AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

Catalyst Heating Modeling, Simulation and Optimization

Rodrigo Peralta Moraes Ruiz
Robert Bosch Ltda

Breno Felippe Coimbra Maia, André Menegon Godoy, Rosalvo Bertolucci Filho, Fabrício Tenório
Vicente
Robert Bosch Ltda

André Pelisser, João Vitor Carrilho Soares de Carvalho

ETAS

[email protected], [email protected] , [email protected] , [email protected] ,

[email protected] , [email protected] , [email protected]

ABSTRACT between emission levels, idling quality and driver comfort.

In this way, finding the parameterization that establishes the
The product development engineering for emission best compromise between these magnitudes becomes
control has become competitive and complex, mainly due to complex due the various factors that permeate the day-to-day
the continuous evolution of the legislation. The catalyst of calibration engineering, such as: System complexity,
heating techniques domain is crucial for vehicle emission project time and cost of tests.
control. However, understanding which heating strategy
promotes the lowest level of emissions is complex in The combination generated by this scenario brings to
proportion to the number of variables to be explored, making light the demand for more analytical and automated methods
the search for the optimal point a real challenge for of data analysis, in order to help calibration engineering to
automakers and systemists. This work aims to develop a find optimized solutions in reduced time intervals, keeping
model-based catalyst heating methodology, which helps in reliability. This type of approach, based on advanced data
the discovery and understanding of possible optimal analysis to find calibration solutions, is called “Model Based
operating points not explored yet regarding the compromise Calibration”.
between emission levels and engine idle speed vibrations. A
Machine Learning model was developed and validated based This work aims to develop and explore the “Model
on real catalyst heating data, produced by an automatic Based Calibration” methodology using catalyst heating
starting system that interacts with the engine calibration calibration as an application scenario. The methodology
parameters via Design of Experiment (DoE). The simulation allows an approach through a simulation tool based on data
results enabled optimizations through an efficient and acquired in a vehicle. Data-based simulation opens horizons
analytical method. This method allows a quality delivery of of prediction and analysis through a virtual interaction
results in a short period of time and enables the search for between control parameters and the engine. This allows
new horizons on catalyst heating strategies. obtaining results in a reduced time interval, in addition to a
better understanding between the dependencies of some
INTRODUCTION phenomena that are involved in catalyst heating.

Engine calibration is composed of several engineering METHODOLOGY

steps, which must converge and contribute to the approval of
a vehicle in terms of emissions [1]. Like any interesting
engineering challenge, these calibration steps deal with The present study was applied to a 1.6 Liter Port Fuel
solutions which are concurrent each other, making the search Injection engine, fueled with ethanol and with variable intake
for the balance point a journey full of obstacles. camshaft control. The methodology is based on an automatic
calibration system, which can interact with the vehicle
One of the most important factors in controlling through servomotors [1]. Through the DoE (Design of
emissions is the catalyst heating strategy. Catalyst heating is Experiment) methodology, the boundary conditions to be
essential for the after-treatment system reach as fast as explored are established and the tests performed. From the
possible the temperature of high conversion efficiency to data obtained, an automated post-processing system is tasked
reduce tailpipe emissions in a vehicle. However, some with obtaining specific information about the variables of
control variables of this strategy, establish competition interest and feeding a Machine Learning model with these

1
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

variables. The Machine Learning model, in turn, has the open the software interface and develop a code in Python to
function of providing a broader view of the system, and execute the FID and perform the HC measurement.
enables interaction and understanding of phenomena, via
simulation. The main purpose is, through the model, to
understand the relationship between the control variables,
with some outputs of interest. The outputs are: Catalyst
temperature, pre-catalyst hydrocarbon emissions, post-
catalyst hydrocarbon emissions, idling vibration and fuel
debit potential. An algorithm based on Fourier transform was
developed to extract, from the engine speed signal, the
magnitude of vibration at idle speed.

