0.03 0.9515 7.03
fat = (6)
100
with: cacao, sugar and fat: pseudocomponent proportions Table 5: Regression statistics for the reduced model
cacao’, sugar’ and fat’: weight in kg Variable Parameter estimate Prob > T V.I.F.
Ca 362.68 0.0027 1.76
The complete combined model (Table 4) represents the data Su 414.89 0.0008 1.76
with an adjusted multiple correlation coefficient of 0.84.
Fa 320.84 0.0069 1.76
The variance inflation factors are the same for all variables
ca*ti 2.99 0.0125 1.76
and equal 7, sufficiently low for mixture experiments. In
su*ti 5.42 0.0006 2.76
the complete combined model only the parameter estimates
of the variables cacao, sugar, cacao*time, sugar*time and fa*ti 4.77 0.0021 2.76
fat*time are significantly different from zero. Replacing in su*em*ti -0.55 0.0029 2.0
this model cacao, sugar and fat by its pseudocomponent fa*em*ti -0.40 0.0256 2.0
proportion, emulsifier, and aroma by its weight in gram and
time by the time in hours, makes it possible to predict the Figure 3 represents for 2 emulsifier and 3 times (major
texture for each cream composition and each point of time horizontal and vertical axis respectively) the response
within the experimental region. surface for cacao, sugar and fat. The right horizontal axis
Using stepwise regression a reduced model with 8 terms is represents the sugar pseudocomponent proportion and the
selected (Table 5), representing the results with an adjusted left horizontal axis represents the cacao pseudocomponent
multiple correlation coefficient of 0.81. proportion. The pseudocomponent proportion of fat equals
The reduced model consists of the main effects cacao, sugar 1 minus the pseudocomponent proportion of cacao and
and fat, their interaction effects with time and the triple sugar.
interactions sugar*emulsifier*time and fat*emulsifier*time. The major changes in force 1 are caused by time.
The emulsifier interacts antagonistically with sugar and Increasing time, increases force 1. Increasing the amount
time and with fat and time. of emulsifier from 0 to 10 grams reduces force 1 and even
The reduced second degree model is used to represent the more as time increases. If no emulsifier is added to the
response surface over the experimental region. The cream, force 1 increases as the sugar proportion increases.
complete experimental region is 5 dimensional (6 factors). If 10 gram of emulsifier is present, the highest force 1 is
Since aroma has no significant effect on the texture of the reached at a low sugar and fat proportion and a high cacao
cream, this factor can be removed from the experimental proportion.
design, reducing the experimental region to 4 dimensions.
CONCLUSIONS
The combination of a {3,1} simplex lattice design in the
mixture components cacao, sugar and fat and a 22 factorial
design in the process variables emulsifier and aroma
resulted in 12 treatments (creams). The textural properties
(intrusion force) of the 12 creams were investigated.
Both the cacao, the sugar, the fat and the emulsifier content
of the cream effected its texture. A combined mixture
model with 24 terms was fitted to the results and reduced to
a model with 8 terms. This model was used to represent the
response surface over the experimental region. Time had a
strong positive effect, while emulsifier had a negative effect
on force 1. If the cream contained no emulsifier the highest
force 1 was found at a high sugar proportion. If the cream
contained emulsifier, a high cacao proportion resulted in a
high force 1.

REFERENCES
Belsley, D.A., Kuh, E. and Welsh, R.E., 1980. Regression
diagnostics: identifying influential data and sources of
collinearity. John Wiley and Sons, New York.
Cornell, J.A., 1990. Experiments with mixtures. Designs,
models and the analysis of mixture data. John Wiley
and sons, New York.
Crosier, R.B., 1984. Mixture experiments: geometry and
pseudocomponents. Technometrics 26, 3:209-216.
Gorman, J.W., 1970. Fitting equations to mixture data
with restraints on the compositions. J. Qual. Technol.
2:186-194.
Scheffé, H., 1958. Experiments with mixtures. J. Roy.
Stat. Soc. B 20:344-360.
 Time (hours)
Force 1 Force 1

1400 1400

168
850 850

1.0 1.0

0.5 0.5
300 Sugar 300 Sugar
1.0 1.0
0.5 0.5
Cacao 0.0 0.0 Cacao 0.0 0.0

Force 1 Force 1

1400 1400

24
850 850

1.0 1.0

0.5 0.5
300 Sugar 300 Sugar
1.0 1.0
0.5 0.5
Cacao 0.0 0.0 Cacao 0.0 0.0

Force 1 Force 1

1400 1400

1
850 850

1.0 1.0

0.5 0.5
300 Sugar 300 Sugar
1.0 1.0
0.5 0.5
Cacao 0.0 0.0 Cacao 0.0 0.0
0
0 10 Emulsifier (g)
Figure 3: Texture (force 1) as a function of time and cream composition
 PROCESS OPTIMIZATION STRATEGIES TO DIMINISH VARIABILITY IN THE QUALITY

OF DISCRET PACKAGED FOODS DURING THERMAL PROCESSING

Philippe Baucour, Kevin Cronin

Department of Process Engineering
University College Cork, Ireland
E-mail: [email protected]

K EYWORDS The aim of this paper is to estimate the dispersion in product

quality induced by a known variability in product thermal behav-
Temperature dispersion, Quality variability, Sterilization, Pro- ior. Variability in other contributing factors such as the quality
cess design, Monte Carlo analysis. reaction kinetics and the initial value of the quality attribute are
ignored. Knowledge of the dispersion or range in a quality pa-
rameter enables the probability that extreme levels will or will
A BSTRACT
not be exceeded to be assessed; this may be significant for cus-
tomer acceptance. The change in food quality with time can be
A distribution in food product thermal properties will produce
classified in terms of chemical changes, vitamin destruction, en-
a distribution in product temperature throughout a thermal pro-
zyme inactivation, color and taste change or desired structure.
cess. This in turn will cause a permanent dispersion in product
For this work a global quality parameter will be analyzed which
quality through the mechanism of the temperature sensitivity of
is a measure of overall quality estimation of the product.
product quality thermal degradation kinetics. To guarantee ac-
Once the dispersion in quality can be estimated, processing
ceptable food safety, the slowest heating product must receive
strategies to diminish this dispersion, whilst maintaining the
adequate thermal treatment, implying that the majority of the
mean quality level within an acceptable band, can be investi-
products are over-cooked. Strategies to ameliorate this problem
gated. In particular the selection of the optimum processing tem-
are suggested by determining the optimum processing tempera-
perature and processing time (collectively known as the process-
ture to minimize the final quality dispersion. As an example, the
ing schedule) to minimize quality variability is analyzed. This
batch sterilization of packaged foods is analyzed and described
work examines the well known problem of the unwanted over-
by a set of three partial differential equations for (i) the heat
processing of most of the product units in a batch heating pro-
transfer, (ii) the quality change and (iii) the microbial reduction.
cess to guarantee that the slowest heating unit receives adequate
In this case process optimization aims to reduce the final stan-
thermal treatment, from a new angle. It must be stated that rela-
dard deviation in quality (without affecting the final safety) by
tively simple models of product temperature, safety and quality
optimizing schedules for the retort temperature.
are employed because the focus is not the accurate determina-
tion of end points for these parameters but rather to identify their
I NTRODUCTION dependence on process settings as a basis for process design op-
timization. Furthermore while the sterilization of canned foods
For any food thermal process, it is not possible to obtain an iden- in a steam retort is selected to illustrate the concept, the intention
tical temperature/time profile and hence a uniform final product is to present a more general framework that could be applied to
temperature between items in the batch (Smout et al., 2001). A a variety of thermal batch processes.
distribution in product initial temperature, variability in prod-
uct thermal diffusivity together with random fluctuations in en- T HEORY
vironmental processing parameters all contribute to dispersion
in the product temperature, (Smout et al., 2000; Varga et al., T EMPERATURE / Q UALITY / S AFETY MODELS
2000). Variability in product temperature will in turn produce
variability in the levels of a number of product quality attributes, In general, theoretical prediction of the temperature at a given
(Lund, 1978; Akterian et al., 1998). It is important to note that point in a packaged food during a heating process is difficult.
the variability in temperature between items in the batch which Analytical solutions are only available for certain special cases.
develops during a thermal process (for instance a heating or cool- For a pure conductive heating model for for a finite cylinder (like
ing schedule) is generally reversible as subsequent conditioning a can) the following differential equation can be applied:
and controlled storage stages will promote temperature unifor-  2
∂2T

∂T ∂ T 1 ∂T
mity. However temperature dispersion combined with temper- =α + + (1)
ature sensitive irreversible quality changes will produce perma- ∂t ∂r2 r ∂r ∂z 2
nent variability in quality levels of the product. The model thus assumes no mass transfer (i.e. the package

1
container acts as an impermeable barrier) and heat flow is solely according to both Zwietering and Hasting (1997) and Skjolde-
in the radial and axial directions. Furthermore assuming the brand and Ohlsson (1993) first order reaction kinetics can be
can surface reaches the temperature (T∞ ) of the surrounding used to describe the dominant quality changes:
medium (steam/hot water) the following boundary conditions
will be appropriate: dQ
= −kQ (5)
dt
Ea
T (r = R, z, t)
= T∞ (t) with k = k∞ e
−R
g Tα (t) (6)
T r, z = + H2 , t
= T∞ (t) (2)
T r, z = − H2 ,t = T∞ (t) where k is the reaction diffusivity and can be sensitive to prod-
T is product temperature, r the coordinate in the radius direc- uct temperature, moisture content, quality property concentra-
tion, z the coordinate in the height direction. T (r, z, t) is the tion, pH, etc. For this work k is restricted to a dependence on
temperature at the position (r, z) at the instant t. Ti is initial temperature alone and it will generally have a relationship with
product temperature and α the thermal diffusivity. The dimen- temperature described by the Arrhenius formula
sions of the can are R (radius) and H (height). In a pure con- Ea is termed the activation energy of the particular quality re-
duction system as described by equation 1 the coldest point in action and is a measure of the temperature sensitivity of the reac-
the cylinder will be the centre of the can. This will be point tion. Rg is the gas constant and Tα (t) is absolute product tem-
of interest and henceforth product temperature is taken to be perature, measured in Kelvin. The term k∞ is a pre-exponential
synonymous with centre temperaure. Therefore the statement factor and quantifies the magnitude of the reaction rate indepen-
T (r = 0, z = 0, t) will be shorten in T (t) or Tα (t) when its dent of temperature.
related to α.
In equations 1 and 2 others parameters such as the product sur- M ODEL PARAMETERS
face area, product thickness and surface heat transfer coefficient
(assumed infinite in this case) are considered to be constant and For the simulation, all the products were taken to have an iden-
time invariant. Depending on the exact formulation of the prod- tical initial temperature Ti of 40◦ C and the processing tempera-
uct thermal model, the thermal diffusivity, α will represent the ture T∞ is investigated within a range from 50◦ C to 200◦ C. The
sensitivity to factors including the product thermal conductivity, physical, thermal and quality properties of the product are cho-
specific heat capacity and density. Thus any random variability sen to represent those of an average food product, (Hallstrom
in these parameters will produce dispersion in the magnitude of et al., 1988). The mean value for thermal diffusivity, µα is
the thermal diffusivity, α and hence in the temperature of the 1.6 × 10−7 m2 .s−1 and this variable is taken to be described by
products throughout processing. a Normal distribution, truncated at the ±3σ limits, with a coef-
If the mean, µα and standard deviation σα in thermal diffusiv- ficient of variation (CV ) of 10% (CV = σα /µα = 0.1). These
ity are known, appropriate estimates of the maximum and mini- values are in line with literature values quoted for food variabil-
mum magnitudes for the thermal diffusivity

can then be assem- ity with the coefficients of variation being in the region of 10%.
bled from αmax/min = µα ± nσα . The dimensions of the pack are a radius (R) of 37.5 mm and
Taking a value for n of 3 implies that approximately 99.7 % height (H) of 108 mm. In the safety model, Tref is 121.1◦ C,
of the items will have a thermal diffusivity that lies within these z is 10◦ C and Fmin is 3 minutes (i.e. 180 seconds) which
limits assuming the underlying distribution is Normal. correspond to the standard values for sterilization. The qual-
Given the dispersion in temperature response, the subsequent ity kinetic parameters are chosen representatively to be a k∞ of
dispersion in safety and quality can then be quantified. Note to 1.36 × 10−8 s−1 and the activation energy Ea is 86 KJ.Kmol−1 .
distinguish the temperature of a product with a certain value of These values should be considered as typical values of quality
thermal diffusivity from another the temperature will be note as parameters.
Tα (t).
Applying the D-z model of microbial destruction, the required
P ROCESS OPTIMIZATION
processing time, tp is related to the sterilization value (F ) by the
following equation. The aim is to diminish the range in product quality at the end
Z tp
of the sterilization process, without affecting the safety of the
Tα (t)−Tref
F = 10 z dt (3) process, by the optimum choice of magnitudes of the process-
0 ing parameters. The range in product quality at the end of the
heating stage arises from the dispersion in product temperature
For a mathematical purpose equation 3 could be written as a that in turn is as a result of the variability in the thermal diffu-
differential equation. sivity. From equation 5 it can be seen that the level of quality at
dF Tα (t)−Tref the end of the heating cycle for a given product is a function of
= 10 z (4) three quality kinetic parameters (Qi , k∞ and Ea ) and four ther-
dt
mal process parameters (Ti , T∞ , α and tp ). A particular quality
Equation 3 or 4 must be solved numerically for each value of reaction rate will fix the magnitudes of the quality parameters.
α in order to ensure that each item can reach a minimum in F Generally the initial product temperature Ti is set by an external
value (i.e. Fmin ). constraint, i.e. the processing / storage regime that the prod-
In general, the order of a quality reaction can vary depending uct has received prior to sterilization. Potentially it is possible
upon the particular quality parameter under study. Nonetheless for the process designer to have some control over the thermal

2
diffusivity α; by adjusting either the surface heat transfer coef- R ESULTS AND DISCUSSION
ficient to the product and/or the product package dimensions, H
and R. In many cases it is not possible to manipulate α and this
will be the situation here. This means that the thermal diffusivity
O PTIMIZATION WITH ONE PROCESSING TEMPERATURE
of each product will be considered fixed and will only vary from
one can to another following a normal distribution N (µα ; σα ).
Hence for this study only the processing temperature T∞ will
In order to validate the numerical model, its output has been
be considered to be adjustable within some acceptable range.
compared to the analytical solution of equation 1 for the case
Because a conduction heating thermal model is employed with
where the processing temperature is constant. A good agree-
negligible surface resistance, the surface of the food product will
ment between the analytical and the numerical solution can be
be at the temperature T∞ and this imposes an upper limit on the
achieved with a 10x10 grid on the can. As an example a treat-
possible range of processing temperature. Two separate cases
ment at 127◦ C for 84 min. will lead to a maximum difference in
will be analyzed.
temperature which is less than 1%.
• The first case is a single set-point schedule where T∞ must
remain at a single value throughout the process. Choos- Using a single set-point allows a reduction in the number of
ing a particular value for T∞ will fix the duration, of the calculations. Applying the D-z model of microbial destruction,
process tp to achieve the minimum F value. An analysis the required processing time (tp ), can be found by considering
will be performed on T∞ to determine the best set (T∞ ; tp ) the time needed for the slowest heating product (with the mini-
in terms of safety and final quality distribution. In this mal thermal diffusivity constant αmin ) to achieve the minimum
relatively straightforward case the criteria to minimized is necessary sterilization value, Fmin that is appropriate for the par-
the quality range ∆Q which is the absolute difference be- ticular food. This will ensure that all the items in the batch are
tween the quality calculated with the minimum and maxi- adequately sterilized. Thus the safety constraint could be rewrit-
mum thermal diffusivity i.e. ∆Q = Qαmin − Qαmax . ten as a function of αmin only. It will link the processing tem-
perature to the processing time and delimit 2 areas and as is well
• The second case is analyzing a schedule where the process- known delimit two areas and an optimal path in the subspace
ing temperature can take two distinct values (T∞1 and T∞2 ) (T∞ ; tp ).
over the procssing time tp . In theory this permits a process-
ing schedule to be developed that allows both a minimum For a single processing temperature two sets of (T∞ ; tp ) were
sterilization effect (set by food safety standards) and pro- selected on the curve based the final quality predicted by equa-
cessing time (set by factory throuput requirements) to be tion 5. Figure 1 presents the relationships between the final qual-
independently satisfied. A processing temperature T∞1 will ity and the processing temperature with 3 different values of dif-
be applied for a duration t1 and a second processing tem- fusivities (αmin , µα , αmax ).
perature T∞2 applied for the remainder of the time. The
piecewise function T∞ (t) will be: 1.0

T∞1 if 0 ≤ t ≤ t1 0.9
T∞ (t) = (7)
T∞2 if t1 < t ≤ tp 0.8

0.7
Calculated with αmin
Final Quality at the centre

So in this case a set of (T∞1 ; T∞2 ; t1 ) must be determined. 0.6

0.5
The mean quality µQ or the standard deviation σQ in quality will
be optimized i.e. a minimization of σQ or (1 − µQ ) with a set of 0.4 Point A
Calculated with µα
constraints depending on the case. To take into account the dis- 0.3
Point B
persion in the thermal diffusivity, the optimization is performed 0.2
with 30 values of α picked within the normal distribution. When 0.1 Calculated with αmax
a optimum is found, an estimation of the final dispersion in tem-
0.0
perature, quality and F values will be performed by using 500 95 105 115 125 135 145 155 165 175 185 195 205

values of α. Table 1 lists the optimizations runs with the con- Retort temperature (°C)
straints in each case. Figure 1: Final quality at the center of the can vs the processing temperature
Table 1: Optimization runs with the associated constraints for 1 or 2 pro- for 3 different diffusivities αmin , µα , αmax . Point A - Case N◦ 1. Point B
cessing stages. Each run is designated by a number N◦ . - Case N◦ 2.
Stages Criteria to minimized Sampling N◦
1∗ Quality range ∆Q 3 1,2
0
2∗∗ Mean quality 1 − µQ 30 / 500 3a, 3b To illustrate the analysis two processing temperatures (T∞
00
Standard deviation σQ 30 / 500 4a, 4b and T∞ ) with the two corresponding process durations (t0p and
∗ For one processing temperature the optimization is not necessary - a sim- t00p ) were selected in the investigated range. In each case, the re-
ple analysis of the final quality range will provide the best schedule time - quired Fmin is achieved at the end. The first arrangement (Case
temperature. N◦ 1) produces a maximum in quality for the mean diffusivity at
∗∗ For 2 processing stages indices a and b correspond to the total processing the end of the process. The second schedule (Case N◦ 2) gives a
time: a for tp =84 min and b for tp =115 min. local minimum in the quality range ∆Q.

3
 Table 2: Optimization results presented in table 1. For cases 3 and 4 the process compared with the single point schedule (Table 2).
temperature and quality profiles are shown on figure 2 and 3.
Case N◦ tp T∞1 T∞2 t1 µQ σQ 200
180
(min.) (◦ C) (◦ C) (min.) - -

Temperature (°C)
160
1 84 127 - - 0.328 0.068 140
2 115 116 - - 0.302 0.040 120
100
3a 84 118 140 46 0.346 0.078 80
3b 115 105 167 87 0.342 0.0714 60
40
4a 84 127 150 82 0.326 0.0677 20
4b 115 109 120 55 0.326 0.044 0
0.0 8.4 16.8 25.2 33.6 42.0 50.4 58.8 67.2 75.6 84.0

Table 2 presents the final quality distribution for the two pos- 1.0
sible schedules (Cases N◦ 1 and N◦ 2). The final mean quality is 0.9
0.8
almost the same between the two schedules; 0.328 for a process- 0.7
0.6
ing temperature of 127◦ C against 0.302 for 116◦ C. The higher

Quality
0.5
0.4
temperature gives the better mean quality as is in accord with the 0.3
0.2
High Temperature - Short Time processing philosophy. By con- 0.1
trast the lower temperature gives a much tighter distribution in 0.0
0.0 8.4 16.8 25.2 33.6 42.0 50.4 58.8 67.2 75.6 84.0
quality; a range in quality of 26% for 116◦ C compared to 42% Time (min.)

for 127◦ C. In other words a small reduction in mean product Figure 2: Temperature and Quality profiles determined after with two
quality is more than compensated for by much improved con- processing temperature (- - Case N◦ 4a; · · · · · · Case N◦ 3a). For each
trol over dispersion in quality. The above analysis indicates that schedule only the extreme values of the thermal diffusvity are represented
batch thermal processes may be more profitably optimised by (αmin , αmax ). The total processing time is 84 minutes.
considering the effect of processing conditions (i.e. temperature)
200
on both mean and dispersion in quality rather than (as has been
160
Temperature (°C)

the case) considering the effect on mean quality only. So the in-
fluence of a variable product thermal diffusivity can lead to the 120

adoption of slower schedule than the one generally selected via 80

the standard sterilization rule High Temperature - Short Time. 40

Ultimately the best schedule will depend on the costs incurred 0

0.0 11.5 23.0 34.5 46.0 57.5 69.0 80.5 92.0 103.5 115.0
by the food manufacturer associated with a trade off between the
mean and range in product quality. 1.0

0.8
Quality

0.6
O PTIMIZATION WITH TWO PROCESSING TEMPERATURES
0.4
By selecting a schedule consisting of two sequential processing 0.2
temperature and an intermediate time (T∞1 ; T∞2 ; t1 ), a reduc- 0.0
tion in the dispersion in Q at the end of the process compared to 0.0 11.5 23.0 34.5 46.0 57.5 69.0 80.5 92.0 103.5 115.0
a single set point temperature may be possible to achieve (obvi- Time (min.)

ously whilst maintaining with the same safety level). From prac- Figure 3: Temperature and Quality profiles determined after with two
tical considerations each processing temperature will be chosen processing temperature (- - Case N◦ 4b; · · · · · · Case N◦ 3b). For each
to lie within a range 50-200◦ C. In order to compare the two ap- schedule only the extreme values of the thermal diffusvity are represented
proaches (a schedule with one versus that with two temperatures) (αmin , αmax ). The total processing time is 115 minutes.
the total processing time, tp will be given a a value equal to 84
or 115 minutes. 36
84 min.
− Case 1
Also to facilitate comparison, another constraint will be im- 32
One single processing temperature
posed on the final mean quality (µQ ) at the centre of the can; i.e. 28
... Case 3a optimization on µQ
− − Case 4a optimization on σ
it must not be less than the quality that can be achieved with only Q

24
one processing temperature (i.e. µQ ≥ 0.328). The optimisation
software found that two combinations of T∞1 and T∞2 gave lo- 20
Count

cal maxima for quality parameters for each of the two process 16
times. For all four schedules the pattern is that a low processing
12
temperature followed by a higher temperature is best for quality
(mean and dispersion) rather than the reverse. The best process- 8
ing conditions, as predicted by the optimisation software, and 4
the resultant mean and standard deviation in final product qual-
0
ity are given in table 2. The data is presented for two separate 0.1 0.2 0.3 0.4 0.5 0.6 0.7
processes, having a total time of 84 min and and 115 min respec- Final mean quality
tively and thus allows comparison with the corresponding single Figure 4: Optimization results. Compared distributions between cases 3a,
set-point schedules. Figures 4 and 5 compare the distribution in 4a and the one determined with only one processing temperature. The total
product quality given by the optimal schedules for the two step processing time is 84 minutes.

4
 50 slower schedule (T∞ =116◦ C, tp =115 min.) will give an ac-
115 min. − Case 2
One single processing temperature ceptable final mean quality and a better standard deviation.
40 ... Case 3b optimization on µQ
−− Case 4b optimization on σQ 2. The 2nd step consists in an optimization of a 2 stages
schedule with a 3 different values of the thermal diffusiv-
30 ity (αmin ; µα ; αmax ). The results show that none of the
Count

schedule selected are suitable for a variable thermal diffu-

20
sivity.
3. Finally, an optimization which takes into account the vari-
10 ability has been performed. It has been shown that the gain
in mean quality is almost negligible from an industrial point
0
of view but a significative reduction of final dispersion is
0.1 0.2 0.3 0.4 0.5 0.6 0.7 possible by adopting a 2 stages schedules. Once again a
Final mean quality slower process provides the best result.
Figure 5: Optimization results. Compared distributions between cases 3b,
4b and the one determined with only one processing temperature. The total This framework could be implemented for any thermal process
processing time is 115 minutes. which involves a quality and / or a safety model. The main dis-
advantage consists in the use of a Monte Carlo analysis which is
For a total processing time of 84 minutes there is a small im- relatively easy to run and provides a good understanding of the
provement in the final mean quality (0.346 against 0.328) but statistical changes during the process. Unfortunately the Monte
a worse standard deviation (0.078 against 0.068) when compar- Carlo is computing time consuming and will not be suitable to
ing the two step schedule of 118◦ C / 140◦ C against the original extend the present study by crossing different sources of vari-
single point schedule. The second two step schedule of 127◦ C ability like the position, the processing temperature or the heat
/ 150◦ C is very close to that of the single point schedule and transfer coefficient.
matches the standard deviation but with a slightly worse mean
value in quality (0.326 compared to 0.328). For the longer pro- R EFERENCES
cessing time of 115 minutes, both two step schedules give an
improvement in mean quality when compared to the single set Akterian, S., Smout, C., Hendrickx, M., Tobback, P., 1998.
point schedule though at the expense of a slightly poorer stan- Application of sensitivity functions for analysing the impact
dard deviation. of temperature non-uniformity in batch sterilizers. Journal of
Overall then, these results indicate that a processing schedule Food Engineering 37 (1), 1–10.
consisting of two set points can only give a marginal improve-
ment in final quality as compared to that achievable with a single Hallstrom, B., Skjoldebrand, S., Tragardh, C., 1988. Heat trans-
set point. Choosing from the four two point schedules the best fer and food products. Elsevier Applied Science, London.
compromise between mean and standard deviation in quality is Lund, D., 1978. Statistical analysis of thermal process calcula-
given by the 109◦ C for 55 minutes followed by 120◦ C for the tions. Food Technology 3, 76–83.
remainder of the 115 minutes. The mean and standard deviation
in quality (0.326 and 0.044) that would result from the adoption Skjoldebrand, C., Ohlsson, T., 1993. A computer simulation pro-
of this schedule are probably marginally preferable to the quality gram for evaluation of the continous heat treatment of partic-
parameters that are available from the single set point schedules. ulate food products. Part 1: design. Journal of Food Engineer-
ing 20, 149–165.

C ONCLUSIONS Smout, C., van Loey, A., Hendrickx, M., 2000. Non-uniformity
of lethality in retort processes based on heat distribution and
A combination of tools to predict the temperature, quality and heat penetration data. Journal of Food Engineering 45 (2),
safety during a thermal processing have been developped. These 103–110.
tools have been combined with a Monte Carlo analysis to take Smout, C., van Loey, A., Hendrickx, M., 2001. Role of temper-
into account the variability of the thermal diffusivity and an op- ature distribution studies in the evaluation and identification
timization routine to determine an acceptable processing sched- of processing conditions for static and rotary water cascading
ule. As an example the sterilization process has been selected retorts. Journal of Food Engineering 48 (1), 61–68.
because it involves thermal processing combined with quality
and safety issues. The schedules tested consist in constant pro- Varga, S., Oliveira, J., Smout, C., Hendrickx, M., 2000. Mod-
cessing temperatures for different period of time (1 or 2 stages). elling temperature variability in batch retorts and its impact
The optimization has been done in 3 steps: on lethality distribution. Journal of Food Engineering 44 (3),
163–174.
1. For a single step process the optimization has been done Zwietering, M., Hasting, A., 1997. Modelling the hygienic pro-
via an analysis of the final quality range determined with cessing of foods - a global process overview. Food and Bio-
the extremal possible values of the thermal diffusivities products Processing 75, 159–167.
(αmin ; αmax ). for which it has been shown that adopting a

5
FLUIDS
 IMPINGEMENT COOLING OF SOLID CYLINDRICAL FOODS USING CFD

Eva E. M. Olsson Lilia M. Ahrné Christian Trägårdh

Environment and Process Environment and Process Food Engineering, Centre for
Engineering Engineering Chemistry and Chemical Engineering
SIK SIK Lund University
P.O. Box 5401 P.O. Box 5401 P.O. Box 124
SE-402 29 Göteborg SE-402 29 Göteborg SE-221 00 Lund
Sweden Sweden Sweden
[email protected] [email protected] [email protected]

KEYWORDS product to temperatures where growth of micro-organisms

will not occur. Rapid chilling reduces the risk of
Impingement cooling, impinging air jet, heat transfer, microbiological growth, increases the product quality and
Nusselt number, CFD, cylindrical food. the production capacity.

ABSTRACT Impinging jets of air are widely used as an effective method

to increase heat and mass transfer in widespread areas, such
Simulations of impingement cooling with a slot air jet in a as cooling of electronics, drying of paper, cooling systems
semi-conefinement area were performed using CFD in gas turbines, anti-ice systems on aircrafts etc. In the food
(computational fluid dynamics). The surface heat transfer industry impingement jet systems is used for heating
coefficient and the Nusselt number distribution on the (Marcroft and Karwé 1999), drying (Moreira 2001) and
cylindrical food product and on the underlying surface were freezing (Soto and Bórques 2001). The high turbulence
studied. The dependency of Nusselt number on Reynolds characteristics enhance the heat transfer and therefore
number was determined under steady state condition. The improve food quality.
Nusselt number increases with Reynolds number, which
was varied between 23000 and 100000. The characteristics Numerous studies, both experimental and numerical, of
of the velocity distribution of the jet impinging on the impingement heat transfer and flow characteristics have
cylinder in the computational domain were also determined. been reported in the literature. Most of the studies are
concerned with the effect of Reynolds number, nozzle-to-
NOMENCLATURE surface distance, nozzle geometry, jet orientation, multiple
jets, cross flow and impinging surface shape. Although
C coefficient of correlation Nu=CRem many different jet characteristics have been considered a
D diameter of cylinder and width of jet (m) great majority of the impinging jet studies are on the flat
Fr Frössling number (Fr=Nu/Re0.5) plate (Baughn and Shimizu 1989; Yan et al. 1992).
h heat transfer coefficient (W/(m2K)) However, some papers are concerned with impinging jets
H/D nozzle-to-cylinder distance on objects such as a curved surface (Lee et al. 1999), a
k thermal conductivity (W/(mK)) cylinder (McDaniel and Webb 2000) and a cylindrical
m exponent of Re in Nu=CRem pedestal (Baughn et al. 1993).
Nu local Nusselt number (Nu=h*D/k)
Nus Nusselt number in the stagnation point Several reviews and summaries on impinging jet heat
Re Reynolds number (Re=vD/ν) transfer can be found in the literature (Downs and James
SST shear stress transport 1987; Jambunathan et al. 1992). Impingement in food
v velocity (m/s) processing has also been reviewed (Ovadia and Walker
x/D distance from the jet stagnation point 1998).

Greek symbols Downs and James (1987) reviewed heat transfer

θ angle on the cylinder characterisitics of impinging jets with round and slot
ν kinematic viscosity (m2/s) nozzles. Geometric and temperature effects, inteference and
crossflow, turbulence levels, incidence, surface curvature,
INTRODUCTION externally applied factors, jet outlet conditions and non-
uniformity of jet arrays were summarised. Jambunathan et
Cooling of pre-cooked solid foods is often the rate al. (1992) reviewed experimental data for heat transfer rates
determining step in the process and therefore often a bottle on circular jet impinging on a flat plate. Jets of Reynolds
neck in food production. With conventional chilling number in the range of 5000-124000 and nozzle-to-plate
methods long chilling times are necessary to cool the surface distance in the range of 1.2-16 were considered.
 250
Wall coarse grid
D Opening 200 fine grid
Inlet
H
150

Nu
Symmetry axis

100

50

0
Cylindrical 0 45 90 135 180
food Wall
θ
x Figure 2. Grid independency. The coarse grid has
approx. 250000 elements and the finer grid
Figure 1. The computational domain. approx. 800000 elements, Re=50000.

In order to understand the heat transfer characteristics of 40 m/s using the nozzle exit width as characteristic length.
impingement cooling of solid foods, the objective of this The initial temperature of the surface of the cylindrical food
study is to determine the Nusselt number on the surface of is 35°C. Using the ambient temperature (the jet exit
the cylindrical food product and on the underlying plate for temperature) and the surface temperature of the cylinder,
varying Reynolds numbers (23000-100000). The goal is the heat transfer coefficient and corresponding Nusselt
also to determine the dependency of Nusselt number on number based on the jet width was then calculated.
Reynolds number, and to study the characteristics of the
airflow from the slot air jet. The jet is impinging on a solid The turbulence model used was the Shear Stress Transport
cylinder in two dimensions in a semi-confinement area. model (SST) found in CFX 5.5. The SST model blends
between a k-ω model near the wall and a k-ε model far
MATERIAL AND METHOD away from the wall. The model uses a low Reynolds
number model if the grid near a wall is very fine (y+<11)
Steady state simulations of a slot air jet impinging on a and a wall function beyond that (AEA Technology).
solid cylindrical food were performed using the CFD-
software CFX 5.5 (AEA Technology). RESULT AND DISCUSSION

The cylindrical food has a diameter of 35 mm and the slot Grid independency
air jet has a width of 35 mm. The depth is infinite. There is
a symmetric boundary condition in the centre of the slot jet The computational domain, see figure 1, uses a tetrahedral
in the computational domain, see figure 1. The length of the mesh of approx. 250000 elements. In the boundary layers
domain is 30 cm, the height is 27.5 cm and the length of the an inflated boundary is used with cumulative prisms from
nozzle is 5 cm. The food is “floating” 5 mm above the the cylinder, y+=0.03-9.7. To make sure that the
underlying surface due to computational difficulties. The simulations are grid size independent a simulation was
nozzle-to-target distance (H/D) is 4. made with a finer mesh consisting of nearly 800000
elements. The finer grid had a similar shape of the
The air inflow has a temperature of 2°C and the operating distribution of local Nusselt number as the coarse grid, but
pressure is 1 atm. The jet Reynolds number is varied especially at θ=135º the finer grid predicted a more
between 23000 and 100000, i.e. a velocity of approx. 10 - pronounced secondary maximum, see figure 2. This could

0.2

0.15

0.1

0.05

−0.2 −0.1 0 0.1 0.2 0.3

0 2 4 6 8 10 12 14 16 18

Figure 3. A contour plot of the velocity of a jet with Reynolds number of 50000
(reflected in the centre).
 0.02 300
0.015
250
0.01
200
0.005 θ

Nus
150
0

−0.005
100
−0.01 50
−0.015 0
−0.02 0 20000 40000 60000 80000 10000
−0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0
Re
Figure 4. Flow pattern of the velocity near the vortex
on the side of the impinging jet. Note: The vectors are Figure 5. Nusselt number in the stagnation point
resampled, the computational grid is not used. as a function of Reynolds number. The fitted
curve corresponds to Nu=0.61·Re0.53.

be due to the difficulty in resolving the vortex behind the described in McDaniel and Webb (2000), along with the
cylinder properly with a coarse grid. Another reason might influence nozzle width has on the heat transfer rates.
be the location of the inner nodes at different dimensionless
distances around the circumference of the cylinder. Stagnation point heat transfer

Flow characteristics The Nusselt number in the stagnation point (Nus) depends
on the Reynolds number, see figure 5. For a fixed H/D
Figure 3 shows a contour plot of the velocity variation in distance the Nusselt number increases with Reynolds
the computational domain. The high velocity of the jet number. This result agrees fairly well with results of other
entrains air of lower velocity from the cavity and an initial authors on a flat plate (Baughn and Shimizu 1989; Yan et
mixing zone is created. Figure 4 shows the velocity vectors al. 1999). Baughn and Shimizu (1989) used a round
of the jet impinging on the cylindrical food. Flow impinging jet with a Reynolds number of 23750 and
separation occurs at about 90º on the cylinder and a vortex various H/D (2, 6, 10 and 14). Yan et al. (1999) used also a
is created at the back (θ=135º) of the cylinder. Just beyond round impinging jet, a Reynolds number of 23000, 50000
the impingement region, the flow structure is significantly and 70000 and an H/D distance of 2, 6 and 10. The Nusselt
characterised by the strong curvature of the streamlines, see number is slightly lower (approx 10%) than on the flat
figure 6. plate. The result of Lee et al. (1999) indicates however that
the Nus should be higher for jet impingement on a curved
The impinging jet consists of two types of flow, one surface. Predicted values are strongly coupled to the
orthogonally that impinge on the surface and a parallel wall features of the turbulence model. Each of the numerous
jet. At the top of the cylinder there is a stagnation point, turbulence models given in the literature has its own
where the jet impinge on the cylinder (θ=0º). The heat strengths and weakness both regarding prediction of the
transfer rate is high in the stagnation point, see figure 5 and flow characteristics as well as for heat transfer.
7. This is naturally a result of the flow predictions as well
as how the analogy between momentum and heat transfer is A linear regression using the logarithms of Nus and Re is
formulated. Optimum nozzle-to-surface distance (H/D) used to obtain the constant C and the exponent m in the
coincides with the length of the potential core (Downs and relation Nus=C·Rem for H/D=4. The result is, C=0.61 and
James 1987). The potential core region for slot jets is m=0.53, i.e. Nu=0.61Re0.53. The Frössling number
H/D<4-6 (McDaniel and Webb 2000). The importance of (Fr=Nu/√Re) in the stagnation point is 0.85-0.9.
the correct distance of nozzle-to-surface target (H/D) is

0.2

0.15

0.1

0.05

−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25

Figure 6. Streamlines of the velocity field, Re=50000 (reflected in the centre).

 350

300

250
Re=23000
200 Re=50000
Nu

150 Re=70000
Re=100000
100

50

0
0 45 90 135 180
θ
Figure 7. Nusselt number variation on the cylinder, θ is the angle on the surface of the
cylinder, θ=0 on top of the cylinder, for varying Reynolds number (23000-100000).

350

300

250
Re=23000
200 Re=50000
Nu

150 Re=70000
Re=100000
100

50

0
0 2 4 6 8
x/D
Figure 8. Nusselt number variation on the underlying surface, for varying Reynolds
number (23000-100000).

