Available online at www.sciencedirect.

com

Expert Systems
with Applications
Expert Systems with Applications 34 (2008) 2360–2369
www.elsevier.com/locate/eswa

Prediction of vehicle reliability performance using artificial

neural networks
S. Lolas, O.A. Olatunbosun *

School of Mechanical and Manufacturing Engineering, University of Birmingham, B15 2TT, UK

Abstract

Product development is an important but also dynamic, lengthy and risky phase in the life of a new product. The optimisation of the
product development phase through extensive knowledge of the involved procedures is believed to reduce the risks and improve the final
product quality. Artificial intelligence and expert systems have been used successfully in optimising the development phase of some new
products as it will be demonstrated by the first sections of this publication. This paper presents the first module of an expert system, a
neural network architecture that could predict the reliability performance of a vehicle at later stages of its life by using only information
from a first inspection after the vehicle’s prototype production. The paper demonstrates how a tool like neural networks can be designed
and optimised for use in reliability performance predictions. Also, this paper presents an optimisation methodology that enabled the
neural network to deal with the limited amount of available training data, common during new product development, and to finally
achieve acceptable prediction performance with small error. A case example is presented to demonstrate the methodology.
Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Neural networks; Product development; Reliability prediction; Knowledge

1. Introduction 1991; Ozer, 1999). However, as Ozer (1999) has commented

new product development is a dynamic, lengthy and risky
It is widely accepted that the modern industries are business.
under immense pressure to deliver high quality and reliable The optimisation of the new product development phase
products at the minimum possible cost. In other words, requires extensive knowledge about all the various proce-
manufacturers must seek to shrink, significantly, the prod- dures that are involved with it. Knowledge, in general, is
uct development phase, from design concept to complete a very important resource for presenting valuable heritage,
product release. Higher quality levels, higher service levels, learning new things, solving problems, creating core com-
customisation and continuous innovation paired with petences and initiating new situations for both individuals
lower production costs have been deemed as necessary and organisations now and in the future (Liao, 2003).
since the 1980s but are no longer enough in the modern The management community, both academic researchers
era. The extra requirements for the industries of today and practitioners, has developed a large number of technol-
and the future are the creation and domination of new ogies and applications to address the issue of knowledge
markets which can be achieved by the development of acquisition, management and application. According to
new products (Cohen & Levinthal, 1994; Hamel & Prahalad, Liao (2003) there is a huge number of publications under
the knowledge management index. All this literature pre-
sents a number of different approaches.
Artificial intelligence and expert systems is the method
*
Corresponding author. of choice for this research project, whose objective is to pro-
E-mail address: [email protected] (O.A. Olatunbosun). vide solutions to issues inherent to the product development

0957-4174/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2007.03.014
 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369 2361

of automotive vehicles. This paper only discusses early parts

of the expert system development, currently under way by Product
Release
the Novel Vehicle Technology and Future Power System
Group of the University of Birmingham.
Expert systems are knowledge based systems that con-