AUTOMATION SYSTEM - The automation system

used in this work was developed in [1].

The system consists of emulating the behavior of a

driver by actuating the brake and clutch pedals, as well as the
ignition key. For this, servomotors were used and
commanded via ETAS ES930, which is programmed using Figure 3. User Interface (Protocol AK).
a script developed in INCA FLOW. Figure 1 shows the
schematic. Source: [2].

The exhaust after-treatment instrumentation used in

this study is composed by a lambda probe, HC pre catalyst
measurement, Catalyst Thermocouple and HC post catalyst.
The figure 4 illustrates the instrumentation.

Figure 1. Automation System Diagram.

Source: [1].

The system is capable to measure hydrocarbons

emissions level, integrated with the automation system. It is Figure 4. After-treatment instrumentation.
possible through the hardware HFR500, which is a Flame
Ionization Detector (FID). It is used to measure hydrocarbon Source: The Authors.
(HC) concentration in a sample gas with a fast response time
[2]. The following figure describes the HFR500 diagram. DoE PLANNING And EXECUTION – Design of
experiment is a method to establish boundaries to get model
from systems based on its data [3]. The tool used to design
and stablish the boundaries in this study was ASCMO Static
Planning as well as the tool used to model the system
behavior was ASCMO Static Modeling. The complete
workflow is described in ETAS ASCMO User guide, as
showed in Figure 5.

Figure 2. HFR500 Diagram.

Source: [2].

For the integration of this HFR500 hardware with the

automation, it was necessary to acquire the AK protocol to

2
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

Figure 5. DoE Workflow.  Monitoring oil temperature and half Catalyst

temperature to reach 25.5℃ and 250℃,
Source: [3]. respectively.
 Set calibration variables according to DoE.
 Start recording.
 Start engine through robot interface (actuator and
In this study, ASCMO Static Planning was used to servomotor).
generate the DoE matrix. The methodology to design the  Wait until engine reach 70℃ (in idle condition).
experiment in this tool, is based on Gaussian processes. It  Stop engine.
allows to explore several numbers of relationships of  Stop recording.
complex phenomena with relatively low effort of measuring  Wait the Soak time with the fan controlled until the
to reach high model accuracy [3]. oil temperature returns to 25.5, then rerun the
previous process with the next DoE line. So
The method used to spread the points to be explored successively until the last line of the DoE.
around the sample space in order to balance good model
accuracy with low effort of measurement, is called “Space The Figure 6 shows the INCA FLOW Script.
Filling Method”. The Table 1 was obtained from ASCMO
User Guide and shows the comparison between the Space
Filling Method” against the conventional one, in terms of
measurement amount.

Table 1. Classic grid measurement versus Space

filling method effort.

Source : [3]

Figure 6. Developed flowchart.

Source: The Authors.

The green Engine Start Counter Iteration block

In this study, de DoE was performed varying 4 control
illustrates the set of decision scripts while the other blocks
variables, which are:
are actions defined by the automation process. The
automated system can measure a variety of DoE scenarios.
 Idle Speed Target: 1200 RPM ~1800 RPM
DATA POST PROCESSING – The data post
 End of injection angle: 50° ~ 250° processing is a script tool based on Matlab® 2020b. Its main
function is to find information in the measurements and
 Intake Valve opening angle: 12° Before Top concatenate this in a structured table. This table will be used
Dead Center (BTDC) ~ 15° After Top Dead later to be both training and validation data in the ASCMO
Center (ATDC) machine learning model.

 Spark efficiency: 40% ~ 100% The information which will be showed in the table, are
both related to the input variables which were defined in the
To cover all the variable range, 70 measurements DoE, as well as the output variables, that can be generated
were performed. internally in the post processing tool. Between the main
outputs to be observed, are:
To execute the DoE, INCA FLOW was used.
the following steps were developed:  Maximum Catalyst temperature: Maximum
 Load DoE block. value reached in the measurement

3
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

 Hydrocarbon concentration: Integrated value

during warm up

 Relative Fuel debit potential: Potential to

remove start and after-start fuel mass
compared with baseline calibration

 Idle Vibration Magnitude

There are many other outputs that can be generated

by the post processing script. In general, the post processing
workflow is described according with the figure 7.