Nusselt number distribution on the cylinder The local minimum at θ=90º corresponds to the separation
of flow at the edge of the cylinder and the local maximum
The Nusselt number varies on the surface of the cylinder, at θ=135º corresponds to the creation of the wake behind
see figure 7. For Reynolds number of 23000, 50000 and the cylinder.
70000 the structure of the curve is identical, but different
values of Nusselt number. The highest heat transfer is Nusselt number distribution on the underlying surface
observed at the top of the cylinder. The Nusselt number on
the upper half of the cylinder is slowly decreasing until Figure 8 shows the distribution of the Nusselt number on
approx. θ≈ 50-55º, where it decreases abruptly to a local the underlying surface (0<x/D<8.5). The Reynolds number
minimum at θ=90º. The Nusselt number increases then to a is varied between 23000 and 100000.
local maximum at 135º and then down to zero at the back of
the cylinder (θ=180º). The Nusselt number is almost zero under the cylinder and
increases to a peak value at x/D=0.5-1. This is in the same
For Reynolds number of 100000, the structure of the curve area where the maximum is on the cylinder. For x/D>1, the
is similar to cases having lower Reynolds number, with Nusselt number decreases slowly. In comparison with
exception of another local maximum at approx. 45º. A results of Lee et al. (2000), the curve after x/D>1 agrees
drastic reduction in boundary layer thickness at about this fairly well, but a round impinging jet on a flat surface is
position, as the flow velocity increases, would be expected used as mention before. The result of Yan et al. (1992) is
suggesting that this local maximum at about θ=45º may also comparable after about r/D=2 where there is a
have some physical explanation. maximum.
CONCLUSIONS Baughn J. and Shimizu, S. 1989 “Heat transfer measurements
from a surface with uniform heat flux and an impinging jet.”
The flow pattern from the jet impinging on the cylindrical Journal of Heat Transfer, No. 111, 1096-1098
food product, the Nusselt number distribution on the Downs, S.J. and James, E.H. 1987. “Jet impingement heat transfer
- a literature survey.” The American Society of Mechanical
surface of the cylinder, the Nusselt number variation on the Engineers, 87-HT-35, 1-11
underlying surface and the Nusselt number dependency on Jambunathan, K.; Lai, E.; Moss, M.A.; and Button, B.L. 1992. “A
the Reynolds number was studied using CFD simulations. review of heat transfer data for single circular jet
impingement.” International Journal of Heat and Fluid Flow,
When the jet impinges on the cylinder, a stagnation point is No. 13(2), 106-115
created and the jet becomes a wall jet. A vortex is created Lee, D.H.; Chung, Y.S.; and Kim, M.G. 1999. “Technical Note
on the side/below the cylinder. Turbulent heat transfer from a convex hemispherical surface
to a round impinging jet.” International Journal of Heat and
The Nusselt number on the surface of the cylinder was Mass Transfer, No. 42, 1147-1156
Marcroft, H.E. and Karwé, M.V. 1999 “Flow field in a hot air jet
estimated. The Nusselt number is highest in the stagnation impingement oven - part I: A single impinging jet.” Journal
point, indicating that the heat transfer rates are largest in the of Food Processing Preservation, No. 23, 217-233
stagnation point. At the upper part of the cylinder (- McDaniel, C.S. and Webb, B.W. 2000 “Slot jet impinging heat
50º<θ<50º) the Nusselt number is approximately the same transfer from circular cylinders.” International Journal of
as in the stagnation point. On the side of the cylinder the Heat and Mass Transfer, No. 43, 1975-1985
Nusselt number drops rapidly to about one third of the Moreira, R.G. 2001 “Impingement drying of foods using hot air
Nusselt number in the stagnation point. The heat transfer and superheated steam.” Journal of Food Engineering, No.
increase slightly to a maximum at θ≈ 135º, it decreases then 49, 291-295
to zero on the back of the cylinder. Ovadia, D.Z. and Walker, C.E. 1998. “Impingement in food
processing.” Food Technology, No. 52(4), 46-50
Soto, V. and Bórquez, R. 2001 “Impingement jet freezing of
The distribution of Nusselt number on the underlying biomaterials.” Food Control, No. 12, 515-522
surface is almost zero below the cylinder, increasing Yan, X.; Baughn, J.W.; and Mesbah, M. 1992 “The effect of
abruptly to a peak value at x/D=0.5-1 and then slowly Reynolds number on the heat transfer distribution from a flat
decreases. plate to an impinging jet.” Fundamental and Applied Heat
Transfer Research for Gas Turbine Engines ASME, No. 226,
The stagnation point heat transfer was determined for 1-7
various Reynolds numbers (23000, 50000, 70000 and
100000). The Nusselt number increases with Reynolds AUTHOR BIOGRAPHY
number as a function of Rem. A linear regression gives the
coefficients C=0.61 and m=0.53 (in Nu=CRem) and the EVA OLSSON is an industrial Ph.D.-student at the
Frössling number is 0.85-0.9. It can be concluded that the Division of Food Engineering, Centre for Chemistry and
heat transfer from the impinging jet is strongly dependent Chemical Engineering, at Lund University, Sweden,
on Reynolds number, i.e. air velocity since (Re=vD/ν). working at the Dept. of Environmental and Process
Engineering at SIK (The Swedish Institute for Food and
FURTHER WORK Biotechnology). Her undergraduate education is from
Chalmers University of Technology, Gothenburg, Sweden,
Further work on cooling of solid foods with impinging jets where she finished her M.Sc. in Chemical Engineering with
is planned. This paper is just the beginning of a Ph.D. Engineering Physics in the end of year 2000.
project and it concludes the first steps of determining the
heat transfer rates from impinging jets on solid foods. LILIA AHRNÉ is research manager for Process
Engineering at SIK. She graduated in Food Engineering in
In future work a more thorough investigation on heat 1989 at Catholic University of Portugal, School of
transfer of impinging jets is planned, along with chilling Biotechnology, and obtained her Ph.D. in Biotechnology,
times necessary to cool the food product to a certain Food Science and Engineering in 1994. Her main areas of
temperature and the influence on texture of foods. research are optimisation of food processes, considering
Experiments for validation of the simulations are going to heat and mass transfer and kinetics of quality changes
be performed in an impingement unit in a recent future. during processing and storage.

ACKNOWLEDGEMENTS CHRISTIAN TRÄGÅRDH is a professor and the head of

the Division of Food Engineering, Lund Institute of
This work has been financed by the Swedish Knowledge Technology at Lund University, Sweden. In research, he is
Foundation and Lecora AB in Vadstena, Sweden. heading the Food Fluids Engineering Group. The core
activities are experiments and modelling of dispersion,
REFERENCES aggregation, coalescence and adhesion of molecules,
particles, drops and bubbles and heat and mass transfer
AEA Technology “CFX 5.5 User manual.” related to fluid flow. He is also theme manager in the
Baughn, J.; Mesbah, M.; and Yan, X. 1993 “Measurements of Future Technologies for Food Production programme in
local heat transfer for an impinging jet on a cylindrical Sweden, assisting prefect at the Centre for Chemistry and
pedestal.” Turbulent Enhanced Heat Transfer ASME, No. Chemical Engineering and educational leader for the
239, 57-62 Biotechnology programme at LTH.
 DETERMINING SHEAR DAMAGE OF MILK IN PROCESS EQUIPMENT
Jesper Sæderup Nielsen
Bo B. B. Jensen
Alan Friis
Food Process Engineering, BioCentrum-DTU
Building 221, DK-2800 Lyngby
Denmark
E-mail: [email protected]

KEYWORDS when exposed to flow fields with a certain shear rate at a

Computational Fluid Dynamics, Milk Fat Globules, Process given temperature. The destabilisation is a consequence of
Modelling, Process Product interaction fat globule deformation by shearing, and depends on fat
globule size, conditions of the fat globule membrane and
ABSTRACT entrapped fat. Disruption of the fat globules may lead to
agglomeration. In milk exposed to a harsh mechanical
Shear rates may result in damage of milk fat globules treatment i.e. high shear rates; agglomeration of fat
depending on the actual process conditions. Using globules could become visible to consumers as insoluble
Computational Fluid Dynamics (CFD) to predict shear rates lumps of fat in the top of the milk container. This is a
in process equipment, it is possible to set-up a food process problem frequently occurring in e.g. organic whole milk, as
model to simulate the fat globule damage in milk during this milk is not homogenised in Denmark.
processing. In extreme situations fat globule damage may In the dairy industry such knowledge has restricted milk
become visible for consumers as insoluble fat lumps. Time, velocities through pipes. These mean velocity
temperature and fat content are the three main factors recommendations have been applied to all pipe diameters
influencing resistant towards shear damage of the fat disregarding that the maximum shear rate decreases with
globules. Result shows that the rate of fat globule damaged increasing pipe diameter at a given velocity. It has now
was increasing during shearing, and that they are most been realised that this approach has its limitations, and the
fragile in the temperature interval between 15 and 35°C. dairy industry has shown great interest in increasing the
Cream containing 38 % fat was damaged at a lower shear capacity of production lines by increasing the velocity.
rate compared to creams with lower fat content. Combining Hinrichs (1998) presented a model for milk and cream
CFD simulations of flow fields in production equipments velocities in pipes, where fat globule damage is avoided.
with results of flow describing parameters effect on food However, processing of milk is performed in other
quality shows to be a powerful tool, which will extent our equipment than pipes e.g. bends, valves, pumps etc.
knowledge of producing good quality foods. Meaning that models based on fat globule damage in
straight pipes are likely to over estimate the allowable
INTRODUCTION velocities as shear rates are higher in complex equipments.
This work aims at developing a model for predicting fat
The quality of food products is determined by the raw globule damage in milk during transport through complex
material quality and the impact of the production processes equipment at different processing conditions. The model is
involved. Previously the applied quality descriptors have based on data from measurements of fat globule damage in
largely been based on bulk parameters and thus lacking milk exposed to a known shear rates at different conditions.
specific information on the influence of local occurrence of Computing shear fields in production equipment using
extreme process conditions in processing plants. Along Computational Fluid Dynamics (CFD) makes it possible to
with the use of continuous processing and design of still predict the total effect from the shear rates in production
larger production facilities where pipe dimensions are on equipment on the milk fat globules. The process model is
the increase the residence time distributions broadens and validated by experiments and is constructed to be
likewise the distributions of derived quality parameters. applicable for scale-up to other production equipments
Hence investigations of local processing conditions and the where the shear rates are known only from CFD analysis.
characteristics of smaller fractions of a bulk can be of great This article presents preliminary results, and illustrates the
importance when determining the quality of a food product. approach.
This paper introduces an approach where process
parameters influencing the quality of milk products have MATERIAL AND METHODS
been studied under controlled conditions in the laboratory.
The distribution in physical characteristics is related to Raw Materials
process models describing the flow of liquids through
pipes. It is shown that the applied model can be used to Pasteurised milk (3,5 % fat) and cream (38 % fat) bought
predict how the process affects the quality distribution of locally was kept refrigerated until use. Required fat content
the final product. in samples were achieved by gently mixing milk and cream.
During processing, milk is exposed to shear rates of
different magnitudes. Fat globules in milk are destabilised
Shear Treatment were applied a sample with a controlled temperature. The
apparent viscosity was recorded during the shear treatment,
The sample temperature was allowed ample time to resulting in a viscosity increase (Figure 1) as fat globules
equilibrate to the specified temperature. Samples were agglomerates in the sample.
sheared in a Haake VT 500 viscometer (Thermo Haake,
Germany) using the NV double cap cup and rotor. Shear treatment
Temperatures were controlled during shearing by a Haake 25
DC30-K10 water bath circulating water to the shear cell.

Viscosity [mPa s]
After treatment, samples were cooled to 5°C and kept for 20
fat globule damage measurements performed the same day.
15
Fat Globule Damage Measurements 10
Fat globule damage was measured by changes in fat globule 5
size distribution measured with a Mastersizer 2000 fitted
with a Hydro 2000 SM sample handling unit (Malvern 0
Instruments Ltd., UK). Mean and 90 % fractal fat globule 0 50 100 150 200
size was calculated with the Mastersizer 2000 software Shear time [s]
(version 4.00) using the general purpose model and a
refractive index of 1,47 for milk fat. Measurements were Figure 1: The increase in apparent viscosity of Cream (38
performed at room temperature. % fat) during Shearing at 20°C with a Shear Rate of 500 s-1.
An increase in viscosity during shearing also indicates fat Enclosed drawing illustrates the ongoing Fat Globule
globule damage. Agglomeration occurring during Shearing

CFD Simulation of Shear Fields The viscosity incline during shearing seem to have three
stages. During the first 10 seconds the viscosity increases
Geometries for process equipment were implemented into rapidly, followed by a stage with slower increase. First
the STAR-CD version 3.100b software package using the stage is properly due to alignment of the fat globules and
mesh generator in Prostar. The mesh was created as an the second stage is due to an ongoing agglomeration. The
unstructured mesh with hexahedral cells using couples to change in fat globule size toward larger globules is
connect un-matching cell faces. For discretisation and probably accelerated by the agglomerated fat globules
modelling of turbulence the second order Central during shearing. The third stage is not shown in Figure 1,
Difference scheme and the RNG k-ε model was used. The but a dramatic viscosity increase usually occurs followed
simulations are iterated until residuals below 10-6 or by a decrease in viscosity. If the last stage occur, churned
constant residuals over a number of 200 iterations. out fat is observed in the shear cell. Stability of fat globules
is influenced by many factors. This project concentrates on
RESULTS AND DISCUSSION determining the influence of temperature, time and fat
content on the resistance of the fat globules to destabilise
At the present state of the project focus are on determining during shearing.
the process conditions that affect milk fat globules. This Figure 2 shows the size distribution, the calculated mean
was done using a shear cell where predefined shear rates and 90 % fractal sizes of fat globules in cream samples

Fat globule Size Distributions Fat Globule Sizes

14 10 50
0s 9 45
12
90 % fractal size [ µ m]

20 s 8 40
40 s
Mean size [µ m]

10 7 35
Volume [%]

60 s
8 100 s 6 30
175 s 5 25
6 350 s 4 20
4 3 15
2 10
2 1 5
0 0 0
1 10 100 0 100 200 300 400
Size [µm] Shearing time [s]
Figure 2: Left Graph shows the Size Distribution of Cream (38 % fat) Sheared at 500 s-1 in 0, 20, 40, 60, 100, 175 and 350
seconds at 20°C. Right Graph shows the Calculated Mean Fat Globule Sizes and 90 % Fractal Sizes for the results
exposed to shearing at different length of time. The size
distribution shows that during shearing the distribution was Fat Globule Size
changed towards larger fat globules due to agglomeration. 30 80
Exposed to the conditions applied in this experiment, the
disruption of globules seems to occur at a constant rate in 70
25
the first 100 seconds of shearing. Further shearing 60

90 % fractal [ µ m]
dramatically changed the size distribution, and fat lumps

Mean size [µ m]
20
can be observed in the sample after shearing. 50
Fat content in the samples has a large effect on the shear
15 40
rates that the sample can withstand without being damaged ,
(Foley et al. 1971; Hinrichs and Kessler 1997). When the 30
10
fat content is increasing, the area where shear can evolve
20
decreases, and the result is an increase in the shear that the
5
surface of the globules exhibits resulting in an even greater 10
risk of disruption. This is show for cream with 30 and 38 %
fat in Figure 3. 0 0
0 10 20 30 40
Temperature [°C]
Fat Globule S izes
Figure 4: Mean (ο) and 90 % Fractal (×) Fat Globule Sizes
16 40
Measured in Cream (38 % fat) Sheared in 30 s with a
14 35 constant Shear Rate of 750 s-1
90 % fractal [ µ m]

12 30
Figure 5 shows predicted shear rates using CFD in a pipe
Mean size [µ m]

10 25 bend and in the vicinity of a butterfly valve. The

calculations were performed with cream flowing at an
8 20
average velocity of 1,5 m/s, which is the typical flow rate
6 15 used in the industry. Using CFD it is possible to estimate
the time particles e.g. fat globules is located in different
4 10
shear regions. To get a proper estimate the CFD simulations
2 5 needs to be done with a large amount of particles to ensure
0 0 that some particles enters the very often small areas of high
shear rates that causes the larger fat globule damage. Such a
0 500 1000 1500 calculation demands large computer power, and has not
-1
Shear rate [s ] been performed in this project so far.
Figure 3: Mean (ο) and 90 % Fractal (×) Fat Globule Sizes
Measured in Cream Samples with 30 % fat (- - - -) and 38
% fat (______) Sheared in 30 s at 20°C at the shown Shear
Rates

Likewise temperatures have an effect on stability of the

milk emulsion during processing (Schaller 1986; Lehmann
1988; Hinrichs 1994; Hinrichs 1998). As shown in Figure 4
fat globules are sensitive towards shearing in the
temperature interval between 15 to 35°C. A critical amount
of crystalline fat exist where fat globules is extremely
sensitive to shearing. Therefore, designing process
equipment should be done with the careful processing of
milk products in mind when process temperatures are
within the critical limits.
Shear rates in pipes with fully developed flow can easily be
calculated from analytical or empirical formulas.
Information on shear rates in complex geometries requires Figure 5: CFD Modelling of Shear Rates in the flow of
the use of computational tools like CFD or advanced Cream through a pipe Bend (left) and through a Butterfly
experimental methods. The information is necessary to Valve opened 10 % (right). The diameter is 22,5 mm and
obtain data for shear rates for all types of equipment used in average Velocity of the Cream 1,5 m s-1. The Arrows show
the dairy industry. Predicted shear rates can be validated by the Direction of Flow
e.g. Laser Sheet Visualisation. The shear rate for cream
flowing at 1.5 m s-1 in a pipe with an internal diameter of CONCULSION
22,5 m s-1, gives a maximal shear rate of above 1000 s-1 at
the internal surface of the pipe. Combining CFD simulations on flow fields in production
equipment with results of flow describing parameters effect
on food quality is shown to be a powerful tool, which will AUTHORS BIOGRAHY
extend our knowledge of producing good quality foods.
Based on the results presented above, it is clear that some JESPER SÆDERUP NIELSEN obtained his M.Sc.
degree of fat globule damage will occur in even simple (Dairy Eng.) degree in 1999 from The Royal Veterinary
processing equipment when processed in the sensitive and Agricultural University, Denmark. Before his studies
temperature interval. From the illustration of shear rates he was employed two years in New Zealand working with
around a slightly opened butterfly valve, it is seen that large design, manufacturing and start up of process equipment for
pressure drops over a valve develop areas with extremely the dairy industry e.g. membrane filtration plants,
high shear rates. To prevent the problems that such high evaporators and spray driers. Presently he is working on his
shear rates can cause, designers of equipment need to avoid Ph.D. study about careful processing of milk products.
unnecessary pressure build up in the production of Understanding how good food can be produced in large
equipment intended for use in dairy industries. This is done scale and teaching students technological aspects of food
by correctly dimensioning process equipment and by production, is the driving force behind his work. Email:
controlling plant capacities by frequency-controlled pumps. [email protected]
Redesign of standard pipe bends towards giving a more
uniform flow pattern, will also decrease total shear in ALAN FRIIS is Associate Professor at BioCentrum-DTU
production equipments. and is heading the research in food process engineering. He
holds a Ph.D. degree in the area of food processing from
FUTURE WORK 1994 and obtained his M.Sc. (Mech. Eng.) degree in 1990
both at the Technical University of Denmark. The main
It is the ultimate goal of this project to set-up a validated research interests are: improvement of quality of
process models capable of estimating fat globule damage in industrially produced food products, catering, aspects of
process equipment. From this model a process care in continuous processing of food products, modeling
understanding can be extracted, which can be used to give and simulation of food process operations, hygienic design
recommendation for appropriate process conditions during of food processing equipment and rheology. Teaching
manufacturing of milk products. Furthermore, it will be activities are concentrated around aspects of food
possible to set-up guidelines for design of process production. Email: [email protected]
equipments. This project is running for another year, which
is manly going to be spent on model development.

ACKNOWLEDGEMENTS

We acknowledge financial support from the Danish

Research Council and the Danish Dairy Board. Professor
Jens Adler-Nissen and Associate Professor Jørgen Risum
both BioCentrum-DTU for many fruitful discussions.

REFERENCES

Foley, J., J. Brady, and P.J. Reynolds. 1971. "The influence of

processing on the emulsion stability of cream." Journal of the Society
of Dairy Technology, 24, 54-58.
Hinrichs, J. 1994. "Grenzbelastung für Fettkügelchen."
Milchwirtshaftliche Berichte, 121, 227-231.
Hinrichs, J. and H.G. Kessler. 1997. "Fat content of milk and cream and
effects on fat globule stability." Journal of food science, 62, 992-
995.
Hinrichs, J. 1998. "Flow rates for milk and cream in pipelines."
Milchwissenschaft, 53, 139-143.
Lehmann, H. 1988. "Rahm- und Fettkugelschädigung." Deusche
Molkerei Zeitung, 109, 671-673.
Schaller, F. 1986. Fettschaedigung bei der Milchbehandlung in
Sammelstellen und Kaesereien." Deutsche milchwirtschaft, 37, 1406-
1410.
 ELABORATION OF TEXTURANT MODEL FLUIDS
FOR FOOD INVESTIGATIONS
Romuald Guérin Michel Lebouché
Guillaume Delaplace LEMTA-CNRS UMR 7563
Gilles Ronse 2 avenue de la forêt de Haye – BP 160 F-54501Vandoeuvre-
Jean-Claude Leuliet les-Nancy Cedex, France
Institut National de la Recherche Agronomique (I.N.R.A.)
Laboratoire de Génie des Procédés et Technologie
Alimentaire
369, rue Jules Guesde, B.P. 39
59651 Villeneuve d'Ascq cedex, France

KEYWORDS
Texturant model fluids, gelatinization, melting, laboratory appear during a heat transfer process like eggs, milk, starch or
scale, pilot scale. gelling agents based products in food industry. To fill this gap,
an experimental program has been started in our laboratory. In
ABSTRACT particular, the aim was to obtain two texturant model fluids
with specific rheological behaviour as shown in figure 1.
In this paper, two texturant model fluids (gelatin or starch
based products) have been elaborated for food investigations
involving heat transfer process. These two model fluids based
on mixtures of water, sucrose and gelatin or starch. Apparent viscosity, (Pa.s)
Solid melting
In this contribution, effect of sucrose content on melting
temperatures of gelatin aqueous solutions and on gelatinization
temperatures of starch aqueous solutions was investigated.
Temperatures were obtained using Differential Scanning
Calorimetry and dynamic viscoelasticity measurements.
It was clearly shown that increasing sucrose content raises both
gelatinization temperatures of starch mixtures and melting Temperature, (°C)

temperatures of gelatin mixtures.

Finally, it was shown that preparation of the two texturant Texturation product
Apparent viscosity, (Pa.s)

model fluids can be achieved in an industrial agitated vessel

(34 litres).

INTRODUCTION

Batch mixing of highly viscous fluids is widely used in the

Temperature, (°C)
food and chemical industries. For these industries, the purpose
of mixing operations could be quite different. For example, the
requirement could be to ensure good homogeneity inside the Figure 1. Evolution of apparent viscosity with temperature for
vessel and/or to improve heat and mass transfer. the two models fluids elaborated in this study.
Efficiency of such operations is related to the geometry of
mixing system and to the operating conditions. From a
practical point of view, optimization of a mixing process in To obtain and select the required texturant products (Figure 1):
food industries still remains nowaday a difficult task. Indeed - first it was chosen to prepare gelatin and starch aqueous
for these industries, most of the agitated fluids are complex solutions containing various sucrose contents;
(non-Newtonian) and properties of raw materials are greatly - then to follow evolution of apparent viscosity of these
influenced by seasons. This causes additional difficulties to the mixtures during heat treatment.
process engineer to select the best operating conditions for the In this contribution, the effect of sucrose content on texturant
batch process. properties of starch and gelatin mixtures, derived from this
In this case, model fluids with controlled and stable rheological experimental program, was presented. In particular, dynamic
properties are required. rheological measurements and Differential Scanning
Unfortunately, from a review of models fluids, it appears that Calorimetry results are reported and discussed. Finally, an
most of them have been elaborated only for isothermal studies. additional trial will be presented concerning elaboration of
One may mention gum solutions to simulate shear-thinning model fluids at pilot scale.
behaviour or glucose syrups to simulate Newtonian liquids. On
contrary, to the authors’ knowledge, less model fluids have
been proposed to simulate drastic changes in the texture which
 Instruments, UK - cone angle 2°, diameter 4 cm) to follow both
MATERIALS AND METHODS the melting process of gelatin/sucrose solutions and the
texturation process of starch/sucrose solutions during heat
Preparation of model fluids. treatment. For that, an increasing temperature ramp varying
from 10°C to 55°C for the gelatin solutions and from 30°C to
Starch/sucrose and gelatin/sucrose aqueous solutions with 95°C for the starch solutions was applied at a rate of 1°C/min.
various sucrose contents (0, 10, 20, 30, 40%, w/w) were In the same time, storage modulus (G'), loss modulus (G") and
prepared. A Waxy corn starch, a pig gelatin and a commercial tan δ (G"/G') were recorded. Note that liquid paraffin was used
sucrose (Beghin-Say firm) were used for this purpose. For all to minimize the water loss during the measurements. These
mixtures, the content of starch and of gelatin were fixed and trials were carried out in triplicate. For all runs, the strain and
equal respectively to 20% (w/w) and 5% (w/w). Solutions were frequency were respectively set at 1% and 1Hz. Finally, G'
achieved by dispersing starch/sucrose and gelatin/sucrose curves versus temperature were analysed (intersection of
respectively at 30°C and at 60°C in distilled water. To increase tangents) in order to obtain representative melting (θ1, θ2, θ3)
the apparent viscosity of these solutions, a shear-thickening has or gelatinization (θ4, θ5) temperatures (Figure 3).
been added. To avoid scatters of viscoelastic properties due to
various time and thermal sample history (Michon and 3500
Gelatin/sucrose mixture Starch/sucrose mixture
al.,1997), all products were vacuumed and maintained in a 3000
θ5
constant temperature at 4°C for 15 hours.
2500

θ1
Measurements of transition enthalpy and determination of 2000

G', (Pa)
representative gelatinization temperatures. 1500

1000
The gelatinization process of the starch suspension was
measured by a DSC apparatus (SETARAM C80). Samples of 500 θ2
θ3 θ4
about 5g were hermetically sealed in metal pans and placed on 0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80
one side. The other, or reference, side held distilled water. The Temperature, (°C)
instrument heated both pans equally from 20 to 90°C at a rate
of 0,1°C/min. Analysis of DSC thermograms of starch mixtures Figure 3. Examples of G' curve obtained for gelatin and starch
allow us to determine three representative temperatures of aqueous solutions. Determination of representative
gelatinization: temperatures for the two models fluids.
- the onset temperature (θ0);
- the temperature of maximum differential heat flow (peak Preparation of model fluids at pilot scale.
temperature, θp);
- the termination temperature (θt) . To establish the influence of mixer scale on texture of model
The peak temperature was easily estimated, but onset and fluids, it was decided to elaborate the product at two scales (a
termination temperatures were generally not sharply defined. In mixing tank of 34 litres equipped with a close clearance
this work, these temperatures were determined by the impeller and a jacketed vessel and a mixing sytem of 0,041
intersections between straight lines drawn through the base line litres equipped with coaxial cylinders submerged in water bath)
and the sides of the DSC peak as proposed by Donovan (1979) and to compare the evolution of their apparent viscosity during
(Figure 2). Finally, DSC thermograms were also used to heat treatment. To carry out measurement in line of the
determine the transition enthalpy. This was done by cutting out apparent viscosity, the two mixing systems have been equipped
and weighting the area of peaks. with a torquemeter and a tachymeter. An increasing
temperature ramp varying from 25°C to 95°C was performed at
1°C/min for the two mixing systems. Note that for the mixing
equipment of 34 litres, it was difficult to maintain this rate at
the end of the trial due to the operating conditions chosen.
θ0 θt Indeed impeller rotational speed and wall temperature were
fixed during all the heat transfer runs.

θp

2
TANK:
1 Vessel diameter, T = 0,346m
Figure 2. Determination of gelatinization temperatures for Vessel heigh = 0,443m
starch/sucrose aqueous solutions from DSC thermogram. 1= inlet AGITATOR (Ekato-PARAVISC®):
2= outlet of cooling diameter, D = 0,320m
or heating water blades width, w = 0,032m
Measurement of Dynamic Viscoelasticity. impeller pitch, p = 0,560m
impeller height, L = 0,340m

Dynamic viscoelasticity measurements were performed in a

Figure 4. Pictures and geometrical parameters of the jacketed
cone plate system viscosimeter (Carrimed AR1000, TA
pilot mixing system used.
RESULTS AND DISCUSSION Using method previously described (Figure 3), it was possible
to obtain from G' curves representative melting temperatures.
Gelatin/sucrose mixtures: Evolutions of texture obtained by These temperatures (θ1, θ2, θ3) are reported in table 1.
dynamic viscoelasticity measurements. Analysis of these data (Figure 6) shows clearly that gel-sol
transition temperatures are influenced by sucrose contents.
Figure 5 reports the results from dynamic viscosity Indeed, for the range of sucrose contents investigated, we can
measurements obtained for all gelatin solutions studied during note that melting temperatures of gelatin increase linearly with
the heating process. It appears that the storage (G') and loss sucrose content.
(G") moduli sharply decrease while in the same time
tan δ increases with temperature. These results prove that the 45

gelatin melting process is induced by increasing temperature. 40

35

Temperature, (°C)
5000 30

4500 25
Symbols Sucrose content
4000 (% w/w)
20
0
3500 Symbols Temperatures
15
10 θ1
3000
G', (Pa)

20 10 θ2
2500
30 5 θ3
2000
40 0
1500 0 10 20 30 40 50

1000 Sucrose content, (% w/w)

500

0
5 10 15 20 25 30 35 40 45 50 55
Figure 6. Effect of sucrose content on melting temperatures of
Temperature, (°C) a pig gelatin determined from G' curves. The bars represent
700 standard deviation of three mesurements.
Symbols Sucrose content
600 (% w/w)

0
500 10
Starch/sucrose mixtures: Evolutions of gelatinization
20
temperatures obtained by DSC measurements.
G", (Pa)

400
30

300 40

Figure 7 reports values of gelatinization temperatures and

200
transition enthalpy obtained by DSC thermograms for the
100
starch/sucrose solutions investigated.
0
5 10 15 20 25 30 35 40 45 50 55
100 20
Temperature, (°C) Symbols Parameters
θo 19
Gelatinization temperatures,

100 90 θp 18
Symbols Sucrose content
(% w/w) θt
17
0 80

Enthalpy, (J/g)
Enthalpy
16
10
(°C)

10 70 15
20
14
tan(δ )

30
60
13
40
1 12
50
11

40 10
0 5 10 15 20 25 30 35 40 45
0,1 Sucrose content, (% w/w)
5 10 15 20 25 30 35 40 45 50 55

Temperature, (°C)

Figure 7. Effect of sucrose content on a corn starch

Figure 5. Change in rheological properties of gelatin/sucrose gelatinization temperatures (θo, the onset temperature; θp, the
mixture with 0, 10, 20, 30 and 40% of sucrose during a heating peak temperature and θt, the termination temperature).
from 10 to 55°C. Conditions: frequency, 1Hz; strain, 1%.
We can note that both the representative gelatinization
Table 1. Melting temperatures determined from G' curves for temperatures (θo, θp and θt) and values of transition enthalpy
gelatin aqueous solutions with various sucrose contents. (from sol to gel) are influenced by sucrose content. For
example, more than 20°C deviation was observed for peak
Sucrose content θ1 θ2 θ3 temperature depending on sucrose content of starch mixtures.
(% w/w) (°C) (°C) (°C) In the same time, transition enthalpy values vary more than
0 13,4 +/- 0,4 23,2 +/- 0,3 30,7 +/- 0,7 70%, depending on sucrose content.
10 16,5 +/- 1,0 26,2 +/- 1,0 33,2 +/- 0,7 These observations are in agreement with those obtained by
20 18,6 +/- 0,5 28,9 +/- 0,6 34,7 +/- 0,6 other authors (Biliaderis and al., 1980; Chinachoti and al.,
1990; Eliasson, 1980; Evans and Haisman, 1982 and Wootton
30 19,7 +/- 0,6 29,4 +/- 1,2 37,0 +/- 0,5
and Bamunuarachchi, 1979) which underlined that
40 23,3 +/- 1,0 31,8 +/- 0,7 38,8 +/- 0,3
gelatinization temperatures greatly depend on the environment
of the starch granules.
Starch/sucrose mixtures: Evolutions of texture determined by Table 2. Gelatinization temperatures of starch/sucrose mixtures
dynamic viscoelasticity measurements. versus sucrose content determined from G' curves.

Figure 8 illustrates the results from dynamic viscoelasticity Sucrose content θ4 θ5

measurements of starch solutions during heating. (% w/w) (°C) (°C)
8000 0 63,6 +/- 0,1 68,1 +/- 0,2
Sucrose content
7000
Symbols
(% w/w)
10 66,6 +/- 0,2 71,0 +/- 0,1
0

6000 10 20 74,6 +/- 0,3 79,1 +/- 0,3

5000
20

30
30 78,5 +/- 0,1 83,0 +/- 0,3
40 87,4 +/- 0,2 91,5 +/- 0,9
G', (Pa)

40
4000

3000

2000
Results show that sucrose content influences gelatinization
temperatures of starch aqueous solutions. Gelatinization
1000
temperatures increase with sucrose content. For instance, corn
starch showed a range of gelatinization temperatures varying
0
30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Temperature, (°C)
from 63,6°C to 68,1°C in pure water and 87,4-91,5°C for
2000

1800
Symbols Sucrose content starch sample with 40% (w/w) of sucrose. These oscillation
(% w/w)

1600
0 tests confirm reasonably well that starch gelatinization
1400
10
temperatures were influenced by changes in the environment of
20

1200 30 the starch granules (Figure 8). This observation was in

G'', (Pa)

1000
40 agreement with various authors (Chang and Liu, 1991; Evans
800
and Haisman, 1982; Oosten, 1982; Singh and al., 2002 and
600
Wootton and Bamunuarachchi, 1980). This can be explain both
400 by reducing the amount of water available to the granules and
200 by adding high levels of hydrophilic solutes.
0 Figure 8 also shown that values of G'max and G"max increased
30 35 40 45 50 55 60 65 70
Temperature, (°C)
75 80 85 90 95 100
with sucrose content. The G' peak was 2879 Pa for pure starch
100 solution against 7447 Pa for starch solution with 40% w/w
sucrose content. These observations clearly indicate that
aqueous starch solution are more viscous with increasing
10
sucrose content. This can be explained simply by the reduction
of water content.
tan(δ )

To compare DSC and dynamic viscoelasticity measurements,

Symbols Sucrose content
(% w/w) representative gelatinization temperatures of aqueous starch
1
0
10
solutions have been plotted on Figure 9. We can note that
20 there is a good agreement between the two set of values.
Indeed, θ4 determined from G' curves correpond to the peak
30
40

temperature (θp) and θ5 correspond to the termination

0,1
30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Temperature, (°C)
temperature (θt). This indicates that the two experimental
devices are accurate enough to detect both the beginning and
Figure 8. Evolution of storage modulus (G'), loss modulus the end of gelatinization process.
(G") and tan δ with temperature (30 – 95°C) for various
aqueous solutions of starch with various sucrose contents: 0, 100
Symbols Temperatures
10, 20, 30 and 40%. Operating conditions: frequency, 1Hz; 95 θ4
Viscosimetric
θ5
strain, 1%. 90
θo
Temperature, (°C)

85
Calorimetric
θp
80
The storage (G') and loss (G") moduli sharply increased with 75
θt

temperature and reached a maximum value which may be 70

attributed to the closely packed network of swollen starch 65

granules (Eliasson, 1986; Hsu and al., 2000; Lii and al., 1985). 60

55
In the same time, tan δ sharply decreases with temperature. 50
From the rheological point of view, this decrease corresponded 0 10 20 30 40 50

Sucrose content, (% w/w)

to a sol-to-gel transition (Hsu and al., 2000).
With further increase in temperature, G' and G" decreased
indicating that the gel structure was destroyed during Figure 9. Gelatinization temperatures of aqueous solutions of
prolonged heating. This destruction was probably due to the starch with various sucrose contents determined by
melting of the crystalline region remaining in the swollen calorimetric (θo, θp and θt) and viscosimetric (θ1 and θ2)
starch granules (Eliasson, 1986). Representative temperatures, methods.
θ4 and θ5, obtained from G' curves and reported in table 2.
Preparation of model fluids at pilot scale. systems. On the contrary, the deviation of apparent viscosity
obtained for gelatin/sucrose mixtures (figure 11) can not be
Figures 10 and 11 show respectively the evolution of the explained by different level of average shear rate but may be
apparent viscosity (ηa) of starch/sucrose and gelatin/sucrose caused by various local shear rate distribution for the two
aqueous solutions obtained at laboratory and at pilot scale. mixing systems (Delaplace et al., 2000).

