Documente Academic
Documente Profesional
Documente Cultură
a r t i c l e i n f o a b s t r a c t
Keywords: This paper proposed an engine fault diagnosis system based on intake manifold pressure signal and arti-
Fault diagnosis system ficial neural network with the Wigner–Ville distribution technique. Traditionally, the engine diagnostic
Intake manifold pressure method depends on the experience of the technician, but some faults might be inaccurately judged by
Wigner–Ville distribution the technician’s experience when the engine is operating. In the present study, an engine platform diag-
Artificial neural network
nosis system using intake manifold pressure was developed. The algorithm of the proposed system con-
sisted of Wigner–Ville distribution (WVD) for feature extraction and the neural network technique for
fault classification. In previous work, the Wigner–Ville distribution was often used to analyze the non-
stationary signal, because it provides a simple and clear energy spectrum diagram both in the time
and frequency domains. This instantaneous energy diagram presented the magnitude of each engine fault
under various operating conditions. The Wigner–Ville distribution extracts these features as database
input to a neural network and the neural network is used to develop the training and testing modules.
To prove the efficiency of the neural network, both the radial basis function neural network and general-
ized regression neural network are used and compared. The experimental results demonstrated the pro-
posed system is effective and the performance is satisfactory.
Ó 2010 Elsevier Ltd. All rights reserved.
1. Introduction ios, 1971), and short time Fourier transform (STFT) (Portnoff,
1980). These methods are often applied in sound and vibration sig-
Owing to the fast growth of the automobile industry, road vehi- nals, but different analysis techniques are not suited to the intake
cles have become a popular mode of transport in people’s daily life. manifold pressure signal. The fast Fourier transform only shows
For electrical control internal-combustion engine, the intake man- the amplitude in the frequency domain. And inspection of the sig-
ifold pressure is an important factor in determining the air–fuel nal magnitude using fast Fourier transform differs in the frequency
mixture of engine combustion. It always demonstrates the engine domain could not determine instantaneous variation. Although
condition and affects the volumetric efficiency, fuel consumption short time Fourier transform provides a fixed-size window to solve
and performance of internal-combustion engines. The intake man- the problem in the time–frequency window (Baydar & Ball, 1996),
ifold pressure is well known to be detrimental to engine system it can not present the instantaneous energy density spectrum of
stability and performance and it should be considered during reg- signals. Meanwhile, the Wigner–Ville distribution technique has
ular maintenance. Traditional fault diagnosis depends on the tech- been widely used in fault diagnosis, and the approach is worth
nician experience and will introduce many subjective senses to using to figure out the energy density from the time–frequency do-
diagnose the fault in the engine. Unfortunately, the traditional pro- main (Loutridis, 2006; Wang, Zhang, & Zhong, 2008). In the study,
cedure is not a precise approach to correctly classify when the en- the Wigner–Ville distribution is used in signals for feature extrac-
gine is operating. An efficient and exact fault diagnosis system tion in the engine fault diagnosis system because the Wigner–Ville
consists of a combination of automatic and computer system. In distribution in the time–frequency domain could simply be used to
the present study, an expert fault diagnostic system using describe the instantaneous energy density spectrum.
Wigner–Ville distribution (WVD) and neural network technology In the fault signal classification, both the radial basis function
is proposed for internal-combustion engine fault diagnosis. network (RBFN) and generalized regression neural network
Several useful signal processing techniques for fault diagnosis (GRNN) are used for comparing the performance in the fault iden-
have been developed, such as fast Fourier transform (FFT) (Corinth- tification process. Generally, the back-propagation neural network
(BPNN) is the basic method in signal analysis (Parvoti & Anandara-
* Corresponding author.
jan, 2002), but the neural network has many defects in the struc-
E-mail address: jdwu@cc.ncue.edu.tw (J.-D. Wu). ture and training process. The training process of the radial basis
0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.06.099
J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544 537
The applications of artificial neural networks have been ex- where ym, the mth element of the y, is the output of the mth node in
tended in recent years, such as in speech recognition and fault the output layer, dlm is the weight from the lth hidden layer neuron
diagnosis. The term RBFN stands for many neural networks that to the mth output layer neuron, and al is the output of the lth unit in
can all be combined within supervised feed-forward networks, the hidden layer.
whose response functions are Gaussian-shaped functions. The To achieve a better recognition rate of the neural network, the
RBF network is comprised of three layers: the input layer, the hid- RBFN needs to be trained. The point of training involves determin-
den layer and the output layer. The network design of RBFN is ing the number of RBF nodes and the output layer weight values.