Figure 9. Integrated HC Concentration.

Source: The authors.

Relative Fuel debit potential– When optimizing catalyst

heating, several combustion phenomena are involved.
Different combinations of calibration parameters can
generate different effects in combustion chamber during the
flame propagation, mixture formation and gases exchanges,
hence, it will affect the exhaust emissions and catalyst
temperature. One of the proposals in executing the DoE
methodology, is to sweep the control variables in regions
which make sense from the calibration perspective in order
Figure 7. Post processing Workflow. to get a wide overview of engine behavior in these regions.
When an optimum dataset, which reduces the post catalyst
emissions, is found, it is interesting to understand its effect
Source: The authors.
in the combustion phenomenon, looking through a variable
that can describe how much the fuel mass is being seized in
Maximum Catalyst temperature– The maximum catalyst
the combustion. In this way, this variable will give the
temperature is obtained by the maximum value reached
information if there is or not, potential to remove fuel mass,
during the measurement, as illustrated in the figure 8.
even after an optimization. It increases the optimization
benefits and gives a deeper vision of the optimization effects
in combustion.

If an optimized calibration dataset, from catalyst

heating perspective, generated a phenomenon in combustion
which seizes better the fuel mass, the availability of oxygen
concentration in the exhaust will be lower, and the lambda
signal will be affected. It indicates that there is a potential to
remove fuel mass.

The figure 10 illustrates a situation that a Catalyst

heating optimized calibration dataset, generated a better
Figure 8. Maximum Catalyst temperature. seizing of fuel mass during the combustion, compared with
a baseline calibration.
Source: The authors.

Hydrocarbon Concentration–The hydrocarbon concentration

both in pre catalyst and post catalyst, is obtained integrating
the value read from FID. The final integrated value is the area
under the HC curve, as illustrated in the figure 9.

4
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

fuel mass when compared with baseline. When the dataset

produced a higher oxygen concentration in the exhaust, the
variable value is negative, showing that there is no
possibility to remove fuel mass when compared with
baseline calibration.

Vibration Modeling – The engine vibration in idle

condition is a very relevant output to be considered when
optimizing catalyst heating. The main idea in catalyst heating
is to provide thermal energy to the exhaust manifold and it
can generate, as a side effect, high levels of vibration in the
engine. The tradeoff between the catalyst heating efficiency
(which is very correlated with the thermal energy availability
Figure 10. Positive fuel debit potential. in exhaust manifold) and the engine vibration magnitude, is
the key to find an emissions calibration solution that can
Source: The authors. balance emissions level and driver comfort. Besides that,
vibration is a “Noise, Vibration and Harshness” (NVH)
Regardless the dataset which generated this constraint and, as well as emissions, it needs to be measured,
behavior in lambda signal, the difference between the controlled, and minimized.
baseline calibration lambda signal (orange curve) and the
optimized calibration lambda signal (brown curve), shows To obtain a variable which can describe the engine
the potential that the optimized calibration contains to vibration’s level, a model was developed to estimate, based
improve mixture, by reducing fuel mass. This is called in this on engine RPM signal, the vibration magnitude. The Matlab
study as “Fuel Debit Potential”, represented in the figure by script, which contains the model, is based on Discrete
the green area. Fourier Transform (DFT).

By the other hands, an optimized dataset, from DFT is applied when there is interest in analyzing
catalyst heating perspective, can generate a worst fuel mass discrete signals in frequency domain [4]. A discrete signal
seizing during the combustion, producing higher oxygen is a signal which is composed by samples. The following
concentration in the exhaust, and the lambda signal will be figure shows a discrete signal composed by “N” samples.
affected as well. The figure 11 illustrates this situation.

