A coustic pseudo spectrum based fault Localization in motorcycles

International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.3, June 2013
DOI:10.5121/ijcsa.2013.3307 79
ACOUSTIC PSEUDOSPECTRUM BASED FAULT
LOCALIZATION IN MOTORCYCLES
Veerappa B. Pagi1
, Basavaraj S. Anami2
and Ramesh S. Wadawadagi3
1, 3
Dept. of Computer Applications, Basaveshwar Engg. College, Bagalkot, India
veereshpagi@gmail.com, rswbec@gmail.com
2
Dept. of Computer Science and Engg., KLE Institute of Technology, Hubli, India
anami_basu@hotmail.com
ABSTRACT
Vehicles generate dissimilar sound patterns under different health conditions. The sound generated by the
vehicles gives a clue of some of the faults. Automotive experts diagnose the faults in vehicles based on the
produced sound. This paper presents a methodology for fault source localization in motorcycles using the
estimated pseudospectra of the sound signals. The pseudospectra are traced to generate a chaincode. The
generated chaincode is transformed into an eigenvector, which is used as the feature vector for
classification using artificial neural network (ANN). The overall classification accuracy is 88%. The
proposed work finds applications in traffic census of the vehicles, traffic rule observance, machine fault
detection and localization, automatic acoustic surveillance and the like.
KEYWORDS
Fault Localization, Chaincode, Pseudospectrum, Eigen Vectors, Neural Classifier
1. INTRODUCTION
The sound patterns produced by the vehicles convey information necessary for fault diagnosis.
Vehicle sound sources include parts of the engine subsystem, timing chain, clutch plate and
exhaust subsystem. We rely on automotive experts for the repair and maintenance of vehicles.
Automated fault diagnostic systems are essential when the vehicle is in remote places and in
places of scarce expertise. Motorcycles dominate the Indian automobile market, with nearly 77%
of total vehicle sales. Society of Indian Automobile Manufacturers (SIAM) has forecast the two-
wheeler segment to register a growth of 6%-8% in 2013-14 [1].
This work is an attempt to localize the fault source in motorcycles based on the produced sound.
The features are extracted from pseudospectra of the sound signals. The chaincode of the
estimated pseudospectra are constructed and transformed into matrices. The eigenvectors [9]
computed over the matrices are considered as features. These features are later subjected to
classification by ANN classifier. The present work leaves scope for further investigation of
localization of combinations of faults. The research contribution is significant since it indicates a
fault well earlier, reducing a possible future accident. The service station experts can use the
findings of this work for preliminary fault diagnosis.
The literature survey is organized into two parts: engine fault diagnosis and gearbox fault
diagnosis. Following is the gist of the studied literature.

80
The 1D central contour moments and their invariants of approximation coefficients of DWT are
used as feature inputs for DTW classifier for determining the health condition of motorcycles [2].
The sound intensity method is used to identify the noise sources of engine front of a diesel engine
[3]. A continuous wavelet transform (CWT) algorithm is combined with an ANN and generalized
regression for analyzing fault signals in a scooter fault diagnosis system [4]. Engine fault
diagnosis uses empirical mode decomposition (EMD) and wavelet packet BP neural network [5].
RMS and Power Spectral Density (PSD) of Massey Ferguson gearbox are calculated to detect
different faults [6]. An adaptive wavelet filter based on Morlet wavelet is used for detection of
symptoms from vibration signals of a gearbox with early fatigue tooth crack [7]. Vibration signals
are decomposed into a finite number of intrinsic mode functions (IMF) and then the
autoregressive (AR) model of each IMF component is established; finally, the corresponding AR
parameters and the variance of remnant are regarded as the characteristic vectors and used as
input of SVM classifier to classify the working condition of gear [8].
From the literature survey, it is evident that reasonable amount of research is reported at the wider
range of applications, ranging from fault classification in machines to gearboxes. Since there is no
work reported on fault source localization of motorcycles based on their sound patterns, we have
taken up a study. The remainder of the paper is organized into 3 sections. The proposed
methodology of the work, along with a brief on tools and techniques, is discussed in Section 2;
the experimental results are discussed in Section 3. Finally, the Section 4 concludes the work.
2. PROPOSED METHODOLOGY
The overview of the methodology is depicted in figure 1. It comprises five stages namely, sound
signal acquisition, segmentation, spectral analysis, feature extraction and classification.
Figure 1. Block diagram of the proposed method
Legend: VS: Valve setting; FC: Faulty crank; CK: Cylinder kit problem; TC: Timing chain fault;
ML: Muffler leakage; SL: Silencer leakage