Reliability
tain expert knowledge, according to the definition given

Problems
by Kasabov (1996). Expert systems have been applied suc-
cessfully to most fields of science including engineering,
science, medicine, agriculture and many more. The main Actual
advantages that make them so attractive for such a large Desired
variety of applications are that they can capture human
knowledge in the form of a set of rules. Expert systems
Models Tested - Time
are not only capable of capturing knowledge but they also
retain this knowledge therefore they form the organisa- Fig. 1. Development of a new vehicle in terms of reliability. Reality versus
tional memory or in other words they store the learning Ideal.
of the organisation. This provides a very attractive benefit
to all the organisations that employ expert systems. These
organisations no longer rely solely on their human work- towards the left side of the figure give a fairly good reliabi-
force to carry and pass the knowledge and the experience lity performance because many of the components installed
gained from older projects to their successors in the com- in them are carried over, without any modifications, from
pany. The organisation’s expert systems incorporate this the older already reliable version. As the new components
knowledge in the form of tools and therefore the acquired are introduced and installed in the test vehicles the reliabi-
knowledge can be retrieved automatically during a future lity problems increase. Then, the reliability performance of
project when it is required by a different employee. Conse- the test vehicles improves gradually after a critical point is
quently expert systems eliminate the danger of wrongly reached, mainly because the engineers react to the reliabi-
understood or forgotten knowledge, something that is very lity problems that are pointed out from the testing proce-
possible when knowledge is retained by humans. Examples dure and correct them. From there the reliability
of more conventional expert systems are presented in performance gradually reaches the required levels and the
Cunningham and Bonzano (1999) and Doyle et al., 1996. vehicle is released to the market. This approach to the
Examples of using artificial intelligence methods such as development phase of a vehicle can be characterised as pas-
genetic algorithms, neural networks and fuzzy logic can sive, because corrective actions are made after the faults
be found in (Al-Thabtabai, 1998; Mohan & Arumugam, have been identified.
1997). More applications of expert systems can be found The ‘‘desired’’ curve in Fig. 1 represents the ideal sce-
in table 5 in (Liao, 2003). nario for the development of a new vehicle. Improvements
This paper presents the first steps towards the construc- in reliability can be achieved through a more pro-active
tion of an expert system. The complete expert system will approach to the procedure. A pro-active approach
combine different types of artificial intelligence methods. through the development phase requires good knowledge
The scope of this paper will concentrate on one of the mod- about it. Knowledge about the particularities of the new
ules of the expert system. This module comprises of a neu- development procedure can be acquired by analysing past
ral network system which was designed to process development phases of predecessors or similar products
reliability data, learn it and then use this knowledge in fore- and also by using the experience of the experts that have
casting and analysis. been involved in the development of the predecessor sys-
tems. A detailed analysis can accentuate trends and effects
2. The problem of the various changes in the design and other parameters
of the process which can then be monitored and learned,
Fig. 1 describes the real situation in the automotive using artificial intelligence tools such as neural networks.
industry for the reliability problems of a vehicle model dur- By using neural networks the engineers would be able to
ing its development phase. The figure is created using the simulate the development procedure for the new product
following methodology. For every problem noticed on a and as a result they can identify the areas which need to
development test vehicle, points are awarded according to receive extra attention and also the areas that do not
the significance, severity and other contributing factors, contribute to the reliability problems of the new product.
creating a final score that quantifies the reliability perfor- Therefore engineers can concentrate the available
mance of the given test vehicle. Models tested represent resources on significant reliability problems that have
time or stage in the development phase. The models closer great contribution to the overall reliability performance,
to the concept vehicle are towards the left side of the axis thereby optimising the development procedure. Another
and the models that are closer to job#1 can be found important benefit of using a pro-active approach for the
towards the right end. In real life (actual curve) vehicles new development procedure is that requirements and
2362 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369