12 2,5 CONCLUSION
10
In this work, prepration of aqueous solutions of gelatin and

Apparent viscosity, (Pa.s)

Apparent viscosity, (Pa.s)

8 starch with various sucrose contents have been carried out to

1,5 elaborate and select two texturant model fluids for food
6
investigations involving heat transfer process. Experimental
4
1
measurements (DSC and dynamic viscoelasticity) with these
solutions have allowed us to underline the effect of sucrose
0,5
2 content on melting temperatures of gelatin and gelatinization
0 0
temperatures of starch..
25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 Finally, an additional trial have been performed to elaborate
Temperature, (°C) and follow apparent viscosity of these model fluids in line at
Figure 10. Evolution of the apparent viscosity for starch based industrial scale. Results obtained prove that it is feasible to
texturant model fluid at the laboratory scale (grey symbols - γ& elaborate such model fluids at this scale.
= 105s-1) and at the pilot scale (black symbols - γ& = 18s-1).
BIBLIOGRAPHY
25 3,5

3 Biliaderis C.G., Maurice T.J. and Vose J.R. 1980. "Starch

20
Apparent viscosity, (Pa.s)

gelatinization phenomena studied by differential scanning

Apparent viscosity (Pa.s)

2,5
calorimetry". J. Food Sci., 45, 1669-1674.
15
2 Chang S.M. and Liu L.C. 1991. "Retrogradation of rice starches
studied by Differential Scanning Calorimetry and influence of
1,5
10 sugars, NaCl and Lipids". J. Food Sci., 56, 564-566.
1 Chinachoti P., Steinberg M.P. and Villota R. 1990. "A model for
5 quantitating energy and degree of starch gelatinization based on
0,5
water, sugar and salt contents". J. Food Sci., 55, 543-546.
0 0 Delaplace G., Torrez C., André C., Leuliet J.-C. and Fillaudeau L.
0 10 20 30 40 50 60 70 80 90 100 2000. "CFD simulation of foodstuff flows in an agitated vessel".
Temperature, (°C)
1st International Conference on simulation in food and
Figure 11. Evolution of the apparent viscosity for gelatin based Bioindustries, edited by Daniel Thiel, 179-186.
texturant model fluid at the laboratory scale (grey symbols) and Donovan J.W. 1979. "Phase transitions of the starch-water system".
at the pilot scale (black symbols). Conditions: γ& = 105 s-1, Biopolymer, 18, 263-275.
heating rate of 1,0°C/min. Doublier J.-L. 1987. "A rheological comparison of wheat, maize, faba
bean and smooth pea starches". J. Cereal. Sci., 5, 247-262.
Eliasson A.-C. 1980. "Effect of water content on the gelatinization of
These figures show that for the two scales there are similar wheat starch". Starch, 32, 270-272.
drastic changes in the texture of the product caused by Eliasson A.-C. 1986. "Viscoelastic behaviour during the gelatinization
increasing temperature. These similar texturation processes at of starch. I: Comparison of wheat, maize, potato and waxy barley
the two scales appear more clearly by comparing for the two starches". J. Texture Studies, 17, 253-265.
mixing systems, activation energy of flow required for sol-to- Evans I.D. and Haisman D.R. 1982. "The effect of solutes on the
gel or gel-to-sol transition in (Table 3). gelatinization temperature range of potato starch". Starch, 34, 224-
231.
Table 3. Activation energy of flow (Eaf, kJ/mol) of models Hsu S., Lu S. and Huang C. 2000. "Viscoelastic changes of rice starch
suspensions during gelatinization". J. Food Sci., 65, 215-220.
fluids at the laboratory scale (L) and at the pilot scale (P) (E – Lii C.Y., Shao Y.Y. and Tseng K.H. 1995. "Gelation Mechanism and
Relative error, %). rheological properties of rice starch". Cereal Chem., Vol.72, 393-
400.
Starch based fluid Gelatin based fluid Michon C., Cuvelier G., Relkin P. and Launay B. 1997. "Influence of
Eaf (kJ/mol) Eaf (kJ/mol) thermal history on the stability of gelatin gels". Int. J. Biol.
L P E (%) L P E (%) Macromolecules, 259-264.
86,0 77,6 10,8 40,1 33,7 19,0 Oosten B. 1982. "Tentative hypothesis to explain how electrolytes
affect the gelatinization temperature of starches in water". Starch,
34, 233-239.
Indeed for the two scales, values of activation energy of flow Singh J., Singh N. and Saxena S.K. 2002. "Effect of fatty acids on the
are same order of magnitude which proves that similar rheological properties of corn and potato starch". J. Food Engin.
texturation process (either gelatinization or melting) has been 52, 9-16.
achieved. Wootton M. and Bamunuarachchi A. 1980. "Application of
Note that on Figure 10, there is a scatter between values of ηa differential scanning calorimetry to starch gelatinization – III.
versus temperature obtained at the two scales for starch/sucrose Effects of sucrose and sodium chloride". Starch, 32, 126-129.
mixture. This may be explained by the different magnitude of
average shear rate during heat treatment in the two mixing
 DIMENSIONLESS ANALYSIS OF FLUID-TO-PARTICLE MASS TRANSFER
COEFFICIENTS IN CONTINUOUS TUBE FLOW
Isabel Abreu Fernanda Oliveira
Escola Superior de Biotecnologia Department of Process Engineering
Universidade Católica Portuguesa University College Cork
Rua Dr. António Bernardino de Almeida
4200-072 Porto, Portugal Cork, Ireland
E-mail: [email protected] E-mail: [email protected]

KEYWORDS information available in literature on fluid-to-particle heat

Fick’s 2nd law, fluid-to-particle heat transfer coefficient, transfer coefficients in continuous flow systems is scarce,
fluid-to-particle mass transfer coefficient, heat/mass transfer often inconsistent and sometimes contradictory. Most of the
analogy. equations found in literature are of the Fröszling type. The
major problem when measuring the fluid-to-particle heat
ABSTRACT transfer coefficient in a continuously flowing suspension of
particles is the measurement of the particles’ temperature
The selection of time-temperature conditions that without disturbing flow conditions (Sastry 1992). Fluid-to-
maximize product quality while ensuring product safety in particle mass transfer coefficients are in general easier to
the aseptic processing of particulate fluid foods depends determine experimentally without disturbing the flow, and,
greatly on the ability to measure or predict fluid-to-particle by using the heat/mass transfer analogy, they may be
heat transfer coefficients. Monitoring the temperature of a converted to fluid-to-particle heat transfer coefficients.
moving particle in two-phase flow without disturbing the
flow conditions is however difficult, which hinders the The heat/mass transfer analogy relies on the similarity
determination of heat transfer coefficients. Mass transfer between the dimensionless relations that govern the thermal
coefficients are often easier to determine experimentally. If and the concentration boundary layers. If, for example, one
correlations between Sherwood (Sh), Reynolds (Re) and has performed a set of mass transfer experiments to
Schmidt (Sc) numbers are developed, they might then be determine the form of a function that relates Sh with Re and
used to predict heat transfer coefficients, by the heat/mass Sc, for a particular surface geometry, the results should hold
transfer analogy, thereby replacing Sh and Sc numbers by in convection heat transfer involving the same geometry,
Nusselt (Nu) and Prandtl (Pr) numbers, respectively. In this simply by replacing Sh with Nu and Sc with Pr, provided the
work fluid-to-particle mass transfer coefficients were range of dimensionless numbers is the same (Incropera and
determined in tube flow. Dimensionless correlations were DeWitt 1996). Conversely, mass transfer coefficients may be
developed and compared with published correlations for heat estimated with correlations developed for heat transfer (e.g.
transfer, although the latter were developed for values of Prg as Califano and Calvelo 1983).The Chilton-Colburn analogy
much smaller than those of Scg determined in the present has proved useful for a number of situations (Chilton and
work. In general, those correlations overpredicted the Sh Colburn 1934, Bird et al. 1960).
numbers, but a fair agreement was observed when using a
correlation reported in literature that considers both the Noordsij and Rotte (1967) developed dimensionless
effect of translational and rotational relative particle-to-fluid correlations for predicting average mass transfer coefficients
velocities. These results show the potential of using for rotating or vibrating spheres immersed in a liquid
heat/mass transfer analogies for estimating fluid-to particle medium. They used different experimental conditions to
heat transfer coefficients in continuous flow. ensure that mass transfer rate was controlled either by
internal diffusion or external convection, which allowed for
separate estimation of diffusivity (D) and mass transfer
INTRODUCTION coefficient (Kc). To the best of our knowledge, fluid-to-
particle mass transfer in continuous flow has only been
The estimation of fluid-to-particle heat transfer studied by Fu et al. (1998). They conducted experiments in
coefficients is one of the key aspects in the design and vertical holding tubes of an aseptic processing system to
optimisation of continuous thermal processing of particulate determine fluid-to-particle mass transfer coefficients, and
fluid foods. Dimensional analysis is often used to correlate reported that the Sherwood number decreased with
heat transfer coefficients with fluid properties, fluid velocity increasing residence time. Those results were then extended
and particle geometry. Correlations have been reported in to predict heat transfer under similar conditions. The slight
literature for fluid flow past single stationary spheres in a increase of fluid-to-particle heat transfer coefficient with
duct (Geankoplis 1983). In the continuous thermal decreasing particulate residence time suggested that fastest
processing of particulate fluid foods the particles flow freely moving particles might not necessarily be the least sterilised.
with the fluid medium, and thus these correlations may not
hold true. Ramaswamy and Zareifard (2000) reported that
 The main objectives of this work were: (i) to measure mass as the tangent of the shear stress versus shear rate curve, for
transfer coefficients in continuous tube flow of a particulate zero shear stress.
fluid in a range of temperatures and flow rate, (ii) to use
these data to develop dimensionless correlations to predict Table 1: The Power Law Model Parameters and Intrinsic
fluid-to-particle mass transfer coefficients, and (iii) to Viscosity of 0.6% CMC Solutions
compare these correlations with others reported in literature
for predicting fluid-to-particle heat transfer coefficients. T (°C) K (Pa.sn) n µs (kg.m-1.s-1)
25 0.54 ± 0.02 0.613 ± 0.003 0.150 ± 0.004
MATERIAL AND METHODS 30 n.a. n.a. n.a.
35 0.37 ± 0.02 0.629 ± 0.002 0.098 ± 0.003
The mass transfer process studied was acetic acid uptake 40 0.25 ± 0.07 0.67 ± 0.02 0.079 ± 0.008
by agarose particles initially free of acid. Particles with 45 0.21 ± 0.02 0.668 ± 0.006 0.071 ± 0.005
different geometry (cubes and spheres) were tested. The 50 0.18 ± 0.01 0.677 ± 0.005 0.061 ± 0.002
carrier fluid was a carboxymethylcellulose solution (0.6%) 55 0.12 ± 0.02 0.72 ± 0.02 0.049 ± 0.005
containing acetic acid (0.2M). The experiments were 65 0.08 ± 0.02 0.75 ± 0.02 0.011 ± 0.007
conducted in a tubular continuous thermal processing unit n.a. – not available, for calculation purposes interpolated
under isothermal conditions. values (k = 0.44 Pa.sn , n = 0.62 and µs = 0.12 kg.m-1.s-1)
were used.
Agarose gel
The density of the acetic acid solutions was determined at
An agarose gel was chosen as a model food because of its different temperatures using a Mettler AM 100 balance and a
stability in acid at high temperatures, without shrinking or density determination kit (Mettler ME-33340). Values
swelling. A 5%w/v agarose gel was prepared by adding ranged from 999.3 to 980.8 kg.m-3 for temperatures between
agarose powder (Nº05068, Fluka Chemika, Buchs, 25 and 65 ºC. The density of the 5% agarose was 1.02 ± 0.01
Switzerland) to deionised water held covered at 95 °C and g.cm-3.
stirred until it was transparent (about 45 min); the mixture
was then poured into a mould and left to set. After gelation, The experiments
detected by the white opacity, the gel was cut into 10 ± 0.1
mm cubes. Agarose spheres with diameter in the range of A continuous tubular processing system was used,
approximately 6 to 9 mm were prepared by dropwise designed by Pinheiro-Torres and Oliveira (1991), and built
injection of the transparent agarose solution into an ice-cold by ARSOPI, S. A. (Vale de Cambra, Portugal). This system
mixture of toluene-chloroform-n-hexane (5:2:1, v/v/v), as includes a solid/liquid mixture feed tank, an alloy gear pump
described by Tahoun (1986). Excess solvent was removed (Eco Gearchem, model no. ECO-G4-KCC-TTQ,
by shortly shaking in 1,4-dioxane, followed by successive Pulsafeeder, New York, USA), an electromagnetic
washing in deionised water. The cubes and spheres were flowmeter (flow transmitter Model PD340, from Process-
stored at 4 °C in deionised water prior to use, for periods up Data) a three way valve to introduce the agarose particles, a
to two weeks. visualization section and a tubular thermal processing unit
with heating, holding and cooling sections (Figure 1). The
The acidifying medium inner tube of the heat exchangers, the holding tube and all
the parts between the inlet and outlet have an internal
Situations of interest to food processing usually involve diameter of 22 mm, and the total length is 30 m. The holding
low Reynolds number, since the carrier fluid often is a section has an upward inclination of 1 in/ft.
viscous solution. Moreover, these solutions usually have
non-Newtonian behaviour. Carboxymethylcellulose (CMC)
solutions are often used as model solutions, as they are non-
Newtonian. CMC is also readily available and convenient
for testing since it is stable, safe to use and clean (Campos et
al. 1994). The 0.6% CMC solutions were prepared by
adding slowly sodium carboxymethylcelulose (Hoechst AG
– Tylose MHB 30000 ypTM) to water at 50 °C, with
continuous stirring maintained for 3 days. Then, acetic acid
(99+% pro analysis, Pronalab, Portugal) was added to the
solution. The rheological characteristics of CMC solutions Figure 1: Schematic representation of the continuous tubular
were determined using a controlled stress rheometer with processing system.
cone-and-plate geometry (Carri-Med CS-50, Surrey, UK). A – stirred tank; B – pump; C- three way valve (inlet);
Rheological measurements were performed at the beginning D – visualization section; E – flowmeter; F1, F2 and F3
and at the end of each set of experiments at the same – heating, holding and cooling sections; G – outlet.
temperature, and no time dependence was found. Table 1
shows the rheological parameters at different temperatures. All the experiments were carried out isothermally, with a
This table also includes the intrinsic viscosity (µs), evaluated difference between the inlet and outlet lower than 1 °C. The
feeding tank is jacketed and was pre-heated and held at the Kcd p (1)
Sh =
selected temperature using steam. The acid solution was Df
pumped from the tank to the unit and recirculated. n
Experiments were started after all the air was purged and  3n + 1  n −3
k  2 (2)
Scg = 
the temperature stabilized. The agarose particles were pre- n 
1− n
heated to the selected temperature. An agarose particle was  v 
introduced after the pump in the three-way valve, and later ρ f Df  
 dp 
removed at the outlet, rinsed and immersed in a flask with  
deionised water. The particle’s residence time was registered
The average acid concentration of the particles was 8 ρ f v 2f − n d np (3)
Re g =
measured by titration with standard 0.02M NaOH with an  3n + 1 
n
automatic titrator (TitralabTM), after dissolving the particle in 2n k  
 n 
50 ml of hot water. Ten 10 particles were run, for each
experiment, and a sample of the acid solution in the tank was
taken to be titrated. The operating variables were: where k and n are the constants of the power law model, ρ is
temperature (T); flow rate (Q); particle geometry and the density, υ is the viscosity, and the subscripts f and p
particle size (dp) (Table 2). stand, respectively, for fluid and particle.

Table 2: Values of the Operating Variables A single correlation was developed for cubes and spheres.
A weight of 10 was attributed to the mass uptake data
Particle regarding cubes, as in this case each Sh corresponds to the
dp (cm) * T (°C) Q (l.h-1) average of 10 particles.
geometry
25; 30; 35; 120; 180;
cubes 1.24 40; 45; 50; 240; 300; RESULTS AND DISCUSSION
55 360
0.63 to 0.69 Fluid-to-particle mass transfer coefficients
spheres 25; 65 180; 360
0.83 to 0.90
* for cubes dp corresponds to the diameter of a sphere The generalized Reynolds numbers based on the tube
with the same volume. diameter ranged from 12.5 to 163 for cubes, and from 21.9
to 221 for spheres, meaning that the regime flow for
MATHEMATICAL CONSIDERATIONS undisturbed liquid flow was laminar. Fluid-to-particle mass
transfer coefficients ranged from 0.55 x 10-5 to 2.14 x 10-5
Joint estimation of D and Kc was avoided because of the m.s-2, increasing both with temperature and flow rate, as
high correlation coefficients between D and Kc (Azevedo et expected. Figure 2 shows the results for cubes (the surface
al. 1998) Kc values were estimated by non-linear regression, represents the dimensionless correlation).
using the Simplex method (Nelder and Mead 1965) for
minimization of the sum of the squares of the errors between
experimental data and the values predicted by the solution of 5
Kc x 10 (m.s )
-1

Fick’s 2nd law (Crank 1975) subject to the appropriate

boundary conditions (uniform initial concentration, 1.6
symmetry condition and a convection boundary condition). 1.4
A Fortran program was specially written for this purpose.
1.2
The D values used in these calculations were estimated using
1.0
data from earlier experiments where the carrier fluid was
0.8
water, so that the external resistance to mass transfer was
0.6
negligible (parameters for the Arrhenius dependence of D on
temperature: Ea (activation energy)= 24 ± 1 kJ.mol-1; D = 380
340 60
300 55
1.88 × 10-9 ± 0.03 × 10-9 m2.s-1 at T = 45 ºC). 260
220 40
45
50

180 35
140 30
25
Analysis of variance and non-linear regression for -1
Q (l.h )
100 20
T (ºC)
developing dimensionless correlations were performed using
the R program (R 1.1.1, A Language and Environment
Figure 2: Effect of Flow Rate and Temperature on Fluid-to-
Copyright 2000). For cubes, Kc was estimated for each
particle Mass Transfer Coefficients during Continuous Flow
group of 10 particles used in a given experiment; for spheres
of Agarose Cube Particles in an Acetic Acid Solution
Kc values were estimated for each particle separately, as
Carrier.
they did not have exactly the same diameter. Sh was then
calculated (Equation (1)) and correlated with the relevant
The analysis of variance (at 95% significance level)
generalized dimensionless numbers, Reg and Scg (Equations
confirmed the influence of temperature and flow rate on Kc
(2) and (3)).
and showed no significant interactive effect of these two
operating variables for particles with cube geometry,
whereas for spheres the effect of temperature was more
pronounced at high flow rates; the diameter of the particles situation (0.33). The exponent of Ψ is not statistically
had no significant effect. The precision of Kc estimates was different from zero, probably owing to insufficient
found to be independent of temperature and flow rate. Biot experimental data (two sphericity factors only were
numbers (Bi=Kcdp/Dp) ranged from 15 to 47 for cubes and evaluated). However, the inclusion of this factor clearly
from 11 to 62 for spheres, increasing with increasing flow improved the fit.
rate and decreasing with increasing temperature, showing
that the diffusion coefficient of the acid in the gel is more Comparison with correlations for heat transfer
sensitive to temperature than the external mass transfer
coefficient. Experimental data were more scattered for Correlations reported on heat transfer studies cover similar
spheres than for cubes, particularly at 25 °C, owing to the ranges of Reg, yet Prg values are much lower than the Scg
smaller size of the particles and consequently lower acid values measured in our work. When comparing our results
uptake. with those correlations, it was found that they tend to
overpredict Sh. A fair agreement was however observed
Dimensional analysis when using a published correlation that considers both the
effect of translational and rotational relative particle-to-fluid
A Fröszling type correlation assuming the theoretical value velocities on the fluid-to-particle heat transfer coefficient in
of 2 for Sh in static conditions (i.e. Reg=0) yielded good fits flow in a straight tube (Baptista et al. 1997b). Using the
to the experimental data (R2=0.91): heat/mass transfer analogy, this equation can be written as:

Sh = 2 + (0.03 ± 0.01) Re g (0.46 ± 0.05) Sc g (0.62 ± 0.04) ψ (0.02 ± 0.30) (4) Sh = 2.0 + 0.025 × Sc 1s / 3 × Gr 1 / 2 + 0.64 × Re 0gvr
.59
× Sc 0gvr
.28
+
(5)
where Ψ is the particle sphericity (1 for spheres and cube
(
0.17 × Re 0gω.71 × Sc 0gω.42 × d p / d t )0.28
edge/diameter of a sphere with the same volume for cubes).
Figure 3 shows the good agreement between experimental This correlation was developed for 7<Reg<284 and
and predicted values and Table 3 summarises the ranges of 144<Prg<1,755. The particles linear and rotational velocities
dimensionless numbers used in this work. were estimated with dimensionless correlations reported by
Baptista et al. (1997a). These correlations were developed
for 62<Reg<4,734, 7.6<Ar<4.34×105, 0.035<Fr<2.29,
300
Sh predicted 0.065<α<0.425 and 0.216<dp/dt <0.649. Thus, the range of
250 Reg, Ar and α do not cover the values found in our work.
Furthermore, while Equation (5) refers to experiments
200
carried out in a straight tube, we have worked with a
150 continuous thermal processing unit that includes a number of
bends. Despite all these limitations, equation (6) yielded
100 predictions of the same order of magnitude, with better
results for cubes, probably due to the better quality of the
50
experimental data for this geometry (Figures 4 and 5).
0
0 50 100 150 200 250 300 300
Sh experimental Sh predicted
250

Figure 3: Prediction of Sh by Dimensional Analysis 200

(Equation (4))
150
Table 3: Range of the dimensionless numbers in the
100
conditions tested
50
Dimensionless Particle geometry
number cubes spheres 0
0 50 100 150 200 250 300
Reg 8.8 - 108 10.3 - 114
Sh experimental
Scg 18,744 - 293,654 9,197 - 221,666
Sh 52 - 266 33 – 284
Ar* 2.09 – 4.90 0.66 – 54.7 Figure 4: Prediction of Sh for cubes using the Correlation
Fr* 0.036 – 0.321 0.080 – 0.321 reported by Baptista et al. (1997b) (Equation (5)).
* Ar - Arquimedes number; Fr - Froude number
CONCLUSIONS
The exponent of Reg is not statistically different from 0.5,
Dimensionless correlations were developed to predict Sh,
the theoretical value for laminar flow over a horizontal
and thus fluid-to particle mass transfer coefficients, as a
surface and the most common figure in published studies,
function of Reg and Scg during flow of particulate fluids. A
whereas Scg is higher than the theoretical value for the same
published correlation for heat transfer was found to predict
Sh values within the same order of magnitude of those Crank, J. 1975. The Mathematics of diffusion. 2nd ed. Oxford
determined in this work, in spite of having been determined University Press, London, England.
in experimental conditions and ranges of dimensionless Fu, W., Sue, Y. and Chang, K.L.B. 1998. “Distribution of liquid-
numbers quite different from those used in our study. This solid heat transfer coefficient among suspended particles in
vertical holding tubes of an aseptic processing system.” Journal
shows the potential of using the heat/mass transfer analogy of Food Science, 63 (2), 189-191.
to estimate fluid-to-particle heat transfer coefficients from Geankoplis, C.J. 1983. Transport processes and unit operations. 2nd
correlations developed in mass transfer studies, which would ed. Allyn and Bacon, Inc., USA.
allow to overcome the hussle in measuring particle Grabowski, S. and Ramaswamy, H.S. 1995. “Incipient carrier fluid
temperature in continuous flow. More studies should be velocity for particulate flow in a holding tube.” Journal of
performed in a wider range of conditions and both heat and Food Engineering, 24, 123-136.
mass fluid-to-particle transfer coefficients should be Incropera, F.P. and DeWitt, D.P. 1996. Fundamentals of heat and
measured, to further assess this approach. mass transfer. 4th ed. John Wiley & Sons, Inc., New York,
USA.
Nelder, J.A. and Mead, R. 1965. “A Simplex method for function
300 minimization.” The Computer Journal, 7, 308-313.
Sh predicted
Noordsij, P. and Rotte, J.W. 1967. “Mass transfer coefficients to a
250
rotating and to a vibrating sphere.” Chemical Engineering
Science, 22, 1475-1481.
200
Pinheiro-Torres, A.C. 1998. “Assessment of continuous thermal
150 processing of liquid foods and its impact on process design.”
Ph.D. thesis, Escola Superior de Biotecnologia, Universidade
100 Católica Portuguesa, Porto, Portugal.
Ramaswamy, H.S. and Zareifard, M.R. 2000. “Evaluation of
50 factors influencing tube-flow fluid-to-particle heat transfer
coefficient using a calorimetric technique.” Journal of Food
0 Engineering, 45, 127-138.
0 50 100 150 200 250 300 Sastry, S.K. 1992. “Modelling the continuous sterilisation of
Sh experimental particulate foods.” In Mathematical modelling of food
processing properties, S. Thorne (Eds.). Elsevier Applied
Figure 4: Prediction of Sh for spheres using the Correlation Science, Cambridge, Great Britain, 71-105.
reported by Baptista et al. (1997b) (Equation (5)). Tahoun, M.K. 1986. “Large agarose-lipase beads for the hydrolysis
of triglycerides.” Food Chemistry, 22, 297-303.
Zitoun, K.B. and Sastry, S.K. 1994. “Determination of convective
ACKNOWLEDGEMENTS
heat transfer coefficient between fluid and cubic particles in
continuous tube flow using non-invasive experimental
The author Isabel Abreu is grateful to Fundação Calouste techniques.” Journal of Food Process Engineering, 17, 209-
Gulbenkian for financial support. 228.

REFERENCES AUTHORS BIOGRAPHY

Azevedo, I.C.A; Oliveira, F.A.R. and Drumond, M.C. 1998. “A ISABEL ABREU was born in Porto, Portugal and obtained
study on the accuracy and precision of external mass transfer an undergraduate degree in Chemical Engineering in 1990 at
and diffusion coefficients jointly estimated from pseudo- the Faculty of Engineering, University of Porto, Portugal.
experimental simulated data.” Mathematics and Computing in She has worked as a teaching assistant at the College of
Simulation, 48 (1), 11-22. Biotechnology, Catholic University of Portugal, where she is
Baptista, P.N.; Oliveira, F.A.R.; Oliveira, J.C. and Sastry, S.K. also completing her PhD. Currently she is teaching assistant
1997a. “Dimensionless Analysis of the flow of spherical
at University Fernando Pessoa, Porto, Portugal. Her research
particles in two phase flow in straight tubes.” Journal of Food
Engineering, 31, 125-136. work focus on the application of statistical techniques to the
Baptista, P.N.; Oliveira, F.A.R.; Oliveira, J.C. and Sastry, S.K. optimisation of the estimation of mass transfer parameters.
1997b. “The effect of translational and rotational relative
velocity components on fluid-to-particle heat transfer FERNANDA OLIVEIRA was born in Viana do Castelo,
coefficients in continuous tube flow.” Food Research Portugal, and obtained an undergraduate degree in Chemical
International, 30 (1), 21-27. Engineering in 1985 at the Faculty of Engineering,
Bird, R.B.; Stewart, W.E. and Lightfoot, E.N. 1960. Transport University of Porto, Portugal, and a PhD degree in Food
phenomena. John Wiley & Sons, Inc., New York, USA. Engineering in 1989 at University of Leeds, UK. She
Califano, A.N. and Calvelo, A. 1983. “Heat and Mass Transfer
worked as Associate Professor and Associate Director at the
During the Warm Water Blanching of Potatoes.” Journal of
Food Science, 48, 220-225. College of Biotechnology, Catholic University of Portugal,
Campos, D.T.; Steffe, J.F. and Ofoli, R.Y. 1994. “Statistical as Associate Professor and Vice-President of the Inter-
method to evaluate the critical Reynolds number for University Institute of Macau, China, and as Senior
pseudoplastic fluids in tubes.” Journal of Food Engineering, 23 Consultant and Director of the Portuguese Society for
(1), 21-32. Innovation (Porto, Portugal) before moving to Ireland,
Chilton, T.H. and Colburn, A.P. 1934. “Mass transfer (absorption) where she is currently Professor and Head of the Department
coefficients.” Industrial and Engineering Chemistry, 26 (11), of Process Engineering, at University College Cork.
1183-1187.
 FLUID EXCHANGE FOR PREDICTING CLEANABILITY USING CFD

Bo Boye Busk Jensen

Alan Friis
Food Process Engineering, BioCentrum-DTU
Build. 221, Soltofts Plads
2800 Kgs. Lyngby
Denmark
E-mail: [email protected]

KEYWORDS In uni-axial flow wall shear stress has been identified as a

Computational Fluid Dynamics, Cleaning, Fluid Exchange, measure of the required flow intensity for removal of
Computer Aided design. material from surfaces (Duddridge et al. 1982). However,
these findings cannot be transferred to true three-
ABSTRACT dimensional flow regimes in complex equipment (Jensen et
al. 2001). Only straight pipes and ducts can be regarded as
This work presents basic studies for future use of flow components with parallel flow as most other types of
simulation tools in the area of hygienic design. By equipment contains zones with recirculating, stagnating or
comparing traditional cleaning test results with state of the swirling flow. This leads us to investigate the influence of
art Computational Fluid Dynamics (CFD) simulations, the fluid exchange rate from near-wall regions in the
investigations are made to find a correlation between fluid spherical shaped valve house (a Mix-proof valve) shown in
exchange and cleanability. Previous work has shown that Figure 1 and an upstand (displayed in Figure 2) to see if
the wall shear stress alone is insufficient to predict the prediction of cleanability is possible from knowledge of
cleanability of components involving complex flow fluid exchange or in a combination with wall shear stresses.
phenomenons. This work is an extension, which covers the
exchange of fluid from near wall zones yielding a A
qualitative correlation for cleanability. The exchange rate is Dp,i
predicted from transient CFD simulations by modelling two
fluids with matching material properties. The investigation
I
shows a direction for future work, as identification of a II
quantitative combination of wall shear stress and exchange
rate is needed to provide a powerful tool for predicting B
cleanability using CFD.
C
INTRODUCTION
As legislation on food processing equipment has been Figure 1: Spherical valve house seen from the end (A), the
extended to include hygienic design during the recent top (B) and the mesh seen from the end (C).
decade (EHEDG 1993) manufacturers of food processing
equipment have seen the importance of evaluating the THEORY
hygienic state of equipment parts. Traditionally, comparing In addition to the wall shear stress, others (i.e. Graβhoff
cleaning tests performed on the component of interest with 1980) have postulated but not proven that the fluid
the cleanability of a straight pipe produce a measure of the exchange rate could be correlated to the degree of
hygienic state of the component. This is implemented by cleanability. Exchange of fluid serves as a transport
the European Hygiene and Engineering Design Group mechanism, due to convection of detergent and heat from
(EHEDG) in their test method (EHEDG 1992). It is the main stream to the near wall region. These two
generally accepted that cleaning effects in closed process components weaken the binding between deposit and
equipment is related to the mechanical action of the flow surface, which allows a lower wall shear stress to remove
(Paulsson 1989). The relevant hydrodynamic parameters the deposit (Eginton et al. 1998). In addition the exchange
are not trivial to identify. However, Computational Fluid of fluid also transports the removed deposit away from the
Dynamics (CFD) offers an attractive alternative to the near wall regions enabling decontamination. The needs for
traditional methods for predicting cleanability (Hall 1998). transporting detergent and heat to surfaces are amplified as
Using CFD requires increased knowledge of both surfaces in areas of low fluid exchange very often are
quantitative and qualitative aspects of the hydrodynamic exposed to low wall shear stress because of recirculation
parameters controlling cleaning. Such knowledge offers zones of low fluid velocity.
equipment designers a tool for evaluating and increasing
the hygienic quality of new components prior to prototype
production for experimental testing.
METHODS discolouration in the valve as the valve house otherwise is
cleaned to such a degree that only very small areas of
Comparing cleaning test and CFD simulations on the
discoloured agar are seen. The critical wall shear stress in
components validates the influence of flow parameters on
tests with decreased cleaning time is not affected
removal of deposit from surfaces. In this work cleaning trial
significantly by the shorter CIP time.
results from an upstand geometry performed by Richardson
et al. 2001 (Figure 2) and a sphere shaped valve house
CFD Model of the Upstand
(Figure 1) are used for evaluating the influence of the
The upstand geometry is modelled using quadratic cells in
exchange rate on the cleanability.
an unstructured mesh with integral and arbitrary couplings
between cells with unmatched faces. Both pipes are
generated from the butterfly method (Figure 3).

Figure 2: Cleaning trail result in upstand. Turbulent flow

from left to right. Dark areas are cleaned zones and bright
areas uncleaned zones (Richardson et al. 2001).

Exchange of fluid can be obtained from experimental

methods such as conductive measurements (Graβhoff 1980)
or by discrete sampling in the product stream, both yielding
results on a macro scale level. An alternative is transient Figure 3: Mesh created to discretizise the flow domain of
CFD simulations that provide a tool for prediction and the upstand.
visualization of the exchange of fluid on both micro- and
macro scale levels. The method used for predicting the Exploiting the geometric plane of symmetry, a model with
exchange of fluid involves two steps: 1) A steady state 51.000 cells is created for one half of the upstand. The
solution to predict the flow field in the geometry using validity of the symmetry plane as a symmetry plane for the
appropriate properties of the CIP fluid of a predefined flow was tested (not shown) and shows only little
colour is carried out. 2) A transient simulation is restarted difference in exchange rates. The mesh is created with a
using the steady state solution as initial conditions. A new bow like shaped mesh of 7×7×40 (width×height×axial)
fluid with the same properties as the old one but with a cells in the core region and encircled by a 5×28×40
different colour is used as inlet fluid from t = 0 s and (radial×tangential×axial) celled cylindrical shaped mesh.
onwards. Stepping through time in small time steps gives The cells closest to the walls are refined radial to obtain y+
the exchange of old with new fluid. ~ 5 as recommended for the use of two-layer models
(Figure 3). Cells part of an arbitrary couple is not refined.
Cleaning Tests on the Spherical Shaped Valve The upstand pipe positioned 1.5⋅D to 2.5⋅D downstream of
The hygiene test used here and in the work of Richardson et the inlet consists of 7×7×18 box like and 28×5×18
al. 2000 is adapted from the guidelines of EHEDG 1992. cylindrical mesh and refinement is made of the near wall
The sterilized test piece and a reference pipe are soiled with cells to y+ ~ 5. 10 rows with double axial distance
a mixture of sour milk and Bacillus stearothermophilus compared to the axial distance in the region of interest are
spores. A gentle cleaning procedure is performed and added to the downstream side of the upstand to position the
cleanability is visualized by incubating the reference pipe outlet boundary sufficiently far downstream. Following this
and test piece in Shapton and Hindes agar (Shapton et al. refinement of the 36 rows centred around the upstand are
1963) for 20 hours at 58°C. As a standard the mean inlet axial refined once followed by an additional axial
velocity is set to 1.5 m/s. The results are visualized by a refinement of the 20 rows of the new mesh centre around
change from purple to yellowish agar due to a pH change as the upstand. The latter of these refinements is only
a result of fermentation by germinating spores still present performed on the cells part of the outer region of the mesh
on the surface (example seen in Figure 2). The sensibility of in the pipe. Finally cells on the main pipe being part of an
the test method requires two or more experiments to ensure arbitrary couple with cells on the upstand are refined
that poor hygienic design exists in areas with discoloured tangential once.
agar. Jensen et al. 2001 found a critical wall shear stress for
the above cleaning test on a 2B surface in a Radial Flowcell CFD Model of the Spherical Valve House
to approximately 3 Pa.
The tests performed on the sphere shaped valve house is A 2D patch (Figure 1) which consists of two areas (I and II)
performed with a decreased cleaning time to obtain is created within the outline of the cross-section of the
 Figure 4: Predicted fluid exchange in an up-stand from CFD simulations. Grey fluid is introduced at time 0 s and
gradually replaces the dark fluid. The pictures are shown at t = 0.05 s, 0.2 s, 0.3 s, 0.8 s and 1.05 s (from left to right).

valve. When the patch is revolved 144 times around then the spherical shaped valve could have been the same as for
centre axis region I and II consists of 16×144×40 cells and the upstand without violating the Courant number of 100.
8×144×46 cells respectively. The in- and outlet pipes are However, with the mesh configuration and models chosen it
made as a butterfly mesh extended to 1½⋅Dp,i up- and was impossible to get a converged transient simulation at a
downstream of the valve house (Figure 1). The pipes time step of 0.005 s. A maximum number of 20 corrector
consist of a 10×10×20 celled quadratic mesh in the centre steps are chosen to obtain a physical solution to the fluid
of the pipe encircled by a mesh of 4×40×20 cells. The outer exchange in the spherical shaped valve compared to only 10
layer of cells is refined once to obtain 30 < y+ < 100 (limits in the upstand.
for law of the wall treatment of the near wall region). The 8 The RNG k-ε model with the two-layer model of Norris
layers of cells closest to the valve house is refined once in and Reynolds is used for turbulence modelling, with a
axial direction. Arbitrary couplings between the pipe and dynamic estimation of the cells located within the near-wall
valve house are used at the interface. A total number 80,000 layer (Rodi 1991) to describe the flow in the near wall
cells are obtained by this method. region.
Two-layer models as the one used in the sphere require y+ In both geometries the inlet conditions are described
values in the range of 5. Therefore, radial refinement is using the Dirichtlet method with a fully developed turbulent
made of all the near wall layer cells in the valve house to velocity profile at an average velocity of 1.5 m/s, turbulent
obtain y+ ~ 5. Cells for refinement are identified from kinetic energy of 5% and turbulent length scale of 1/10 of
simulations using previous meshes. Near wall flow in the the inlet pipe diameter. The outlet is described as a pressure
in- and outlet pipe is modelled by use of the “law of the outlet with constant pressure on the boundary. Walls are
wall” method that requires y+ > 30 and as a consequence a prescribed with non-slip conditions and as hydraulic
less number of cells. The total number of cells in the model smooth.
with two-layer treatment of the near wall flow is 180,000.
RESULTS
SIMULATION SET-UP Fluid Exchange in Upstand
The simulations are performed using the finite volume In the short upstand investigated the flow is more
based code Star-CD v. 3.100b. The Navier-Stokes complicated than in a longer upstand i.e. H/D = 0.5. A
equations set-up transport equations for momentum, energy recirculation zone is seen in the longer upstand on a plane
and scalars by use of appropriate parameters (for more coinciding with the symmetry plane. As the space for this
information see i.e. Versteeg et al. 1995). Spatial and recirculation zone is reduced in the short upstand it is
temporal discretisation of the transport equations are suppressed and almost non-existing. Instead a recirculation
performed using the second-order accurate Self Filtered zone in the plane perpendicular to the axial direction is
Central Difference and Crank-Nicholson schemes formed in the entire upstand and a recirculation zone is also
respectively. Casey et al. 2000 showed the importance of seen in the horizontal plane. The effect of the non-existing
choosing second-order discretisation for both dimensions of recirculation zone is seen in the flow downstream of the
progress as numerical diffusion is significant in lower order upstand as it is pushed away from the wall producing a
schemes. zone of low exchange rate (Figure 4). This is not seen in the
A time step of 0.005 s and 0.0005 s for the upstand and long upstand, as the horizontal recirculation zone does not
spherical shaped valve respectively is chosen to obtain influence the main flow as much as in the short upstand. In
Courant numbers below 100 to capture features of the flow. the downstream zone exchange of fluid is approximately 3
If the Courant number is too high, the risk of information and in the upstand 6 times slower than in the part affected
convected across cells in the domain is large and thereby by the bulk flow. Investigations of the magnitude of the
information from one cell is likely to be transferred across wall shear stress (not presented) shows wall shear stress
neighbouring cells. The time step used in the simulations of
 A B C

Figure 5: Vector plot of flow in spherical shaped valve seen from the inlet towards the outlet (A). The plane for
visualisation is located perpendicular to the axial direction of the pipes and in the centre of the valve house. Fluid
exchange at t = 0.13 (B) and 0.23 s (C) are shown in the valve with flow from left to right.