Figure 11. Negative fuel debit potential.

Figure 12. Discrete signal.
Source: The authors.
Source: The authors.
In this case, the optimized calibration from catalyst
heating perspective (brown curve) , does not show potential The Fourier series exponential representation of a
to remove fuel mass when compared with baseline discrete signal, can be described following the equation
calibration ( orange curve). This is called in this study as a below [4] :
negative Fuel debit potential, represented in the figure by the
red area.
𝑥[𝑛] = 𝑎 𝑒 (1)
Therefore, the variable “Relative Fuel Debit Potential” 〈 〉
can be constructed. And it uses the measurement produced
by the baseline calibration. When the dataset produced a
lower oxygen concentration in the exhaust, the variable value
is positive, showing that there is the possibility to remove

5
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

Where 𝑎 is a complex phasor which weights the In frequency domain, it is possible to understand that
periodic component in the specific frequency. The frequency the roughness in engine speed signals is reflected in the
will be always a multiple of the natural frequency, 𝜔 . spectrum through higher magnitude component modules 𝑎
, at low frequencies. In this way, the following equation was
To build the frequency spectrum signal representation, used to correlate the engine speed roughness with a scalar
it is needed to find each 𝑎 component and match it with its variable:
respective frequency. The following equation describes how
𝑵
to obtain the components.
𝑽𝒊𝒃𝒓𝒂𝒕𝒊𝒐𝒏 𝑴𝒂𝒈𝒏𝒊𝒕𝒖𝒅𝒆 ≅ |𝒂𝒌 |𝒅𝝎 (3)
1 𝟏
𝑎 = 𝑥[𝑛]𝑒 (2)
𝑁
〈 〉 Where “N” is the amount of magnitude
components, and 𝜔 is the frequency. This approach allows
This methodology can be applied in engine rpm signal to transform the spectrum magnitude in a single variable
in order to obtain the magnitude of the periodic components which can represent all engine speed roughness level in idle.
which composes the signal. This value will be a direct The following figure shows the spectrum and the level
information about vibration. Three different engine vibration associated. The level is extracted from the last value of
levels were used to develop the vibration level model. The integral curve indicated the following figure.
figure below, shows these three signals.

Figure 15. Integrated value of spectrum

Figure 13. Engine Roughness in time domain. magnitude.

Source: The authors. Source: The authors.

Using DFT, it is possible to estimate the coefficients MACHINE LEARNING MODEL USING
𝑎 , and correlate them with their respective frequencies to ASCMO
build the spectrum visualization. The coefficients 𝑎
represent the amplitude module of each periodic component ETAS ASCMO modelling software was chosen as
which describes the signal. The following figure shows the the platform for generating the catalyst heating behavior
spectrum visualization of the three engine RPM signals estimation. This software relies on supervised machine
showed before. learning algorithms to generate statistical models based on
the Bayesian modelling method.

The Bayesian method consists on defining a prior

distribution based on the expected system behavior, and then
combining this prior distribution with sampled real data,
resulting in a posterior distribution. This posterior
distribution can be interpreted as a set of functions that, when
combined, may define the observed phenomena. Figure 16
shows an example of the Bayesian method for a 2-point
dataset. The first panel show the prior distribution, solely
based on assumptions and expectations of the phenomena.
The second panel shows the posterior distribution - the
combination of the prior distribution and the 2-point dataset.
Figure 14. Engine Roughness in frequency domain.

Source: The Authors.

6
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

RESULTS AND DISCUSSIONS

MODEL SIMULATION – In ASCMO the model is

represented in a 2D screen which contains all the
dependencies profiles between the inputs explored in the
DoE and the outputs of interest. In the horizontal axis are the
inputs and in the vertical axis, the outputs. The figure 18
shows the model.