reliability targets can be imposed on the sub-system and By training the neural network, implicit knowledge can be
component suppliers. The ‘‘everybody plays’’ concept that built into it (Pham & Pham, 1999). The training of a super-
has been created by General Electric Appliances (Antony vised neural network occurs by presenting typical input
& Banuelas, 2002) explains the positive effects the supplier patterns and the corresponding expected output patterns.
involvement can have in the reliability improvement The strength or otherwise the weights of the connections
efforts. between the neurons are modified by using an error mea-
There are many examples of world leading companies in surement between the actual and the expected results, until
manufacturing and engineering fields that have realised the results of the network are satisfactory. For this proce-
that the improvement of reliability and quality of their dure a backpropagation algorithm is used. It propagates
products can only be achieved through optimisation of the error, between the expected and actual results, back-
all of their operations. This is the reason why a number wards through the structure of the network and then it
of leading companies world wide use techniques like total computes the weight modifications necessary to improve
quality control and six sigma to improve all aspects of their the actual results of the network’s outputs in order to pro-
structure management and operation (McClusky, 2000; vide the most correct solution to the problem. Multilayer
Antony & Banuelas, 2002; Catherwood, 2002; Johnson & perceptron is the most popular neural network. A multi-
Swisher, 2003; Kwak & Anbari, 2004). Despite the commit- layer perceptron is a feed forward network, which can per-
ment of many companies to such programs the difficulty form static mapping between input space and an output
arises from the fact that reliability improvement of the space. It consists of neurons organised in a number of lay-
product during the product development requires continu- ers that can be categorised into three parts. The first part is
ous reliability estimation and updating (Yadav, Singh, the input layer that allows that network to communicate
Chinnam, & Goel, 2003), that requires detailed knowledge with the environment, the second part is commonly known
about all of the specific details of the procedure. The lack as the hidden part, where one, two or more layers of neu-
of knowledge creates what is known as epistemic uncer- rons exist depending on the problem’s demands and gener-
tainty, which according to Cizelj, Mavko, and Kljenak alisation requirements. The last part of the multilayer
(2001, Cizelj) is defined as the lack of knowledge about a perceptron is the output layer through which the network
system or phenomenon. presents its answers to the user. A schematic diagram of
a basic multi-layer perceptron and its learning procedure
3. Neural networks for reliability prediction can be seen in Fig. 2.
Some examples of ANN applications in various indus-
3.1. Background and applications tries are Zhang and Friedrich (Zhang, 2003), who used
an ANN in the composite material science in order to assist
The use of ANNs has grown in popularity during the in the selection of suitable composites according to the
last few years. The reason for this is that neural networks requirements of the application. Yin, Rosendahl, and
represent a novel and modern approach that can provide Luo (2003) used ANN in the coal industry, to model
solutions in problems for which conventional mathematics, the relationship between the ultimate analysis data and
algorithms and methodologies are unable to find a satisfac- its proximate analysis data, which are both used in order
tory and acceptable solution. These problems are usually to classify coals. Other interesting ANN applications of
very complex and some of the mechanisms involved have multi-layer feed forward backpropagation neural networks
not been fully understood by the researchers dealing with include (Nascimento et al., 2000; Pleune and Chopra, 2000;
them. Venkatesh and Rack, 1999; Zhang, 2003). All these appli-
ANNs are inspired by the human brain functionality cations demonstrate the superiority of the backpropaga-
and structure, which can be imagined as a network that tion ANNs in solving difficult and complex problems,
comprises of densely interconnected elements called neu- which could not be solved at all, or could be solved only
rons. Despite this fact the ANNs’ objective is not to model by expensive and time consuming testing.
it. Instead their purpose is to be useful models that can be
used for problem solving and knowledge engineering, in a
way that resembles the human process for problem solving
and knowledge acquisition. Both biological and artificial
networks have the following main and important features;
learning adaptation, generalisation, massive parallelism,
robustness, associative storage information and spatiotem-
poral information processing (Kasabov, 1996).
The operational manner of ANNs is that when inputs
are applied to the input neurons the network performs a
summation of the weighting factors and then it activates
one or more specific output neurons that are capable of
providing the most suitable answer for the given problem. Fig. 2. Multi-layer perceptron learning example (Kasabov, 1996).
 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369 2363