below the critical value of 3 Pa in the upstand and in a thin Cleaning of Sphere Shaped Valve
streak downstream of the upstand. The zone of low
Figure 6 illustrates cleaning results on one side of the
exchange rate downstream of the upstand extents to more
sphere from a single cleaning procedure with decreased
than 1.5⋅D downstream of the upstand. cleaning time. A common feature of the results is that
discolouring is present just next to the inlet of the sphere.
Fluid Exchange in Sphere Shaped Valve Further downstream the discoloured zone decreases in
width towards the outlet. In one experiment the problem
The flow pattern in a spherical valve house with in- and
zone was present slightly below and in another slightly
outlet located as shown in Figure 1 is very complex because above the equator.
of transition from a small circular cross sectional area to a
larger of non-circular shape produces two swirl zones on
top of each other in the zone seen closest to the outer
surfaces (Figure 5). Near the cone the flow is almost uni-
directional towards the outlet. These swirl zones have a
large influence on the exchange of fluid from areas close to Equator
the outer walls. As new fluid enters the sphere most of it
flows directly through the centre of the valve and towards
the outlet pipe. Within 0.23 s most of the old fluid is
exchanged in the main stream. As time progresses the old
fluid is exchanged in the remaining part of the sphere. The
zones with lowest exchange rate are near the equator of the
sphere (light areas in figure 5 (C)) and especially in the area
just next to the inlet. The total exchange takes 0.95 s. Wall
shear stress is below 3 Pa in large zones on the surfaces Figure 6: Illustration of results from cleaning trails with
(Figure 6). Only near the outlet pipe, close to the top and reduced cleaning time compared to the standard
bottom and near the inlet, close to the bottom are wall shear EHEDG test method. Black is uncleaned zones. Dotted
stresses above 3 Pa (looking only at the outer surface of the lines mark the zones predicted as cleanable from a
valve). critical wall shear stress of 2.5 – 3 Pa.
Detailed information on the hygienic state in other parts of
Cleaning of Upstand a sphere shaped valve house is under preparation for
The potential problems of cleanability in an upstand of H/D publication.
= 0.16 with a mean inlet velocity of 1.5 m/s is seen in the
upstand itself as expected (Figure 2). Upstream of the DISCUSSION
upstand the main pipe is cleaned to a degree where Comparing cleaning results, CFD simulations and the
discoloration of the agar is not seen. On the contrary a great critical wall shear stress shows that wall shear stress alone
part of the pipe surface downstream and on the same side of is insufficient for approximating problematic areas in the
the upstand is with attached spores. This zone is located upstand and the spherical shaped valve used in this study.
from the upstand and until a downstream position of Areas with low wall shear stresses are cleaned (valve) and
approximately 1.5⋅D. areas of high wall shear stress are not cleaned (upstand).
In the upstand zones with low exchange rate is shown as
problematic areas. Especially in the downstream shadow
zone of the upstand problems with hygiene was seen. This Conditions." Biotechnology and Bioengineering, 24,
was not expected from the wall shear stress but could be 153-164.
foreseen by investigations of the exchange rate. In this area Eginton, P.J., J. Holah, D.G. Allison, P.S. Handley, and P.
the exchange rate is lower than in the main stream, while Gilbert. 1998. "Changes in the Strength of Attachment
other areas in the upstand with the same wall shear stress as of Micro-organisms to Surfaces Following Treatment
in the downstream problem zone was totally cleaned. with Disinfectants and Cleansing Agents." Letters in
In the valve areas with wall shear stress below the critical Applied Microbiology, 27, 101-105.
value of 3 Pa is cleaned. The expected area of hygienic EHEDG. 1992. "A Method for Assessing the In-place
problems according to the wall shear stress is outlined in Cleanability of Food-processing Equipment." Trends in
Figure 6 as the dotted lines and shows a large area around Food Science & Technology, 3, 325-328.
the equator expected to be uncleaned. Comparing Figure 6 EHEDG. 1993. "The EC Machinery Directive and Food
to predicted exchange (Figure 5) shows that the zone of low Processing Equipment." Trends in Food Science &
exchange rate is decreased into a smaller and smaller zone Technology, 4, 153-154.
just above the equator as time progresses. The areas cleaned Graβhoff, A.V. 1980. "Untersuchungen zum
having low wall shear stress, has a relative high exchange Stromungsverhalten von Flussigkeiten in zylindrischen
rate compared to the exchange in i.e. the upstand and other Totraumen von Rorhleitungssystemen." Die Kieler
places in the valve with the same wall shear stress. From Milchwirtshaftlichen Forschungsberichte, 32, 273-298.
this, zones with low wall shear stress but clean can be Hall, J. 1998. "Computational Fluid Dynamics: A Tool for
explained by the relative low exchange time (below 0.3 s). Hygienic Design." Fouling and Cleaning in Food
Processing,
Jensen, B.B.B. and A. Friis. 2001."Flow Parameters in
CONCLUSION
Hygienic Design." in The 32nd R3 Nordic Symposium,
Comparing the EHEDG tests to simulation results show Reklamproducenterne, Enköping, Sweden, 347-353.
that wall shear stress is insufficient to predict the outcome Paulsson, B.-O. 1989. Removal of Wall Deposits in
of the hygiene tests. Areas with wall shear stress lower than Turbulent Pipe Flows, Lund University - Department of
the critical value found experimentally using a Radial Food Engineering,
Flowcell could be cleaned and areas with higher wall shear Richardson, P.S., R.M. George, and R.D. Thorn.
stress not cleaned. It is shown that this can be explained 2001."Application of Computational Fluid Dynamics
from the rate of fluid exchange. With the gained knowledge Simulation to the Modelling of Cleanability of Food
of hydrodynamic cleaning effect it is possible to Processing Equipment." in Proceedings of the Eighth
supplement i.e. the EHEDG test with CFD simulation. This International Congress on engineering and Food
allows manufacturers of food processing equipment to test (ICEF8), Technomic Publishing co., Lancaster
different designs using CFD, which could lead to improved Pennsylvania, 1854-1858.
quality of the hygienic state of closed equipment in the Rodi, W. 1991. "Experience with Two-Layer Models
future. To accomplish this, future work is needed to identify Combining the k-e Model with a One-equation Model
quantitative values for a combination of wall shear stress Near the Wall." AIAA-91-0216, 1-12.
and fluid exchange to be present to achieve a satisfactory Shapton, D.A. and W.R. Hindes. 1963. "The
EHEDG test results. Standardisation of a spore Count Technique." Chemistry
and Industry, 9, 230-234.
ACKNOWLEDGMENTS Versteeg, H.K. and W. Malalasekera. 1995. An Introduction
The authors whish to thank Mr. Jens Folkmar Andersen and to Computational Fluid Dynamics - The finite Volume
Mr. Preben Esbensen from Alfa Laval LKM a/s, Kolding, Method, Longman Scientific & Technical, Essex,
England
Denmark, Mr. Erik-Ole Jensen from Arla Foods Amba,
Brabrand, Denmark and the Biotechnological Institute,
Kolding, Denmark for their participations in fruitful AUTHOR BIOGRAPHY
discussions. The Centre for Hygienic Design and Cleaning BO BOYE BUSK JENSEN was born in Hou, Denmark
of Food Process Equipment is acknowledged for the and went to the Aalborg University, Department of Energy
financial support of this work. Technology, where he studied Mechanical Engineering in
the area of energy consuming/producing machinery and
REFERENCES fluid dynamics. He obtained his degree in 1998 after which
he worked as a Research Assistant using Computational
Casey, M. and T. Wintergerste. 2000. ERCOFTAC Special
Interest Group on "Quality and Trust in Industrial CFD Fluid Dynamic at the same department as he graduated
- Best Practice Guide", ERCOFTAC, from. In 1999 he enrolled as a Ph.D. Student at
BioCentrum-DTU, Technical University of Denmark
www.ercoftac.org, Hampshire
Duddridge, J.E., C.A. Kent, and J.F. Laws. 1982. "Effect of working in the field of hygienic design using
Surface shear stress on the Attachment of Pseudomonas Computational Fluid Dynamics, which is still ongoing.
fluorescence to Stainless Steel under Defined Flow
 SLIDING MESH METHOD TO SIMULATE FLOW OF HIGHLY VISCOUS FLUID IN
STIRRED TANK AGITATED WITH HELICAL RIBBON IMPELLER.
Guillaume Delaplace Claude Torrez
Romuald Guérin Christophe André
Jean-Claude Leuliet
Institut National de la Recherche Agronomique Hautes Etudes Industrielles (H.E.I.)
Laboratoire de Génie des Procédés et Technologie Alimentaires Laboratoire de Génie des Procédés
369 rue Jules Guesde, B.P. 39, 13 rue de Toul,
59651 Villeneuve d’Ascq Cédex, France 59046 Lille Cedex, France
E-mail: [email protected] E-mail: [email protected]

data can be measured and consequently it is impossible to

obtain the velocity profile at whole stage during a cycle (a
KEYWORDS cycle is the time taken for impeller to rotate one revolution).
CFD, helical Ribbon agitator, laminar, particle tracks, sliding This lack of information about the velocity field is very
mesh annoying since a detailed characterization of the three
dimensional flow pattern inside the vessel is of vital
ABSTRACT importance to gain a better understanding of the state of the
In this paper, steady and unsteady particle tracks into a stirred flow of the fluids and of the mechanisms responsible for
tank was conducted by means of a commercial available fluid homogenization and transport processes.
dynamic (CFD) software package (FLUENT). In this area, the perspective offered by advanced computational
A rounded bottom vessel equipped with an atypical helical fluid dynamics (CFD) seems very interesting. The use of CFD
ribbon impeller was used as mixing equipment and the viscous for the investigation of mixing flow phenomena is not new and
media was highly viscous fluids. demonstrable benefits gained from its application have been
The reliability of the numerical procedure was demonstrated obtained for classical agitators such as Rushton turbine
on the basis of a comparison of the numerical results with (Ranade et Joshi 1990; Luo et al. 1994; Jaworski et al. 1997;
tracer experiments. Torrez and André 1998; Micale et al. 1999; Ranade et al.
It is shown that using a sliding mesh technique, computational 2001) and pitched blade turbine (Armenante et al. 1994;
tools are able to predict accurately trajectory of a particle in Jaworski et al. 2001; Aubin et al. 2001) in the turbulent flow
the vessel. regime. Development of unstructured and hybrid meshing
Even if CPU times values are still high, these approaches are techniques has allowed the representation of more complicated
likely to have a significant impact in the close future to get an geometries often found in laminar flow mixing [anchor (
idea of mixing efficiency of various mixing systems. Bertrand and al. 1996; Delaplace et al. 2000b), helical ribbon
agitators with or without screw (Tanguy et al. 1992
Kaminoyama and Kamiwano 1994; Tanguy et al. 1997; De la
INTRODUCTION Villéon et al. 1998; Delaplace et al. 2000c), helical screw
impeller (Aubin et al. 2000)
Helical ribbon agitators have been widely used for mixing Although the numerical knowledge of velocity field under
liquids of high viscosity. There are many experimental studies laminar regime have represented an important step forward,
providing basic information of helical ribbon impeller these works is a shortcoming to investigate homogenisation
concerning power consumption (Delaplace et Leuliet 2000) of process in agitated vessel.
agitators and /or mixing characteristics (Delaplace et al. 2000a) Indeed, most of the particle tracks have been performed under
Unfortunately there were only few experimental works devoted steady flow. This means that the velocity field used by the
to a detailed description of the hydrodynamics inside the vessel software to compute the trajectory of a fluid element was not
(Bourne et Butler 1969; Carreau et al. 1976; Soliman 1985) time-dependent but steady. Consequently, it is not possible to
From a practical point of view, this can be explained easily predict exactly particle track in the tank since fluctuations of
since there are not a lot of experimental techniques available velocity relative to the revolution of the impeller are
for the local hydrodynamic characterization of flow pattern disregarded. These modelling simplifications can be explained
induced by such complex mixing systems. easily: the design of these agitators are not trivial and therefore
On one hand, Pitot tubes or Hot-wire Anemometry are correctly capturing the characteristic flow features require large
intrusive techniques (Mavros 2001). amount of CPU times.
On the other hand, Laser Doppler Velocimetry (LDV) and The aim of this paper is to show that, using a sliding mesh
Particle Image Velocimetry (PIV), required the need for the method and a commercial code (FLUENT 5.4 then FLUENT
fluid to be transparent. This is seldom the case in foodstuff 6.0), succeeds in predicting the real trajectory of a fluid
industry. Moreover, even if real fluids are transparent, these element subjected to laminar flow in a three dimensional
techniques required a penetrating laser beam which are limited mixing tank.
in such high viscosity fluids due to light diffusion (Aubin et al. The reliability of the numerical unsteady particle tracks was
2000). Other inherent difficulties also explained the few compared with flow pattern visualization obtained by tracer
experimental data. For instance, when blades of the close experiments. For that, a colour indicator was injected in the
clearance impeller going by the laser beam, no experimental transparent liquid and its flow was captured by a video camera.
MATERIALS AND METHODS The sliding mesh modelling approach involves two mesh
regions, one (inner grid) attached to a rotating geometry
Mixing system (impeller) and one (outer grid) attached to stationary
A sketch of the mixing system investigated is shown in Figure boundaries of the flow (tank walls). During the calculation, the
1. The dimensions of the mixing system are also reported in cell zones slide (rotate) relative to one another along the grid
Figure 1. interface in discrete steps. To compute the interface flux, the
This non standard helical impeller has two particularities: intersection between the interface zones is determined at each
firstly, it is equipped with an anchor at the bottom in order to time step. In this study, the time step was equal to 0. 2 seconds.
avoid the formation of a stagnant zone; secondly the size of the
pitch ratio p/d (p/d=1.7) is greater than with a typical double A view of the outer and inner grids used in this work are shown
helical ribbon mixer (0.5 <p/d<1.2). During all the in Figure 2. The grid interface is located at the middle of the
experiments, the volume of agitated liquid was kept constant wall clearance.
and equal to 34 .10-3 m3 which corresponds to a liquid depth,
HL of 0.409 m.

Figure 1. Picture of the agitator (PARAVISC®, Ekato) and

scheme of the mixing system used. (vessel diameter, t =0.346
m, impeller diameter, d =0.320 m; blades width, w =0.032 m,
impeller pitch, p = 0.560 m; impeller height, L= 0.340 m, tank Figure 2: View of the grids used to discretize the
height, Hc = 0.443 m, Liquid height HL = 0.409 m). domain.(left) Inner grid - (right) Outer grid

Fluids Note that to obtain a reasonable mesh size, it was decided to

The Newtonian fluids used for tracer experiments were various mesh the concentric zone around the shaft with hexahedral
mixtures of glucose syrup with various water contents. The cells while regions close to the helical blades of the impeller
Rheological properties and densities of these fluids were consist of tetrahedral cells. (Figure 3)
obtained at the same temperature as that encountered in the
mixing equipment.
Numerical simulation of steady and unsteady particle tracks
were carried out with Newtonian liquids. For these CFD
simulations, viscosity and density of the agitated fluid were
respectively set equal to 2.82 Pa.s and to 1380 kg/m3.

Numerical simulations of steady and unsteady particle

tracks
In this work, the trajectory of a fluid element subjected to
laminar flow in an industrial agitated vessel equipped with an
atypical helical ribbon impeller was computed numerically
(particle track).
The trajectory of a fluid element has been computed using two
approaches. The first one used a steady velocity field obtained
Figure 3: View of the cell types containing in the inner grids.
by a rotating reference frame technique. For this case, particle
track neglect velocity fluctuations due to the revolutions of the
Actually, an unstructured mesh of 218,164 cells has been
agitator blade. Consequently, the motion of the fluid element
required to represent the whole mixing system (agitator +
is only an approximation of real trajectory.
tank).
The second one used an unsteady velocity field obtained by a
Note that the geometry of the mixing system used for
sliding mesh technique.
simulation is very similar to the experimental equipment
(figures 4 and 5). The rounded bottom of the vessel has been
The numerical tools and procedure solution used by the
accurately duplicated (figure 5). Only additional parts of the
commercial CFD finite volume software (FLUENT) to
impeller used to fix scrapers haven’t been meshed
simulate the velocity field inside the vessel using a rotating
reference have already been described in details (Delaplace et
al. 2001b , 2001c). and will not be repeated in details here.
Only specific details about the sliding mesh modelling options
will be given.
 For this test, various colors refers to various residence time. In
this figure, six particles at various locations in the vessel were
launched to obtain a view of the flow pattern in the vessel.
Theses results agree in a certain extent with the experimental
observations of tracer particles under same flow conditions.
Indeed, rotating counter clockwise direction, it has been
observed that tracer particles go down around the stirring shaft
with an helical trajectory and subsequently go up along the
vessel wall.
However, analysis of time to achieve a close loop (axial
circulation time) are disagree with previous studies (Delaplace
Figure 4. Mesh of the agitator [PARAVISC®, Ekato]. et al. 2000c). This may be explained by the fact that the
The initial position of the fluid element when using sliding velocity field used by the software to compute the trajectory of
mesh technique is also reported. the fluid element are steady and consequently do not take into
account velocity fluctuations due to impeller revolutions.
Unfortunately these data are required for accurate computation
of fluid trajectory. Consequently, important errors are
generated. For example, we can detect recycle zones which are
not observed experimentally by recording flow of freely
suspended particles. Consequently, such approach fails to
predict precisely the trajectory of fluid element in the tank.

Unsteady particle tracks

An example of a particle trajectory obtained using sliding
Figure 5. Mesh of the rounded bottom vessel studied. mesh technique for the close clearance agitator studied is
shown on Figure 7 (N=35 rev/min).
The boundary conditions at the impeller shaft and at the vessel
walls were those derived assuming the no-slip condition. At the
free surface, the boundary conditions were modelling by
requiring that there is no normal velocity and zero normal
gradient for all variables.

Resolution of the algebraic equations was carried out using the

Semi-Implicit algorithm Pressure Linked equation (SIMPLE)
with a first order upwind discretisation scheme.

RESULTS AND DISCUSSION

Steady particle tracks

An example of trajectories results obtained for steady particle
tracks with the close clearance agitator studied is presented in
Figure 6 (N=35 rev/min, µ =2.8 Pa.s). Figure 7: Particle trajectories obtained numerically when
sliding mesh technique is performed.

Trials 1 and 2 refers to a similar injection which has been

duplicated with two different computers. For the two
simulations, tracer injection has been launched at the upper part
of the vessel as shown in Figure 4.
As expected,, for the two computers, the trajectory of the fluid
element are similar.
Note that on figure 7, only 4 seconds of the real mixing
process has been simulated. This is explained by the large
amount of CPU times required to achieve such simulations.
For instance, at the beginning (trial 1), the computations were
performed on a Sillicon Graphics computer with 300 MHz
Octane R12000 processor and required more than 20 days of
CPU times to simulate only 4 seconds of the real mixing
process. For a Hp workstation X4000 computer with a 1.8 GHz
Figure 6: Particle traces (colored by residence time(seconds)) P4 Xeon processor , CPU times decrease. 48 hours of CPU
obtained numerically when steady flow simulation is
performed.
times is required to simulate 4 seconds of real homogenisation BIBLIOGRAPHY
process (trial 2).
Armenante P.M., C-C. Chou and R.R. Hemrajani 1994.
Tracer experiments with a coloured fluid show that the "Comparison Of Experimental And Numerical Fluid
computation of the fluid element trajectory is in agreement Velocity Distribution Profiles In A Unbaffled Mixing
with those observed experimentally as indicated on figure 8. Vessel Provided With A Pitched-Blade Turbine." I.
Chem.E.Symp. Series n°136, 349-362.
Figure 8 was obtained when a colour indicator was injected in Aubin J, Naude I, Bertrand J., Xuereb C. 2000. "Blending Of
the transparent liquid and its flow was captured by a video Newtonian And Shear-Thinning Fluid In A Tank Stirred
camera. It appears that the red dye follow an helical trajectory With A Helical Screw Agitator." Trans.I.Chem.E., Vol 78,
along the shaft. This is also the case for numerical results Part A, 1105-1113.
(Figure 7). Aubin J., Mavros P., Fletcher D.F., Xuereb C., Bertrand J.
2001. "Effect Of Axial Agitator Configuration (Up
Pumping, Down Pumping, Reverse Rotation) On Flow
Patterns Generated In Stirred Vessels." Proc.
Trans.I.Chem.E., Part A, Chem. Eng. Res. Des. 79, A8,
845-856.
Bertrand F., Tanguy P.A., Brito-de La Fuente E. 1996. "A New
Perspective For The Mixing Of Yield Stress Fluids With
Anchor Impellers." Chem.Eng.Res. and Des., Vol 29, N°1,
51-58
Bourne J.R., Butler H. 1969. "An Analysis Of The Flow
Produced By Helical Ribbon Impellers." Trans. Instn.
Chem. Engrs., 47, 11-17.
Carreau P.J., Patterson I., Yap C.Y. 1976. "Mixing Of
Viscoelastic Fluids With Helical-Ribbon Agitators: Mixing
Time And Flow Patterns." The Can. J. of Chem. Eng. 54,
135-142.
Figure 8: Flow visualization in stirred vessel using a red colour De la Villéon J., Bertrand F., Tanguy P.A., Labrie R., Bousquet
tracer with similar rheological and physical properties than the J., Lebouvier D. 1998. "Numerical Investigation Of Mixing
viscous media. Efficiency Of Helical Ribbons." AIChE Journal, Vol 44,
N°4, 972-977.
Moreover, although it is two early to announce it Delaplace G., J.C. Leuliet 2000. "Power Consumption of
(computations are not yet finished), it seems that there are less Helical Ribbon Impellers in Highly Viscous Liquids – A
scatters between experimental and numerical particle tracks Review. " Entropie, 227, 10-21.
when unsteady velocity field is computed. Delaplace G.; J.C. Leuliet and V. Relandeau. 2000a.
Indeed preliminary estimation of numerical average times "Circulation and mixing times for helical ribbon impellers -
taken by a fluid element to cross a given reference plane two review and experiments." Experiments in Fluids; 28-2; 170-
times in the same direction (axial circulation time) seems in 182.
better concordance with experimental data. Delaplace G. , Torrez C., Gradeck M., Leuliet J.C. and C.
André 2000b. "Flow Of Newtonian And Non-Newtonian
Fluids In An Agitated Vessel Equipped With A Non-
CONCLUSION Standard Anchor Impeller " Proc. 10th European
Conference on Mixing, Delft, The Netherlands, H.E.A. Van
In this work, first a three-dimensional resolution of the flow den Akker and J.J. Derksen (editors), Elsevier Science B.V.
equations in a mixing vessel has been conducted by CFD , 321-328.
approach using two approaches : a sliding mesh and a rotating Delaplace G., Torrez C., Andre C., Leuliet J.-C., Fillaudeau L
reference frame technique. 2000c. "CFD Simulation of Foodstuff Flows in an Agitated
Then trajectory of a fluid element obtained by the two Vessel. " Proc. 1st International Conference on Simulation
approaches (steady and unsteady flow field) have been in Food and Bio Industries, Nantes, France, Daniel Thiel
computed and compared to those obtained by tracer (editors), Publication of the Society for Computer
experiments. Simulation international: 289-291.
It has been shown that sliding mesh method succeeds in Delaplace G. , Torrez C., C. André, J.C. Leuliet 2001a. "Tracer
predicting the trajectory of a fluid element subjected to laminar Experiments – a way to validate computational fluid
flow in a three dimensional mixing tank. On the contrary, dynamic simulations in an agitated vessel" Proc. First
Steady particle tracks are not adapted for this close clearance International Congress and Second French Congress on
impeller. tracers and tracing methods, Nancy, Vol 15, N°79 , 77-84.
It is now clear that the science of mixing will benefit a lot Delaplace G. , Torrez C., J.C. Leuliet, Belaubre N, C André
from numerical simulation in a close future. Nevertheless, at 2001b. "Experimental and CFD simulation of heat transfer to
the state of the art, it is indispensable that the amount of CPU highly viscous fluids in an agitated vessel equipped with a non
times required for these simulations decreases. standard helical ribbon impeller " Chem. Eng. Res. and Des.
Vol 79: A8 , 2001, 927-937
Jaworski Z., Dyster K.N., Moore I.P.T. Nienow
A.W.,Wyszynski M.L. 1997. "The Use of Angle Resolved
LDA Data To Compare Two Differential Turbulence
Models Applied To Sliding Mesh Cfd Flow Simulations In
A Stirred Tank." Récents Progrès En Génie Des Procédés,
Vol 11, N°51, 187-194
Jaworski Z., Dyster K.N., Nienow A.W. 2001. "The Effect of
Size, Location And Pumping Direction Of Pitched Blade
Turbine Impellers On Flow Patterns LDA Measurements
And CFD Predictions." TransIChemE, Part A, Chem. Eng.
Res. Des. 79, A8,887-894.
Kaminoyama M.and M. Kamiwano 1994. "Numerical analysis
of reaction process of highly viscous liquids in a stirred
vessel equipped with a double helical ribbon impeller."
IChemE.Symp. Series n°136, 541
Luo J.Y. R.I. Issa and A.D. Cosman 1994. "Prediction Of
Impeller Induced Flows In Mixing Vessel Using Multiple
Frames Of Reference." IChemE.Symp. Series n°136, 549-
556
Mavros. 2001. "Flow Visualization In Stirred Vessels A review
of experimental Techniques, Trans.I.Chem.E., Vol 79, Part
A, 113-128.
Micale G., Brucato A., Grisafi F., Ciofalo M. 1999. "Prediction
of flow fields in a dual-Impeller Stirred vessel. " AIChE J.,
vol 45, N°3, 446-464
Ranade V.V., Joshi J.B. 1990. "Flow Generated By A Disc
Turbine: Part II Mathematical Modelling And Comparison
With Experimental Data." TransIChemE, Part A, Chem.
Eng. Res. Des 68, A1, 34-50.
Ranade V.V., Perrard M., Le Sauze N., Xuereb C., Bertrand J.
2001. "Influence of Gas Flow Rate on Structure of Trailing
Vortices of Rushton Turbine: PIV Measurements and CFD
Simulations." TransIChemE, Part A, Chem. Eng. Res. Des.
79, A8,957-964.
Soliman M. G. 1985. "Agitation de fluides visqueux
pseudoplastiques par un double ruban hélicoïdal. " Thèse de
Docteur Ingénieur, INP Toulouse.
Tanguy P.A., Lacroix R., Bertrand F., Choplin L.,Brito-De La
Fuente E. 1992. "Mixing Of Rheogically Complex Fluids
With An Helical Ribbon-Screw Impeller: Experimental
And 3D Numerical Studies.", AIChE Symp. Ser. N°286,
Process Mixing-Chemical and biochemicals Applications,
G. Tatterson and R. Calabrese eds, 33-39
Tanguy P.A., Thibault F, Brito-De La Fuente E., Espinosa-
Solares T., Tecante A. 1997. "Mixing Performanceinduced
By Coaxial Flat Blade –Helical Ribbon Impellers Rotating
At Different Speeds " Chem. Eng. Sci., Vol 52, N°11, 1733-
1741.
Torrez C. and André C 1998. "Power Consumption of a
Rushton Turbine Mixing Viscous Newtonian and shear-
thinning fluids: Comparison between experimental and
numerical results" Chem. Eng. Technol. 21, 7, 15-20.

ACKNOWLEDGEMENTS
The authors are grateful to C. DELAHAYE and F. GAIGHER
for their valuable work in obtaining some of the data
presented here.
SIMULATION
 A DSS for forecasting the Performance Indicators in Food Industries
Mahdi Movahedkhah Daniel Thiel
ENITIAA, Nantes, LARGECIA E.N.I.T.I.A.A. Nantes, LARGECIA
et Ecole Centrale de Paris Rue de la Géraudière - BP 82225
Laboratoire de Productique – Logistique 44322 Nantes Cedex 3 (France)
E-Mail : [email protected] E-Mail : [email protected]

SUMMARY: industrial firms. The performance cannot be only based on

quantitative data, but also on more qualitative variables like
The objective of this paper is to present the principle of an for example, the delays and the production cycles times.
Interactive Decision Support System (DSS) applied to the According to many authors, this requires a "global" and
management and the control of food production systems. A "multi-criteria" approach, which is reflected as well on the
preliminary survey based on a systemic approach, enabled us performance control methods as on its evaluation tools
to define the causal links between various industrial (Ecosip, 1990) , (Lorino, 1991), (Kaplan & Norton, 1992).
performance indicators. This first work permits us to Thanks to an intensive survey in different firms, our research
conceptualize this empirical knowledge by an influence brought in light various explanatory factors of the
diagram and by a generic simulation model based on performance and on the causal relationships which link the
Forrester’s System Dynamics Theory. The DSS which we processes and the actions to the obtained results. We firstly
propose, is composed by this continuous simulation model present the results of this survey. From these empirical bases,
which extracts the data from a spreadsheet (or an EIS we formalized a causal diagram showing the
Executive Information System) which directly extracts its interrelationships between these performance indicators.
information from an ERP software (Enterprise Resource This generic representation allowed us to build a continuous
Planning). simulation model (flows-stocks model, cf. Forrester) which
The objective of this DSS is to forecast the evolution of the has been the heard of our DSS.
performance indicators and its originality, comparing to
traditional statistical methods based on past chronological 2. THE RESULTS OF THE SURVEY
series, is to take into account the causal relationships
between these indicators. A first interview-based inquiry was necessary to make easier
the construction of a questionnaire concerning the structure
KEYWORDS: Performance Indicators, DSS, continuous of the production systems, their problematic at different time
simulation, System Dynamics, fresh food products. scales, their inputs-outputs and their behavior faced to
uncertainty. This investigation was based on the system
1. INTRODUCTION thinking paradigm (according to Forrester 1958 ; 1961 ; 1994
; Richardson 1984 ; 1992). A cybernetic vision highlighted
During the twenty last years, different transformations in the the regulation mechanisms of some industrial organizations
principles of the production organization in food industries, faced to different fluctuations of their essential components
have been observed (cf. Treillon et al., 1991 ; 1995 ; Nakhla, (cf. their performance indicators). Nineteen firms have been
1994 ; Lecomte et al.,1996). This mutation was vital for a asked. This study was limited to “simple” mechanisms easy
large number of companies and can be explained by the to identify and was based on 1292 answers given by these 19
evolution of the markets and by an instability of the firms and corresponding to the decisions faced to
environment. The new production systems have often to dysfunctions of the food production systems.
react to different disturbances either external or internal to
the firm (Jacot et al., 1996). To implement new control Representation of the cybernetic mechanisms
systems, there have to adapt there performance measurement
and forecasting tools. Our research focuses on this We qualitatively identified the most significant Indicators of
problematic and aims to propose the structure of a new Performance (IPs). From the different answers given by the
decision support system (DSS) based on industrial managers, we analyzed the existence of some cybernetics
performance indicators. In this research field, we have control loops, we conceptualized an original knowledge of
observed that no synthetic performance indicators have been the industrial organizations causal working. In our sample,
find out according to these new production systems. The we have had three groups of food industries:
industrial performance concept is still often limited to a
quantitative evaluation of the industrial productivity. By an Group 1: dairy products manufacturers (like yogurts,
accountancy point of view, the performance evaluation is cheeses,…)
thus often reduced to the control of individual performances Group 2: elaborated products manufacturers (like flat
of each resource of the firm. But we know today that other cooked, sausages,…)
forms of performance were imposed in response to the Group 3: poultry cutting industries.
evolutions of the technology and the environment of the
After an aggregation of the variables and a qualitative preliminary negotiation between the sales managers and the
analysis of the loops, we have chosen the following variables scheduling managers.
as performance indicators:
In case of too longs manufacturing lead times (IP2a).
Indicators of Performance : The problem of the too long manufacturing delays for the
IP 1: Sales fluctuations group 2, can be solved by purchasing new manufacturing
IP 2: Quality of service or service rate tools.
IP 3: Internal Reliability rate
IP 4: External Reliability rate In case of non-quality problems and customers complaints
IP 5: Information System Efficiency (IP2b).
IP 6: Inventory Level rate Customers complaints due to quality problems, can be
IP 7: Real manufacturing costs reduced in short/middle terms, by increasing employees
motivation and staff training. In the short term, a statistical
Now, we will describe the causal relationships which quality control is often carried out. The groups 1 and 3
formalize the production systems behaviors observed in case propose to improve the reliability of their production
of IP variations. equipments.

Consequences of customer orders fluctuations (IP1) Consequences of the variations of the internal reliability
of production equipments (IP3)
In case of very urgent customer orders
In the case of sales increasing in a sudden and instantaneous In case of fluctuating production capacities, productivity
way (pulses), the systems are controlled in short term, by a decreasing and absenteeism increasing (IP3a).
re-scheduling of the orders in process and by overtime or The diminution of the production capacities can be caused by
temporary employment. productivity decrease, an absenteeism increase or frequent
stoppages of work. In these cases, an active regulation of the
In case of a sudden and permanent sales increase systems can be carried out by a rigorous analytic cost
In the case of sales increasing in a sudden and permanent accounting, by automation of the production equipment, by
way (steps), the systems of the three groups, are controlled in the " re-motivation " and the responsibility of the personnel,
the middle term, by temporary employment. Moreover, the by the re-organization of the tasks or by the implementation
group 1 uses frequently overtime, the group 2 personnel of flexible schedules in the middle term like by tighter and
hiring and the group 3 personnel temporary employment. rigorous controls of the short-term activity.
In case of work stoppages caused by strikes (IP3b).The
In case of occasional and very significant decreasing of sales problem of the stoppages of work in the case of strike for the
In the case of occasional declining of the customer orders group 1, can be regularized by external subcontracting.
which implies a loading decrease, an active regulation (in the In case of frequent dysfunctions of the production
short term) is carried out by production anticipation. equipments (IP3c), the machines stoppages and too long
set-up times (IP3b).
In case of sudden decreasing of sales Frequent dysfunctions in production due to machines
In the case of sudden decreasing of the sales, only the breakdowns, are controlled in the middle term in an active
response of the group 1 is significant : the solution is the way, by preventive maintenance and personnel training.
regulation by production anticipation. Moreover for the groups 1 and 3, the long-term action will
be done by new machines investment, automation or
In case of seasonal variations of the sales specialization of the factories by products. Nevertheless, the
Seasonal variations in sales often involve the need of group 1 has in short-term, tendency to gather several
flexibility by following the sales profile. customer of production orders.

In case of random variations of the sales Consequences of the variations of the external reliability
A random variation of the sales can involve flexibility by rate (IP4)
following the sales pattern. For the group 1, another solution
consists in scheduling by priorities. In case of deliveries problems with providers
The implementation of an ERP or a JIT system can imply a
Consequences of the Quality of service variations (IP2) reduction of the missed items in production. If there are too
much missed items, some companies have diversified their
In case of non- delivered orders and frequent customers providers or have increased their stock levels.
complaints (IP2a).
The customers complaints for delivery problems, are often In case of non-respect of the upstream workshops planning
treated in the short term by a strict and more frequent In this case, groups 1 and 2 often centralize the scheduling
production flow control ; for the group 1, by a management system and the group 3 sub-contracts its work.
of the orders by priorities and for the groups 1 and 3, by a
Consequences of the variations of the information system the additional costs caused
reliability (IP5) by dysfunctions and various
readjustments.
In case of frequent dysfunctions of the orders and Supplying time Delivery time of raw
information transmissions materials and components.
In this case, the companies control the production systems in Readjustment time Following manpower and
the medium term, by a revision, an improvement and by capacity readjustments, this
procedures automation. time corresponds to the
response time of the
Consequences of the variations of the inventory levels organization for setting new
(IP6) solutions.
CLD PF Consumption Limit Date of
In case of a too high level of stock of raw materials (IP6a) the fresh Finished Products.
and finished products (IP6b). RM Variations This variable calculates the
In theses cases, a physical control will be done by a recurrent difference between the real
inspection of the inventories levels. A real time control of the stock of row materials (or
production flows can also be ensured. In the long term, a components) and the
"Just In Time" system can be implemented to reduce the desired inventory level
stocks. MPS Variations This variable calculates the
difference between the
Consequences of the real cost variations of the forecasts based on the MPS
manufactured products (IP7) and the real sales.
Work-in-process The production work-in-
In case of real costs increases process.
The increase of the real manufacturing costs can be treated in Shipping FP The flow of finished
an active way, by a reinforcement of controls of the material products sending to the
expenses, of the production rejections rate, and by analyzing customers.
the labor over-costs. Moreover, the group 3 sub-contracts in Training Rate measuring the
the middle term, a part of its production. In the same way, it intensity of the staff
develops automation or improves its cost control system. training.
Nominal motivation Rate of the desired
3. The structure of the generic model motivation level.
Real motivation Real rate of motivation of
Of this preliminary analysis of the decision mechanisms, we
the employees.
propose to formalize this first qualitative knowledge by an
MPS Master Production
influence diagram which shows the causal interactions
Schedule.
between the characteristic variables of these systems. Figure
Information flow Quality Quality of the orders and
1 describes the causal diagram of our generic model. The
glossary of the variables of the model is presented in table 1. information transmission.
Overtime ratio The maximum rate of
overtime possible to realize
Table 1 : Generic model variables listing in short time
Temporary staff ratio Maximum rate of
Nominal capacity Theoretical capacity of the temporary staff which can
production system. be recruited in short term.
Real capacity Production capacity to Rejections FP Rejection rate of finished
answer to the commercial products caused by various
needs or following internal factors of production and by
or external disturbances. products goings beyond the
Coef. of working hour Coefficient indicating the dates of consumption.
variation rate of overtime or Rupture RM and Average percentage of the
temporary staff hiring. components raw material out-of-stock
Quality control Rate representing the level.
intensity of the quality
control of the products. Stock RM and components Level of the raw material
inventory
Production flow Control Rate representing the
intensity of the quantitative Objective stock MP and Level of the desired
and qualitative production components inventory of raw materials
controls. and components.
Additional costs This variable determines
Stock FP Real level of the stock of Production rate The manufacturing rate
finished products. (political choice).
Lot Size The size of the production Service FP rate The Finished Products
batches (political choice). customer service rate
Rate provision The raw materials rates (varying from 0 to1)
supplying (political choice). Variation of capacity The coefficient of capacity
Scrap rate The average material scrap variation.
rate. Productivity variation The coefficient of
Stoppages rate The average machines productivity variation.
breakdown rate. Sales This variable corresponds
DP Non-quality rate The rates of non-conform to the volume of the
Delivered Products. customer orders.

Figure 1. Influence diagram

The figure 1 can be interpreted in this following way : the The diagram in the figure 1 clearly shows the causal
nodes of the graph represent the variables, the arrows interdependence between the different performance
indicate the causal relations between these variables, the indicators, object of this research. This representation is a
signs of the arrows correspond to the polarity of the causal good qualitative base for behaviors analysis of the
relations : a sign (+) indicates a influence relation between production systems after fluctuations of their IP.
two variables going in the same direction and the sign (-), in After this first work, we develop a model represented by a
the opposite direction. Forrester's flow-stock diagram and transcribed it in a finite
differences equations system.
The double-arrows translate the delay effects corresponding
to an exponential duration between the cause and effect (cf.
Forrester, 1961). The encircled variables are the Performance 4. PROPOSAL OF A DSS
Indicators and are numbered IPi according to values of the
index i described in the previous paragraph. By implementing this continuous simulation model, different
experimentations from various scenarios of fluctuations of
the different indicators have be done according to a given
experimental plan. These simulations allowed us to study the of our research consists in validation of these models and
stability of these systems faced to different perturbations. We attractors by comparison with real observations in different
will not present here the results of these simulations but we companies.
will show how to complete this first analysis by simulating
an interactive neural network. This connectionist model was
based on the previous influence diagram (see fig. 1) which 5. REFERENCES
formalized the causal relationships between the performance.
ECOSIP, 1990, Gestion industrielle et mesure économique,
An influence diagram was built according to the systemic approches et applications nouvelles, éditions Economica.
and cybernetics principles. It was essentially based, taking FORRESTER J.W., 1958, "Industrial Dynamics : A major
into account the complexity of the food production systems, breackthrough fo decision makers", Harvard Business
on preliminary observations which highlighted the regulation Review, 36, n°4,, July-August, pp. 37-66.
and industrial management systems. From this first cognitive FORRESTER J.W., 1961, Industrial Dynamics, Cambridge,
representation, we built a generic simulation model which Massachusetts, MIT Press.
will is the base of an operational Interactive Decision FORRESTER J.W., 1994, "Policies, Decisions and Information
Support System (DSS). This support can simulate, starting Sources of modeling" in Modeling for learning
from initial values performance indicators, the possible organizations, Productivity Press, System Dynamics Series.
evolution of these same indicators. JACOT J.H., MICAELLI J.P. (coordinateurs), 1996, La
performance économique en entreprise, édition Hermès.
More concretely, the production managers regularly controls KAPLAN R. S., NORTON D. P., 1992, The balanced scorecard
the flows of production and the level of the indicators like - measures that drive performance Harvard Business
the quality level, the stocks level, the returns rate, etc. These Review, January-February, p 71-79.
data have to be entered in the DSS which will simulate the LECOMTE C., MAROLDA A., THIEL D., 1996 “La flexibilité
evolution of these indicators from these initial conditions industrielle dans la production et la logistique alimentaires :
thanks to our system dynamics model. The logistic and quelques réflexions », Revue IAA, octobre, p. 763-769.
production managers can than propose to modify some LORINO P., 1991, Le contrôle de gestion stratégique : la
production policies after analyzing the causal structure of the gestion par les activités, éditions Dunod.
model. The input data of the DSS will come from an SIMON H.A., 1960, The new science of management
Executive Information System (EIS) which directly imports decisions, Evanston, Harper and Row, New York,
its data from an ERP software (see figure 2). Publishers.
NAKHLA, 1994, "Les systèmes de gestion de production
DSS
assistée par ordinateur dans les industries agro-alimentaires",
Revue Française de Gestion Industrielle, p. 267-285.
IDSS RICHARDSON G.P., 1984, "Loop polarity, Loop dominance,
Real system System and the concept of dominant polarity", Proceedings of the
(ERP, SCM …) EIS
Dynamics 1984 International System Dynamics Conference, Oslo,
Norway.
RICHARDSON G. P., 1992, Feedback Thought in social
Figure 2. Environment of the DSS Science and Systems Theory, UPP, University of
Pennsylvania Press.
This DSS which permits to make causal analysis showing the TREILLON R., 1991, "Agro-industrie et flux de production :
interdependence between the performance indicators, proves quels systèmes de gestion ?", Revue Française de Gestion
its benefit confronting to traditional statistical methods of Industrielle, , p. 318-338.
search for correlation between variables. Indeed, these TREILLON R., C. LECOMTE, 1995, "La gestion industrielle
forecast methods based on the past values can only dans les entreprises alimentaires : perspectives et réalités",
independently predict the evolution of each indicator. Congrès International de Génie Industriel, Montréal, p. 171-
176.
However, in many companies, taking into account the
significant number of performance indicators, the frequent
question is : starting from the forecasts given by this
statistical model, how to analyze the tendencies being able to
be contradictory, between various performance indicators ?
We also implemented a prototype of DSS which has been
validated by two companies partners of our research.