shown in Fig. 1, which has n input signals composing an input vec- The method to represent the total square errors Es can be defined
tor that are sent to a hidden layer composed of RBFN neural units as:
538 J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544
Pn dðx;x Þ R 1
1X s X
i 2
b ðxÞ ¼ Pi¼1 e
i yedðy;yi Þdy
Es ¼ t m yim ðxi Þ ð9Þ Y n R1
dðx;xi Þ 1 edðy;yi Þdy
ð12Þ
2 i¼1 m i¼1 e 1
R1 2
where t im are the prospective values of the network, and S is the measuring the two indicated integrations, using 1
zez dz ¼ 0,
number of training samples. The number of hidden neurons is auto- yields the following:
matically calculated until the training value converges to prospec- Pn i
tive Es in the procedure. In the RBF network, the hidden layer is b ðXÞ ¼ Pi¼1 Y expðdðx; xi ÞÞ
Y ð13Þ
n
the most important layer providing most of the bunching power i¼1 expðdðx; xi ÞÞ
1 X
n
2
^f ðX; YÞ ¼ edðx;xi Þ e½ðyyi Þ=ðry Þ ð11Þ
ðpþ1Þ=2
nð2pÞ r1 r2 rp ry i¼1
Pattern layer
Pp 2
where dðx; xi Þ ¼ j¼1 ½ðxj xij Þ=rj , and a physical explication of the
possibility evaluation ^f ðX; YÞ is it assigns the sample probability of
width r for each sample Xi and Yi. Here r is the parameter repre-
∑ ∑ Summation layer
senting the width of the ‘‘kernel function”, and is called the smooth-
ing factor (Parzen, 1962).
Connecting a smoothing factor with each feature is an implicit
way to sketch the features to depend on their predictive signifi- Output layer
cance. The probability estimation is the sum of these sample prob-
abilities. Replacing the joint probability evaluate in Eq. (9) into the y( x)
condition mean Eq. (8), and exchanging the order of integration
and summation yields Fig. 2. Architecture of generalized regression neural network.
J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544 539
experimental work, the engine was operated in an idle condition gine platform was used in this experimental work. A pressure sen-
(800 rpm), 1000 rpm, 1500 rpm, 2000 rpm and 3000 rpm. sor was used to measure the intake manifold pressure of the
engine platform. The pressure signals were transited using pres-
4.2. Experimental process and signals analysis sure sensors and recorded using a data acquisition system, and
then used the Wigner–Ville distribution to display the different
In the present research, the engine intake manifold pressure diagram from the intake manifold pressure signals. The features
signals were analyzed to verify the proposed faulted diagnosis sys- were normalized using Eq. (4) to simplify these data. Then, the fea-
tem. After the experimental work, the intake manifold pressure tures were fed into the RBFN and GRNN for fault identification.
signals of the engine were recorded by a pressure sensor with a In the classification, both RBFN and GRNN were evaluated for
data acquisition system. Fig. 3 shows the experimental procedure the proposed system. The purpose of the neural network is to find
of the intake manifold fault diagnosis system. The fuel injection en- the same engine fault under different operation conditions. All
Pressure signal
Data Acquisition System
Pressure sensor
Wigner-Ville distribution
Training Data
GRNN
Classifiers
Engine RBFN
Result
Fig. 4. Intake manifold pressure signal of idle in normal condition using Wigner–Ville distribution.
540 J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544
experiment data are separated into two parts because RBFN and the experimental results. In the engine platform, there are 200 data
GRNN are the supervised learning networks. The experiment data in each operation, 40 data sets are used to train faults in the neural
are divided into the testing and training parts. The networks use network, and 160 testing data sets are used to classify the result of
the database to train data and employ the testing data to verify the training database.
Fig. 5. Intake manifold pressure signal of idle with one injector fault condition using Wigner–Ville distribution.
Fig. 6. Intake manifold pressure signal of idle with two injector fault condition using Wigner–Ville distribution.
J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544 541
4.3. Experimental results and classification 13. Clearly, the energy amplitudes of the intake manifold pressure
signal shown in the figures differ from one another. After feature
According to the present study, the intake manifold pressure extraction, the instantaneous energy of various faults is fed into
signals are feature extraction by Wigner–Ville distribution. The the proposed intelligent fault diagnosis system for fault classifica-
diagrams of instantaneous energy density are shown in Figs. 4– tion. The recognition rate of fault classification using RBF and
Fig. 7. Intake manifold pressure signal of idle with intake air-leak condition using Wigner–Ville distribution.
Fig. 8. Intake manifold pressure signal of idle with intake plugged condition using Wigner–Ville distribution.
542 J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544
GRNN are summarized in Tables 1 and 2. In the result, the pro- the RBF has a feed-forward stage during the training procedure,
posed system has an acceptable recognition rate for fault diagnosis it took much more training time to achieve the convergence goal.
under several engine operating conditions. Despite all the evalua- Thus, the GRNN network can quickly complete the training process
tion results having high recognition rates, however, the results and obtain an optimal identification rate.
show the classification time of GRNN is less than RBFN. Because
Fig. 9. Intake manifold pressure signal in 3000 rpm in normal condition using Wigner–Ville distribution.