3.2. Objectives and description of the proposed ANN network. A small amount of this data was reserved to cre-
ate validation and testing data, whose use is described later.
During the current project an ANN was created in order Thus a very thorough network optimisation procedure,
to assist the engineers in predicting the reliability behaviour including training, backpropagation, neuron numbers, hid-
of an automotive vehicle during its development phase. den layer numbers and other was undertaken in order to
Due to the complexity of the procedure due to the high achieve the best possible results.
number of parameters and influencing factors that are
involved, an accurate mathematical model without a great 5. Network design and optimisation method
number of assumptions is not possible. Many assumptions
usually result in reduced accuracy. 5.1. Overview
In more detail the proposed network was designed in
such a way that it could predict the behaviour of the vehicle The main objective of the work described in this section
at 6000 km by using only information about the vehicle’s was carried out to discover which one of the many different
reliability issues at 0 km. This practically means that an feed-forward backpropagation neural network combina-
initial inspection, during which the initial reliability tions and architectures would be most suitable for the task
and quality issues of vehicle are recorded, will provide and also the network architecture that would be able to
enough information to simulate this vehicle’s performance deal with the small amount of available training data.
at 6000 km. If data for higher mileage ranges is included The design and optimisation of the network was carried
then the trained network will be able to predict the out in three main stages. In brief, during the first stage
reliability degradation for the vehicle for a later stage of the fundamental parameters for the network, like the trans-
its life. fer functions and the training functions have been tested.
During the second stage the type of pre-processing method
4. Data collection and data nature more suitable for the network has been tested. Finally, the
third stage of the network optimisation has been carried
One of the most important stages in the design of a out to discover the most suitable number of hidden layers
supervised ANN is the data collection and data prepara- and the number of suitable neurons in each. It has been
tion, thus the examples for training must be representative suggested by a number of researchers that have used ANNs
of all the possibilities concerning the application. that the selection of all the parameters that have been
Researches that have used ANNs with supervised learning described above is a trial and error procedure (Lee,
support the previous statement (Yin et al., 2003; Zhang, Almond, & Harris, 1999).
2003).
The data used for this work has been extracted from a 5.2. Network optimisation – stage 1
number of databases, which record reliability and quality
issues. The context of the data used was in terms of reliabil- During the first stage of network optimisation the fun-
ity evaluation audit (REA), which records problems on the damental parameters of the network have been tested to
vehicle’s systems and assesses points according the severity investigate which one gives the best performance. The fun-
and importance of the problems. Four inputs into the net- damental parameters for the network are the transfer func-
work which represent the behaviour of the vehicle at 0 km tions for each layer and the training functions.
were initially extracted from the databases. More precisely The transfer functions that have been used for this
the inputs into the system are the code number of the vehi- group of networks were the tansig function, for the hidden
cle, which could preserve the chronological point in the layer and the purelin function, for the output layer. Both
development phase at which the vehicle was tested, the are very common choices for this type of ANN (Vogl,
total number of issue score points for this vehicle, the num- Mangis, Rigler, Zink, & Alkon, 1988; Hagan, Demuth, &
ber of issues that appeared on this vehicle and finally the Beale, 1996; MathWorks, 2004).
average score for each issue. The training algorithm selection is a more complicated
Three target parameters or in other words three desired task than the selection of the transfer functions. There is
answers will be used to demonstrate to the network what a large number of training algorithms that can produce a
the correct answer is to any given input parameter. These faster, more accurate and more reliable network. The main
parameters are the total REA points, the number of issues complication in the selection of an appropriate training
and the average number of points per issue all at 6000 km. algorithm arises from the fact that the performance of all
At the end of the data collection there were 29 vehicles the available training algorithms depends on the specific
that have been tested and had satisfactory information problem under consideration (Hagan et al., 1996; Math-
available in terms of the parameters stated above. The Works, 2004). A training algorithm that performs well in
amount of the available data could be judged to be very a function approximation problem might not perform very
small and could potentially cause network generalisation well when applied to a pattern recognition problem. There-
problems. Adding to the problems was the fact that not fore, all the main training algorithms that, according to ref-
all the data from the 29 cars could be used to train the erence (MathWorks, 2004), have the best performance over
2364 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369

Table 1 Performance of the Six Networks-Mean Absolute Error Performance

Network code names and the corresponding training algorithm name 50
Net1
Network code Training algorithm full name MATLAB Acronym 45 Net2
name code Net3
Net4
Net1 Levenberg–Marguardt Trainlm LM 40

Mean Absolute Error (mae)

Net5
Net2 Baayesian Reguralisation Trainbr BR Net6
35
Net3 Resilient Backpropagation Trainrp RP
Net4 Scaled conjugate Gradient Trainscg SCG
30
Net5 BFGS Quassi-Newton Trainbfg BFG
Net6 Variable Learning Rate Traingdx GDX
25
Backpropagation
20