This paper opens a new way for studying production system THANKS
behaviors. The first simulation results of this connectionist This work is included in the framework of a public Contract
model based on our system dynamics model, has permitted 2000-2006 between the State and the Region “Pays de la
to underline trends which some industrial companies could Loire” and is supported by the French Ministry of
meet in the proximity of these attractor states. The next step Agriculture and Fisheries.
 EFFICIENT SIMULATION OF DIFFUSIONAL TRANSPORT IN COMPOSITE
MATERIALS – APPLICATION TO THE IDENTIFICATION OF DISTRIBUTED
DIFFUSION COEFFICIENT IN FOOD PACKAGING
Olivier Vitrac
Laboratory of biotechnology of polymer and food safety
INRA – UMR FARE
CPCB, Moulin de la Housse
51687 Reims cedex, France
E-mail: [email protected]

KEYWORDS within food items (e.g. off-flavors, toxic substances).

chemical engineering, computer aided engineering, Consequently, there is considerable interest in the
pseudospectral methods, parameter identification, image development of robust and efficient numerical methods for
analysis the prediction of cross mass transfer between i) food, ii) food
packaging materials and iii) the surrounding, which takes
ABSTRACT into account the conditions of process, storage of both food
and packaging. Such tools or methodologies would find
In a first part, Chebyshev pseudo-spectrals methods were immediate applications in diverse fields covering industrial,
investigated for modeling diffusional transport through research and regulatory purposes. These include:
arbitrary structure including large jump in physico-chemical - problem 1) computer aided development of food
properties (diffusion coefficient, partition coefficient) such packaging via simulation (forward problem),
as those encountered in multilayer packaging materials. - problem 2) numerical identification of mass transfer
Stable and accurate results are obtained by incorporating properties in complex configurations (i.e. structured
Robin boundary conditions through a weak formulation of materials) (inverse problem)
local transport equation. In a second part, the method The purpose of this paper is to propose an enough general
combining spectral and multi-domain approach was used to mathematical formulation of diffusional transport so as to
derive a new numerical inverse procedure for retrieving handle efficiently both previous technological problems. The
unknown distributed diffusion coefficients from forward problem presents several difficulties associated with
concentration history within structured samples. Finally, the the numerical solution of partial differential equations (PDE)
whole method was tested to assess diffusion coefficients of a with i) a discontinuous initial condition, ii) large jumps in
dye in a thin controlled 3 layers polymeric material. Reliable properties (e.g. diffusivities, solubilities) on small scales
and consistent results were obtained for each uncoupled (e.g. thin layers, defects). Indeed, non-smooth solutions also
layer by combining approximations of spatial and time first called “multiscale solutions” would required high frequency
derivatives on local Chebyschev basis through a variational information that is only tractable via fine/adaptative spatial
formulation independent of unknown boundary and interface discretization and/or optimal interpolation (Funaro 1991).
conditions. Among various concurrent numerical strategies, pseudo-
spectral (PS) techniques, so called “spectral domain
INTRODUCTION decomposition”, appeared as one of the most natural to map
with high accuracy an heterogeneous domain into
Food packaging is nowadays an integrating part of food “smoother” sub-regions and to formulate implicitly transport
items and offers a large source of potential innovation for phenomena close to discontinuity as interfacial conditions
both food and packaging industries. As a result, new (Boyd 2000). There are several further motives to use PS
packaging materials are always more sophisticated and they methods. One is that the extension of differentiation
are changed from passive and waste barriers into active, operators to 2D is competitive with tensor-product-based of
intelligent, recyclable and reusable materials. They are 1D ones (Funaro 1991). The second is that PS methods can
usually designed as composite materials including handle automatically and individually periodic
hierarchical organization by means of the assembling of discontinuities by means of a suitable choice of functions
multi-layers, multicomponents, nanostructure (e.g. basis (i.e. Fourier series). The third is that they reduce the
dispersion, coatings, surface deposits or treatments…). From computation cost owing to sparse block matrix, scalable
the technological point of view, the optimization of induced algorithms and possible replacement of matrix multiplication
properties (e.g. barrier, selectivity, reactivity, retention) by FFT algorithm (for Fourier basis, but also for Chebyschev
required an accurate control of diffusional transport through expansion, which is merely a Fourier cosine series). A last
such complex structure. From the point of view of regulatory reason is that the calculated analytical form of the solution
authorities and consumer, both modifications of the life has meanings within both and frequency spaces, and makes
cycle of food packaging and rapid diversification of it therefore possible to devise analogies between operators
formulations and combination of materials could potentially involved in numerical resolution of PDE and signal analysis,
increase the risk of i) uncontrolled release of undesirable such as “filtering”, “homogenization», “aliasing” properties.
substances (additives, contaminants) or ii) higher reactivities
The paper is organized into three sections. First, we focus on where u is the concentration in a particular diffusive species.
devising a flexible and efficient PS technique for the 1-D The segregation of diffusing species at the boundary ∂Ω
diffusional transport within food packaging and at their between regions ΩA and ΩB (macroscopic heterogeneity)
interfaces with food. A general weak form of the PDE according to their affinity is described assuming sharply
accounting for more common boundary conditions (BC) was discontinuous Dirichlet and (possibly discontinuous) rate-
formulated and used to derive a flexible multidomain limited Neumann constraints:
version. The advocated mathematical formulation was an u (s, t ) ∂Ω = K u (s, t ) ∂Ω
extension to M domains (M≥2) of the one proposed by  B AB A
 ∂Ω r ∂Ω r
(2)
Faccioli et al. (1996) for the propagation of waves with  D B (s, t )∇u (s, t ) B ⋅ n ∂Ω = D A (s, t )∇u (s, t ) A ⋅ n ∂Ω
absorbing BC, which presents analogy with partition BC. In
order to decrease the computation cost, the Legendre grid where KAB is the partition coefficient between structures A
was replaced by the Chebyschev-Gauss-Lobatto grid and B. In the case where B is a liquid, the set of equations
relevant for FFT algorithms. Possible variations within an (2) is replaced by a similar mechanism assuming the
element – but possibly discontinuous at elements interfaces – conservation of diffusive and convective fluxes, written as a
was assumed smooth and treated pseudospectrally by means Robin boundary condition:
of a special norm involving the variable property (Frutos and
Muñoz-Sola 1995). Differentiation matrices were finally
∂Ω r
[
− D A (s, t )∇u (s, t ) A ⋅ n ∂Ω = hm K AB u (s, t ) A − u B (t )
∂Ω
]
(3)
derived from stable algorithms provided by Weideman and r
where hm is the convective mass transfer coefficient and n is
Reddy (2000). The second section provides some numerical
experiments of time dependent transport through multilayer the normal vector to ∂Ω. Equation (3) requires an additional
materials, and are discussed regarding the accuracy and mass balance for the diffusive species in region ΩB. In case
conditioning of operators. The third section proposes a fast of a constant volume VB and surface contact SB for region
numerical tool based on previous operators of interpolation ΩB, and without limiting hydrodynamic conditions (i.e. with
and differentiation dedicated to the interpretation of 2D uniform concentration in ΩB), equation (3) can be
images of time dependent transport processes of marked conveniently replaced by an instantaneous and local mass
molecules within food packaging. The case of tracer balance at the boundary ∂Ω:
∂u (s, t ) A
∂Ω
diffusion within a three-layer material is more particularly ∂Ω r V
detailed. − D A (s, t )∇u (s, t ) A ⋅ n∂Ω = B K AB (4)
SB ∂t
MATHEMATICAL FORMULATION OF THE Macro-heterogeneities/structure splitting
FORWARD PROBLEM
In one dimension (s = [x 0 0]), the calculation domain was
Several schemes have been proposed for exploiting the split into M (M≥2) non-overlapping macro-elements Ωi=1..M
structure of scale space where diffusional transport occurred. (i.e. macro-heterogeneities or macro-structure), 2 external
Most approaches need a very dense sampling of the
boundaries, ∂Ω1 and ∂ΩM, and M-1 internal interfaces, noted
hierarchical structure and are therefore computationally
γi=1..M-1 for the interface between Ωi and Ωi+1 (with 1≤i≤M-1).
expensive. Bi-modal description including possible large-
The way we include the discontinuous solution to equation
scale rough macrostructure and smooth microstructure would
(1) that are necessary in our multiscale approach is to
be a good compromise. Indeed, the microstructure
consider a weak coupling between the solutions ui to PDE
contribution may be described by a set of space- and –time
(1) obtained on different subdomains. Since we do not need
field variables and simulated by the phase field-method or
any of the difficult results concerning weak solutions, only
assessed from a digitization of an experimental one (Zhu et
the basic feature of variational approach is here outlined.
al., 2001). As a result, this coarse grained description leads
Let v=V(Y) be a smooth test function. If we multiply both
to model diffusional transport through an arbitrary
heterogeneous structure by a diffusion equation with a sides of (1) by v and integrate over Ω1∩..Ωi..∩ΩM, we obtain
smooth variable diffusion coefficient and suitable boundaries after integration by parts:
M
∂ui (t ) M
∂  ∂ui (t ) 
conditions at the interfaces of macro-heterogeneities. In the
case of a material with no microstructural evolution, the field ∑∫
i =1 Ω i ∂ t
v dx = ∑∫∂
 Di
i =1 Ω i x  ∂x 
 v dx
variables are only function of spatial coordinates s and the M
∂ui (t ) ∂v ∂u (t ) ∂u (t )
diffusion coefficient is mapped for the entire calculation
domain as D(s).
=− ∑∫D
i =1 Ω i
i
∂x ∂x
dx − D1 1 v
∂x ∂Ω1
+ DM M v
∂x ∂Ω M

M
∂ui (t )v ∂u (t )v
Transport equations through an arbitrary heterogeneous
structure
+ ∑D
i =1
i
∂x γ i
− Di +1 i +1
∂x γi
(5)

According to the previous hypothesis and neglecting the According to continuity of flux at interfaces γi imposed by
interaction between diffusing species, the temporal diffusive (2) the last summation term vanishes. Moreover, by
transport process through the structure is finally written in a imposing (4) and an impervious/symmetry boundary layer
common form as: ( ∂ui (t ) ∂x = 0 ) respectively at ∂Ω1 and ∂ΩM, we conclude
∂u (s, t )
= ∇ ⋅ D (s )∇u (s, t ) (1) that at each t, the set {ui}1..M satisfies the variational
∂t problem:
 i i The final linear system for the multidomain is obtained from
M ∏ Kj
∂wi (t ) M ∏K j
∂wi (t ) ∂v
an integration parts of the weighted residual form:
∑∫ ∑∫
j =1 j =1
v dx = −
∑ (R ( ) w( ) , v )
Di dx M
i
∂w N(i ) ∂  (i ) ∂w N(i ) 
i =1 Ω i K0 ∂t i =1 Ω K0 ∂x ∂x N
i
N
i
N N , with R N(i ) w N(i ) = k i −  J N k i Di 
i
i =1 ∂t ∂x  ∂x 
V ∂ w1 (t ) ∂Ω1 and by applying the Lagrangian polynomial basis
− 0 v ∂Ω constructed with GLT nodes as test functions vN. Thus, a
S0 ∂t 1

close set of equations is obtained:

wi (t ) γ = wi +1 (t ) γ i = 1..M − 1 (6) a) Internal equations
i i

where {wi }i =1.. M is a piecewise solution (with same units of RN(i ) = 0 at x = {xn }n=1.. N −1
(i =1.. M )
(9)
concentration) continuous at all interfaces γI and derived b) BC equations
from mass transfer equilibrium considerations so that:  ∂w(i ) V ∂wN(1) 
u (t ) k1  D1 N + 0  = ω0(1) RN(1)wN(1) at x = x0(1) (10)
wi (t ) = i i (7)  ∂x S 0 ∂t 
∏ j K
j =1
 ∂w (i ) 
k M  D1 N  = −ω N( M ) RN( M ) wN( M ) at x = x N( M ) (11)
 ∂x 
Formally speaking, equation (6) has been obtained from (1)
after taking at each time the variation respectively to all c) Interface conditions (IC)
admissible functions v : Ωi → ℜ , which are precisely square ∂w (i +1) ∂wN(i )
ki +1 Di +1 N − ki Di = ω0(1) RN(1) wN(1) + ω N( M ) RN( M ) wN( M )
(Lebesgue) integrable along their first derivative. ∂x ∂x
wN(i ) = wN(i +1) at x = {γ i }1≤i ≤ M −1 (12)
Spectral approximation of local transport
The main advantage of this variational approach over
classical collocation lies in its higher accuracy, since it
In order to complete the discretization within each element
entails a linear combination between the PDE (1) and the
(i.e. to account for local transport), one has to choose an
physically consistent BC or IC.
approximation subspace with finite basis {Tn }n −0.. N for
each wi : Ω i → ℜ and a quadrature for the evaluation of the Time marching
integrals in (6). If {Tn }n = 0.. N are the Chebyschev polynomial
The matrix assembling of (10)-(12) leads to an ordinary
of degree ≤ N or its equivalent definition as weighted differential equation (ODE) in the usual form: Mw& = Dw ,
interpolant: w (Ni ) (x n ) = TN ( x n ) for a suitable set of collocation where M is diagonal and D a sparse matrix with almost all
1 real and negative eigenvalues. In practice, D contains
points {x n }n = 0.. N on the canonical interval [–1, +1], ∫ w(x )dx diffusional transport information through not shape regular
−1 macro elements and with variable equivalent transport length
can be approximated for any continuous function by the or capacity. As a result, D is partially ill-conditioned with
N conditioning number typically between 10-15 and 10-19 and
truncated series ∑ w ( x )ω N n n . The truncation error is in increasing with the order of the approximation. For
n =0 economical and reliability reasons, only fully implicit
particular zero for all w(x) which are polynomials of at most schemes must be used to solve the ODE set for any initial
degree 2N+1 and 2N-1, for respectively Gauss (G) and condition : w(t = 0) = w0 , such as backward differentiation
Gauss-Lobatto-Chebyschev (GLT) quadrature. Best formulas (BDF) or numerical differentiation formulas
performance for both interpolation and integration are (NDF). Best results are obtained with a quasi-constant step
achieved by choosing the latter relatively to the roots (GLT size implementation of NDF up to order 5 (ode15s function
( )
nodes) of 1 − x 2 TN −1 ( x n ) and their corresponding weights of Matlab 6.x or from Matlab ODE suite). Due to the
ωn (Boyd 1999). It follows that a finite version of equation sparsity of D, this technique is highly competitive whatever
(6) including a very high accurate second order differential the chosen BC even for large system (500×500) comparing
operator can be provided: with the introduction of the finite-element based
i preconditionner recently proposed by Shen et al. (2000).
 ∂wN(i )
i
M
 M
 ∂v 
∑  ki
i =1  ∂t
, v N  = −  H (i ) , N 
i =1  ∂x  N
∑ It is worth to notice that BC (4) or Neumann BC coupled to
an initial solution with zero gradient at the boundary must be
N
(i ) initialized by including an external collocation point or by
V0 ∂ wN ∂Ω1 suitably modifying the initial solution (e.g. by replacing a
− vN (8) constant profile with an Heaviside function).
S0 ∂t ∂Ω1

i
 ∂wN(i )  NUMERICAL EXPERIMENTS
with ki = ∏ K j K 0 , H (i ) = J N(i )  k i Di , J N(i ) the
j =1  ∂x 
Figure 1 provides three dimensionless examples of our PS
Chebyschev interpolation operator at GLT nodes, (ϕ ,ψ )N the
i
method for the contamination propagation from one external
1 N layer to the others within a 4 layer material including large
inner product defined by ∫ ϕ (x )ψ (x )dx ≈ ∑ ϕ (xn )ψ (xn )ω n .n =0
jumps in ki and Di values at each interface (and constant
−1 within each layer). Unidirectional and bidirectional transfer
are considered with different BC to illustrate the pp and ldpe (k1..3=1). Moreover, it was assumed that i)
performance of the numerical scheme with respect to the significant dye transport occurred only along direction x, ii)
partitioning and computation time. In particular, full boundary of A and C verified impervious conditions. Let
impervious BC is a severe test to assess the simulation { }
J N(i ) i =1..M be an interpolant of Di(x) in the coefficient form:
consistency regarding mass conservation laws. As an N
indication of the typical required CPU time, the whole figure
1 was generated on Matlab 6.1 within 2 s on a P4, 2.2 GHz.
∑ a ( )φ ( ) for
i =0
n
i
n
i
some suitable basis functions φ n(i ) . The so-

called inverse problem consists in finding

{ }
A = a n(i ) i =1..M ,n =0.. N .

Figure 1: Numerical Experiments (M=4, N=30) for Various BC.

INVERSE PROBLEM: IDENTIFICATION OF

DISTRIBUTED DIFFUSION COEFFICIENTS

Typical example

In order to rapidly assess barrier properties of packaging

composite (e.g. multilayer), the basic idea was to measure
the migration trough the whole structure of a non-volatile
UV dye (2',5'-dimethoxyacetophenone, maximum Figure 2: Variation of Local Concentration with time.
absorbance at 330 nm) initially present in one layer.
Versatility and robustness of the method require that Identification procedure
diffusion properties may be retrieved from concentration
profiles acquired for any of consecutive discrete times, in Equation (8) combined with a subdomain weak formulation
other words independently of the knowledge of initial leads to a natural approach for solving the latter problem.
conditions. Local dye concentrations were inferred from Indeed, since either wi or ui are assumed to be included
several microscopic observations (UV CCD camera model within the coarse concentration picture series, noted Pxyt,
Hamatsu-C4742-95812E5 coupled with an UV microscope matrix A remains the only unknown. By remarking that Di
model Karl Zeiss -MPM800MCS) of a same cross-section values are independent between Ωi layers, “finding A” can
area for different transmitting light wavelength (330 and 420 be lumped into a common and direct (i.e. non iterative)
nm). In particular, the ratio log10(max(I330)/I330) over minimization problem of mean weighted residuals (MWR).
log10(max(I420)/I420) was shown to be directly related to the For a given φ n(i ) basis, MWR determines A by imposing for
local dye concentration in common polyolefins and
each Ωi (N+1) conditions:
independent of surface irregularities, where Iλ is the 2D
intensity image at the wavelength λ. ( )
v j , R N(i ) wN(i ) [x, a0 ..a N ] Ω = 0, for j = 0...N
i
(13)
Figure 2 presents the 2D concentration results (1024×1248 for some suitable “test functions” vj(x) where the inner
pixels2) for 7 consecutive times (each 5 hours, after 4 days of
migration at 40°C) within a M=3 layer packaging ABC
product (v, w)Ω i
= ∫ ω ( x) v( x ) w( x) dx for a given non-
Ωi
(thickness: 155+60+85 µm) coated with an epoxy resin,
where A and B are a low density polyethylene (ldpe) and C a negative weighted function ω (x) and any v(x) and w(x).
polypropylene (pp). The dye was initially only located in Equation (8) requires an accurate estimation of w(x) and its
layer A and presented no significant partition effect between first derivatives in time and space from series Pxyt,. Since
Pxyt (1024×1280×7) contained noisy data along each (N=0). As crude approximation, each Di (i=1,2,3) was
dimension: x (subspace of diffusion path), y (redundant identified from equation (13) with constant weight and v as
subspace) and t (temporal subspace), Pxyt was subsequently “tent” or “hat” function. The latter made it possible to get an
“statistically/likely” projected/collocated onto suitable inner product formulation independent of boundary
subspaces. Firstly, Pxyt was factorized along in the conditions. As a result, this approach yielded to consistent D
frequency domain of y so as to reduce the whole estimation for layers A and C, with as expected DC<DA, but
computation cost and low pass filter crude data. The led to a difference between A and B up to one magnitude
obtained 1024×40×7 analytical series P’xyt was secondly order, while same or at least similar values were needed.
interpolated in the L2 sense by M=3 piece-wise Chebyschev More robust estimation was achieved by using test functions,
polynomials (8≤N≤10) respectively to each layer along x. the first derivative of which decreased rapidly down to 0
Entailed BC (zero slope) and IC (continuity) were enforced close to interfaces, so as to reduce the influence of poor
owing to the introduction of corresponding constrains estimation of ∂u (x ) ∂x close to interfaces in (13). In the
(relatively to Lagrange multipliers) in the assembled same manner, significant improvement (i.e. DA≈DB and
regularized least square problem. No condition on the first DC/DB≈12 as expected) were obtained by choosing
derivatives at interfaces were introduced since there are weighting functions, which take into account that regions
depending on the ratio of the unknown Di. The obtained were ∂u (x ) ∂x (or to some extent ∂u (x ) ∂x ) almost vanished
48×7 coefficients were finally used to compute spectrally a
did provide no pertinent information about the behavior of
7th order centered approximation of ∂w(x ) ∂t . w(x ) and D.
∂w(x ) ∂x were calculated for the same corresponding time. It is clearly of interest to compare the a posteriori numerical
solution, which can be calculated from identified diffusion
Reconstructed and identified results coefficient, with the reconstructed measures, assuming
reasonably a step initial profile. The comparison is analyzed
Profiles of u(x), ∂u (x ) ∂x , and ∂u (x ) ∂t were a priori as relative local error in figure 3. Errors in u was negligible
reconstructed with M=3 and N=8 (figure 3). As expected, the whereas errors in ∂u (x ) ∂x and ∂u (x ) ∂t in the range of
dye was smoothly distributed between A and B and a sharp ±20% were accordingly partially correlated and related to
change in u was detected between B and C. In one hand, the regions with weak signal or close to discontinuities. These
whole reconstruction procedure led to enough noise results confirm the interest of our direct technique of
reduction to make both first time and spatial derivative identification that make it possible to assess and interpret
almost as exploitable as the primitive signals u. On the other accurately 7 concentration pictures in less than 30 s on a P4
hand, enough high frequency along x remained in u to 2.2 GHz.
account for possible discontinuities between materials with
different properties, such as ldpe and and pp (commonly one CONCLUSION
magnitude order). Close to external boundaries, zero
gradient conditions regularized conveniently all profiles. The proposed pseudo-spectral method appears to provide
accurate numerical results for both direct and inverse
problems of diffusional transport within complex structure.
Indeed, the multidomain approach combined with weak
solutions allows effective and robust calculations in presence
of strong jumps in transport properties. Further work appears
desirable to validate both relative and absolute determination
of diffusion coefficients in more general cases.

REFERENCES

Boyd, P. 2000. Chebyschev and Fourier Spectral Methods. 2ed.

Dover Publications, Mineola, N.Y.
Faccioli, E.; Maggio F.; Quarteroni A.; and A. Taghan. 1996.
“Spectral-domain Decomposition Methods for the Solution of
Acoustic and Elastic Wave Equations.” Geophysics 61, No.4,
1160-1174.
Funaro, D. 1997. Spectral Elements for Transport-Dominated
Equations. Springer-Verlag, Berlin.
de Frutos, J. and R. Muñoz-Sola. 1995. “Chebyschev Pseudospectal
Collocation for Parabolic Problems with non Constant
Coefficients”. In Proceedings of ICOSAHOM’95, Houston.
Figure 3: Reconstructed Profiles of u(x), ∂u (x ) ∂x , and Shen, J.; Wang F.; and J. Xu. 2000. “A Finite Element Multigrid
∂u (x ) ∂t from Concentration Images corresponding to Preconditioner for Chebyschev-collocation Methods.” Applied
Numerical Mathematics 33, 471-477.
figure 2 (a priori reconstruction) and from identified Di
Weideman, J.A.C and S.C. Reddy. 2000. “A Matlab Differentiation
Values (a posteriori reconstruction). Matrix Suite”. Technical report. http://ucs.orst.edu/
~weidemag/differ.html.
Because no significant variation of Di was expected within
each layer, the basis for Di was reduced to an unique scalar
 NUMERICAL SIMULATION OF THERMAL AND FLUIDDYNAMICAL
TRANSPORT EFFECTS ON HIGH PRESSURE INDUCED
INACTIVATION

Christoph Hartmann and Antonio Delgado

Chair of Fluidmechanics and Process Control
Technische Universität München
Weihenstephaner Steig 23
D-85350 Freising
Germany
E-mail: [email protected]

KEYWORDS PROCESS MODEL

High hydrostatic pressure, pressure induced inactivation,
transport phenomena, thermo-fluid-dynamics, process non- Two different configurations are considered. The first one,
uniformity subsequently denoted as Process 1, represents the inacti-
vation of B.subtilis α-amylase dissolved in TRIS-buffer in
ABSTRACT a cylindrical high pressure chamber of approximately 0.8 L
volume. The application of a pressure of 500 MPa at a
The current paper deals with the effects of convective temperature of 313 K is achieved by mass augmentation in
transport and heterogeneous heat transfer on high pressure the chamber, i. e. the solution is pumped directly into the
induced inactivation. While the pressure within a high chamber until the required pressure level is reached. The
pressure chamber can be considered as spatially constant, inlet velocity is chosen such that the level of 500 MPa is
the temperature, in general, is a heterogeneously distributed reached within a finite time interval of 25 sec. The pressure
field quantity. In the present paper it is shown up to which is held during 25 min. The pressure chamber is shown in
extend thermal heterogeneities and convective transport figure 1. Its diameter and its height are 0.1 m leading to a
lead to non-uniform inactivation results. As examples, the volume of 0.8 litres. The grey shaded area represents a
inactivation of B.subtilis α-amylase dissolved in TRIS- plane of axissymmetry. This plane will be used
buffer and the inactivation of E.coli suspended in a liquid subsequently for graphical representation of field
similar to water in a model pressure chamber at pressures distributions.
between 400 MPa and 500 MPa are considered. Aspects of D
the viscosity of the matrix as well as the heat transfer
properties of the packaging material are considered. In both
cases, significant process inhomogeneities can be z
determined. D x
y
INTRODUCTION

The treatment food or food ingredients with high pressures

between 100 MPa and 1000 MPa is always accompanied 0.1D
by an increase of the temperature in the treated volume.
This is due to the change of the thermodynamic state of the
system. Finite transport velocities of heat and suspended
substances lead to the fact that temperature as well as
substance concentration have to be considered as
distributed balance quantities. In the present paper, the U
influence of hydrodynamic and thermodynamic transport Fig. 1: Pressure chamber of process 1, D = 0.1 m
on the process uniformity is investigated taking into
account the above mentioned aspects. The second process, subsequently denoted as Process 2, is
The inactivation of a dissolved enzyme (alpha-amylase) and an indirect one. This means, the processed samples are
of suspended micro-organisms (E. coli) in matrix liquids packed in containers which are placed in the chamber at
similar to water is simulated as a model process. During the distinct locations. The pressure is applied by mass increase
high pressure processing a fluid flow develops which is of a liquid pressure medium leading to a compression of the
interacting with the temperature field. Therefore, the medium, the package material and the contained suspension
inactivation takes place in a moving matrix and in transient of E. coli in UHT milk. The pressure of 400 MPa is applied
thermal conditions. at a temperature of 303 K over 16 minutes. In process 2, the
pressure application is assumed to be instantaneous, i. e. the
change of volume of the packages as well as the induced modelling techniques have been validated by comparison to
forced convection are neglected in order to circumvent experimental data (Hartmann, 2002).
numerical difficulties related to moving interfaces and
fluid-structure interaction problems. The assumption is METHODS
justified in the sense that the pressure increase phase is
short compared to the pressure holding phase and forced The finite volume method is used for the solution of the set
convection is quickly dominated by natural convection of equations. Therefore, the commercial solver CFX-4 has
(Hartmann, 2002 and Hartmann, 2002a). The geometry of been enhanced by own software extensions providing
the chamber is shown in figure 2. The volume amounts up solution algorithms for the above-mentioned equations.
to 6.3 litres. Again the grey shaded area represents the
plane of graphical distribution of the field quantities. RESULTS AND DISCUSSION

It can be shown, that the heterogeneity of the inactivation

∅D
process depends strongly on the time scales of the
hydrodynamic compensation, thermal compensation and
the inactivation.
Package 5 In process 1, significant heterogeneities can be observed for
z the case that the hydrodynamic compensation time scale is
Package 4 much smaller than the inactivation time scale and the
D x
y Package 3 thermal time scale is much larger than the inactivation time
scale. Figure 3 shows the distribution of the relative activity
Package 2 (instant activity scaled with its initial value) of alpha-
Package 1 amylase obtained at the end of the high pressure process in
a matrix fluid which of the viscosity is hundred times larger
than that of water. The graphical representation shows a
cross section through the pressure vessel (grey-shaded area
in figure 1).

Fig. 2: Chamber for Process 2, D = 0.2 m

MODEL EQUATIONS

In order to simulate these processes, the basic equations of

fluid dynamics, i. e. the continuity equation, the Navier-
Stokes-equation, the energy equation and a balance
equation for the activity of the suspended substances have
to be solved. The latter contains a kinetic model for the
inactivation, which is incorporated as a sink term in the
equation. The kinetic model is taken from (Denys et al.
2000) for the enzyme and from (Hinrichs 2000) for the
micro-organism.
An equation of state for water (Saul and Wagner 1989)
valid in the considered pressure and temperature range is
implemented as well as pressure and temperature dependent
relations for the viscosity (Först et al. 2000) of water and
the specific heat capacity.
These data are taken for both processes, since the specified
liquids are similar to water and generic state data are not Fig. 3: Relative activity retention in a 0.8-litre volume after
available for the considered range. 1200 sec at 500 MPa and at 313 K
As thermal boundary condition constant temperature is
prescribed at the housing. Fluid adhesion on the wall is The activity retention is a distributed quantity that varies in
assumed. In process 1, a constant velocity U is prescribed the range between 28 % and 48 %. In the core zone and in
and a priori adapted such that the desired pressure increase the inlet, the highest activity retention can be observed
within the compression phase is reached. In process 2, the while minimum values are obtained in the outer region of
pressure is set instantaneously to the desired value of the chamber. This is due to the cumulative effect of thermal
400 MPa leading to a temperature increase equal to that of a heterogeneities. In the core and at the inlet, the temperature
purely adiabatic process. is comparatively low due to forced convection of cold fluid
All liquids are supposed to be newtonian, compressible, and into the chamber. The outer region is at a higher
chemically inert. The flow is laminar. The assumptions and temperature due to compressive heating. Therefore, the
inactivation rate is higher in this region leading to a more
efficient inactivation rate. 2.5E-06
Furthermore, it has been observed that the inactivation

rel. cfu-concentration
leads to smaller values of the inactivation if the size of the 2.0E-06 average value
chamber is increased (scale-up). This is due to the larger standard deviation
1.5E-06
thermal compensation time scale leading to the effect that
the inactivation is running at a higher average temperature. 1.0E-06

Concerning process 2, the following results have been 5.0E-07

obtained. Again the process homogeneity depends strongly
0.0E+00
on the temperature decay time scales. An additional aspect
1 2 3 4 5
enters into the discussion: the thickness and the thermal
properties of the package material. Up to a certain extent package number
the package material represents a heat barrier which Fig. 5: Cfu-concentration in 6.3 L-volume after 960 sec,
provides almost homogeneous thermal conditions in all five improved heat transfer
packages. This is the case, when the package material is
poly propylene at about 2 mm thickness. Figure 4 shows the The differences between the inactivation in the individual
average values of the colony forming units (cfu) per ml packages can be explained by the characteristics of the
scaled by the initial concentration of 108 cfu per ml in each average temperatures in the packages shown in figure 6.
of the five packages at the end of the process. The temperature rises up to 314 K during the adiabatic
compression in all packages and then decreases to 305 K in
package 1 and 307 K in package 4 within 960 sec. A
1.0E-08 temperature difference of about 2 K is established between
package 1 and 4 after 220 sec and persists during the
average value
rel. cfu-concentration

8.0E-09 complete process. The temperature difference leads to an

standard deviation
inactivation which runs at different rates in both packages,
6.0E-09
which yields the considerable difference in the
4.0E-09 cfu-concentration of more than one order of magnitude.

2.0E-09 315

0.0E+00 313
Temperature [K]

1 2 3 4 5 311
package number 309
Fig. 4: Cfu-concentration in 6.3 L-volume after 960 sec 307 package 4
305 package 1
It can be seen that an inactivation of at least eight orders of
303
magnitude has been reached within the considered time
0 200 400 600 800 1000
period. In addition to this, the inactivation is maximum in Time [sec]
package 4 and minimum in package 1. The standard
deviation amounts up to 30 %, which can give rise to the
assumption, that a significant process non-uniformity might Fig. 6: Average temperature in packages 1 and 4 at reduced
exist in the present case. Nevertheless, spatial non- package material thickness
uniformities at such low concentrations of surviving cells
are of minor importance. Therefore, the inactivation can be CONLUSION AND PERSPECTIVES
supposed to be highly efficient in this case.
Reducing the thickness to about 0.65 mm leads to thermal In the present analysis, thermally induced process
heterogeneities which induce a strongly heterogeneous heterogeneities are shown for two inactivation processes. It
inactivation result. Figure 5 shows the average values of the can be stated, that process heterogeneities have to be
scaled cfu-concentration and their standard deviation for expected, when the fluid motion is of low intensity and heat
this case. The differences between the inactivation results conduction dominates thermal compensation. Furthermore,
obtained for the different packages are much more heat conduction has to be of the same time scale than the
pronounced in the present case. They are larger than one inactivation itself.
order of magnitude. The standard deviation in the Futhermore, the package material represents a heat barrier
individual packages amounts up to 30 %. In this case, the which significantly influences thermal conditions of packed
non-uniformity between the individual packages is food. The more the material is insulating the more the
intensified by the non-uniformity within each package. temperature is homogeneous leading to a homogeneous
inactivation in all packages.
Principal subject of further analysis is the modelling of the
finite compression phase for configurations with packed
foods. Here, a special treatment of the moving package pressures up to 25000 MPa. J Phys Chem Ref Data 18:1537-
interface becomes necessary. 1564.

ACKNOWLEDGMENTS

This work has been partially supported by the Max-

Buchner-Forschungsstiftung (Max-Buchner Science
Foundation) within the project 2238 and by the German
Science Foundation (DFG) within the project HA 2228/1.

AUTHOR BIOGRAPHY

CHRISTOPH HARTMANN went to the Technische

Hochschule Darmstadt, Germany, where he studied
Mechanical Engineering. After a one year stay in the Ecole
Centrale de Lyon, France, he obtained his diploma in 1992.
He continued his studies preparing a PhD in Computational
Fluid Dynamics within the Mechanics Department of the
Technische Hochschule Darmstadt. He defended his PhD in
1996. After several month of post-doctoral activities he
joined the Engineering Systems International (ESI)
Company in 1997 as a research and development engineer
for Finite-Element-Software (PAM-CRASH) for the
simulation of occupants' safety in the automotive industry.
In 1998, he joined the department of Fluid Dynamics and
Process Control of the Technische Hochschule München as
the head of the fluid dynamics research group. His global
field of interest are interaction processes between
thermofluiddynamics and food- and biotechnology
preferably solved by simulation techniques.
E-mail: [email protected]
http: www.wzw.tum.de/lfp

REFERENCES

Denys S, van Loey AM, Hendrickx ME. 2000. A modeling

approach for evaluating process uniformity during batch high
hydrostatic pressure processing: combination of a numerical
heat transfer model and enzyme inactivation kinetics,
Innovative Food Science & Emerging Technologies 1:5-19
Först P, Werner F, Delgado A. 2000. The viscosity of water at
high-pressures - especially at subzero degrees centigrade,
Rheologica Acta 39:566-573
Hartmann C. 2002. Numerical Simulation of Thermodynamic and
Fluiddynamic Processes during the High Pressure Treatment
of Fluid Food Systems, Innovative Food Science and
Emerging Technologies 3 (1), pp. 11-18
Hartmann C, Delgado A. 2002a. Numerical Simulation of
Thermofluiddynamics and Enzyme Inactivation in a Fluid
Food System under High Hydrostatic Pressure, In: R. Hayashi
(ed.), Trends in High Pressure Bioscience and Biotechnology
(pp. 533-540), Elsevier, Amsterdam.
Hartmann C, Delgado A. 2002b. Numerical Simulation of
Convective and Diffusive Transport Effects on a High
Pressure Induced Inactivation Process, accepted in
Biotechnology and Bioengineering
Hinrichs J. 2000. Ultrahochdruckbehandlung von Lebensmitteln
mit Schwerpunkt Milch und Milchprodukte – Phänomene,
Kinetik und Methodik -. Fortschritt-Berichte VDI, Reihe 3,
Nr. 656, Düsseldorf: VDI
Saul A, Wagner W. 1989. A fundamental equation for water
covering the range from the melting line to 1273 K at
 A PETRI NET BASED SIMULATION APPROACH FOR
EVALUATING BENEFITS OF TIME TEMPERATURE
INDICATOR AND WIRELESS TECHNOLOGIES IN
PERISHABLE GOODS RETAIL MANAGEMENT
Navneet Bhushan and Kishor Gummaraju
Infosys Technologies Limited,
Electronics City, Hosur Road,
Bangalore –561229,
India
E-mail: [email protected]
http://www.infy.com

KETWORDS By-Date basis to A Dynamic Shelf Life Based Inventory

Time Temperature Indicators, Petri Nets, Wireless LAN, Management or the Least Shelf Life First Out (LSFO) basis.
Perishable Goods, Retail Management
In this paper, we propose a solution to inventory, scheduling
ABSTRACT and supply chain management of perishable items
combining the TTI and Wireless technologies. The proposed
Effective management of perishable goods is a major solution is compared with the existing situation, using a
problem faced by the retailers of perishable goods. Time- discrete event simulation methodology based on the
Temperature-Indicator (TTI) labels available today can Generalized Stochastic Petri Nets (GSPN).
indicate the remaining shelf life of these perishable products.
This technology can change the way perishable goods In the next section we describe the existing system and
inventory is managed today, from a First-In-First-Out proposed solution. Thereafter a brief introduction to Petri
(FIFO) and Fixed Sell-By-Date basis to Least Shelf Life First Nets and GSPN is provided. The GSPN based simulation
Out (LSFO). In this paper a futuristic solution combining TTI approach, the results of the simulation, and an analysis of
and wireless technologies is proposed. Although the TTI and these results to evaluate the benefits of the proposed solution
wireless based solution proposed in the paper is considered are described thereafter. Finally, conclusions and the scope
better, the benefits of the proposed solution have to be for further work have been provided.
analyzed in quantitative terms. These have been analyzed
using a Petri Net based approach for system simulation. The
analysis shows that a retailer of perishable goods can reduce SHELF LIFE BASED PERISHABLE INVENTORY
the losses suffered due to the expiry of perishable items MANAGEMENT
before they can be sold, by more than 90% and increase the
net profits by more than 10% in the test scenario. Thus we Every perishable item has a certain date before which it
can conclude that adoption of the proposed system will needs to be sold to the consumer. This date is computed
benefit the retailer to a great extent. assuming that the required cold chain is maintained. Any
breakage in the cold chain, where the product is subjected to
INTRODUCTION periods of higher temperature, would affect the longevity of
the product. The current scenario in the life of a perishable
The life of perishable products such as fruits, meat products, good is shown in the figure below.
medicines and certain chemicals is a function of the
product’s characteristics and the environment in which the Retailers today are unable to determine the history of the
product is stored. An efficient cold chain is essential to “Time Temperature Exposure” of the product. This has the
ensure that the product remains fresh for the expected following business implications:
duration. However uncertainty in the cold chain, while the
product is in transit to the retail store, adds to many • The product purchased by the consumer could be unsafe
complexities. The retailer could run the risk of selling an although it is well within the sell-by-date.
unsafe product to the consumer or might need to discard
items that have perished before their expected date of
• The product could perish in the warehouse itself, well
before the printed sell-by-date. The product could have
expiry.
been pushed out to the store in time, had the warehouse
been aware of the remaining shelf life.
Time-Temperature-Indicator (TTI) labels (Taoukis et. al.
1998) available today can indicate the remaining shelf life of • The product could also perish in the retail shelf, well
these perishable products. This technology could before the printed sell-by-date, resulting in high shrink.
revolutionize the way in which perishable inventory is • It could be possible to retain some of the products longer
managed in the future. Inventory management would move on the shelf (even beyond the sell-by-date), if the cold
from the current First-In-First-Out (FIFO) and Fixed Sell- chain was maintained very well. However, the retailer
 (unaware of this) may markdown the price in an attempt
to sell the product before the sell-by-date. PETRI NETS

The ideal Inventory Management System would require the System studies require system representation and system
sell-by-date to be revised to reflect the actual shelf life of the analysis. To represent a system for its complete depiction
product. and analysis many abstract methods (models) have been
developed. Petri Nets (PN) is one of the most powerful
The Opportunity formal models of Information flow. Concept of Petri Nets
has its origin in work done by Carl Adam Petri in 1962.
Time Temperature Integrators, are products, which are Since late 70’s PNs have found increasing usage in design
capable of measuring the life of temperature sensitive and analysis of a large number of systems (Ajmone and
products. This is a small adhesive label consisting of an Chiola 1987; Ajmone et al. 1987; Bhushan 1997; Haas and
enzyme and a substrate filled ampoule separated by a Shelder 1989; Peterson 1977; Trivedi and Sun 2001).
breakable seal. Pressure on the ampoules breaks the seal and
allows the two liquids to mix. This activates the time-
temperature clock. The color of the label changes from A Petri Net is a particular type of digraph with an initial
green at the start to yellow at the end of the product life. state called the initial marking, M0. The underlying graph,
N, of a PN is a directed weighted bipartite graph with two
kinds of nodes, called places and transitions, where arcs are
Any increase in temperature beyond the stipulated either from a place to a transition or a transition to a place.
temperature would hasten the color change, indicating a In graphical representation, places are drawn as circles and
reduction in life of the product. Since the process is transitions as bars or boxes. Arcs are labeled with their
cumulative in nature, it enables recording the entire history weights (positive integers), where a k-weighted arc can be
of temperature exposure. Thus the change in color, if interpreted as the set of k-parallel arcs. A marking assign a
captured, could help indicate the extent to which the shelf non-negative integer to each place. If the integer assigned to
life of the product has been impacted and subsequent action a place is j we say that the place has j tokens.
to be taken.