Figure 16. Example of the Bayesian method application for

a 2-point dataset.

On the right panel, the dashed lines represent 4

possible distributions, and the continuous line the mean
value of those 4 distributions, which ends up being the model
prediction. As an interesting benefit, the Bayesian Method
also estimates the models in this case represented by the
shaded area, accounting for twice the local standard
deviation for each input x.

ASCMO implements the classic approach of the Figure 18. ASCMO Model representation.
Bayesian Method using a Gaussian Process as the basis for
the prior distributions. Since this methodology derives from Source: The Authors.
the Gaussian Probability Distribution, a stochastic process
governs the properties of those functions [5]. It is also noted For each input and output, there is a distribution curve
that this approach dispenses parametric models and instead, based on the measured data, which represent the dependency
defines a prior probability distribution over functions between the couple of input and output, it is, how each
directly [6]. specific input affects each output. For example, the green
square marked on the figure 18, represents the dependency
Detailed mathematical expansions on both the profile of the relationship between Vibration Magnitude and
Bayesian Method and the Gaussian Regression Process can Spark Efficiency. The red square, marked on the figure 18,
be found at various sources, such as Rasmussen X, Bishop represents the dependency profile of the relationship
Y and Pelisser et al Z. between the Intake Camshaft Valve Opening Angle and the
HC concentration pre catalyst. This logic is valid for all the
As a result of those given statistical and data-based distribution curves plotted on the screen.
approaches, the modelling method applied by ASCMO suits
perfectly the needs for this study regarding the expected high In the simulation tool, the Leave-One-Out method is
complexity of the correlations, model self-validation and used to make a pre analysis of the prediction quality. The
good model quality even when big datatsets are not method consists in estimating n models, each with n-1
available. training data. Afterwards, the precision is determined based
In this study, after the data post processing, a post on the point which is not involved in the model estimation
processed matrix is available to feed the ASCMO on order [3].
to estimate the virtual model of engine behavior. The
following figure illustrates this process. The model quality is quantified by the Root Mean
Square Error (RMSE) and the 𝑅 .

The RMSE is defined as the following equation [3].

∑ (𝑋 −𝑋 )
𝑅𝑀𝑆𝐸 = (4)
𝑛

Where 𝑛 is the number of measured points. 𝑋 is

the predicted value and 𝑋 is the measured value.

Figure 17. ASCMO modeling workflow. The 𝑅 is obtained from the comparison between the
variance that remains after the model training and the
Source: The Authors. variance concerning the mean value of all measuring data
[3].

7
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

∑ (𝑋 −𝑋 ) To obtain the vibration absorbed by the accelerometer,

𝑅 =1− (5) a specific signal processing method was used, which is the
∑ (𝑋 −𝑋 )
Spectrum Energy Estimation, the method is the spectrum
Root Mean Square (RMS). To get the RMS value in a
The figure 19 shows the RMSE and 𝑅 for each output. spectrum representation, the root sum square of all spectral
magnitude components (𝒂𝒌 ) is calculated [5]. The following
equation describes the RMS value of a spectrum

𝑎 𝑎
𝑅𝑀𝑆 = + 𝑎 + (6)
2 2

The 𝑎 term is the same presented before. RMS is a

single number which represents the level of spectrum energy
[7]. The figure 21 shows the relationship between each axis
of accelerometer signal (vertical axis) and the RPM vibration
Figure 19. 𝑅 and RMSE of each output prediction. magnitude estimator (horizontal axis).

Source: The Authors.

The Vibration Magnitude, Catalyst Temperature and

HC Concentration Post Catalyst, showed higher precision in
prediction using Leave-One-Out method when compared
with HC pre catalyst and Relative Fuel Debit Potential. One
possible reason for this is due these variables are very
connected directly with the combustion residues, hence, very
susceptible to combustion variations. In this study, a
validation will be presented later to understand the reasons
why there is difference in the prediction quality between the
outputs.