15

a number of different problems, were tested in order to 10

discover the ones that performs better for this specific
application. 5
0 5 10 15
Six training algorithms were tested. The different net- Trial Number
works created with each different algorithm were given Fig. 3. Performance of Networks in terms of mean absolute error.
code names such as net1, net2. . . net6. The training algo-
rithms with their corresponding code names are presented
in Table 1. Initially, for the testing of all the networks
the same number of neurons in the hidden layer has been Consequently, there was the need for improvement in the
used. There is no specific rule for the selection of the num- networks that would increase their prediction quality and
ber of neurons in the hidden layer or the number of hidden would reduce the error values.
layers that will produce the optimum results for a given
problem. The number of neurons in the hidden layer is usu-
5.3. Network optimisation – stage 2
ally determined empirically, despite the fact that there have
been numerous methods proposed by various researchers
Suggestions in a number of publications such as (Hagan
(Kaminskyj & Materka, 1995; Kang & Song, 1997; Lee
et al., 1996; Al-Assaf & El Kadi, 2001) reinforced the opi-
et al., 1999). Due to the obvious variance in the opinions
nion that the instability and the poor performance of the
regarding the number of neurons in the hidden layer and
networks tested in stage 1 was due to the fact that the input
the number of hidden layers these numbers will be decided
values were high, ranging from 0 to 500. Al-Assaf and Kadi
after trials. For stage 1 the selection of the neurons in the
used a combination of normalisation techniques for the
hidden layer was decided by using the rule of thumb, there-
inputs. For this project a technique called minimum and
fore twelve neurons in the hidden layer were selected. The
maximum technique, has been used as in (Hagan et al.,
initial network configuration will be 4-[12]-3. This code
1996; MathWorks, 2004). In the later two references
means four neurons in the input layer, twelve in the hidden
another two data pre-processing techniques have been pro-
layer and three in the output layer. This configuration has
posed. These are the mean and standard deviation method
been used for the trials of the six different training algo-
(MMM) and the principal component analysis (PCA)
rithms. The results of the trials are depicted in Fig. 3, in
(Ishihara, Ishihara, Nagamachi, & Matsubara, 1997; Di
terms of mean absolute error. X-axis represents the number
Natale et al., 1999; Briqueu, Gottlib-Zeh, Ramadan, &
of trials that have been carried out for each one of the six
Brulhet, 2002; Lacassie, Roser, Ruiz Del Solar, & Herve,
different networks. Fifteen trials for each network have
2004).
been carried out because the responses of the networks
Net1 and net2 have been selected for training with the
were not stable, meaning that the performance of the same
pre-processed data from stage 1 because they performed
network was varying from one training session to another.
better and more reliably during this first trial session of
In Fig. 3 net1 appears to be the one that produces the more
stage 1. Net2 with the Bayesian regularisation algorithm
consistent results when compared with the other networks.
was expected to work much better with pre-processed data
It also gave the lowest error values. The second best perfor-
mance in terms of singular minimum error has been
achieved by net3, but this network appeared to have a great
difficulty in achieving this performance consistently. Net2 Table 2
appeared to be fairly consistent and was very close in terms Network code description for stage 2
of performance with net1. Overall, the values of error in all Pre-processing method Net1 Net2
the three assessing parameters were not very good and even Min–Mix Method (MMM) Net1_1 Net2_1
the best performing networks were not able to predict the Mean and standard deviation method (MSDM) Net1_2 Net2_2
behaviour of the vehicles at 6000 km at a satisfactory level. Principal component analysis (PCA) Net1_3 Net2_3
 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369 2365

according to (MathWorks, 2004). A new coding system The effects of the three pre-processing methods on the
was created to facilitate the separation of the different Bayesian Regularisation method, namely net2, are illus-
pre-processing methods. This coding system is presented trated in Fig. 5. It is apparent that the data manipulation
in Table 2. It must be mentioned that the networks pre- prior to its application to the network has a very positive
sented in Table 2 preserved the main characteristics that effect on all aspects of its performance. The error values
have been explained in Table 1. The networks have the dropped and in almost all the trials the response of the net-
same training algorithms and their structure remained the work was stable and consistent with small variations. The
same, 4-[12]-3. best performance in terms of lowest error values and con-
In Fig. 4 the response of the Levenberg–Marguardt sistency have been achieved by the MMM. Therefore this
network or net1 to the three pre-processing methods is method has been chosen for the rest of the trials in the fol-
demonstrated. The network responded positively to the lowing stages.
pre-processed data. The performance of the network has Fig. 6 depicts the comparison between the two net-
been greatly improved. The MMM pre-processing method works, net1 and net2, when the MMM is applied. The low-
has resulted in a much better performance, in terms of con- est error value is achieved by net1 but its response it terms
sistency and the lowest error performance. of consistency is not as good as that of net2.
Summarising the results from stage 2, the two networks,
net1 and net2 have responded very well to the introduction
of the pre-processed data. The error values have been
mae-Performance of Net1
40
Min Max
greatly reduced in both cases and also the consistency of
Mean and Std Deviation the networks with respect to the network response has been
35 Principal Comp Analysis improved. Despite the great improvements that have been
Original
achieved by the introduction of the pre-processed data
Mean Absolute Error

30
there is the need for further improvements in order to
reduce the error further.
25

5.4. Network optimisation – stage 3

20

15
This optimisation stage has focused on testing different
neural network architecture structures by altering the num-
10 ber of neurons in the hidden layer and the number of the
hidden layers.
5 The network that was used to assess the differences in
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Trial Number the performance of the network with respect to its different
architectures was net2_1 selected from stage 2. The reason
Fig. 4. Mean Absolute Error performance of net1 with all the pre- behind this selection is that this network had a very consis-
processing methods.
tent and stable response at every training trial.