81
The following subsections briefly explain each of the important stages in the methodology.
2.1. Acquisition of Sound samples
The automotive guidelines suggest the recording frequency in the range of 9 kHz to 30 kHz. The
sound signals of the motorcycles are recorded using Sony ICD-PX720 digital voice recorder, with
sampling frequency of 44.1 kHz and quantized with 16 bits. The recording of the sound samples
is carried out in service stations under the supervision of service experts. The selected healthy
motorcycles are less than one year old, not run more than 6000 km, well-maintained and regularly
serviced in authorized service stations. The motorcycles having faulty crank, damaged timing
chain, unset valve, muffler leakage, silencer leakage, and faulty cylinder kit are considered in this
work. The scope of the work is limited to the fault source localization in motorcycles with single
fault. Figure 2 depicts the recording environment.
The recording environment has disturbances from human speech, sound of other vehicles being
serviced, air-compressor and auto-repair tools. The recorder is held 500 mm from the centerline
of the exhaust end, and at the angle of 45° measured from the centerline of the exhaust end, and at
the height of the exhaust pipe. The motorcycle is in neutral state, and stationary. The engine runs
in idle state and the throttle is controlled by the expert, while recording the sound signals.
Figure 2. Recording environment
A brief description of the faults considered in this work is given as under.
Timing Chain: Timing chain operates the valves. A loose chain vibrates and results in improper
timing for valve operation. This results in abnormal operation of engine leading to noise.

82
Valve setting: For smooth functioning of engine, correct operation of valves is necessary. Any
deviation of 5 to 10 degrees in valve opening/closing will cause considerable rise in peak
combustion chamber pressures, leading to change in sound.
Crank fault: It may occur due to wear and tear of either oil ring, first ring or second ring.
Muffler Leakage: The exhaust gases coming out of combustion chamber passes through muffler.
The main function of muffler is to reduce the noise and filter exhaust gases. Due to the reactive
gases in the residual exhaust, mixed with water vapour, result in corrosion reactions. This creates
minute holes in the muffler and changes the firing sound coming out of engine.
Silencer leakage: If there is a hole inside the silencer filter pipe or the damaged gasket, it causes
silencer leakage.
2.2. Segmentation
The acquired sound samples are segmented into samples of one-second each for uniformity in
processing. The portion of the signal of duration one second, beginning from the local maxima is
considered as a segment. The next segment begins at local maxima in the next 50 ms duration
from the end of the previous segment.
2.3. Feature Extraction
The chaincode of the pseudospectrum is constructed by tracing the gradient changes at each
point. The eigenvectors computed based on the constructed chaincodes are used as feature
vectors.
2.3.1. Pseudospectrum estimation
The multiple signal classification (MUSIC) algorithm [10] is used for estimating the
pseudospectrum of the sound signals. The pseudospectrum is calculated using estimates of the
eigenvectors of a correlation matrix associated with the input data. The MUSIC estimate is given
by the Equation (1).
( )1
)(
1
)()(
1
)(
1
2
1