VIBRATION MODEL VALIDATION – To validate

the vibration model based in the Engine RPM signal, a three-
axes accelerometer signal was used to be compared with the
RPM Vibration Magnitude estimation via ASCMO model.
Figure 21. Accelerometer axes versus RPM Vibration
The accelerometer is located under the handbrake to Magnitude Estimator.
get the vibration information in the vehicle interior.
Source: The Authors.
The x-axis is measuring vibrations in longitudinal
direction. The y-axis is measuring in transversal direction, The figure 21 demonstrates clearly that Vibration
the z-axis is measuring in the normal direction. The figure 20 Magnitude Estimator has a directly proportional relationship
illustrates the accelerometer axes configuration in vehicle. with the vibration absorbed by the accelerometer in all axes.
For this reason, in this study, the RPM Vibration Magnitude
Estimator was used as reference to measure and qualify the
Engine RPM vibration level, once it is a information present
in any Engine Control Unit and turns the methodology more
flexible.

GAUSSIAN MODEL VALIDATION – To validate

the Gaussian model, three different tests were performed in
three different model regions. The first boundary condition
is according to the following figure:
Figure 20. Accelerometer Axes.

Source: The authors.

8
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

Figure 22. Test 1: Boundary Condition. Figure 24. Test 3: Boundary Condition.

Source: The Authors. Source: The Authors.

The boundary condition in the first test (Test) is : The boundary condition in the third test is:

 Idle Speed Target: 1500 RPM  Idle Speed Target: 1500 RPM

 End of injection angle: 242,9 (°)  End of injection angle: 228 (°)

 Intake Valve opening angle: 6° BTDC  Intake Valve opening angle: 9° ATDC

 Spark efficiency: 50%  Spark efficiency: 90%

The second boundary condition (Test 2) is showed in These three conditions were reproduced in vehicle to
the figure 23. compare the in-vehicle results with the model prediction. For
each condition, 5 measurements were performed, totalizing
15 measurements. The purpose is to cover repeatability
between measurement, once that all the phenomena which
are being analyzed are related with combustion, a stochastic
phenomenon.

The table 2 shows the results of the first test.

Table 2. Test 1 Results.

Source : The authors

Figure 23. Test 2: Boundary Condition.
Test 1

Source: The Authors. Vibration Magnitude Temp Cat HC pos Cat HC pre Cat Relative Fuel Debit Potential(%)
Model 4,73 649,24 18,80 297900,00 13,25
Mean 4,03 631,79 21,74 283134,35 5,74

The boundary condition in the second test is: Error

Error(%)
0,70
17,37
17,45
2,76
2,94
13,54
14765,65
5,22
7,51
130,79
Std 0,05 2,21 2,02 13298,33 4,88

 Idle Speed Target: 1500 RPM The class “Model” is the ASCMO model prediction.
“Mean”, is the mean result of each variable between the 5
 End of injection angle: 240 (°) measurement which were performed in that test. “Error” is
de absolute difference between the model and the mean.
 Intake Valve opening angle: 1° BTDC “Error (%)” is how much percent the model is distant from
the mean, and “Std” is the standard deviation of each
 Spark efficiency: 75% variable between the 5 measurements which were performed
in that test.
The third boundary condition (Test 3) is showed in the
following figure. The table 3 shows the results of the second test.

Table 3. Test 2 Results.