5.4.1. One hidden layer

50
mae-Performance of Net2 The default error assessment parameter for the Bayesian
Min Max Regularisation training algorithm is the sum squared error
Mean and Std Deviation
45
Principal Comp Analysis
(sse). Net2 uses this training algorithm and as a result the
Original assessment parameter that has been used in this stage is
40
the sse. There have been two different values for sse. The
Mean Absolute Error

35 first was the minimum sse value achieved during the train-
ing, according to which the training is stopped. The second
30
was the total sse value, which calculates all the results col-
25 lectively. One new assessment parameter has also been used
in this stage. It is called regression performance of the net-
20 work which is the result of data post-processing, where a
15
regression analysis between the network’s response and
the desired response is performed. The regression results
10 are represented by values between 0 and 1 where a value
of 1 indicates perfect correlation between the targets and
5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 the actual outputs of the networks and 0 indicates
Trial Number the opposite. After the end of the trials it appeared
Fig. 5. Mean Absolute Error performance of net2 with all the pre- that the regression values for two of the output parameters,
processing methods. the total number of REA points and the number of issues,
2366 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369

Min Max Method- mae

40
net11
net2 1
35

Mean Absolute Error

30

25

20

15

10

5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Trial Number

Fig. 6. Comparison of the error performance between net1 and net2 when using the MMM.

were always above 0.9 and more precisely 0.95 and above. 5.4.2. Two hidden layers
The regression value for the third parameter, which is the The networks have been modified accordingly to incor-
average number of REA points per issue, was not very porate two hidden layers in their structure. Some random
good taking values of 0.4–0.6. Therefore, the regression trials to specify approximately the best combination of
value for the third output parameter is the important one neurons in each hidden layer were carried out. The results
that indicated the best performing combination. of the different layer and neurons combinations are pre-
Starting from the network with one hidden layer the sented in Figs. 9 and 10. In the figures the three final layers
number of neurons has been changed taking values from of the networks are represented because the number of
1 up to 30 neurons. In Fig. 7 the lowest min sse value is inputs is fixed. For example 5_2_3 means that the network
achieved by the network with 10 neurons. In terms of is a 4-[5-2]2-3 network. The figures indicate that the
regression performance, Fig. 8, the network with fifteen network 4-[5-5]2-3, (5_5_3), with five neurons in the first
neurons in the hidden layer manages to correlate the values hidden layer and five neurons in the second hidden layer,
of the third parameter better than any other. The difference had the best performance with all the assessment parame-
in the sse values of the networks that have between 6 and ters i.e. both min sse value and high regression. Overall it
15 neurons in the hidden layer is not as great, which means was apparent that the error values were reduced by a great
that overall the network with a number of neurons in its amount and the regression values, especially for the third
hidden layer within the range of 10–15 will give much bet- output parameter, was improved and reached values over
ter results. These results suggest that the rule of thumb for 0.8. In the one layer networks the regression values for
the number of neurons in the hidden layer was correct. this parameter was not more than 0.78. These results

4.5

4
Sum Squarred Error (min)

3.5

2.5

1.5

0.5

0
1 2 3 4 5 6 10 15 20 25 30
Number of Neurons

Fig. 7. The performance of the network for a range of neurons in the hidden layer.
 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369 2367

0.8

0.7

Regression Performance
0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 10 15 20 25 30
Number of Neurons

Fig. 8. Regression performances for the third network output value (Number of average points).

0.8

0.7
Minimum Sum Squarred Error

0.6

0.5

0.4

0.3

0.2

0.1

0
5_2_3 10_2_3 15_2_3 5_3_3 10_3_3 15_3_3 5_4_3 10_4_3 15_4_3 5_5_3 10_5_3 15_5_3

Network Architectures

Fig. 9. Collective results for the minimum sse performance of networks with two hidden layers.