Fig. 10. Intake manifold pressure signal in 3000 rpm one injector fault condition using Wigner–Ville distribution.
J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544 543
Fig. 11. Intake manifold pressure signal in 3000 rpm two injector fault condition using Wigner–Ville distribution.
Fig. 12. Intake manifold pressure signal in 3000 rpm intake air-leak condition using Wigner–Ville distribution.
5. Conclusions using Wigner–Ville distribution for feature extraction and the arti-
ficial neural network for classification was performed. The Wig-
In the present study, an investigation of an expert system based ner–Ville distribution extracts the features as input for a neural
on intake manifold pressure fault diagnosis in the engine platform network in the proposed system. In the fault identification, the
544 J.-D. Wu, C.-K. Huang / Expert Systems with Applications 38 (2011) 536–544
Fig. 13. Intake manifold pressure signal in 3000 rpm intake plugged condition using Wigner–Ville distribution.
References
Table 1
Performance of fault recognition using RBF in various faults and operation conditions. Baydar, N., & Ball, A. (1996). A comparative study of acoustic and vibrations in
detection of gear failures using Wigner–Ville distribution. Mechanical System
Idle 1000 rpm 1500 rpm 2000 rpm 3000 rpm and Signal Processing, 15(6), 1091–1107.
Normal (%) 100 100 100 100 75 Corinthios, M. J. (1971). A fast Fourier transform for high-speed signal processing.
One injector fault (%) 100 100 100 100 100 IEEE Transactions on Computers, C20, 843–846.
Two injectors fault 100 100 100 100 94 Loutridis, S. J. (2006). Instantaneous energy density as a feature for gear fault
detection. Mechanical System and Signal Processing, 20, 1239–1253.
(%)
Min, H., & Jianhui, X. (2004). Efficient clustering of radial basis perceptron neural
Intake air-leak (%) 100 100 100 100 77.5
network for pattern recognition. Pattern Recognition, 37, 2059–2067.
Intake plugged (%) 100 100 100 100 100
Parvoti, F. Y., & Anandarajan, M. (2002). Classifying inventory using an artificial
Average (%) 97.86 neural network approach. Computer and Industrial Engineering, 40, 389–404.
Time (s) 23.4 Parzen, E. (1962). An estimation of a probability density function and mode. The
Annals of Mathematical Statistics, 33, 1065–1076.
Portnoff, M. (1980). Time–frequency representation of digital signal and system
based on short-time Fourier analysis. IEEE Transactions on Acoustic, Speech and
Table 2 Signal Processing, 28, 55–69.
Performance of fault recognition using GRNN in various faults and operation Pulido, A., Ruisanchez, I., & Rius, F. X. (1999). Radial basis functions applied to
conditions. classification of UV–visible spectra. Analytica Chimica Acta, 388, 273–281.
Sarker, M., Yegnanarayana, B., & Khemani, D. (1998). Backpropagation learning
idle 1000 rpm 1500 rpm 2000 rpm 3000 rpm algorithms for classification with fuzzy mean square error. Pattern Recognition
Normal (%) 99.38 100 100 99.38 100 Letters, 19, 43–51.
One injector fault 100 99.38 100 100 100 Simon, D. (2002). Training radial basis neural networks with the extended Kalman
(%) filter. Neurocomputing, 48, 455–475.
Specht, D. F. (1991). A general regression neural network. IEEE Transactions on
Two injectors fault 100 100 100 100 100
Neural Networks, 2, 568–576.
(%)
Staszewski, W. J., Worden, K., & Tomlinson, G. R. (1997). Time–frequency analysis in
Intake air-leak (%) 100 100 100 96.25 96.25 gear box fault detection using the Wigner–Ville distribution and pattern
Intake plugged (%) 100 100 100 99.38 97.5 recognition. Mechanical System and Signal Processing, 11(5), 673–692.
Average (%) 99.5 Wang, C., Zhang, Y., & Zhong, Z. (2008). Fault diagnosis for diesel valve trains based
Time (s) 17.5 on time–frequency images. Mechanical System and Signal Processing, 22,
1981–1993.
Wu, J. D., & Chiang, P. H. (2009). Application of Wigner–Ville distribution and
probability neural network for scooter engine fault diagnosis. Expert Systems
artificial neural network both using RBFN and GRNN were carried with Applications, 36, 2187–2199.
Wu, J. D., Wang, Y. H., Chiang, P. H., & Bai, M. R. (2009). A study of fault diagnosis in
out and compared in the fault diagnosis system. The experimental scooter using adaptive order tracking technique and neural network. Expert
results showed the GRNN is better than RBFN both for the fault rec- Systems with Applications, 36, 49–56.
ognition rate and for classification time.