∑∑ +=+=
=








= N
pk
H
k
N
pk
H
kk
H
music
fevfevvfe
fP
where N is the dimension of the eigenvectors and vk is the kth eigenvector of the correlation
matrix of the input signal. The integer p is the dimension of the signal subspace, so the
eigenvectors kv used in the sum correspond to the smallest eigenvalues. The vector ( )fe consists
of complex exponentials, so the inner product
)( fevH
k amounts to a Fourier transform. In the
eigenvector method, the summation is weighted by the eigenvalues k of the correlation matrix,
as shown in Equation (2).
( )2
/)(
1
)(
1
2

k
N
pk
H
k
ev
fev
fP








=
∑+=

83
Figure 3 shows the spectra of sound signals of healthy and faulty motorcycles.
Figure 3. Spectra of sound signatures of healthy and faulty motorcycles
In case of the spectra of healthy samples the spectral peaks decrease monotonically and exhibit
regular variations. But, in case of faulty motorcycles, the degraded harmonicity, non-monotonous
decrease in spectral peaks and spurious peaks at higher frequencies, are observed. Figure 4
combines the spectra of different faults under consideration. The variations in peaks and
harmonicity are larger at low frequencies. The features based on this part of the spectrum aid for
better classification.
Figure 4. Spectra of different faults
2.3.2 Chaincode of a spectral segment
The chaincode of a spectral segment is generated based on the gradient changes, tracing the
pseudospectrum from left to right.
The chaincode vector of a spectral trace is computed as:
)3(
)1()(
)1()(
)1()(
1
7
0
)( 
+<
+>
+=





=
ixix
ixix
ixix
iChCode

84
Where, 1281 ≤≤ i
Figure 5 shows the directions in which the spectral gradients change.
Figure 5. Chaincode directions
Since the gradient changes can be observed only in three directions, the constructed chaincodes
comprise 0 (right), 1 (top-right), and 7 (bottom-right) as shown in Figure 5. The reference feature
vector for healthy motorcycles is constructed by taking the mode of the respective chaincodes.
Similarly the reference feature vector for faulty motorcycles is constructed.
2.3.3. Dimensionality reduction
Figure 6 summarizes the dimensionality reduction process. The 128-digit chaincode is
transformed into a matrix of size 16 × 8. The adjacent values in each row are averaged,
converting the matrix into a square matrix of order 8 × 8. The eigenvector of the reduced square
matrix is computed, resulting in a vector of length eight. The eigenvector is considered as the
feature vector input to the neural classifier with eight inputs.
Figure 6. Dimensionality reduction
2.3.4. Eigenvectors
Let A be a complex square matrix. Then if λ is a complex number and X a non–zero complex
column vector satisfying AX = λX, we call X an eigenvector of A, while λ is called an eigenvalue
of A. We also say that X is an eigenvector corresponding to the eigenvalue λ. Table 1 summarizes
the mean eigenvectors for different types of faults.

85
Table 1. Mean eigenvector values for different types of faults.
2.3.5. ANN classifier
Figure 7 shows the overview of the architecture of the ANN. The eight features, extracted from
the chaincode, are input to the neural network with 8 input nodes. The six output nodes
correspond to the six-bit output vector indicating the type of the fault in the motorcycle. The
hidden layer contains 10 nodes. The neural network is trained using backpropagation-learning
algorithm [11]. The stabilized weights are reloaded and test vectors are input during testing. The
optimal number of hidden layer neurons is chosen using the criterion [12]:
( )4
log