0.9

0.8

0.7
Regression Value

0.6

0.5

0.4

0.3

0.2

0.1

0
5_2_3 10_2_3 15_2_3 5_3_3 10_3_3 15_3_3 5_4_3 10_4_3 15_4_3 5_5_3 10_5_3 15_5_ 3
Network Architecture

Fig. 10. Collective results for regression performance of networks with two hidden layers.

demonstrate that the target of achieving a very good the future. Figs. 11–13 present the predicted results for this
prediction level has been reached. This means that the network, which gave results that are much closer to the
4-[5-5]2-3 network will be used to perform similar tasks in expected behaviour of that of the real vehicle.
2368 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369

500 6. Conclusions
Targets
Simulation Results
400 Errors In the work presented here it has been demonstrated
how an ANN system can be developed in order to predict
REA Points/Errors

300
accurately enough the reliability behaviour of an automo-
tive vehicle at 6000 km by only knowing information from
the initial vehicle’s inspection at 0 km. The ANN managed
200
to learn the associations between the input parameters at
0 km and the target values of the reliability performance
100
at 6000 km. The ANN has therefore managed to recognise
and identify the degradation mechanism from 0 to 6000 km
0 and has used this knowledge in order to predict the reliabi-
lity behaviour of the test vehicle models through their
-100 development phase. The overall error for the whole net-
0 5 10 15 20 25 30
work and all the three output parameters was almost 9%
Vehicle Code Number
which is quite acceptable. The reason for this error value
Fig. 11. Prediction results for the total REA points at 6000 km. was due to the fact that there was a limited amount of data
and paradigms to train the network and improve its ability
to generalise well. The immediate effect of this system is
that a smaller number of vehicles can be tested at 6000 km.
100 An initial inspection at 0 km can indicate what its reliabi-
Targets lity behaviour will be at higher mileage. The system can
Simulation Results
80 Errors be easily adapted to calculate reliability performance for
a number of higher mileages, given that tests to this mileage
have been carried out.
Number of Issues

60
The first step after this stage is to enrich the system with
other parameters that describe changes during the develop-
40
ment stage of the vehicles. These new parameters can be,
for example, in the form of improvement factors from
20 one version of the model to another. This will assist in
the investigation and prediction of the vehicle’s reliability
0 performance when changes and improvements are intro-
duced to it. For example the engineers will be able to visu-
-20
alise what will happen to the vehicle if a certain type of
0 5 10 15 20 25 30 improvement is introduced to it. This way they will be able
Vehicle Code Number to apply to the vehicle those improvements that will have
Fig. 12. Prediction results for the number of issues at 6000 km. the best effect on the reliability performance and also those
improvements that will be the cheapest. Resource manage-
ment could be another application of the system.

7 References
Targets
6 Simulation Results
Errors
Al-Assaf, Y., & El Kadi, H. (2001). Fatigue life prediction of unidirec-
tional glass fiber/epoxy composite laminate using neural networks.
Average REA Points

5
Composite Structures, 53, 65–71.
Al-Thabtabai, H. (1998). A framework for developing an expert analysis
4
and forecasting system for construction projects. Expert Systems with
Applications, 14(2), 259–273.
3
Antony, J., & Banuelas, R. (2002). Key ingredients for the effective
implementation for six sigma program. Measuring Business Excellence,
2
6(4), 20–27.
Briqueu, L., Gottlib-Zeh, S., Ramadan, M., & Brulhet, J. (2002). Inferring
1
lithology from downhole measurements unsing an unsupervised self-
organizing neural network: study of the Marcoule silty clayish unit.
0
C.R. Geoscience, 334, 331–337.
Catherwood, P. (2002). What’s different about six sigma. Manufacturing
-1
0 5 10 15 20 25 30 Engineer, 81(8), 186–189.
Vehicle Code Number Cizelj, R. J., Mavko, B., & Kljenak, I. (2001). Component reliability
assesment using quantitative and qualitative data. Reliability Engi-
Fig. 13. Prediction results for the average REA points at 6000 km. neering and System Safety, 71, 81–95.
 S. Lolas, O.A. Olatunbosun / Expert Systems with Applications 34 (2008) 2360–2369 2369