Source : The authors

9
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

Test 2 example, the temperature. The colder the engine coolant

Vibration Magnitude
Model 3,33
Temp Cat
538,09
HC pos Cat
25,13
HC pre Cat
293690,00
Relative Fuel Debit Potential(%)
39,93
temperature, or the intake air temperature, more
Mean
Error
2,64
0,69
525,41
12,68
26,04
0,92
322308,13
28618,13
36,07
3,86 unpredictable the combustion result. If considered that the
Error(%) 26,10 2,41 3,52 8,88 10,70
Std 0,04 4,56 0,73 10484,64 4,48 present model is proposed to predict from the cold phase,
until the warmed-up engine (the whole warm up phase), a lot
The table 4 shows the results of the third test. of randomness is involved. For this reason, a good qualitative
prediction is acceptable in this model application.
Table 4. Test 3 Results.
The HC pre catalyst concentration is being validated,
Source : The authors but it will not be considered in the model optimization. The
purpose to have HC pre catalyst in the analysis is to
Test 3 understand the effects in combustion phenomena before the
Vibration Magnitude
Model 2,61
Temp Cat
496,33
HC pos Cat
33,60
HC pre Cat
328840,00
Relative Fuel Debit Potential(%)
40,34
effect of catalyst conversion. But the main focus in
Mean
Error
2,41
0,20
522,17
25,84
28,67
4,93
320765,87
8074,13
38,05
2,29
optimization will be the HC post catalyst.
Error(%) 8,10 4,95 17,21 2,52 6,02
Std 0,02 1,13 0,63 4879,67 3,82

In the first test, the Relative Fuel Debit Potential did

The purpose of the tests is to sweep the model in not fit in a quantitative quality. The reason is due the test
three significantly different regions to explore the prediction condition, especially regarding the Spark Efficiency value
capacity and the results robustness. With this kind of (50%). It promotes instability in the combustion, turning
analysis, it is possible to understand if the amount of data is harder the repeatability of the test and increasing the standard
being enough to provide, or not, a good confidence interval deviation between the measurements. As the Fuel Debit
in all regions. Potential is based in the lambda signal, which depends on the
oxygen concentration in the exhaust manifold, it
The prediction robustness can be classified in 2 repeatability is directly related with the combustion residues,
classes according with the capacity to minimize the error hence, if the combustion is unstable the Fuel Debit Potential
between the model and the real engine behavior. The first will be as well.
one is the qualitative prediction capacity, which is related to
the model capacity in always hit the result trend, but not the Based on the validation results analysis and the
absolute value. The second, is the quantitative prediction previous considerations related to the combustion
capacity, which is related to the model capacity in hit the phenomenon, the model demonstrated high capacity in
result values with a high accuracy. prediction robustness in all regions for all variables of
interest.
In the first test, the prediction robustness of Catalyst
temperature and HC post catalyst, were excellent, and it fits OPTIMIZATION POTENTIAL – Once the model
in a quantitative prediction class. The prediction robustness prediction is reliable, the optimization through simulation is
of Vibration Magnitude was good, and it still has some level possible and reliable as well.
of confidence to be considered a quantitative prediction, but
not with the same quality of HC post catalyst and Catalyst ASCMO has different optimization options. In the
Temperature. The prediction of Relative Fuel Debit Potential single criteria optimization, a variable is optimized
does not show robustness to be a quantitative prediction class according to the gradient descent method. When it comes
in this specific test, but it can still be considered as a about gradient, it is related with functions which contains
qualitative prediction, once the model predicted that there multiples inputs and a single output.
was potential to remove fuel (positive Fuel Debit Potential)
and it fits with the reality. In the first optimization simulation, the single
criteria was used to minimize HC post catalyst, establishing
In the second test, the prediction of Catalyst a maximum value to engine speed idle target in 1400 RPM.
Temperature, HC post catalyst, Relative Fuel Debit potential
were excellent and can be classified as a quantitative Compared with the baseline calibration, the
prediction. The Vibration Magnitude prediction was good optimization simulation reduced nearly in 50% the HC post
and can be understood as well as a quantitative prediction. catalyst concentration, but it increased nearly 24% the
vibration magnitude. The fuel debit potential after the first
In the third test all the predictions value are very well optimization simulation is estimated nearly in 34%
fitted with the real engine behavior, and all can be considered compared with the baseline calibration.
as a high quality of quantitative prediction.
To explore the tool potential in emission reducing,
When it comes about combustion, quantitative capacity keeping the vibration magnitude a second optimization
is something very sensible. Combustion is a stochastic simulation was carried out. A a hard upper bound was
phenomenon, it means, the result is influenced by a lot of established in vibration magnitude in the same value of
random and unpredictable combinations of inputs. Some baseline calibration. In the second optimization with
environmental conditions, increase this randomness, for vibration magnitude hard upper bound, the HC post catalyst