The Proposed Solution Formally a PN is defined as a 5 tuple PN ={P,T,F,W,M0},

The figure below shows the way in which the technology where,
could be exploited to secure the benefits envisaged. The
warehouse will manage the inventory based on the revised P = {p1, p2,…pm} is a finite set of places
sell-by-date, which in turn is a function of the remaining T = {t1,t2,…,tn} is a finite set of transitions
life. In addition the warehouse will also use this information F⊆ {P x T} ∪ {T x P} is a set of arcs (flow relation)
to provide a feedback to the vendor regarding possible mis- W: F → {0,1,2,3,…..} is a weight function
handling, resulting in shrink. This would enable the vendor M0: P → {0,1,2,3,…} is the initial marking
and the transporter to better manage the cold chain. The P∩T = φ ; P∪ T ≠ φ.
retailer will manage the inventory based on the revised sell-
by-date (recomputed at the retail receiving point). He will The behavior of many systems can be described in terms of
also establish decision support systems to determine the system states and their changes. In order to simulate the
extent of markdown required to push the perishable product dynamic behavior of a system, a state or marking in a PN is
out before actual expiry. changed according to the following transition (Firing) Rule:

Wireless technology could be deployed effectively as shown • A transition t is said to be enabled if each input place p of
in figure 3 below, to enable capture of the present sell-by- t is marked with at least W(p,t) tokens, where, W(p,t) is
date and the actual available shelf life while interfacing with the weight of the arc from p to t.
the back-end store systems. Wireless Local Area Network • An enabled transition may or may not fire (depending
(WLAN) is one technology that can create the infrastructure upon whether or not the event actually takes place)
for enabling this system. The hand held device would need
to read the TTI label as well as the sell-by-date on the • A firing of an enabled transition t removes W(p,t) tokens
barcode. It would further need to match the TTI indicator from each input place p of t, and adds W(t,p) tokens to
with the lookup tables for calculating the remaining shelf each output place of t, where W(t,p) is the weight of arc
life. This would enable an informed decision to generate the from t to p.
appropriate label. A small portable printer could also
generate the required labels with revised prices. Timed Transition Petri Net (TTPN) is a special class of PNs
in which time is associated with each transition. These
To analyze the benefits of the proposed solution vis-à-vis transition times may be deterministic or stochastic. Further
existing situation, a simulation study was carried out. The PN Transitions may have a probability of firing associate
simulation is based on the well-established methodology of with them. These are called Generalized Stochastic Petri
Petri Nets. In Section 4 we describe the PN based simulation Nets (GSPN).
model and results of the analysis. However, first we give a
brief overview of Petri Nets in the next section.
In next section we describe how the representation boxes having the least Sell by Date are picked in a FIFO
capability of PNs can be used to simulate a system using the manner.
discrete event simulation methodology. • The Vendor: The vendor sends on an average 2000 boxes
every 10th day. This is assuming that retail stores have a
System Simulation using Petri Nets capability of storing 2000 boxes for 10 days after which
the goods expire. These boxes are sent in 20 trucks each
The simulation proceeds in the following steps: carrying 100 boxes. These trucks reach the respective
stores on an average in 2 days starting from the vendor to
• Modeling of the system (states and transitions) as a Timed the stores. On an average 10% of these trucks per 10 days
Transition Petri Net (TTPN). In most of the cases the do not meet the cold storage requirements. 10% actually
transition times are stochastic variables. Hence the TTPN exceed the cold storage required by Sell by Date. And
becomes a GSPN (Generalized Stochastic Petri Net). The 80% meet the cold conditions required by Sell by Date.
Petri Net has a set of places called sink or collector places Also, we assume that the goods not meeting the required
where all the end states of the system are captured. These cold conditions (we call them Category C) perish on an
collector places are used to evaluate the system average in 2 days of reaching the retail outlets. The goods
parameters of interest. meeting the cold storage (Category B) perish on an
• Event driven Simulation of the Petri Net using Next-Event
average in 8 days of reaching the retail outlets. While
goods exceeding the cold storage (Category A) perish on
simulation approach. The firing of a transition
representing the occurrence of an event involves removing an average in 10 days of reaching the retail outlet.
as many number of tokens from the input places as • Present Scenario: There is no way presently to
specified by the arcs weights, and putting as many tokens distinguish between the three categories at the retail
in the output places as per the weights. outlets. Hence the Boxes are randomly picked (average
• Count the number of tokens in each of the collector
200 boxes per day) from Category A, Category B or
Category C. This leads to possible customer
places.
dissatisfaction if they buy Category C and loss by the
• Repeat the simulation till the desired confidence level and store if they sell Category A goods much before their
accuracy is achieved. actual Sell by Date.
• The average number of tokens in the collector places is • Proposed Solution: In the Proposed Solution because of
used for computing the relevant parameter, e.g., TTI labels, it is possible to distinguish between the three
throughput, response time, etc. categories and schedule their selling based on Least Shelf
Life First Out (LSFO) scheduling
The two situations described above are analyzed using a
discrete event simulation methodology developed. The PN MODEL OF THE SYSTEM
values of the parameters to describe the two scenarios are
described in the Test Scenario below. The GSPN corresponding to the system described above is
shown in figure 4. There are five collector places in the
THE PETRI NET BASED SIMULATION OF THE GSPN shown above - Requested, CategoryASold,
PERISHABLE GOODS RETAIL MANAGEMENT CategoryBSold, CategoryCSold and TotalPerished. Two
input places called BoxesWithVendor and DemandCreated
The approach starts with modeling of the system under represent the number of boxes sent per day from the vendor
study as a Generalized Stochastic Petri Net (GSPN) and to various stores and demand at the stores per day
studying the dynamic state of the system as time progresses respectively. Dashed Circles represent the initial conditions.
using the discrete event simulation methodology. For the Tokens in these input places are zero. The bars represent the
purpose of evaluating the value of the proposed solution a transitions indicating the occurrence of events. It should be
simulation scenario as described below has been assumed. noted that the arrowhead from place BoxesWithVendor to
transition BoxesSent is filled circle. Similarly the arrowhead
The Simulation Scenario from the place DemanCreated to transition Demand is a
filled circle. This represent that the condition to fire the
transition is that the place should not have any tokens.
• The Demand: In multiple Retail Stores there has been Since, the input places do not have any tokens, the transition
observed an average demand for 200 Boxes of Perishable will keep on firing at the regular interval of time associated
Goods per day. Further this demand has been observed to with the transitions.
follow a Poisson probability distribution, i.e., the inter
arrival time between two demands is exponentially
distributed with mean 200 boxes per day. The three layered bar indicates that this is a transition with
stochastic firing time. The associated probability distribution
• The Perishable Goods: The Perishable Goods are labeled with stochastic transitions is exponential. The single layer
with Sell by Date at the Vendor Place. This Sell by Date is bars represent that the transitions have deterministic
computed assuming average cold storage conditions on associated time periods. In the present case, the time period
the way from the vendor to the store. There is no way at associated with these transitions is zero, i.e., the transitions
the retail store presently to find out the boxes that haven’t are immediate. However, there can be an associate
met the average cold storage on the way. Therefore the probability of firing of a particular transition. As the
transitions, VeryWellMaintained, ColdChainMaintained and number of Category B boxes sold is much more in the
ColdChainNotMaintained are three transitions with proposed solution. This leads to the increase in net profit by
associated probabilities of 0.1, 0.8 and 0.1. This represent 10%. The total loss in $ terms is down by more than 90%
input variables where it is assumed that in 80% of the cases with the proposed system compared to the existing situation.
the cold chain is maintained as per the standard. Whereas, It must be stressed here that these results are for the assumed
10% of the cases the cold chain exceeds the required level, scenario within the ideal conditions defined.
and in 10% cases the cold chain is not maintained.

Simulation Results Table 2: Detailed Analysis of the Simulation Results

The modeled system was simulated for 365 days (1 year) of Parameter Present Proposed Change
operations. The results are shown in Table 1. These are at (%)
95% confidence level and 5% error. The simulation time of Boxes Sold 61032 61512 0.786
a year was considered enough to take care of any (S) (Category A, (Category B
fluctuations in the assumed scenario. Table 1 also shows the B and C) only)
relative change in various parameters in the two cases. As Revenue (S * 61032 * 120 = 61512 * 120 0.786
can be seen there is a considerable increase in the sale of $ 120) 7323840 = 7381440
Category B Boxes in the proposed scenario (more than Profit (S * 61032 * 20 = 61512 * 20 = 0.786
25%). Category B goods sold as percentage of total goods $20) (P) 1220640 1230240
sent also increases by 22%. Also the number of goods Loss Due to 5999 * $10 = 0
perished as percentage of total goods is decreased by more selling 59990
than 2.5 times. Category C
(1)
Table 1: Simulation Results and Relative Change in Loss Due to 6061 * $5 = 0
Parameters Selling 30305
Category A
Parameters Present Proposed Relative (2)
Case (E) Solution Change Loss Due to 220 * $100 = 87 * $100 = -60.455
(N) (N/E) Perished 22000 8700
Total Boxes Sent 69231 71224 1.029 Boxes (3)
by Vendor (X) Loss Due to 220* $20 = 87 * $20 = -60.455
Sold Category B 48972 61512 1.256 Loss in Profit 4400 1740
(Y) because of
Sold Category C 5999 - - Perished
Sold Category A 6061 - - Goods (4)
Perished (P) 220 87 0.395 Total Loss 116695 10440 -91.054
Left (L) 7979 9625 1.206 (L=
% Category B sold 70.7 86.4 1.222 1+2+3+4)
(100 * Y/X) Net Profit (P 1103945 1219800 10.495
% Perished (100 * 0.318 0.122 0.384 –L)
P/X)
Category B Sales (% terms) have increased by 22%,
perished goods have decreased by 60%. CONCLUSIONS

Analysis of Simulation Results The proposed solution based on TTI technology to manage
Perishable goods inventory will result in considerable
As can be seen from the results in Table 1 above, the savings. In the assumed scenario in this paper, the results
proposed solution decreases the amount of goods perished indicate an increase in net profit of more than 10% and a
by 60% over the existing scenario. In addition it increases reduction in loss of perishable goods by more than 90%.
the sale of Category B goods by 22%, while there are no Hence we propose that the retail stores should seriously
Category C or Category A sales. consider the proposed solutions to increase profits and
decrease enormous amount of wastage that occurs due to the
Assuming the cost price of a box at the retail to be $100 and existing supply chain management of perishable goods.
selling price to be $120, i.e., on each box $20 profit is
assumed. Also let the loss of customer goodwill for selling REFERENCES
each category C box is (say) $10. Further, assume the loss
due to selling of each Category A Box to be $5. Based on Ajmone M. and G. Chiola. 1987. “On Petri Nets with Deterministic
these assumptions Table 2 gives the benefits of the proposed and Exponential Firing Times”, LNCS 266, Springer Verlag.
solution over existing solution in Net Profit Terms for the Ajmone M., G. Chiola and A. Fumagalli. 1987. “An Accurate
assumed scenario. As can be seen that although the total Perfromance Model of CSMa./CD Bus LAN”, LNCS 266,
number of boxes sold in the two cases is almost same, the Springer Verlag.
Bhushan N. 1997. “Performance Evaluation of an Air Defence Engineering from Indian Institute of Technology (IIT),
System Modelled as a Petri Net”, Proceedings of Symposium Madras, India and other in Software Science from Allahabad
on Systems Analyses for Defence, CASSA, Bangalore, India. University, India. He has 12 years experience in research
Haas P.J. and G.S. Shelder, 1989. “Stochastic Petri Net and software development related to strategic decision
Representation of Discrete Event Simulation”, IEEE
Transactions on Software Engineering, SE-15(4).
making, modeling & simulation, and networking. His
Peterson J.L. 1977. “Petri Nets”, ACM Computing Surveys, Vol. 9, research interests include Performance Evaluation, Mobile
No. 3. Computing, Software Engineering and Artificial
Taoukis P.S., M. Bili and M. Giannakourou. 1998, “Application of Intelligence. His email id is [email protected].
Shelf Life Modelling Of Chilled Salad Products to a TTI
Based Distribution And Stock Rotation System”,
International Symposium on Applications of Modelling as
KISHOR GUMMARAJU is a senior consultant with the
InnovativeTechnique in the Agri-Food Chain. (Ed.) L.M.M. Domain Competency Group (DCG) of Infosys Technologies
Tijskens and M.L.A.T.M. Hertog, NetherLands. Limited and has over eight years of professional experience
Trivedi, K.S., and H. Sun. 2001. “Stochastic Petri Nets and Their in the areas of supply chain management, retail and finance
Applications to Performance Analysis of Computer across a range of industries such as Chemicals, FMCG
Networks”, Recent Developments in Operational Research, and Foods. Kishor is a post-graduate in Management Studies
(Eds) Agarwal M.L. and Sen K., Narosa Publishing House, from Indian Institute of Management (IIM), Lucknow. His
New Delhi, India. interests include Performance Management in
Supply Chains and Supply Chain Optimization. He may be
reached at [email protected]
BIOGRAPHIES

NAVNEET BHUSHAN is currently working with Software

Engineering and Technologies Labs (SETLabs), Infosys
Technologies Limited as Senior Research Associate. He
holds two Masters degrees- one in Computer Science and
 The Current Scenario Apply a The Future
Apply a Barcode with
Vendor Barcode with Vendor “Sell By Date” Find cause &
“Sell By Date” and a TTI take action
Maintain the Manage the Find cause &
Transportation cold chain
Transportation cold chain take action
Potential Vendor rating Update the Inform Vendor
breakage in Manage inventory systems to Sell By date & transporter
Warehousing the cold chain Warehousing monitor vendor based on in case of
based on FIFO or
performance remaining life reduced life
the “Sell-By-Date
Maintain the Manage the
Distribution Distribution cold chain
cold chain
Manage inventory
Manage inventory Retail Category based revised “Sell-
Retail Higher shrink based on FIFO or Retail Management By-Date”
Sale of unsafe product systems to help
the “Sell-By-Date”
Sub Optimal use of available life decisions on
pricing.
Figure 1: The Current Scenario in the Supply chain of
Perishable Goods Figure 2: The Futuristic Proposed Scenario
Update
Inventory Inventory
Information
movement based
on remaining
life.
New label printed
to enable sale at
New price and
an optimal price
sell by date
information
Category
Management System
Figure 3: Integrating TTI and Wireless LAN Technologies
 CategoryASol CategoryBSol
BoxesWithVendor CategoryCSol
GoodT Perished
oSell PutOn
BoxesSent
Shelf
Requested 20 ExceedSellByD
Box G
BoxesN
eeded
VeryWellMaintained p = 0.1
Transportation
SellByDate 20
ReachingStor G Demand
ColdChainMaintained p = 0.8
CheckColdCh TotalPerish
10 ActuallyPerishe
G
ColdChain ColdChainNotMaintained
p= 0.1
DemandCreated
Figure 4: Petri Net Model of the Perishable Goods Supply Chain

 
 
  

   

 

 

   
  

 


 

 
 
  
  
 
     
 ! "#$ %"&&
'
 ( (

)
*

 +

 (
 

   . 3
  !
 
(
    (
 
  ,



  , 
  )(
 (
 

 -.. 
 /
   

  3  
 , ,
 (

     (
   ,
 
,  
:
 
 (
,
 
 
     (   
,

( 
:(


 
,
       )   ,(     2,
 
 (   )(
 (
 

 ( 
,  ,
 


, 


(  .
  ( 
 (
  
 (
  ( ,
     . 

0
 ,Æ  1 ( , 
 
, 
 0( (
 (
 , , ( )


(
 2,., 
 .
,    (
 
  ,
 
,


 
 (
 
 . .
  

 )   ,Æ   . (   
)
 .

  ( ,   ( 

   
 
 

. ,(
( 
 
, 

. 
  0 !
 

 ( 1  !)
  (, ( ,
  ,  ,

     ( (

 
    )

 2,
  (  0 

.  4

 0

 ,
 (  !
  . (  ,   0((  2,


)

,, / 3,(  
(  
  


  , ( ,
  (
( 
  ( 3  4(  (
 (
 , ,
 ,  ,

  (
.
 0( (   )(
 (
 
)    ( ,
  (
   )

   .  
 ,  5

  0( (
 (
  , ,
 2, ,.  ,
0 ,  ( .
    (
 /  (    6   2,
  )
, . 
  6
 
 7
  (
  (   0(   
8 79 
,
 (-


   2,
 
( 
 
,   
.  ( ,)
8-
9 
 , ,  
 0(
 

,  2, 3,(    .
     0( (
 (
 , ,
 2,



,  , 

  ,(
1 (

4     (  ,  

 
 ( (
  
 (
  (
 (
 

)
 (

:  , 
  ( ,    .

 (   

7 !

 
,
  ) .
,( 6
 ( ,
 

 /
  , ( 
 
   (





(   ,  
0  (  
7)
 8
;)  (  . 
<
 <9  (   (   
8=>
)
 ?@@A9 ( 
  ,)     
 ,  . (  , 
  ,  .  
, 
  ,

  (
  .    .

 .,  ,
  


  ( 
 )(
 (
   0( (
 (
 



(
 

  ,   
 ( 
,


(   (

(  .
  (

 

 .,  0( ,
 (
 
 (
  
 (

0
 ,Æ  1
   0( (
 
   3  ( ) ( , 
  ( , 




(     ,  ,. 2, (
 2,., 
 .
, (, 
( 
 ( ,   (   )(
 
,  ,
 
 .,  (
(
 
 . .

,   

(
 ( (
    
 . ,

3
10
L

.      2,

  (  ← Critical point

  
 (   0( .0 
2
10

 2,
  4
   (
 
  ( 
,
 
 . .




1
10 G+L+S


  (  ,  (, 
 (

ρ (kg/m )
  (     , 

 

3
0
10 G+L

 
 (
 
 . .
 .  
−1
10 G

  .

 
(


   )
2,  (  ,  
,  ,
 −2
10 G+S


 .,   
 (
  )
 (   , 
 . .
 (,( ) −3
10
−500 0 500 1000 1500 2000 2500 3000


 /
  ( 
 -..  
h (kJ/kg)

,  
,  ,
 
 .,)

   
 (
 ,  (
, 3, ?* ,
    )(
 (

 

.
  ,  (  
 ) 

  , 0
E - * 
E D * 2,E * 
(

,  
,  ,
 
,
 ., ( /
  ,. 
 
 ,  
 (
 .  ) 0 

.   . 
,
   

   

.  6 
  (  (
  
 
 (
 ( 

.

) (   (  )(
 ( 
,
  ,  
  0 (

(
 

   (  .   
 ( 

.
 ( 
,   ,


 )(
 (
 

     ) ( 
   (
 3
  
  (
.  
,
 ( 
,  ,
 (
 (
 ( 
,
  ,
 0


  (  
 (
  (   ( 

.
 ( 
  0  
( ,  
,
 ( 2,.,
)
         1 . , -..    

 ., 
     
  ,  (   )(
 (
 
)
( (  (  )(
 (
 

  
  , 0

   1, $ , F 
.
   (  )(
 
  6) 
  (  1,    .  
( ( )
 /  (  /    ( , (
 
,
 


 6 (


/  (
 
  
 (           .  .  ( .,)
( 

  ,)
, (
 

 (
  1, ?  (  1, 
 ,  (   (
 (
 

 
 .  ( -..    0

,)
  
,
  , 3 0

 

8
= ?@@B9  8
= ?@@%9 
 . ,   
8'

 =! ?@A"9  ,  ( .,
 
( 6 /
 
 .0  
 )
700

(
 (  .,

 
,
 .
  (
600

500

(
 ,
 
 ,. 2, . ,  400

    0(( / (   ,

T (°C)

300

    
 (
 3, ?  ( .,) 200

  ( 6 /  , 0

 ( 
  100


 ( 

    2, 0
  2,) 0

., 0( 
 8-CD)

9 ( , 
 ( 
 3
2

(
 

:
  0
 


3000
1 2500
0 2000

.

 (
  (

   3 !

1500
−1 1000
−2 500

( .,
  ( -CD)

0( ( -)

)
0
10 3 −3 −500
log ρ (kg/m ) h (kJ/kg)

  (  

,  ( (
 
,
,
 

 ( / (
.  
,
 3, $* 
,
,   
 )
 , 
,
 
 .,  
() (
  , 0

0

 , (
 2,., 
 1 1

0.8 0.8

XS & Xsc (kg/kg)

0.6 0.6
XG (kg/kg)

0.4 0.4

0.2 0.2

0 0
3 3
2 2
3000 3000
1 2500 1 2500
0 2000 0 2000
1500 1500
−1 1000 −1 1000
−2 500 −2 500
0 0
10 −3 −500 −3 −500
log ρ (kg/m3) h (kJ/kg)
10
log ρ (kg/m3) h (kJ/kg)

3, "* -



,   
 3, F* 
 , 
 


,)
(
  , 0
    
 (
  , 0



 2,., 4     (  , 
 0( ( 2, ( .,
 .0
1 (

 ( 

 0   0
  
(
  ,
   (,
 1
,.   )(
 

 
 . 
0.8

0.6
  

 
, 
 
 
 
)
XL (kg/kg)

0.4 ,
 ( .,
  ( 

   (
 
 (
   ( 

0.2
,  1 (
 


 (
 !
)
0
 Æ 0 
,
  (



  ( 


3
2
3000
1 2500
0 2000
1500

   
−1 1000
−2 500
0
10 3 −3 −500
log ρ (kg/m ) h (kJ/kg)

   
 
 .  ,
   ( 


3, #* D2, 


,    ( (
 
0)! 
 ,   


 (
  , 0
 

( (/
   ( 

  ,


  (    
 (
 
 ,  . 
,
 3 (, , 
         ( 
 .  (,( 
  

 
    .

  3  
 , (
,)
  (
 2,., 0  . 
 

3  , ,
 

,  

 (,
   

 6
 2,
 )

 0


, ( 0
 
 ! 
  

  
  6


 G, 0

  ( , 
 !   ,)  ,  .   .
  
 (


 0( 0



 ( 
 .0    ( , (  6
 2,

0

 , 
,  
  ( .  , ,
 
  
 (
    ,  (

    ( /   ( 

 

 ( 1  ( 
 . 

,  . 
 0   
 ,.)   (  2,
  (   (
 (


 (
  ( ,
 ,Æ   . 

  

3
 ,., (
( 0(   3 )
    ( ) ,  ( ,.,  ( 
 0
   


 
 .     (
: (
 

 0
 8
 (
 9
 

/ ,  (  
!  ( 

  ) 
 ,  ( , 0(( 0  (
   ( 0
   0( ( (

! 

  .

  ( ,,  (  ,
   ( (
  ( 
 , 80

 ( 
,
 2,., ( 


,
 9 (,  
 .
, (
 ( ) 
 2, 0

  ( 

  )

0 ( (
  ( 
 ,  .  () , .
  ( 
 
 , ( 


 ,  ( ,
 ( 

 , 
.
  ( 
 
  0 !
 
) 100


 0 .  ,   ( 0
80

60

40

T (°C)
   !"# "$"% 20

−20

( 1  !
     ( (
 
  −1.22 −1.2 −1.18 −1.16 −1.14 −1.12 −1.1 −1.08 −1.06

  
 0

 ,
 ( ,.

h (J/kg) x 10
7

 (
  ( ,

 ( 6  ( ,
 10

 ( .
 /   ( 0

  8

 ( 

   2,  (  
 ) 6

P (mbar)
 ( (
  (     
() 4

/
   ( 

   . (   ( 2

0 6
 2,
 ,Æ* 0
−1.22 −1.2 −1.18 −1.16 −1.14
h (J/kg)
−1.12 −1.1 −1.08 −1.06
7
x 10



H & 8?9

 3, B* 
,
  ,
,  )


Û (




H  8$9
=( 5 8¿ 9  2,
  ( ,  (   < 1
vapour

8:9  2,
  ( 
 0
 


 2
liquid
0.8 ice

8=9  2,
  ( (
 , (  ,
   X (kg/kg) 0.6

1, B
 % (  



,  0.4

)$F7
 ( 

, ( 0
       0.2

2,., 0( 0
 
, ( 
  ( ,
  0
−1.22 −1.2 −1.18 −1.16 −1.14 −1.12 −1.1 −1.08 −1.06

 (     


 , 


) h (J/kg) 7
x 10

  ! 0(  ,

 
 ( (
 1

 , 

,
 
,  ,   ) 0.8

  ( ,.

 

 2, ( 0.6
X (kg/kg)

,   

  =( ( , 
) 0.4

, 8)@F79  
( ( ,  
   0.2
solid sugar

(  ,
    ( 2, 0
 , 0
−1.22 −1.2 −1.18 −1.16 −1.14 −1.12 −1.1 −1.08
pure ice

−1.06

(  ( 

,
 
,  , ) h (J/kg) x 10
7


  
 =(
 (  ,
    (
   ( ,  
    , 
, 3, %* 4


,  (

 ( ,
 , ( , 
, 
,  8
, 2,
  
 .
  

,
( /     , 

, 0
E  ,
 
 .
  

, ,
E
  , ( 
 /  ( 0(  ,  
 .
  

, 9
 ( ,  , ( 
 / 

 
 
   


 
(
 

 3,( 
  ( (
  ,  ( 0 6
 2,
 ,Æ*
 

  ( 2, 0


 
 ) 

  ( ,
 , (  

H  8"9
( 0

 =( ( 
!, ,.  (

,
  
( ( ,
 
   
,  

Û
 ( 
 

,  0
  (  



H  8#9
( 0
  




 0 
,

 
  (
  , 
 
 =(  8: 9  2,
  ( 
G0  ,)
 
, ,  ( 0 (
 

  0
  
,,
 ( 8:9  2,
  ( (
 

,  (  (       0
 
, (  
 (  ,
   1, A


  ,
 @ (  



,
, #F7
  
 
 
  (   

 &&% '( (
  (     ( 
 (
 
   
 ( 
,  ( , )
(  !
     
,, /  ( 
 ,  ( 

  0

 ( 


  
( 1  !
   . (  ) / 
, ,  
   , (
, 
,  
( (   

 .  . . 0 6
 2,
  ( 
( 
, ,
 ( 0

 ,
   2,

   ( !  (
 
( 
,  ,( ,  ( ,.
 (   (   )(
 

 (

 ( 
 (   (  ??"I  ( 
 .   ( 3  4( 

,  0
 (
.   .  
 ,  0    .
   , 0(    ,
.
    ( 
 ,, (   ,
0 . ,. ( 5

 0 ,  ( .

50

40

30
   (
 / , . 
 
20
6
 
 7
 87 9 
,
T (°C)

 (-


  8-
9
10

−10

−20
0 50 100 150 200 250
t (−)
300 350 400 450 500

J?K '
 =

 =! ?@A" L3,

50

 ( (

   (
,)
40

30

 (
  '$  ?%"?F ;  #%"?F ;L
P (mbar)

20

10


 
 
 5 A@  $
 F&&)F?@
0 J$K 
=  ?@@B 
   
  
       
  
0 50 100 150 200 250 300 350 400 450 500
t (−)

 
   !       " #
$
3, A* 
,
  ,
,   
= 

 3
 

, 
,, / J"K 
=  ?@@% 
   
  
 
  %      &
 
    $ 
= 

 )
1
vapour

 -

liquid

J#K =>
)
 E 7 5
E  

0.8 ice

 =E
 7 ; )  ?@@A L4 
0.6
X (kg/kg)


  D
  
 ,
0.4

0.2

0
L
 $ #%B  '  $&F)$?$
0 50 100 150 200 250 300 350 400 450 500
t (−)

0.8

0.6
X (kg/kg)

0.4

0.2
solid sugar
pure ice
0
0 50 100 150 200 250 300 350 400 450 500
t (−)

3, @* 4


,   8
, 2)
,
  
 .
  

, 0
E 
,
 
 .
  

, ,
E , 

 .
  

, 9

     

 




. 7)

 8
;<
 <9  ( ,

 (
    0( (
 
   (
,   (   )(
 (
 

 )

 (
,
 
 
  

 
,
   (    (


!
 (0 (
 
 !   

 Changing The Game: Systematic Innovation in Food
Engineering Using TRIZ and Function Simulation Tools

Barry Winkless
AMT Ireland, University College Cork,
Tel: 00 353 (0) 87 9720544 Email: [email protected]

Darrell Mann
University of Bath, UK
Tel: +44 1225 826465 Email: [email protected]

Dr. Barry O’Connor

AMT Ireland, University College Cork,
Tel: 00 353 (0) 21 4902549 Email: [email protected]

“Developing technological The Russian initiated Theory of

competence through process Inventive Problem Solving, TRIZ, is a
innovation is more than a shopping philosophy, a series of tools, methods
expedition for new equipment; it and strategies developed through over
requires the systematic assessment, 1500 person years of research and the
exploration and development of study of over two million of the
capabilities” Bessant (1994:57) world’s most successful patents.

Introduction: The key findings of TRIZ research

are:-
The following paper demonstrates the • that all innovations emerge
application of TRIZ (Theory of from the application of a
Inventive Problem Solving) in very small number of
improving food processing inventive principles and
technologies. The paper begins with a strategies
description of TRIZ and it’s five • that technology evolution
‘pillars’. A functional simulation of a trends are highly
convection-based French fry process predictable
then follows. The TRIZ tools of • that the strongest solutions
Functional Analysis, Contradictions, transform the unwanted or
and Technology Evolution are utilised harmful elements of a
within this analysis (using system into useful
CREATRIZ, a computer-based TRIZ resources.
application). The paper concludes with • that the strongest solutions
a description of viable concepts and also actively seek out and
improvements for the selected food destroy the conflicts and
process, and highlights a need for trade-offs most design
continued research in this area. practices assume to be
fundamental.
TRIZ: The theory of Inventive
Problem Solving TRIZ offers users access to the
knowledge and experiences of the
world’s finest inventive minds. It is
intended to complement and add unnatural at first. The next section
structure to our natural creativity rather provides an overview of the major
than replace it. elements of TRIZ, before the paper
TRIZ is both simple and complex. To moves on to examine the use of TRIZ
learn and gather a working knowledge to a pair of food industry related
of the whole structure will probably problems.
take six months. Some people are
prepared to make this investment, and THE FIVE PILLARS OF TRIZ
others are not. Those that are not can
take comfort from the fact that they 1500 person years of TRIZ research
will be able to learn and realise have produced a significant number of
significant benefit from just a short innovation tools and methods. Figure 1
exposure to individual elements of the illustrates the five main elements of
overall structure. TRIZ that distinguish it from other
innovation and problem solving
TRIZ is different to most other strategies.
creativity aids, and quite appear a little

TRIZ

Space and Time

Contradictions

Functionality

Resources
Ideality

Figure 1: The five pillars of Triz:

Contradictions contradiction part of TRIZ is that,

TRIZ researchers have identified the given there are ways of ‘eliminating’
fact that the world’s strongest solutions contradictions’, problem solvers
have emerged from situations in which should actively look for them.
the problem solver has successfully
sought to avoid the conventional trade- Ideality
offs that everyone else has taken for While studying the patent database,
granted. More importantly they have TRIZ founder Genrich Altshuller
offered systematic tools through which identified a trend in which systems
problem solvers can tap into and use always evolve towards increasing
the strategies employed by previous ‘ideality’ and that this evolution
successes. The most commonly applied process takes place through a series of
tool in this regard is the Contradiction evolutionary S-curve characteristics. A
Matrix – a39x39 matrix containing the key finding of TRIZ is that the steps
three or four most likely strategies for denoting a shift from one S-curve to
solving design problems involving the the next are predictable. A number of
1482 most common contradiction underlying technology evolution trends
types. Probably the most important consistent with the ideality concept
philosophical aspect of the have been identified during the course
of research on the global patent harmful. In a heat
database. Used as a problem definition exchanger, for example, the
tool, the ideality part of TRIZ MUF is to transfer heat to
encourages problem solvers to break the working medium;
out of the traditional ‘start from the everything else in the
current situation’ type of thinking, and system is there solely
start instead from what is described as because we don’t yet know
the Ideal Final Result (IFR). The how to achieve the MUF
simple definition of IFR is that the without the support of the
solution contains all of the benefits and ancillary components.
none of the costs or ‘harms’ (Systems may of course
(environmental impact, adverse side- perform several additional
effects, etc). Although there are many useful functions according
instances where systems have been to the requirements of the
seen to evolve all the way to their Ideal customer.)
Final Result, many have not. The 2) In traditional function
method gets users to think about these mapping, the emphasis is
situations by working back from the very much on the
IFR to something which is practicably establishment of positive
realisable. Generally speaking these functional relationships
solutions incorporate the concept of between components. TRIZ
systems solving problems ‘by places considerable
themselves’. The key word is ‘self’; emphasis on plotting both
things that achieve functions by the positive and the
themselves – self-cleaning, self- negative relationships
balancing, self-heating, self-aerating, contained in a system, and,
etc – all represent, when incorporated more importantly, on using
in a true TRIZ fashion, very powerful the function analysis as a
and resource-efficient solutions. means of identifying the
contradictions, in-effective,
excessive and harmful
Functionality relationships in and around
Although the functionality aspects of a system. Function and
TRIZ owe a significant debt to the attribute analysis of systems
pioneering work on Value thus becomes a very
Engineering, the method of defining powerful problem
and using functionality data is definition tool.
markedly different; sufficient at the 3) Functionality is the
very least to merit discussion as a common thread by which it
distinct paradigm shift in thinking becomes possible to share
relative to traditional occidental knowledge between widely
thought processes. Three aspects are differing industries. A
worthy of particular note:- motor car is a specific
1) The idea that a system solution to the generic
possesses a Main Useful function ‘move people’,
Function (MUF) and that just as a washing powder is
any system component a specific solution to the
which does not contribute generic function ‘remove
towards the achievement of solid object’. By classifying
this function is ultimately and arranging knowledge
 by function, it becomes being used to generate beneficial
possible for manufacturers effects in a number of new product
of washing powder to developments from vacuum cleaners,
examine how other commercial bottle-washing systems,
industries have achieved the and in helping to empty vessels
same basic ‘remove solid carrying powder-based substances
object’ function. ‘Solutions more quickly.
change, functions stay the
same’ is a message forming
a central thread in the TRIZ Thinking in SPACE and TIME
methodology: People want While not strictly speaking a TRIZ
a hole not a drill. development, TRIZ researchers have
also recognized the enormous
importance of thinking about situations
Use Of Resources from all angles. TRIZ users are
The Resources part of TRIZ relates to continuously changing their
the unprecedented emphasis placed on perspective on problems – zooming in
the maximisation of use of everything to look at the fine details, zooming out
contained within a system. In TRIZ to see the bigger picture, and thinking
terms, a resource is anything in the about how the situation is affected by
system which is not being used to its changing time – whether that be nano-
maximum potential. TRIZ demands an seconds or decades – in both the past
aggressive and seemingly relentless and future. This is not a natural process
pursuit of things in (and around) a for most people – our brains aren’t
system which are not being used to wired that way – and so TRIZ contains
their absolute maximum potential. tools to help in the process of thinking
Discovery of such resources then in TIME and SPACE.
reveals opportunities through which
the design of a system may be All of these TRIZ pillars are brought
improved. In addition to this relentless together from a problem definition
pursuit of resources, TRIZ demands perspective during a function and
that the search for resources also take attribute analysis (FAA) of a system.
due account of negative as well as the Several attempts have been made by
traditionally positive resources in a TRIZ researchers and suppliers to
system. Thus the pressures and forces integrate function analysis/value
we typically attempt to fight when we engineering techniques into TRIZ
are designing systems, are actually software-based problem definition
resources. By way of an example of processes. To date, these attempts have
this ‘turning lemons into lemonade’ largely failed to deliver much by way
concept, TRIZ users often think of of user benefit over non-software
resonance as a resource. This is in based methods. This has largely been
direct contradiction to most Western due to the fact that software has been
practice, where resonance is commonly written to do little more than mimic the
viewed as something to be avoided at manual methods. In effect, the
all costs. TRIZ says that somewhere, potentially enormous leverage to be
somehow, resonance in a system can gained by harnessing the power of the
be used to beneficial effect. In effect, computer has remained largely
resonance is a potent force lever untapped. Problem solvers have seen
capable of amplifying small inputs into the consequences of this failure in
large outputs. Resonance is currently three main areas:-
 above problems and to offer users
1) an inability to take adequate powerful new ways of managing the
account of the attributes complexity contained within the large
(for example, weight, majority of problem situations. The
volume, surface finish, paper now examines application of the
corrosion resistance, etc) of system to the first of two case studies:
a system component. The
consequence of this is that
main often key functional 1. Towards a better convection
relationships are not drier:
identified in models.
2) an inability to take due The following case study focuses on
account of the time issues the problems encountered in drying
surrounding a problem. The French fry potato chips without the
essential consequence of extra addition of oil, in a high velocity
this is that the user does not convection cavity. Machine X
have any value-adding consisted of an air blower, a heating
means of defining how a element, a cavity, a filter a time-
functional model changes dependent release mechanism, and a
with time-based elements of sensor feedback system. The major
the problem, and thus problem with Machine X was the
receives no guidance on relatively frequent need to have it
how time-based problem cleaned. Cleaning was also difficult
solving strategies can be due to the design of the inner cavity in
applied. This is particularly which the fries were subjected to the
Important from the high velocity convection stream. It was
perspective of decided at an early stage that any
3) a total failure of the changes should focus on improving the
software to direct a user cleanliness of the machine and if
from the function analysis possible require no addition of
model to the appropriate expensive or even cheap parts!
TRIZ solving tools.
The first stage in the project involved a
The paper has employed the novel functional analysis, using CreaTriz
function analysis/value engineering software in order to ascertain the key
modelling technique to be found in the problem areas. Figures 2 and 3, below,
CreaTRIZ problem management illustrates the functional analysis
software. The new model has been diagrams:
designed to overcome each of the
 Figure 2: Functional diagram A

Figure 3: Functional Diagram B:

Functionality: Harmful and Useful surfaces. The oil itself carries out the
functions: useful function of improving heat
transfer and mouth-feel of the chip but
What the function analysis makes clear incurs the harmful function of being
is that the forced air within the system the major cause of dirt in the machine.
carries out both harmful and useful
functions. Examination of the useful From a system analysis perspective,
functions highlights the presence of 1) the functional analysis diagram can be
heating of the oil by the air, 2) heating used not only highlight the problems
of the fries, and 3) agitation of the fries contained within the system, but also
in order to improve heat transfer and categorise them into different types in
cooking properties. It also incurs order to facilitate the selection of the
harmful functions- predominantly the most appropriate TRIZ (or other) tool
spraying of oil from the French Fries to help tackle that problem. In this
onto machine walls and the over particular case, for example, the model
agitation of chips (shown as an highlights the existence of a number of
‘excessive’ action) causing more contradictions in the system. As
widespread distribution of oil and previously mentioned, the description
potato crumb debris, which in turn of problems as contradictions is a
increases the difficulty of cleaning the powerful tool in TRIZ as it enables the
machine after use. engineer to extract powerful
contradiction-breaking solutions from
The temperature within the system the most successful inventive solutions
obviously has the useful function of of others who have successfully
heating and cooking the fries, however overcome similar problems. The
there is also the harmful function of contradictions and suggested solutions
causing the oil to liquefy and again for the fry-heating machine were
cause deposition of oil onto machine identified as follows:
 Figure 4: Possible Solutions to Hot air flow problems:

Contradiction 1: (with reference to Where Worsening Factor is: Loss of

the functional analysis diagrams) substance (where the oil is being
Results for Improving Factor: heated and therefore spraying from the
Temperature (Hot Air is Heating the French fries).
French Fries) Results: Refer Figure: 1.2:

Figure 5: Principle 21-Hurrying

Many other patents, for example US the constraints placed upon design
Patent 5, 997, 938 ‘process for changes it was felt that such a drastic
preparing improved oven-finished change would incur too many time and
French fries’ conduct French fry cost penalties whilst only offering
heating at very high temperatures over small benefits to the system as a
a very short time in order to minimise whole. Possible harmful effects from
oil spray and potential damage to the hurrying the process also included: the
French fry. Present cooking time in need for better containment, powerful
Machine X was 60 seconds and due to heating mechanisms, better insulation.