Cohen, W. M., & Levinthal, D. A. (1994). Fortune Favors the Prepared Liao, S.-H. (2003). Knowledge Management Technologies and Applica-
Firm. Management Science, 40(2), 227–251. tions – Literature Review from 1995 to 2002. Expert Systems with
Cunningham, P., & Bonzano, A. (1999). Knowledge engineering issues in Applications, 25, 155–164.
developing a case-based reasoning application. Knowledge-Based MathWorks, I. (2004). Neural Network Toolbox User’s Guide. MATLAB
Sytems, 12(7), 371–379. 7.0, Release 14, MATLAB Manual.
Di Natale, C., Paolesse, R., Macagnamo, A., Troitsky, V. I., Berzina, T. McClusky, R. (2000). The Rise, Fall, and Revival of Six Sigma. Measuring
S., & D’ Amico, A. (1999). Pattern recognition approach to the study Business Excellence, 4(2), 6–17.
of the interactions between metalloporphyrin Langmuir–Blodgett films Mohan, S., & Arumugam, N. (1997). Expert system applications in
and volatile organic compounds. Analytica Chimica Acta, 384, irrigation management: an overview. Computers and Electronics in
249–259. Agriculture, 17, 263–280.
Doyle, M. D., Ang, C. S., Martin, D. C., & Noe, A. (1996). The visible Nascimento, C. A. O., Giudici, R., & Guardani, R. (2000). Neural
embryo project: embedded program objects for knowledge access, network based approach for optimization of industrial chemical
creation and management through the world wide web. Computerized processes. Computers and Chemical Engineering, 24, 2303–2314.
Medical Imaging and Graphics, 20(6), 423–431. Ozer, M. (1999). A survey of new product evaluation models. Journal of
Hagan, M. T., Demuth, H. B., & Beale, M. H. (1996). Neural Network Product Innovation Management, 16, 77–94.
Design. PWS Publishing Company. Pham, D. T., & Pham, P. N. T. (1999). Artificial intelligence in
Hamel, G., & Prahalad, C. K. (1991). Corporate imagination and engineering. International Journal of Machine Tools and Manufacture,
expeditionary marketing. Harvard Business Review, 69(4), 81–92. 39, 937–949.
Ishihara, S., Ishihara, K., Nagamachi, M., & Matsubara, Y. (1997). An Pleune, T. T., & Chopra, O. K. (2000). Using artificial neural networks to
analysis of Kansei structure on shoes using self-organizing neural predict the fatigue life of the carbon and low-alloy steels. Nuclear
networks. International Journal of Industrial Ergonomics, 19, 93–104. Engineering and Design, 197, 1–12.
Johnson, A., & Swisher, B. (2003). How Six Sigma Improves R& D. Venkatesh, V., & Rack, H. J. (1999). A neural network approach to
Research Technology Management, 46(2), 6–17. elevated temperature creep-fatigue life prediction. International Journal
Kaminskyj, I., & Materka, A. (1995). Automatic source identification of of Fatigue, 21, 225–234.
monophonic musical instrument sounds. In IEEE International Vogl, T. P., Mangis, J. K., Rigler, A. K., Zink, W. T., & Alkon, D. L.
Conference on Neural Networks. (1988). Accelerating the convergence of the backpropagation method.
Kang, J. Y., & Song, J. H. (1997). Neural network applications in Biological Cybernetics, 59, 257–263.
determining the fatigue crack opening load. International Journal of Yadav, O. P., Singh, N., Chinnam, R. B., & Goel, P. S. (2003). A fuzzy
Fatigue, 20, 57–69. logic based approach to reliability improvement estimation during
Kasabov, N. K. (1996). Foundations of Neural Networks, Fuzzy Systems, product development. Reliability Engineering and System Safety, 80,
and Knowledge Engineering. MIT Press. 63–74.
Kwak, Y. H., & Anbari, F. T. (2004). Benefits obstacles and future of six Yin, C., Rosendahl, L., & Luo, Z. (2003). Methods to improve prediction
sigma approach. Technovation, 1–8. performance of ANN models. Simulation Modelling Practice and
Lacassie, J. P., Roser, B., Ruiz Del Solar, J., & Herve, F. (2004). Theory, 11, 211–222.
Discovering geochemical patterns using self-organising neural net- Zhang, Z., Barkoula, N. M., Karger-Kocsis, J., & Friedrich, K. (2003).
works: a new perspective for sedimentary provenance analysis. Artificial neural network predictions on erosive wear polymers. Wear,
Sedimentary Geology, 165, 175–191. 255, 708–713.
Lee, J. A., Almond, D. P., & Harris, B. (1999). The use of neural networks Zhang, Z., & Friedrich, K. (2003). Artificial neural networks applied to
for prediction of fatigue lives of composite materials. Composites: Part polymer composites: a review. Composites Science and Technology, 63,
A, 30, 1159–1169. 2029–2044.