10
AEA – Brazilian Society of Automotive Engineering - SIMEA 2023

concentration was reduced nearly in 42% compared with the development of data products with high excellency, most
baseline calibration. The vibration magnitude is the same of efficient and cost-effective.
baseline and the fuel debit potential is 81% compared with
the baseline. The initial results open a promising window for
application of the method for higher complex systems with
Although the HC post catalyst has a higher value in the different engine configurations. This method flexibility
second optimization compared with the first one, the fuel characteristic is one of the points of interest from this
debit potential increased significantly, and the vibration working group and some future activities are already planned
magnitude was kept if compared with the baseline. With a as next steps.
fuel debit potential in this in a higher value, the fuel mass
could be reduced in the calibration and the final proposal It is also an important point of interest the
could be even more optimized than in the first optimization consolidation of this work as a different method (with
attempt. It can indicate that, in the optimization criteria, it is potentially superior quality) compared to the current and
needed to pay attention not only in the HC post catalyst but traditional calibration tasks for specific workpackages. It
in fuel debit potential as well. opens a real possibility of inclusion of a specific calibration
service into the supplier portfolio.
A third optimization simulation was performed. The
criteria was, minimize HC post catalyst, keep de vibration REFERENCES
magnitude in a hard upper bound with the same value of the
baseline calibration and maximize fuel debit potential. In the [1] Alexandre Esteves, Alexandre Kawamoto, André
third optimization, HC post catalyst was reduced in 31% Pelisser, Davi Carmelutti. “Automated in-vehicle engine
compared with baseline calibration. Vibration energy was calibration to optimize emissions levels using machine
reduced in 18% and the fuel debit potential increased to learning” – SIMEA 2021. São Paulo.
162%. Besides that, in this optimization the engine speed idle
target was calculated in 850 RPM, which could bring an [2] Cambustion. HFR500 Fast Response FID Hydrocarbon
effect in fuel consumption as well. Measurement System – User Manual version 2.9.

It is important to find the tradeoff between maximizing [3] ETAS GmbH. ETAS ASCMO Static V5.2 - User’s
fuel debit potential and reducing HC post catalyst to Guide, 2018.
understand the how it affects the final emissions values in
emission cycles. In this study, these optimizations were not [4] Alan Oppenheim, Alan Wilsky. Signals and Systems:
validated in emission cycle, and it will be considered in Second Edition. Prentice Hall, New Jersey.
future studies.
[5] Rasmussen, C. E. and C. K. I Williams (2006).
CONCLUSION Gaussian Process for Machine Learning. MIT Press.

The method of Model Based Calibration allows the [6] Bishop, C. M (2006). Pattern Recognition and Machine
optimization of complex systems by comparing the model Learning. Springer.
predictions and real measurements in a very large operation
range. The most significative consequence is the gain of [7] SIEMENS. Root Mean Square (RMS) Overall Level.
efficiency directly related to costs reduction (reduction of August 2019.
resources demand and development time). A previous https://community.sw.siemens.com/s/article/root-mean-
estimative shows a reduction about 60% in costs and 25% in square-rms-and-overall-level
time for a catalyst heating workpackage calibration for a
medium complexity system (PFI, monofuel and one side
VVT). There is also a quantitative non-mensurable gain for
know-how by the possibility of visual observation of
innumerous variables correlation in a very large spectrum via
systematic way.

With those considerations, it is clear the powerful

competitive advantage of the described method. It allows the

11