Figure 6: Principle 36-Phase transitions

The application of this principle to the process for finishing French fries’
solution of the problem may at first highlights- ‘after cooking the fries are
seem obscure however if focus turns to optionally coated with oil’.
the French fry-oil could be added after Alternatively focus could fall to the
the chip has been heated- the heat of oil- the type used (sunflower oil at
the French fry (i.e. loss of heat of the present) and how the temperature of
French fry) could heat the oil to a the air affects the oil. Possible harmful
certain extent and therefore increase effects from phase transitions: the need
the homemade/just fried taste and to add oil, expensive research into oils,
texture of the fry, as US Patent 6, 013, different flavour profiles.
296 ‘forced air convection oven

Figure 7: Principle 29-Fluid

Many convection-based patents use and would involve costly

moisture within a heat stream (such as redevelopment. Also texture and taste
US 6229131, 6188045, and 5942142) profiles differ from traditional or
in order to improve heat transfer to impingement ovens without moisture
food products. For the purposes of used in the process.
Machine X, this is too major a change

Figure 8: Principle 31-Hole

An interesting solution to the oil spray hold the oil better in process. Research
problem could be the introduction of would have to be carried out in order
small holes to the chip as a means to to test the viability of this concept and
results may not be satisfactory. For the Results for improving factor:
purposes of the Version 1 machine this Force/Pressure: Agitation of chips
may be moving away from an ideal increases heat exchange and overall
solution. heating effect.
Worsening factor: Duration of moving
Contradiction 2: object (the chips are moving around
too much and for too long).

Figure 9: Results for contradiction 2:

One of the major problems within . However a cage would probably only
Machine X was the propagation of oil have a limited effect on oil spray
in the machine making it difficult to within the system- cleaning would then
clean. Periodic action could solve this also include the cage. ‘Take Out’ also
problem. By pulsating the air in the highlights how the system could pulse
system, the French fries will not move air- the taking out of parts of the
around within the system as much as blower in order to create a pulsed air-
they would with a steady flow, nor will flow. ‘Take Out’, could also refer to
they cause as much oil build up at the removing oil completely from the
filter. This is due mainly to the fact system..
that pulsed air will not push the French
Fries to such a great height. This The results of the TRIZ study focused
solution in itself moves the machine development efforts on moving
towards a more ideal system because it towards a more ideal system, i.e. the
will not add expense to the pulsing of the air, which in turn had the
manufacture of the machine, and will benefit of decreasing the size and cost
negate to a large extent the need for an of the filter. This solution is also in
expensive filter. ‘Take Out’ highlights agreement with one of the TRIZ
solutions that have been utilised in technology trends1 which highlights a
other French fry cooking patents, general trend towards pulsation as a
where a small cage or basket is used to more ideal system:
contain the movement of the chips

1
Triz technology trends refer to a number of
trends technological systems follow in the
course of their development. Many of these
trends were identified by Geinrich Alsthuller,
creator of the Triz methodology.
 Figure 10: From Jet to Pulsating Jet-One of the Triz technology trends

Conclusions

This case has demonstrated a small Application of TRIZ to food

proportion of the capabilities of the engineering type issues – whether they
TRIZ systematic innovation be improvement of existing processes
methodology. The simulation of or conceptualisation of new products –
systems through analysis of function is is still at a relatively early stage of
shown to offer new perspectives on evolution. The opportunities for
both the management of complexity achieving tangible benefit in both areas
within systems, and the application of are believed to be considerable.
problem solving tools aimed at
providing much stronger solutions than
would emerge from traditional
optimisation and trade-off design
strategies.
BIOTECHNOLOGY
 MODELISATION OF KILLER YEAST SYSTEMS: DETERMINATION OF MIXED
CULTURE DYNAMICS WITH A SPECIFIC MEMBRANE BIOREACTOR

Sébastien Pommier
Claire Albasi
Pierre Strehaiano
Marie-Line Delia.
Laboratoire de Génie Chimique, CNRS UMR 5503, Equipe Fermentations & Bioréacteurs.
INP ENSIACET, 118 route de Narbonne, 31062 Toulouse Cedex, France.
E-mail: [email protected]

KEYWORDS Actually, biomass quantification of each of the strains is hard

Killer Yeast, Mixed Culture Dynamics, Membrane to carry out, especially when microorganisms from the same
Bioreactor, Fermentation Modeling. species are involved. Though it is quite easy to estimate the
global biomass concentration (by spectrophotometry,
ABSTRACT microscopic cell counting, etc.), the main difficulty is to
differentiate one strain from another. Usual techniques
A reactor specifically designed to carry out studies of required for such a differentiation are microscopic
interacting populations was developed. Its basic principle observations, agar plate counting and DNA analysis. These
consists in keeping both strains of the mixed culture separate techniques are very fastidious and sometimes expensive. A
by the way of a porous membrane, allowing substrates and specific membrane bioreactor made it possible to pick up
metabolites to pass through. As the different strains grow new fundamental information for a better understanding and
separately, they can be sampled and analyzed distinctly. The a modeling of the killer interaction and its consequences on
culture medium is shared by both strains, so far as population dynamics in Saccharomyces cerevisiae cultures.
interactions between them occur as if they were mixed in the
same vessel. The membrane bioreactor made it possible to
MATERIAL AND METHODS
carry out mixed killer/sensitive fermentations with a double
physiological characterization of both strains and thus to
The New Membrane Tool
compare a mathematical model for killer yeast interaction
dynamics to reliable experimental data.
The experimental membrane bioreactor was designed in
INTRODUCTION order to make it easier to get instantaneous information
about each of the microorganisms involved in two-strain
Many fermentation processes are carried out by various mixed fermentations. The two strains grow separately but
microorganisms, in particular those taking place under non- share the same culture medium. Therefore it is composed of
sterile conditions such as winemaking. Interactions between two jars interconnected by a hollow fiber-membrane module,
these microorganisms is a key parameter for the control of immersed in the liquid of one of the jars (see Figure 1).
the bioreactions. One of the most famous exemples of Fibers are U-shaped and they are hold together in their upper
interactions between two microorganisms is the killer part with an epoxy resin. The bundle is contained in a
phenomenon. This phenomenon was described for the first stainless steel receptacle.
time about 40 years ago (Bevan and Makover 1963). It The pore size of the membrane fibers (0,1 µm) allows
occurs between yeast species sharing the same culture metabolites and substrates to pass through while keeping
medium: one strain called "killer" has the property to excrete each of the microorganisms separate. Medium flow and
a toxin that affects another strain called "sensitive" by mixing between the two jars are induced by alternatively
damaging its cell membrane and in the end killing it. This applying pressure into the headspace of each vessel. Level
phenomenon can influence the quality of the final product as control by conductivity probes drives electrovalves defining
killer strains tend to become dominant and thus to set alcohol the medium way of circulation. An extra benefit of the
level and specific flavors. alternate flow is the control of fouling by backflushing.
In order to model, control, and optimize mixed fermentation The main constraint of the bioreactor validity is that
processes involving killer interactions, it is necessary to be dynamics are governed by the reaction rates and not by the
able to perform kinetic and physiological characterization of mixing rates. The device performances have been checked
each of the protagonists. Modeling is mainly restrained by by comparison with classical fermentation methods (Salgado
the lack of experimental data about mixed culture dynamics. Manjarrez et al. 2000). Neither problem due to mixing times
nor incident linked to fouling has been noticed.
 Numerical Method

Smoothing of measured populations was established by

spline functions (Reinsch 1967) in order to estimate the
Air 1 dX
specific growth rate µ = .
level X dt
electrovalve controller
The proposed mathematical model was solved by an explicit
manometer
Runge-Kutta formula, and parameter identification was
sterile bridge
air filter obtained using a simplex method developed on Matlab.
sampler
level
probe RESULTS AND DISCUSSION

The physiological characterization of each strain was

performed thanks to the membrane bioreactor and to the
hollow
fibers separation between the sensitive and killer cells. It was then
A
possible to assess the mixed culture growth kinetics and in
magnetic B
stirrer the end to compare a mathematical model to reliable
experimental data.
Figure 1 : Membrane Bioreactor Scheme
Physiological Characterization of Yeast Cells

Cultures The physiological characterization of the yeasts growing in

the mixed culture (dead cells or viable cells) is fundamental
Commercial yeast strains Saccharomyces cerevisiae K1 for chasing the killer effect. Two methods were tested :
(killer strain) and Saccharomyces cerevisiae S6 (sensitive intracellular ATP measurements and methylene blue
strain), both obtained from Lallemand Inc™, were used. staining.
Mixed fermentations were carried out in the previously As only viable cells contain ATP, it is possible to numerate
described membrane bioreactor. viable cells directly from intracellular ATP measurements.
A synthetic culture medium was used (Strehaiano 1984). This analytical method begins to be widespread and can be
Glucose was the carbon source, and ammonium sulfate the automated in plant-scale fermentation processes.
nitrogen source. Temperature level was set to 25°C, which is The traditional method for the assessment of yeast viability
optimal for the involved toxin activity. Inoculation of each is methylene blue staining and numeration in a Thoma
strain was performed in the separated vessels of the device. counting chamber. This technique is cheaper, but not
Total fermentation volume was 4 liters. appropriate for on-line measurements.
Population estimation was carried out through O.D. (620 nm, A good correlation between intracellular ATP concentration
Anthelie Advanced Spectrophotometer) and cell and viable population measured by methylene blue staining
numeration with a Thoma counting chamber. was established for each of the yeast strains (Figures 2 and
Methylene blue staining was used for the estimation of yeast 3). One can notice the ATP level per cell is 3 times higher in
viability. the killer strain K1 than in the sensitive strain S6.

Mixed cultures were carried out with an initial ratio of 10% According to these observations, methylene blue staining
killer cells for a total seeding at 3.3 million viable cells per was chosen for the physiological characterization of the
milliliter referred to the total working volume (4 L) . yeasts in the mixed culture further described. Thus both
viable and dead cells were counted from a single observation
through a counting chamber.
ATP Measurements

Intracellular ATP measurements have been performed, in

order to characterize biomass physiological state. ATP
quantification was assessed through bioluminescence.
A luciferine-luciferase enzymatic complex is used to react
with free ATP in the samples. Light emission by the
ATP/luciferin-luciferase reaction is measured through a
luminometer (LUCY Cδ life). Yeast cells are hydrolyzed
(Biofax A, Yelen) before adding the enzymatic complex .
 most of the sensitive population was characterized as dead.
4E-05
The end of the fermentation was performed by the single
killer strain.
intracellular ATP (mol/L)

4E-05
3E-05
3E-05 160 100%
2E-05 140
y = 3E-07x 80%
2E-05

Population (*1e6 cells/mL)

R2 = 0,973 120
1E-05

Viability (%)
100 60%
5E-06
80
0E+00 40%
60
0 20 40 60 80 100 120 140
40
20%
Viable cells(*1e6 cells/mL) 20
0 0%
Figure 2 : : Relation between viable cell counting (Thoma 0 5 10 15 20 25 30 35
t (h)
chamber) and intracellular ATP concentration for
total sensitive cells total killer cells
S.cerevisiae K1 % viable sensitive cells % viable killer cells

Figure 4 : Evolution of total cells population and percentage

6E-06 of viable cells for both sensitive and killer strains growing in
the mixed culture.
intracellular ATP (mol/L)

5E-06

4E-06

3E-06
y = 1E-07x
2E-06 100%
R2 = 0,9616
1E-06 80%
% killer cells

0E+00
60%
0 10 20 30 40 50
40%
Viable cells(*1e6 cells/mL)
20%

0%
0 5 10 15 20 25 30 35
Figure 3 : Relation between viable cell counting (Thoma t (h)
chamber) and intracellular ATP concentration for
total cells viable cells
S.cerevisiae S6

Figure 5 : Evolution of killer/sensitive proportion in the

Mixed Culture Growth Kinetics
mixed culture
Microbial growth kinetics of each strain of the mixed culture
was established by simple analytical methods thanks to the
separation between killer yeasts and the sensitive ones.
Figure 4 shows the evolution of the killer and sensitive The killer toxin involved in the observed phenomenon is a
populations in the mixed culture. protein which has been identified as a glycoprotein of about
16 kDa (Pfeiffer and Radler 1984). The pore size of the
Estimation of biomass viability for each strain by methylene porous membrane of the bioreactor (0,1 µm) makes it
blue coloration made it possible to determine the killing possible for such molecules to pass through. The death of the
effect kinetics. The device made it possible to perform an sensitive cells in one of the vessels was due to the action of
accurate analysis of the ratio of viable and dead cells for the K2 toxin, that was produced by the killer yeasts in the
each population all along the fermentation. The biomass other vessel, but which concentration was homogenous in
evolution underlined a break in the sensitive strain growth, the whole reactor volume.
as the killer population quickly became dominant (Figure 5).
In the first 10 hours of the mixed culture, sensitive cells grew
at maximum viability level, and the killer/sensitive ratio Modeling
remained almost constant. During this period, both strains
grew just as if they were separate, without any lethal A mathematical model expressing the dynamic
interaction. After about 12 hours, the viability of the characteristics of killer/sensitive yeast mixed cultures was
sensitive cells began to decline. Less than 15 hours later, proposed (Ramon-Portugal et al. 1997, Alfenore 1999). It is
based on the quantification of the medium culture killer dX d , s
toxicity. The physiological state of both killer and sensitive = rd , s (11)
dt
cells is considered, thus biomass is segregated into 4
variables: viable killer cells (Xv,k), viable sensitive cells (Xv,s),
dT
dead killer cells (Xd,k), and dead sensitive cells (Xd,s). = rPT − rCT (12)
dt
An expression of the growth rate (rv,k ; rv,s) and of the death
rate (rd,k ; rd,s) is given for each of the strains, as expressed in dI
Equations (1), (2), (3), and (4) below. The terms µd,k and µd,s = rI (13)
dt
stand for “natural” death rate of killer and sensitive cells.
The growth rate is supposed to be limited by inhibition, and
is expressed by a logistic-like expression. The variable I is The new device made it possible to obtain the evolution of
standing for an unidentified inhibitor, which is common to both killer and sensitive strain growing together in the mixed
both killer and sensitive strains. culture, as well as a their physiological characterization
(viable or dead) by methylene blue staining or/and ATP
measurements. The collected data were confronted to the
rv , k = µmax , k . X v , k .(1 − Ak .I ) (1)
model.
Experimental values of Xv,k , Xd,k , Xv,s , and Xd,s were used for
rv ,s = µmax ,s . X v,s .(1 − As .I ) (2)
the identification of parameters in Equations (1), (2), (3), (4).
Variables T and I cannot be experimentally measured, thus
rd , k = µd , k . X v , k (3) parameters α, and a in Equations (5) and (7) have been
arbitrary set to 1. Any changes in this value only induces a
rd , s = µd , s . X v, s + K . X v , s .T (4) proportional factor in T and I evaluation.
As shown in Figures 6 and 7, the proposed mathematical
The death rate of the sensitive strain is divided into two expressions properly fit experimental data.
terms : natural death and death caused by the killer effect,
which is linked to the medium toxicity (T). The toxin
production rPT was shown to be proportional to the growth 140
Population (*1e6 cell/mL)

120
rate of the killer strain (Ramon-Portugal et al. 1994) as 100
expressed in Equation (5). The toxin consumption rate, rCT , 80
60
due to adsorption on sensitive cells, is supposed to be almost
40
instantaneous and proportional to both toxicity and viable
20
sensitive population (Equation (6)) .
0

rPT , = α .rv , k (5) 0 5 10 15 20 25 30 35

t (h)
Viable killer cells - Model Dead killer cells - Model
Viable killer cells - Experimentals Dead killer cells - Experimentals
rCT = −W . X v , s .T (6)

Inhibitor I is assumed to be produced during both killer and Figure 6 : killer strain growth kinetics in mixed culture
sensitive yeast growth, with a production rate given by
Equation (7) :
100
rI = a.(rv , k + rv , s ) (7) 90
Population (*1e6 cell/mL)

80
70
The dynamic evolution of each of the six variables of the 60
50
model is given by Equations (8) to (13) : 40
30
20
dX v , k 10
= rv , k − rd , k (8) 0
dt 0 5 10 15 20 25 30 35
t (h)
dX v , s Viable sensitive cells - Model Dead sensitive cells - Model
= rv , s − rd , s (9) Viable sensitive cells - Experimentals Dead sensitive cells - Experimentals
dt

dX d , k Figure 7 : sensitive strain growth kinetics in mixed culture

= rd , k (10)
dt
This curve fitting is however incomplete for a good NOMENCLATURE
appreciation and a possible improvement of the model.
Mixed fermentations with different initial conditions must be a Inhibitor production constant (g.cell-1)
studied in order to check whether the model can be Braking factor for logistic-like
A (L.g-1)
predictive or not. expression
I Growth inhibitor (g.L-1)
K Toxin lethal action constant (L.g-1.h)
t Time (h)
CONCLUSION T Toxin concentration (g.L-1)
W Toxin consumption constant (L.cell-1.h)
A new experimental device has been used in order to get
fundamental data for the modeling of mixed cultures X Yeast population (cell.mL-1)
including killer interactions. Biomass segregation thanks to a
porous membrane made it possible characterize the α Toxin production constant (g.cell-1)
physiological state of both strains of the co-culture. This µmax Maximum specific growth rate (h-1)
characterization was performed through two different µd Specific natural death rate (h-1)
techniques: methylene blue staining and ATP measurements.
The results showed these methods were matching.
Subscripts
An existing mathematical model was confronted to collected
data. This confrontation validated the proposed mathematical
k killer
expressions for the description of the killer phenomenon.
s sensitive
Further experiments have to be carried out in order to v viable
confront the model to different starting conditions. The d dead
model can also be improved, especially by introducing pH
and temperature factors as they happen to be crucial for toxin
activity. AUTHOR BIOGRAPHY

REFERENCES SEBASTIEN POMMIER was born in Tulle, France and

went to the Institut National Polytechnique de Toulouse
Alfenore, S. 1999. “Interaction du type killer chez Saccharomyces where he studied Chemical Engineering at ENSIGC. He is a
cerevisiae. Etudes physiologiques et cinétiques. Quantification PhD student in the Laboratoire de Génie Chimique in
et modélisation”. Thèse, Institut National Polytechnique de Toulouse. His work is focused on experimental study and
Toulouse, FRA. modeling of interactions in yeast mixed cultures.
Bevan, E.A. and M. Makover. 1963. “The physiological basis of the
killer character in yeast”. Genetics Today. XIth International
Congress of Genetics. Geers S.G. Editor. Pergamon Press,
Oxford. 1 , 202-203.
Pfeiffer, P. and F. Radler. 1984. “Comparison of the killer toxin of
several yeast and the purification of a toxin of type K2”. Arch.
Microbiol. 137, 357-361.
Ramon-Portugal, F.; M.L. Delia-Dupuy; G. Schneider; and P.
Strehaiano 1994. “Yeast killer activity: a quantitative study”.
Biotechnol. Techniques, 8, 797-804
Ramon-Portugal, F.; M.L. Delia-Dupuy; H. Pingaud, G.A. Carrillo-
Leroux; and J-P. Riba 1997. “Kinetics study and mathematical
modelling of killer and sensitive Saccharomyces cerevisiae
strains growing in mixed culture”. Bioprocess Engineering, 17,
375-381
Reinsch, C.-H. 1967. “Smoothing by spline functions”. Numerische
Mathematik, 10, 177-183
Salgado Manjarrez, E.; C. Albasi; and J-P. Riba. 2000. “A two-
reservoir, hollow-fiber bioreactor for the study of mixed-
population dynamics: design aspects and validation of the
approach”. Biotechnol. Bioeng., 69(4), 401-408
Strehaiano, P. 1984. “Phénomènes d'inhibition en fermentation
alcoolique”. Thèse Dr ès Sciences, Institut National
Polytechnique de Toulouse, FRA.
 A NEW APPROACH FOR THE ENZYMATIC ACTIVITY
TEMPERATURE DEPENDENCY MODELLING.
APPLICATION TO THE BEER MASHING PROCESS
Cédric Brandam Hervé Pingaud Jacques Proth
Xuân-Mi Meyer Centre de Génie Industriel TEPRAL, Brasseries Kronenbourg
Pierre Strehaiano Ecole des Mines d’Albi Carmaux 68 route d’Oberhausbergen
Laboratoire de Génie Chimique Campus Jarlard - Route de Teillet 67037 Strasbourg (France)
(LGC UMR - CNRS 5503), 81013 Albi CT Cedex 09 (France)
BP 1301, 5 Rue Paulin Talabot
31106 TOULOUSE (France)
Email: [email protected]

KEYWORDS isothermal reactive medium must take into account both

effects of temperature and the effect of time. Yet, classical
Enzymatic activity, Modelling, Amylases, Mashing. enzyme activity measurements, as done in brewery, do not
always enable to characterise all these effects.
ABSTRACT

This paper deals with a new approach for the enzymatic Activity
activity temperature dependency modelling. The global Kinetic contribution
enzymatic activity is modelled by the contribution of two
Global activity
terms: thermal denaturation function of couple time-
temperature and specific activity only temperature
dependant. The approach and the modelling has been Thermal
applied with success to the beer mashing process. denaturation
Measurements on pilot-plant experiments have been
realised to determine the thermal denaturation during
mashing for α-amylase, β-amylase and β-glucanase.
Specific activities temperature dependency of these Temperature
enzymes have been obtained on laboratory experiments.
Figure 1: Temperature Influence on Enzymatic Activity
INTRODUCTION
Several enzymes are involved in brewing and more
Enzymes are protein compounds whose biological particularly in mashing. During this step, malt endogenous
function is to catalyse reactions occurring in organisms enzymes catalyse the hydrolysis reactions of the malt
(Wallach,1997). Their action is based on the more or less compounds (starch, β-glucans and proteins). Final beer
specific recognition between an enzyme region (active site) characteristics (organoleptic quality and ethanol content)
and the molecule concerned by the catalysed transformation are greatly dependent of mashing insofar as it determines
(substrate). Many parameters can affect this recognition the wort composition before fermentation. Mastering the
ability and so increase or decrease the enzyme action. More enzymatic activity is an essential step to further model the
than the enzyme concentration, difficult and rarely production of fermentable carbohydrates, dextrins and β-
determined, it is important to quantify this recognition glucans during mashing and to improve the control of this
ability. The notion of enzymatic activity was then process. As far as mashing is concerned, the most important
introduced and defined as the speed at which a reaction enzymes are the α- and β-amylases involved in the starch
catalysed by the enzyme is realised. hydrolysis into fermentable carbohydrates and dextrins, the
Temperature is one of the most important parameter β-glucanases which catalyse β-glucan hydrolysis into
influencing enzymatic activity. It has a double effect : on residual β-glucan chains and the proteases necessary for the
one hand, it increases the reaction rate by providing the protein hydrolysis into free amino acids and peptides. As
energy necessary to overcome the activation barrier of the each enzyme reacts differently to temperature, mashing
reaction, on the other hand, it acts on the tri-dimensional must be performed according to a temperature diagram
structure of enzymes, leading to the denaturation and the drawn up to favour each enzyme optimal temperature
deactivation of these proteins. These two antagonist effects range. Therefore, it is essential to have a good knowledge
result in an optimal temperature for the enzymatic activity of the temperature dependency of the activity of each
as shown on figure 1. involved enzyme.
Moreover, time also contributes to the enzyme
denaturation. Modelling the enzymatic activity in a non
 In the brewing literature, the measurement method of speed of specific substrate consumption for enzymes gives
specific substrate consumption at a reference temperature is the specific activity with respect to the temperature.
always used to evaluate enzymatic activity evolutions
during mashing (Moll et al., 1981; Marc et al., 1983; MODEL
Muller, 1990; Koljonen et al., 1995). For instance, in
breweries, α-amylase activity is evaluated by measuring the Enzyme denaturation
consumption of a specific substrate (azurine-crosslinked
amylose) at 40°C, in pH 5.6 buffer, during 20min Thermal denaturation kinetic is determined at a
(MEGAZYME method). The values obtained performing reference temperature and it is usually represented by a first
this method compare the residual activity of different order rate process:

( )
samples at a reference temperature. They can be considered
as a relative activity measurement and are interesting to d[E] −E
rdE (Tref ) = = k dE . exp( dE )[E ] (1)
determine and compare malt initial enzymatic potential dt Tref RT
when the grains are in the same initial conditions. But this
method is not well adapted to follow activity evolutions
when temperatures vary (as during mashing), because it rdE (Tref) is the denaturation rate (U/kg.s), [E] is the active
only takes into account the thermal denaturation site concentration measured by Megazyme method (U/kg of
contribution. Until now, no authors have reported global maïsche), EdE is the activation energy for the denaturation
enzymatic activity values corresponding to mashing (J/mol), kdE is the preexponential factor (s-1), T is the
processing conditions though it is the most interesting temperature (K) and R is the gas constant (8.31 J/mol.K).
measure to determine the optimal mashing temperature
policy. Temperature kinetic contribution
To cover this lack, a new approach is proposed to
determine and model the global enzyme activity. It is A relative specific activity as(T) has been defined as the
illustrated on α-, β-amylase and β-glucanase activity in a ratio of the initial speed of specific substrate consumption
mashing tank. v(T), measured at the temperature T, on the initial speed
measured at the reference temperature v(Tref) :
ENZYME ACTIVITY MEASUREMENT
v (T )
a s (T ) = (2)
The two temperature contributions involved in the v(T ref )
global activity are determined separately.
Relative specific activity of sites, only temperature
Enzyme denaturation dependent, is modelled by polynomial laws.
To measure enzyme denaturation during mashing, Global activity
mashes were realised in the 100 litre microbrewery of
TEPRAL Research (KRONENBOURG breweries research The evolution rate of the global activity
center). All experiments were done with the same raw
material: malt from Scarlett variety. 20kg of malt were ra (T) expressed in U/kg.s, can be represented by the
mixed with 70kg of water. Ten experiments have been product of thermal inactivation rate by the relative specific
realised with different temperature policies. activity of these remaining sites at the mashing
temperatures:
Samples were withdrawn at the beginning and the end ra (T ) = rdE (Tref ) * a s (T ) (3)
of each temperature rest. They were rapidly cooled at 4°C So, the enzymatic activity can be globally represented by
in carbo-ice to stop enzymatic activities and then equation:
centrifuged for 12 min at 4000g. Sample enzymatic activity −E
is then measured in liquid phase by the Megazyme ra (T ) = −k dE . exp( dE )[E ] × a s (T ) (4)
RT
methods, based on specific coloured substrate consumption.
These methods measure enzyme activities at a reference This mathematic representation takes into account both
temperature: 30°C for β-amylases and β-glucanases and temperature and time effects on the enzymatic activity.
40°C for α-amylases. The experimental relative error for
each enzyme denaturation method has been determined to
be about 7%. RESULTS and DISCUSSION

Temperature kinetic contribution Experimental results for denaturation

From the same active enzyme quantity, isothermal The experimental results of α-amylase, β-amylase and
laboratory experiments have been carried out at β-glucanase denaturation for the classical four temperature
temperatures ranging between 30 and 75°C. The initial rests mashing diagram are represented on figure 2.
 In table 1 are reported for each tested malt enzyme the
150 76°C 80 denaturation threshold Tdenat, the temperature corresponding
to the maximal specific activity Tmax and temperature
70
beyond which no activity remains Tdis.

Temperature (°C)
60°C
Activity (U/kg)

100 Beta 60
50°C Table 1 : Characteristic Temperatures of malt α-amylase, β-
Alpha 50
amylase and β-glucanase
50 37°C Beta-gluc 40
Tdenat Tmax Tdis
30 α-amylase 60°C 63°C 75°C
0 20 β-amylase 55°C 63°C 70°C
0 20 40 60 80 100 β-glucanase 40°C 45°C 55°C
T ime (min)

Figure 2: α-amylase, β-amylase and β-glucanase Thermal

Global enzymatic activity
Denaturation (measured by Megazyme method).
From these experimental values, global enzyme activity
As long as the threshold temperature of denaturation is
evolutions can be calculated by equation (4). Figure 4
not reached, the enzymatic potential remains constant and,
shows the re-built values for a classical temperature
as the activity are measured at the reference temperature,
diagram.
the activities keep constant, even if temperature increases in
the mashing tank. For instance, the denaturation threshold 35 0 76°C 80
is about 50-60°C for α-amylase and β-amylase and about
30 0 70
40°C for β-glucanase. Then, as soon as the denaturation
temperature is exceeded, activities decreases rapidly for β-

Temperature (°C)
25 0 60°C
Activity (U/kg)

60
amylases and β-glucanases, slowlier for α-amylases. 20 0 50°C
Beta 50
Experimental results for the temperature kinetic 15 0
Alpha 40
contribution 10 0 37°C
Beta-gluc
50 30
On figure 3 are plotted the relative specific activities of
the α−amylase, β-amylase and β-glucanase versus 0 20
temperature. 0 20 40 60 80 100
T ime (min)
8 Figure 4: α-amylase, β-amylase and β-glucanase
7 Beta-amylase Global Activity.
Relative specific activity

6 Alpha-amylase
5
The importance of taking into account the temperature
Beta-glucanase
kinetic contribution term appears clearly by comparing
4 figures 2 and 4. In one case (fig 2) the enzymatic activity
3 remains constant between 50 and 60°C whereas it increases
2 clearly if the thermal contribution is taken into account.
1 The global amylase activity increases with the temperature
until 60-63°C. The best efficiency is located between 60-
0
63°C for amylases, between 37-45°C for β-glucanases.
20 30 40 50 60 70 80
T emperature (°C)
Model parameter estimation
Figure 3: α-amylase, β-amylase and β-glucanase Relative
Specific Activity versus Temperature. Model parameters have been estimated by a least-square
method from the experimental values collected from the ten
The relative specific activities reach a maximum before experiments realised with different mashing diagrams. The
decreasing more or less rapidly according to the obtained values are:
temperature sensitivity. β-glucanase are the most heat
sensitive of the three enzymes and α-amylase the less. α-amylase: kdα = 6,94 1030s-1 Edα = 224,2 kJ/mol
At 63°C, relative specific activity is 7 times higher than
activity measured at 40°C for α-amylases, and 2.5 times
β-amylase: kdβ = 3.06 1068s-1 Edβ = 460,1 kJ/mol
higher than the activity measured at 30°C for β-amylases.
For β-glucanases, relative specific activity is the same
between 30°C and 50°C, but it is 2.3 times higher at 45°C. β-glucanase: kdβg = 3,89 1031s-1 Edβg = 246,1 kJ/mol
 These parameter values are in accordance with the 160 Experimental values 76°C 80
habitual bibliography values. 140 Model
70

Temperature (°C)
120 63°C
To model the temperature kinetic contribution, the

Activity (U/kg)
relative specific activities were modelled by four 100 60
polynomial laws for each enzyme (Temperature must be 80
expressed in Kelvins in polynomes) : 60 50
42°C
40
α-amylase: 37°C 30min 40
20
as = 1 for T < 40°C
as = -0,0011 T3 + 1,091 T2 –352,9 T + 38008 0 30
for 40°C < T < 63°C 0 20 40 60 80 100
T ime (min)
as = -0,0253 T2 + 16,553 T + 2702,1
for 63°C < T < 74°C Figure 6: Comparison of β-glucanase global activity
as = 0 for T > 74°C experimental values and model prediction for a four rests
mashing temperature policy (37°C – 42°C – 63°C – 76°C)
β-amylase:
as = 1 for T < 37°C
as = 0,0495 T – 13,993 for 37°C < T < 63°C The tested mashing diagrams cover all the usual
as = -0,3829 T + 131,22 for 63°C< T < 70°C temperature mashing scale. The representation of the α-
as = 0 for T > 70°C amylase, β-amylase and β-glucanase activities by the model
is good for all experiments.
β-glucanase:
as = 1 for T < 30°C
as = -0,0028 T2 +1,8566 T – 299,91 Model validation
for 30°C < T < 45°C
as = 0,015 T2 –9,9014 T + 1636,3 The model was applied to the mashing of malts different
for 45°C < T < 56°C from the one used for the parameter estimation.
as = 0 for T > 56°C Figures 7a and 7b show the good agreement between model
and experiment for these new experiments and so validate
Global α-amylase and β-amylase activity calculated the model.
values are compared to the experimental values on figure 5.
The mashing temperature diagrams is a classical mashing
policy with three rests. On figure 6, the global β-glucanase Kinetic rate of enzymatic catalysed reaction
activity is represented for a four rests temperature diagrams.
Our works questioned the way to model the enzymatic
activity evolution back. Until now, kinetics of enzymatic
reactions introduced activation energies and pre-
exponential factor to represent temperature effect on the
1600 Experimental values 76°C 80
production rates:
Model −E
1200
70 r = k. exp( )[enzyme][substrate]
63°C RT
Temperature (°C)
Activity (U/kg)

180min 60
800 With our model, the temperature effect is already taken
50 into account in the enzymatic activity modelling. Then, the
50°C
Alpha enzymatic reaction rate can be simply written:
400
Beta 40 r = k.[enzyme][substrate]

0 30 The kinetic of products need to identify only one kinetic

0 50 100 150 200 250 parameter, not temperature dependant. The identification
T ime (min)
problem is so easier. This is a very interesting advantage to
model the real enzymatic activity at each times and
Figure 5: Comparison of α-amylase and β-amylase global temperatures of mashing.
activity experimental values and model prediction for a
three rests mashing temperature policy
(50°C – 63°C – 76°C)
 1600 76°C 80 concentrations during mashing". Journal of Food
Experimental values
Engineering, 26, 329-350.
Model
70 Marc A., Engasser J.M. 1983. "A kinetic model of starch
1200 63°C hydrolysis by alpha- and beta-amylase during mashing".

Temperature (°C)
Activity (U/kg)

15min 60 Biotechnology and Bioengineering, Vol.XXV, 481-496.

800 Moll M., Flayeux R., Lipus G., Marc A. 1981.
50 "Biochemistry of mashing". MBAA technical quaterly,
50°C Alpha
400
18, 166-173
37°C Beta 40 Muller R.E. 1990. "The effects of mashing temperature and
mash thickness on wort carbohydrate composition".
0 30 Journal of the Institute of Brewing, 97, pp.85-92
0 20 40 60 80 100 120 Wallach J. 1997. "Les enzymes". Coll.128, Nathan
T ime (min)
University
Figure 7a: Model validation for α-amylase and β-amylase.
Comparison between model and experiment values for a
mashing realised with an other malt.

200 76°C 80
Experimental values
Model
70
150 63°C
Temperature (°C)
Activity (U/kg)

60
100
50
45°C
50 30min
40

0 30
0 20 40 60 80
T ime (min)

Figure 7b: Model validation for β-glucanase. Comparison

between model and experiment values for a mashing
realised with an other malt.

CONCLUSIONS

The classical way to measure enzymatic activity is not

sufficient to characterize its temperature dependency. The
original modelling of global amylase activities presented in
this paper takes into account not only the denaturation term
but also the kinetic contribution influencing enzymatic
activity and enable to determine the global enzymatic
activity at effective mashing temperatures. This approach
will allow to determine the optimal temperature policy to
apply during mashing to obtain the best enzyme efficiency
and to reach the wort specification necessary to get a
reliable beer organoleptic quality.

REFERENCES

Brandam C., Meyer X.M., Proth J., Strehaiano P., Pingaud

H. 2002. "A new reaction scheme for the starch
hydrolysis and temperature policy influence during
mashing". Food Science and Biotechnology, Vol.11,
No.1, 40-47.
Koljonen T., Hämäläinen J.J., Sjöholm K., Pietilä K. 1995.
"A model for the prediction of fermentable sugar
 A new strategy for production of Saccharomyces biomass
by fed-batch cultures
Audrey Serra, Isabelle Masneuf
Patricia Taillandier, Enita
Pierre Strehaiano, Cours du général de Gaulle,
Laboratoire de Génie Chimique BP 201, 33175 Gradignan cedex
UMR 5503, INP-ENSIACET, [email protected]
118 route de Narbonne,
31062 Toulouse Cedex, France
[email protected]

Jean François Laffort

Laffort Œnologie,
7 rue Franc Sanson,
BP 17, 33015 Bordeaux, France
www.laffort.fr

KEYWORDS INTRODUCTION

Production, fed-batch, yeast biomass, optimization In biological experiments, a response of a cell culture has to
be analysed in relation to the environment. There is an
ABSTRACT exchange between the environment and the cells, and
environmental parameters may influence their physiology. An
For some industrial uses (baking, wine-making, brewing) important factor in studies dealing with the regulation of
great amounts of yeast viable biomass are needed. Due to the metabolism is the nutritional limitation which may change the
Crabtree effect which Saccharomyces cerevisiae are sensitive metabolism of the cells drastically. The regulation of the
to, cultures with excess oxygen supply and glucose limitation process is so strictly connected to this regulation of
are generally carried out in order to get a maximal biomass metabolism.
yield of 0,5 g/g. To achieve such conditions, fed-batch
cultures are usually run. In order to optimise aerobic fed batch During growth of Saccharomyces cerevisiae in a batch culture
cultures of glucose-sensitive yeasts with regard to under aerobic conditions with glucose as sole energy and
productivity and cell yield, it is necessary to use a feed-back carbon source, ethanol is always released into the culture
feeding control devices. The most efficient one is based on medium as a by-product. This fact leads to a fall in biomass
the Respiratory Quotient (RQ) measurement using fermentor yield. However, if the yeast is cultivated in a chemostat at low
exhaust gas stream. dilution rates, a strictly oxidative metabolism can be
observed ; this state results in exclusive formation of biomass
The objective of this work was the establishment of a general and carbon dioxide. When the dilution rate is increased (i.e.
model of sugar feeding rates for the production of different S. the glucose concentration) up to a particular critical value, the
cerevisiae wine strains, under different conditions. This yeast metabolism becomes fermentative and ethanol appears
glucose feeding rate was first obtained experimentally for one in the medium because of the catabolic repression by glucose.
typical strain before being validated with others. An average So, when the glucose concentration is high, even with excess
glucose feeding profile was established as grams of sugar oxygen, some yeasts consume the glucose mainly by the
versus time. This mathematical expression modelled the fermentative way. This phenomenon is known as Crabtree
glucose feeding kinetics to obtain an optimal biomass effect (Fiechter et al., 1981) and these yeasts are called
production. In the second step, this profile was extrapolated glucose-sensitive yeasts.
successfully to others strains, without used the RQ
measurement. For some industrial uses (baking, wine-making, brewing)
great amounts of yeast viable biomass are needed. So, due to
The advantage of this work is double. First, the profile allows the Crabtree effect which Saccharomyces cerevisiae are
the biomass production of the yeast strain Saccharomyces sensitive to, aerobic cultures with glucose limitation are
cerevisiae for any volume of fermentor and any sugar generally carried out in order to get a maximal biomass yield
concentration in the feeding medium. Second, thanks to this of 0,5 g/g.
model the yeast biomass production can be achieved without
the RQ on-line measurement so, with a simpler device. To achieve such conditions, fed-batch cultures are usually run
thus permitting the circumvention of the problem of aerobic
fermentation due to the glucose effect (Yamane and Shimizu,
1984). In order to optimise aerobic fed batch cultures of
glucose-sensitive yeasts with regard to productivity and cell
yield, it is necessary to use a feed-back feeding control stirring rate. The pH was kept constant at 5.0 ± 0.1 by
devices. The most efficient one is based on the Respiratory automatic addition of 1 M NaOH with a pH controller
Quotient (RQ) measurement using fermentor exhaust gas (Inceltech LH PPV3, Toulouse, France or Applikon
stream (Aiba et al., 1976 ; Cooney et al., 1977). Thus, the Biocontroller ADI 1030, Schiedan, Holland). The temperature
yeast always undergoes a respiratory metabolism and none was kept constant at 30°C.
ethanol is accumulated.
Gas analysis.
The objective of this work was the establishment of a general
model of sugar feeding rates for the production of different S. Gas analysis was performed on-line with a Servomex type
cerevisiae wine strains and different Saccharomyces species 4100 analyser (Servomex, Crowborough, England). The
which are Crabtree positive. This glucose feeding rate was exhaust gas was cooled in a condenser, dried with a
first obtained experimentally for one typical strain before Hygrostop (Environnement S.A., France) and subjected to gas
being validated and extrapolated with others. analysis in a paramagnetic O2 and infrared CO2 gas analyzer
sampling every 5 min along the fed batch culture. The gas
MATERIAL AND METHODS analyzer was calibrated with air and nitrogen and a mixture of
5% CO2 and 95% synthetic air.
Yeast strain.
Feeding of reactor.
Three strains of Saccharomyces cerevisiae was used in this
work : 71 B (Lallemand, Montreal, Canada), VLC3 and L5 The on-line calculation of RQ was made automatically by a
(Laffort Oenologie, Bordeaux, France). program using the data given by the exhaust gas analysis. The
reactor was feeding by a peristaltic pump (Minipuls 3, Gilson,
Culture media. Villiers-Le-Bel, France), which the flow rate was adjusted
according to either the RQ value or the program value. To
It was a synthetic medium containing [mg.(g glucose)-1] : maintain a strictly respiratory metabolism, the RQ value had
ammonium sulfate, 300 ; KH2PO4, 140 ; CaCl2,2H2O, 5 ; to be equal to 1, for the studied yeast strains. By this way, the
NaCl, 10 ; MgCl2,6H2O, 120 ; Yeast Extract (Oxoid) 20. glucose was completely transformed in biomass (strictly
Ammonium sulfate and the other products of the medium oxidative metabolism).
were separately heat-sterilized (121°C, 20 min) to avoid
Maillard reaction, then mixed and added of 0.2 mL.(g Biomass evaluation.
glucose)-1 of each sterile vitamins and minerals solution
which had the following composition [mg.L-1] : “vitamins Three methods were used to estimate the active biomass in
solution” : biotin, 3 ; calcium pantothenate, 40 ; inositol, 250 ; the cultures : cell count / viability, optical density and dry
pyridoxine HCl, 50 ; thiamine HCl, 100 ; “minerals weight.
solution” : FeSO4,7H2O, 556 ; ZnSO4,7H2O, 576 ;
CuSO4,5H2O, 14 ; Na2MoO4,2H2O, 50 ; CoCl2,6H2O, 50 ; Cell count / viability.
MnSO4H2O, 338 ; H2SO4, 10 drops. This synthetic medium If needed cell suspensions were diluted to a maximal
was established by Aranda Barradas (1999). The use of a concentration of 50 x 106 cell.mL-1 in distilled water.
defined media makes it possible to study the influence of the Methylen blue solution (0.1 g methylen blue in 1 liter sodium
process conditions on the yeast physiological state tricitrate 2%) was added volume to volume to the culture
independently of variations due to the medium composition. sample. Microscopic count was carried out 10 minutes later,
However, the culture on the synthetic medium has been magnification × 400, using a Thoma hematory-meter.
compared to a culture on molasses medium, which is used in
the industrial process, and we obtained same results for the Optical density.
growth kinetics and the yield (data not shown). So the Samples were poured in 1 cm optical length spectrometric
synthetic medium has been validated. tubs from Alliance Concept. Absorbance of yeast suspensions
was measured with Secoman Anthelic Advanced (Toulouse,
Culture conditions. France) spectrophotometer, at 620 nm wavelength. A linear
regression between absorbance and biomass concentration is
Precultures were grown to stationary phase in shake-flasks observed for absorbance range of 0.1 to 0.8. For higher
cultures on synthetic medium (glucose 50 g.L-1), at 30°C, 250 values, samples were diluted with distilled water. We define
rpm. These cultures were used to inoculate batch cultivation. optical density as the red absorbance multiplied by dilution
This step was used to form a sufficient quantity of fresh factor.
biomass for the fed-batch cultivation. The fed batch cultures
were conducted in two bioreactors of 4 (Inceltech Series 210, Dry weight.
Toulouse, France) and 14 working volume liter (Applikon Z8 For measuring the biomass concentration, 10 mL of culture
1315 M607, Schiedan, Holland), equipped with a condenser sample were centrifuged, washed twice with distilled water to
to avoid loss by evaporation during the culture. The reactor eliminate particles of medium culture and then dried by a
was aerated with sterile air. The dissolved O2 concentration humidity analyzer (Precisa HA60, Suisse). Figure 1 shows the
was monitored with an O2 electrode (model 34 10 3003, linear regression of biomass concentration versus optical
Ingold (Mettler Toledo AG, Suisse)) and remained density.
automatically above 20% of air saturation by varying the
 30
ethanol present in the medium was formed in majority during
the batch culture. During the fed batch culture the ethanol
y = 0,3615x concentration did not increased, indicating that little ethanol
[Biomass] (g.L-1)

20 R2 = 0,999 was produced. So, for these cultures, we obtained an optimal

production.

10 The feeding of reactor profile is presented on figure 3. The

feeding flow had increased exponentially and we established
an average glucose feeding profile as grams of sugar versus
0 time. Experimental data are represented as symbols and the
0 20 40 60 80 curve is a polynomial regression obtained with the “Statlets”
Optical Density (A) software, expressed by the equation (1) :

Figure 1. Linear regression between biomass concentration Sugar = 0.371 + 1.857×t + 0.799×t2 - 0.086×t3 + 0.005×t4 (1)
and optical density for S. cerevisiae L5 growth on synthetic
medium.
400
Determination of glucose and ethanol.
300
Glucose was determined with an enzymatic glucose analyzer

Sugar (g)
(YSI 2700 Select, Yellow Springs, Ohio, USA). Ethanol was 200
analyzed by gas chromatography (Chrompack 437A, Delf,
The Netherlands). Prior to injection, the cells were removed 100
by centrifugation and a known amount of isopropanol 1%
solution was added as internal standard. The column was
0
Chrompack Poraplot Q wide-bore (0.53 mm × 25 m), the
0 5 10 15 20
injector and detector temperatures were 200°C and the vector
gas was nitrogen (8 mL.min-1). Time (h)

Figure 3. Feeding of glucose during a 4 L fed batch culture.

RESULTS AND DISCUSSION
This mathematical expression modelled the glucose feeding
Establishment of a general model of sugar feeding rate. kinetics to obtain an optimal biomass production for the strain
Saccharomyces cerevisiae and a feeding volume of 2.5 L.
The first step was the production of the wine strain S. This optimal biomass can be obtained since an oxydative
cerevisiae 71B by fed-batch cultures regulated by RQ on-line pathway is followed all along the growth. The model takes
measurement. The experiment has been repeated several into account the oxygen demand for growth. With this
times. These cultures were conducted in the 4 L bioreactor relation, it is easy to determine the feeding flow according to
with an initial volume of 1.5 L at 7.5 g.L-1 of biomass. Two the volume of reactor and the sugar concentration of the
examples are showed on figure 2. alimentation medium.
60
Model validation.
50
[Ethanol] (g.L-1)

40 In the second step, this profile was validated with the

[Biomass]

production of an other strain (VLC3) in the same conditions

30 (volume, sugar quantity), but without used the RQ
measurement. The feeding of reactor was controlled by an
20
informatic program which calculated the flow rate each five
10 minutes, according to the equation (1). The theoretical and
experimental flow profiles were showed on figure 4.
0
0 5 10 15 20 The feeding of the reactor was done as planned by the
Time (h) previous polynomial regression. The results of this culture are
Figure 2. Evolvement of biomass (●,▲) and ethanol (■, ♦) presented in the table 1. The sugar concentration was
concentration during two fed-batch cultures regulated by RQ measured in the reactor, so not consumed by the yeasts.
measurement.
The residual sugar concentration in the reactor was very
The behaviour was the same for the two cultures. The weak, indicating that the yeasts have consumed all the carbon
biomass was accumulated until 45 g.L-1 and 49 g.L-1 for 350 g source. So, the sugar feeding according to the model was not
consumed glucose. We obtained a maximal yield (0.5 g in excess and fitted the growth of this strain.
biomass / g sugar) and a great final concentration of cells
(2.6×109 cells per mL, with a viability superior to 97%). The
The final biomass concentration appeared lower than 35
previously, but in fact, the final volume was more important,
due to the pH regulation. Therefore, we obtained a good 30
production of biomass, with a yield of 0.485 g biomass / g

Flow (mL.min )
25

-1
sugar consumed and an excellent viability of cells. The 20
growth of cells was not limited by the sugar feeding.
15

10 10
5
8
Flow (mL.min -1 )

0
6 0 5 10 15 20
Time (h)
4
Figure 5. Feeding flow profiles for the 14 L (▲) and the 4 L
2
(●) fed batch cultures.
0
0 5 10 15 20 Table 2. Biomass production during the 14 L fed batch
Time (h) culture.

Figure 4. Theoretical (◊) and experimental (—) flow of Initial Final

feeding for the fed batch culture of VLC3 without RQ time time
measurement.
[Biomass] (g.L-1) 6.11 24.2
Table 1. Biomass production during a fed batch culture Cells number (.106/mL) 250 1139
controlled by the model of sugar feeding rate.
Viability (%) 99.7 98.5
Initial Final Volume (L) 6.8 11
time time
[Biomass] (g.L-1) 6.84 38.32 In this experiment, the maximal volume was not reached. The
6
Cells number (.10 /mL ) 330 2530 culture was limited by the oxygen transfer and the foam
formation, so only 2.8 L of medium was bring in the
Viability (%) 97.2 98.1
fermentor. The production was stopped after 12 hours instead
Sugar (g) 0 0.068 of 18 hours, at half of exponential phase. Therefore, the
Volume (L) 1.5 4.7 quantity of biomass produced was lower, but, in fact, as the
sugar quantity bring was lower too, the yield was good (0.505
g biomass / g sugar consumed) and the productivity almost
the same (2.1 g.L-1.h-1).
The productions of biomass, with and without the RQ
measurement, were very similar for the two strains
For evaluate the quality of the production, the biomass
(productivity of 2.3 and 2.25 g.L-1.h-1, respectively). The
formed was dried like Active Dried Yeasts (A.D.Y. ). We
exponential profile allowed to get optimal growth. So, we
could validate the model of sugar feeding rate established for obtained 220 g of dried yeast with 19.3×109 viable cells per g
a fed batch culture of a Saccharomyces cerevisiae yeast strain of dried yeast. This quantity allows to inoculate 10 to 12 hL
with a medium feeding of 2.5 L. of musts. Some experiments in flasks showed a good
fermentation of sugar by the biomass produced (data not
Model extrapolation. presented), indicating a high quality for the production.

In the last step, we used this model to produce others yeasts CONCLUSION
strains, without RQ measurement and with an other volume of
bioreactor. In this work, we have determined a model of sugar feeding
rate for the biomass production. The advantage of this work is
The fed batch culture was conducted in the 14 L bioreactor double. First, the profile allows the biomass production of the
with the yeast L5. In this experiment, we changed the volume yeast strain Saccharomyces cerevisiae for any volume of
of reactor but not the growth characteristic, because we used a fermentor and any sugar concentration in the feeding
Saccharomyces cerevisiae strain like previously. Like the medium. So we have a good flexibility of the correlation. This
specific growth rate was the same for all strain profile enables an optimal conversion of sugar in biomass. In
Saccharomyces cerevisiae, the feeding sugar profile could be fact, we obtained a productivity of 2.1-2.3 g.L-1.h-1and in
obtained by a translation of the volume. Previously, the initial industrial process, the usually value is 2 g.L-1.h-1. These
volume was 1.5 L and the feeding volume was 2.5 L, now, results were achieved if the best conditions culture were
these volumes were, respectively, 6 L and 8 L. The results of present, in particular, if the culture was not limited by
the culture were presented on figure 5, for the comparison of aeration.
the two flow profiles and on table 2.
Second, thanks to this model the yeast biomass production
can be achieved without the RQ on-line measurement so, with
a simpler device. Like this, the production is easier.

We are now testing it for others Saccharomyces species with

a correcting factor according the specific growth rate of each
wine strain. These strains must have the same metabolic
characteristics, i.e., the yeast must be Crabtree-sensitive.

REFERENCES

Aiba S. ; Nagai S. ; Nishizawa Y. ; “Fed batch culture of

Saccharomyces cerevisiae: a perspective of computer control to
enhance the productivity in baker's yeast cultivation”,
Biotechnology and Bioengineering, 1976, 18, 1001-1016.
Aranda Barradas J.S. 1999. “Production de biomasse levurienne :
influence du procédé sur les potentialités fermentaires des
levures ”, Thèse Institut National Polytechnique de Toulouse,
France.
Cooney C.L. ; Wang H.Y. ; Wang D.I.C. ; “Computer-aided material
balancing for prediction of fermentation parameters”,
Biotechnology and Bioengineering, 1977, 19, 55-67.
Fiechter A. ; Fuhrmann G.F. ; Kappeli O. ; “Regulation of glucose
metabolism in growing yeast cells“, Advanced in microbiology
and physiology, 1981, 22, 123-183.
Yamane T. ; Shimizu S. ; « Fed-batch techniques in microbial
processes », Advanced in biochemical engineering and
biotechnology, 1984, 30, 147-194.

AUTHOR BIOGRAPHY

AUDREY SERRA was born in Nice (East-South of France)

in 1978. In 2000, she obtained a diploma of process engineer,
with a specialisation in biotechnology. Now, she is doing a
PhD thesis in the Laboratoire de Génie Chimique, in the team
“Fermentations et Bioréacteurs” (Toulouse, France). The aim
of her work is the production of wine yeast biomass to be
used as starters.
 OPTIMAL DYNAMIC EXPERIMENT DESIGN FOR ESTIMATION
OF MICROBIAL GROWTH KINETICS AT SUB-OPTIMAL
TEMPERATURES: MODES OF IMPLEMENTATION

Bernaerts, K. and Van Impe, J.F.

BioTeC-Bioprocess Technology and Control
Katholieke Universiteit Leuven
W. De Croylaan 46, B-3001 Leuven, Belgium
Tel. +32-16.32.19.47 Fax. +32-16-32.19.60
E-mail: [email protected]

KEYWORDS In this work, optimal dynamic input design for param-

optimal experiment design, dynamic conditions, predic- eter estimation is applied in the field of predictive mi-
tive microbiology, bioreactors, constrained design, ad- crobiology. This research discipline aims at modelling
vanced data processing. the outgrowth, inactivation or survival of pathogenic
and spoilage micro-organisms in food products that are
characterised by various product properties and exposed
ABSTRACT
to various processing, distribution or storage conditions
Application of optimal experiment design for parameter (see, e.g., Whiting and Buchanan, 1994). These models
estimation enhances the informative character of exper- aid in food safety management systems and risk assess-
iments in view of maximising the parameter estimation ment studies which makes that their (predictive) quality
quality. In this work, dynamic temperature inputs are is of great importance.
applied to optimally estimate two kinetic parameters Nowadays, great –but not always efficient– effort is spent
describing the temperature-dependence of the specific on the determination of kinetic parameters (appearing
growth rate of micro-organisms. Validity of the model in so-called secondary models) which relate microbial
structure is constrained to (i ) the sub-optimal growth dynamics to environmental influences. The classical ap-
temperature range, and to (ii ) relatively small temper- proach comprises a labor-intensive and time-consuming
ature gradients. This paper presents different ways to data generation step. At different levels of the environ-
deal with these model validity constraints during opti- mental factor(s) under study, microbial growth is mon-
mal experiment design. itored. Subsequently, the relationship between treat-
ment combinations and microbial kinetics is identified.
Instead of applying static experiments, optimally de-
INTRODUCTION
signed dynamic conditions may engender a maximum
Within mathematical modelling exercises, experimen- information within a minimum number of experiments.
tal data generation, model structure characterisation, In Bernaerts et al. (2000), time-varying temperature
parameter estimation and model validation make up profiles are designed for optimal estimation of the
the major –though closely intertwined– steps. Exper- Square Root model parameters based upon (a single
imental data can largely contribute to the effectiveness set of) cell density measurements. The so-called Square
(and efficiency) of model structure selection and param- Root model is known as a prototype model describing
eter estimation. Assuming the model structure is per- the influence of sub-optimal growth temperatures on
fectly known, the remaining parameter estimation prob- the maximum specific growth rate of micro-organisms
lem can reasonably benefit from the application of opti- (Ratkowsky et al., 1982). It can be shown that struc-
mal experiment design techniques for parameter estima- tural parameter identifiability requests for dynamic tem-
tion. Practical parameter identifiability requires persis- perature conditions. Amongst different input parame-
tently exciting inputs yielding observations that contain terisations, an optimised step input encompassing an
enough information such that the model parameters can abrupt temperature shift spanning the complete sub-
be estimated uniquely and with the greatest statistical optimal growth temperature domain proves to be most
confidence, i.e., the estimated values approach the true informative. Experimental validation using Escherichia
parameter values and carry a minimum variance. coli K12 (MG1655), however, reveals a transient distur-
bance of the exponential growth regime by this large and the following set of differential equations:
sudden temperature increase (Bernaerts et al., 2002).
This delayed response (intermediate lag phase) violates dn Q(t)
= · [b · (T (t) − Tmin )]2 ·[1 − exp (n(t) − nmax )]
the model validity –implicitly presumed during opti- dt 1 + Q(t) | {z }
µmax (T (t))
mal experiment design– and thus has implications to- dQ
wards optimal design, parameter estimation and their = [b · (T (t) − Tmin )]2 ·Q(t).
dt | {z }
efficiency. µmax (T (t))

At this point, three strategies complying with this tran- (3)

sient growth behaviour yet striving for optimal param- In (3), n(t) denotes the natural logarithm of the cell
eter estimation can be put forward. density [ln(CFU/mL)], nmax is the natural logarithm of
(i ) The model structure is adapted to accurately de- the maximum population density, and Q(t) is a mea-
scribe the intermediate lag phase induced by the large sure for the physiological state of the cells [-], the initial
temperature jump. value of which affects the lag time after inoculation. The
(ii ) The optimal design is recalculated based on the maximum specific growth rate µmax [h−1 ] is a (square
original model yet constraining the admissible temper- root) function of the temperature with b [◦ C−1 ·h−1/2 ]
ature shifts to explicitly guarantee model validity. a regression coefficient, and Tmin [◦ C] the (theoreti-
(iii ) Available experimental data are subjected to ad- cal) minimum temperature for growth. The maximum
vanced data processing. cell density can be assumed temperature-independent
within the temperature region of interest. Note that
in this model –combination of the growth model of
Baranyi and Roberts (1994) and the Square Root model
DESIGN PROBLEM AND METHODOLOGY of Ratkowsky et al. (1982)–, it is assumed that the max-
imum specific growth rate µmax instantaneously adapts
In general, optimal experiment design for parameter es- to the (possibly time-varying) temperature.
timation relies on the minimisation or maximisation of
The sensitivity vector ∂n/∂p is scaled with respect
some scalar function of the Fisher information matrix.
to the (nominal) model parameters and the (nominal)
Under the assumption of uncorrelated normal random
model output, yielding fully relative sensitivity func-
errors with zero-mean and constant variance, the inverse
tions (Munack, 1989). The scaling of the sensitivity
of this matrix defines the parameter variance-covariance
functions by the model parameters takes into account
matrix (see, e.g., Walter and Pronzato, 1997).
the significant difference in order of magnitude between
In this particular study, the modified E-criterion is con- the Square Root model parameters. Dividing by the
sidered for optimal design of a dynamic temperature in- nominal output (no (t)) tends to equalise the weight of
put [T (t)] which yields cell density data [n(t)] enabling small and large cell density counts within the Fisher
efficient estimation of the Square Root model parame- information matrix. Apart from these weighting fac-
ters, i.e., p = [b Tmin ]T . This criterion produces designs tors, no further weighting is needed for this single out-
aiming at circular contours of the identification cost put case. The sensitivity functions are calculated by
within the parameter space (b, Tmin ). In other words, numerical integration of the associated differential equa-
the cost surface is shaped such that an optimisation al- tions. Nominal model parameters (po ) represent an ini-
gorithm can easily find unique parameter values during tial guess for the unknown Square Root model parame-
parameter estimation. Concurrently, the diameter of ters. Due to the model non-linearity, the Fisher informa-
the confidence contours corresponding with the optimal tion matrix depends on the unknown model parameters
design may be minimised as well. The design problem themselves which makes the resulting designs only lo-
can be mathematically formulated as cally valid. Realistic values for the nominal parameters
of E. coli K12 (MG1655) (grown aerobically in Brain
min [Λ(F(t, po ))] (1) Heart infusion broth at pH 7.55), initial states and other
admissible T (t), t∈[0,tf ]
model parameters are: bo = 5.261 × 10−2◦ C−1 ·h−1/2 ,
o
Tmin = 8.689 ◦ C, n(0) = ln (103 ), Q(0) = 200 (no ini-
with Λ(F(t, po )) the condition number of the Fisher in- tial lag), nmax = ln (3 × 109 ) (Bernaerts et al., 2002).
formation matrix F evaluated at the nominal parameter The minimisation of Λ(F(po )) is established by optimal
T
vector po , [bo Tmin
o
] . The Fisher information matrix choice of the degrees of freedom of the parameterised
F is defined as follows: temperature input (, T (t)). Here, temperature step in-
puts containing the following degrees of freedom: the
Z Ã ! Ã !T
tf
∂n ¯¯
g ∂n ¯¯
g initial temperature T1 [◦ C], the time of the temperature
o
F(t, p ) , ¯ · ¯ dt (2) step tshift [h], and the final temperature T2 [◦ C]. Gen-
0 ∂p p=po ∂p p=po
eral constraints are the lower and upper temperature
bound of the sub-optimal growth temperature region,
where n(t) denotes the model output which is given by i.e., [Tlow , Thigh ].
Optimisation of a step function according to the afore- e.g., Bernaerts et al., 2002). When the domain of ad-
mentioned design strategy converges to a large temper- missible temperature inputs is extended to include T (t)
ature jump from the lower to the upper boundary of the inputs encompassing several temperature –but ampli-
sub-optimal growth temperature range. Experimental tude limited– steps, it can be seen that the information
implementation yields colony counts (n(t)) exhibiting content can be significantly increased by introduction of
an intermediate lag phase (Bernaerts et al., 2002). a series of (small ) incremental temperature shifts (see,
e.g., Bernaerts et al., 2002).
MODES OF IMPLEMENTATION Based upon practical feasibility limitations and the in-
formative character of larger temperature changes, a se-
To cope with this intermediate lag, interventions take quence of three temperature shifts of 5◦ C, is selected.
place at the level of the model structure, the design, and
the experimental implementation and data processing. Advanced data processing

Model adaptation The starting point of this novel approach is to maintain

the original problem formulation (without any adapta-
In general, microbial lag is observed at inoculation (t = tion/constraint at the level of the model structure nor
0) and can be appointed to the biochemical adaptation at the level of the temperature input). Available ex-
of cells to the new environmental conditions. In (3), the perimental data are then subjected to advanced data
initial lag phase is modelled by the adjustment function processing. Key idea is that segments of (on themselves)
non-informative (static) data are concatenated such that
Q(t) dQ optimally computed informative data are obtained.
α(t) , with = µmax · Q(t).
1 + Q(t) dt Optimised step temperature inputs based on the mod-
ified E-criterion (1) are all characterised by T1 = Tlow
Assuming that Q(t = 0) is small, the adjustment func- and T2 = Thigh . Instead of performing a dynamic ex-
tion will be smaller than one causing the specific growth periment, the experimental plan consists of two static
rate of the population to be smaller than the (actual) experiments, one at either temperature level. The tem-
maximum specific growth rate. As time proceeds, Q(t) perature levels are now fixed, and the remaining degree
increases exponentially, and α(t) will tend to one. At of freedom, i.e., tshift , can be optimised. In contrast
that point, the population starts exponential growth with previous computations, initial conditions and other
(µ = µmax ). Obviously, the variable Q(t) looses any model parameters are here chosen in perfect agreement
biological interpretation from this point on. Once α(t) with the experimental observations. Based upon the
is approximately equal to one, a change in the actual resulting information, the computed optimal dynamic
environment (causing a change in µmax ) cannot evoke experiment is (re-)constructed through concatenation
an additional lag phase in the growth curve. of (informative) data windows from the static growth
To introduce an intermediate lag within the exponential curves. The methodology is explained in full detail in
growth phase, Q(t) is re-initialised to a small value at Van Impe and Bernaerts (2002).
the time of the temperature shift tshift . This results in
an additional degree of freedom in the fitting procedure.
RESULTS AND DISCUSSION
Doing so, the experimental data can be described well
(i.e., low mean sum of squared errors and no trend in Modes of implementation are validated for E. coli K12
the residuals) which also yields adequate estimates for (MG1655). Material and methods are described in, e.g.
the Square Root model parameters (see below). Bernaerts et al. (2002). Figure 1 summarises the exper-
imental data, the identified model, as well as the con-
Constrained input designs fidence on the estimated model parameters. For full
details reference is made to Bernaerts (2002).
Dynamic temperature inputs are (re-)parameterised and
constrained in view of model validity (i.e., no interme- Model adaptation
diate lag induction). When defining admissible temper-
ature inputs, a trade-off between the optimality of the Using the additional fitting parameter Q(tshift ) yields a
experiment design (in terms of the value of the design good description of the dynamic growth data as well as
criterion) on the one hand, and model validity on the satisfactory confidence levels on the parameter estimates
other, needs to be made. To avoid intermediate lag, the (see Figure 1, left column).
temperature change ∆T must be limited to (at most) Yet, the modelling methodology becomes reasonably im-
5◦ C (based on experimental knowledge). Doing so, the practical within the more general framework of growth
simple step profile including a single temperature shift modelling and prediction making under changing tem-
cannot induce sufficient information to allow accurate peratures as the intermediate lag phenomena are insuf-
estimation of the Square Root model parameters (see, ficiently characterised at present. Moreover, the inter-
 Model adaptation Constrained input design Advanced data processing
24 40 24 40 24 40

21 35 21 35 21 35

18 30 18 30 18 30
lnN (− or …)

T [°C] (−−)

T [°C] (−−)
lnN (−)
T [°C]
lnN
15 25 15 25 15 25

12 20 12 20 12 20

9 15 9 15 9 15

6 10 6 10 6 10
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25
time [h] time [h] time [h]

1.1 1.1 1.1

1.08 1.08 1.08

1.06 1.06 1.06

1.04 1.04 1.04

1.02 1.02 1.02

b/b

b/b

b/b
∧

∧
1 1 1

0.98 0.98 0.98

0.96 0.96 0.96

0.94 0.94 0.94

0.92 0.92 0.92

0.9 0.9 0.9

0.9 0.92 0.94 0.96 0.98 1 ∧ 1.02 1.04 1.06 1.08 1.1 0.9 0.92 0.94 0.96 0.98 1 ∧ 1.02 1.04 1.06 1.08 1.1 0.9 0.92 0.94 0.96 0.98 1 ∧ 1.02 1.04 1.06 1.08 1.1
T /T Tmin / Tmin Tmin / Tmin
min min

Figure 1: Upper plots: Cell density data of E. coli K12 (MG1655) (o) corresponding with the optimally designed T (t)-
conditions (−−). In case of advanced data processing, these data result from appropriate concatenation of windowed
static data. The full line depicts the fitted growth model (3). In case of intermediate lag, Q(t) is re-initialised at
t = tshift . Lower plots: Joint confidence region on b̂ and T̂min (at 99% confidence level) (Bates and Watts, 1988). The
dashed lines delimit the 95% individual confidence interval on b̂ and T̂min . The circle (o) depicts po .

mediate lag cannot be accounted for in the computed Advanced data processing
optimal experiment designs for parameter estimation.
The proposed model adaptation remains a pragmatic The optimal design framework –without additional
solution. A founded model structure adjustment would constraints– is retained but the data generation and
represent a study on its own. data processing step are modified. Realisation of this
procedure (see Figure 1, right column) yields accurate
parameter estimates as expected from the theoretical
optimal experiment design (Van Impe and Bernaerts,
Constrained input designs
2002).

Putting constraints on the abrupt temperature shift Although looking rather artificial, the methodology can
drastically reduces the information content when rely- be justified by the original goal, i.e., exploring the effec-
ing on temperature inputs with a single step. Inclusion tiveness of (theoretical) optimal experiment design for
of a series of small temperature shifts can augment the practical parameter identification. The imposed tem-
information content significantly. As few as three lim- perature input encompasses an instantaneous temper-
ited –yet optimally designed– temperature steps yield ature change at tshift which is usually not feasible in
acceptable parameter estimation results. This is illus- practice (due to, e.g., limited heating capacity). In com-
trated in Figure 1 (middle column). The experimental parison to the preceding methods, model validity under
data reveal that the temperature input does not vio- time-varying temperatures is not put to trial. Moreover,
late model validity (no lagged growth) and suffices to the model remains not valid under the simulated tem-
enable (reasonably) accurate estimation of the Square perature profile. Such temperature step would induce
Root model parameters. an intermediate lag (see above).
Despite the more involved experimental implementa- The data windowing technique provides an attractive
tion, the time-varying temperature yields an excellent alternative to deal with (circumvent) intermediate lag
validation test for the applicability of the growth model phenomena. Static experiments can be easily generated
under (more realistic) dynamic temperature conditions. using common laboratory equipment. The optimisation
of T (t) and subsequent data processing step can be eas- ity restrictions. Int. J. Food Microbiol., Special issue
ily implemented into a userfriendly software program. 73(2-3), 147–159.
Munack, A. (1989). Optimal feeding strategy for iden-
tification of Monod-type models by fed-batch exper-
CONCLUSIONS
iments. In: Computer applications in fermentation
Large temperature jumps initiate intermediate lag and technology, modelling and control of biotechnological
thus cause model validity problems when considering the processes (N.M. Fish, R.I. Fox and N.F. Thornhill,
proposed growth model. Because the model structure is Eds.). pp. 195–204. Elsevier. Amsterdam.
assumed correct within the framework of optimal ex- Ratkowsky, D.A., J. Olley, T.A. McMeekin and A. Ball
periment design for parameter estimation, the lag phe- (1982). Relationship between temperature and growth
nomenon should be dealt with in one or another way. rate of bacterial cultures. J. Bacteriol. 149, 1–5.
Different solutions coping with this model validity lim- Van Impe, J.F. and K. Bernaerts (2002). (submitted).
itation have been put forward. Overall, the extracted Walter, E. and L. Pronzato (1997). Identification of
parameter estimation accuracy is in all cases highly ac- parametric models from experimental data. Springer.
ceptable. The selection of either approach will depend Masson.
on user-specific preferences, a priori knowledge on the Whiting, R.C. and R.L. Buchanan (1994). Microbial
micro-organism, and objectives. modeling. Food Technol. 48(6), 113–120.

ACKNOWLEDGEMENTS BIOGRAPHY
KRISTEL BERNAERTS was born in Bonheiden, Belgium,
This research is supported by the Research Council of
in 1974. She received her Master’s degree Engineering in
the Katholieke Universiteit Leuven as part of projects Chemistry and Biochemistry with specialisation in indus-
OT/99/24 and IDO/00/008, the Institute for the Pro- trial microbiology from the Katholieke Universiteit Leuven
motion of Innovation by Science and Technology (IWT), in 1997. Since 1997, she is working as a research assistant at
the Fund for Scientific Research - Flanders (FWO) as BioTeC - Bioprocess Technology and Control, K.U.Leuven.
part of project G.0213.02, the Belgian Program on In- She received a doctorate in applied biological sciences from
teruniversity Poles of Attraction and the Second Multi- the Katholieke Universiteit Leuven in 2002. Her main re-
annual Scientific Support Plan for a Sustainable Devel- search interests are in the field of predictive microbiology,
opment Policy, initiated by the Belgian State, Prime more particularly, the design of optimal experiments aiming
Minister’s Office for Science, Technology and Culture. at accurate model building.
JAN F. VAN IMPE was born in Varsenare, Belgium, in 1965.
The scientific responsibility is assumed by its authors.
He received his master’s degree in electrical and mechani-
cal engineering from the University of Ghent in 1988, and
REFERENCES a doctorate in applied sciences from the Katholieke Univer-
siteit Leuven in 1993 (ESAT Laboratory). Thereafter, he
Baranyi, J. and T.A. Roberts (1994). A dynamic ap- joined the Department of Food and Microbial Technology of
proach to predicting bacterial growth in food. Int. J. the same university (since 1995 as an associate professor).
Food Microbiol. 23, 277–294. Since 2001, he is professor at the Department of Chemical
Bates, D.M. and D.G. Watts (1988). Nonlinear regres- Engineering. During that period he started the BioTeC re-
sion analysis and its applications. John Wiley and search group, which has nowadays about 20 members. He
Sons, Inc.. New York. teaches courses in system analysis, bioreactor engineering,
and process engineering and control. His research interests
Bernaerts, K. (2002). Dynamic experiment design and
are mainly in nonlinear systems and identification, nonlinear
validation in predictive microbiology: optimal esti- and adaptive control, and its applications in model-based
mation of two-parameter microbial growth kinetics as optimization and control of (bio-)chemical conversion pro-
function of temperature. Phd Thesis 519. Department cesses. He has (co-)authored about 250 papers in these ar-
of Food and Microbial Technology, Faculty of Agri- eas. In 1995 he received the Barco Award of the Fund for
cultural and Applied Biological Sciences. Katholieke Scientific Research, in 1996 he was laureate of the Belgian
Universiteit Leuven. Royal Academy, and in 1997 he was awarded the KIWANIS
Bernaerts, K., K.J. Versyck and J.F. Van Impe (2000). prize of the KULeuven Research Council. In 2002 he became
On the design of optimal dynamic experiments for a member of the board of the EUCA, the European Union
parameter estimation of a ratkowsky-type growth ki- Control Association.
netics at suboptimal temperatures. Int. J. Food Mi-
crobiol. 54(1-2), 27–38.
Bernaerts, K., R.D. Servaes, S. Kooyman, K.J. Versyck
and J.F. Van Impe (2002). Optimal temperature in-
put design for estimation of the Square Root model
parameters: parameter accuracy and model valid-
AUTHOR
LISTING
 AUTHOR LISTING

Abreu I....................................... 141 Gummaraju K............................ 173

Agioux L. .................................. 99
Ahrné L.M................................. 127 Hamdami N. .............................. 109
Albasi C..................................... 197 Hartmann C. .............................. 169
André C. .................................... 151 Herlihy D................................... 65
Arendsen A.R.J. ........................ 180
Arnold S. ................................... 86 Ioannou I. .................................. 99

Baucour P. ................................. 119 Jensen B.B.B. ............................ 132/146

Becker T. ................................... 86
Bernaerts K................................ 212 Kurz T. ...................................... 86
Bhushan N. ................................ 173
Brandam C................................. 202 Laffort J.-F. ............................... 207
Brecht J.K.................................. 35 Le Bail A................................... 109
Butler F...................................... 18 Lebouché M. ............................. 136
Leuliet J.-C................................ 136/151
Carney E.................................... 18 Lian G. ...................................... 75
Chevallereau V. ......................... 99 Luna R....................................... 35
Coppenolle H............................. 22/114
Corrieu G................................... 99 Mallikarjunan P......................... 91
Cronin K.................................... 81/119 Mann D. .................................... 185
Cummins E. ............................... 18 Martin D.R. ............................... 75
Masneuf I. ................................ 207
Darius P.L.................................. 31 Massonnet G. ............................ 31
de Jong P. .................................. 13/45 Mauris G. .................................. 99
De Rijck G................................. 114 Melrose J.R. .............................. 75
Delaplace G. .............................. 136/151 Meyer X.M................................ 202
Delgado A. ................................ 86/169 Mitzscherling M........................ 86
Delia M.-L................................. 197 Montañez J.C. ........................... 35
Demir A.D................................. 81 Monteau J.-Y............................. 109
Downing C.J.............................. 50/55 Movahedkhah M. ...................... 159
Duffy G. .................................... 18 Murnleitner E. ........................... 86

Eder H.H.................................... 5 Nally P. ..................................... 18

Neutelings T.............................. 114
Fellner M. .................................. 86 Nielsen J.S................................. 132
Fernandez P. .............................. 35
Fonseca S.C............................... 35 O'Connor B. .............................. 103/185
Frias J.M.................................... 35 O'Brien S................................... 18
Friis A........................................ 132/146 O'Connor B. .............................. 65
O'Dwyer A. ............................... 60
Griffin B. ................................... 50 Oliveira F.A.R........................... 35/141
Guérin R. ................................... 136/151 Olsson E.E.M. ........................... 127
 AUTHOR LISTING

O'Mahony T............................... 55 Strehaiano P. ............................. 197/202/207

Perrot N. .................................... 99 Taillandier P.............................. 207

Pingaud H.................................. 202 Thiel D. ..................................... 159
Pommier S. ................................ 197 Torrez C. ................................... 151
Portier K.M................................ 31 Trägårdh C. ............................... 127
Proth J........................................ 202 Trystram G. ............................... 99

Riverol C. .................................. 103 van Berkel A.I. .......................... 180

Ronse G. .................................... 136 van de Wiel J............................. 45
Roudot A.-C. ............................. 27 van der Knaap H.C.M. .............. 31
Van Impe J.F. ............................ 212
Schrevens E. .............................. 22/31/114 Verschueren M.......................... 13/45
Serra A....................................... 207 Vissers M.M.M. ........................ 13/45
Sheridan J. ................................. 18 Vitrac O..................................... 164
Sheridan P. ................................ 91
Shilton N. .................................. 91 Weerts A.H................................ 75
Smit E........................................ 13/45 Winkless B. ............................... 185
Straatsma H. .............................. 13/45