Nd
N
Cn =
where, n=number of hidden layer neurons, C=constant to yield optimal performance, d=number
of features, and N=number of rows in the training sample matrix.
The mean squared error (MSE) is computed for the sample sets with 70% of the samples used for
training, 15% for validation and 15% for testing.
Figure 7. Architecture of ANN
Figure 8 shows a typical learning process of ANN where the goal is met. In this case, goal for the
error is set as 0.00001, the neural network is trained with learning parameter of 0.3, and the
algorithm halted after 731 epochs. If the goal is met and the same feature vectors are used for
testing, the classification accuracy will be 100%. If the user terminates the testing process before
the goal is met, normally for large test sets, the classification accuracy suffers. As the number of
training samples is increased, the learning rate is increased accordingly.
Eig1 Eig2 Eig3 Eig4 Eig5 Eig6 Eig7 Eig8
F1 (VS) 9.1235 0.8387 0.4893 0.4535 0.2013 0.0940 0.0210 0.0272
F2 (FC) 9.3850 0.7107 0.6083 0.4536 0.1943 0.0585 0.0000 0.1897
F3(CK) 8.2913 1.4210 0.4033 0.2424 0.1008 0.0883 0.0313 0.0000
F4 (ML) 9.3886 0.4682 0.2182 0.1493 0.0000 0.0000 0.0000 0.0000
F5 (SL) 7.7834 1.8341 0.3541 0.2058 0.1021 0.0925 0.0085 0.2258
F6 (TC) 8.0524 1.3658 0.4363 0.2718 0.0597 0.0261 0.0000 0.0167
I H
O
I H
O
I H
Output Fault
[100000] Mis-set Valve
[010000] Faulty crank
[001000] Faulty
cylinder kit
[000100] Muffler
leakage
[000010] Silencer
leakage
[000001 Faulty timing
chain
Input vector
[8 bit
eigenvector]

86
Figure 8: Learning process by ANN
Figure 9 shows the validation MSE plotted for varying number of neurons in the hidden layer.
The minimum MSE is observed when 10 neurons are present in the hidden layer. Hence, the
neural network is designed with eight input nodes, 10 hidden layer nodes and six output nodes.
The performance of the built neural network is analyzed for different combinations of training,
validation and testing sets.
Figure 9. Validation MSE for varying number of hidden layer neurons
3. RESULTS AND DISCUSSION
The faults present in motorcycles of different models are considered. Models considered are
Passion plus, Extreme, Street 100, CD 100, CD Dawn, CD Deluxe, CBZ, Splendour, Splendour
plus, Super splendour, all from Hero Honda (Presently known as Hero MotoCorp). The training
sample set contains 360 sound samples of motorcycles. The samples are partitioned into disjoint
sets for training and testing based on random selection. The size of the training and test sample
sets is varied from 10 to 60, for each type of fault. Confusion matrix for a total of 360 test
samples, containing 60 samples of each type of fault, is given in Table 2.

87
Table 2. Confusion matrix showing classification performance
From the Table 2 it the false acceptance rate is observed to be 0.1167, which is admissible in real-
world fault diagnosis. Hence, the overall classification accuracy of the approach is 0.8834.
Classification accuracy for different types of faults is shown in Table 3, when the number of test
samples for each type of fault is varied from 10 to 60. The classification accuracy is 100% when
10 and 20 samples of each type of fault are used for testing.
Table 3. Classification accuracy for different faults on varying sizes of inputs
It is evident from Table 3 that the classification accuracy suffers with increase in the size of the
samples sets.
4. CONCLUSION
The presented work identifies the source of the fault in motorcycles, based on the acoustic
signals. Eigenvectors of the transformed chaincodes generated from the pseudospectra are used as
feature vectors. The extracted features are classified into one of the six types of faults using ANN
classifier. The classification accuracy is over 88% for testing with 60 samples of each type of
fault. Since these results are obtained without denoising, the approach is suitable for real-world
implementation. The method is not apt for on-ride fault diagnosis since it requires considerable
No. of test
samples
Mis-set
Valve
Faulty
crank
Faulty
cylinder
kit
Muffler
leakage
Silencer
leakage
Faulty
timing
chain
60 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
120 1.0000 1.0000 0.9000 0.9500 0.9000 1.0000
180 1.0000 1.0000 0.8000 0.9667 0.8667 0.9667
240 0.8750 1.0000 0.7750 0.9750 0.8750 0.9250
300 0.7600 1.0000 0.8000 0.9800 0.9000 0.8400
360 0.8167 1.0000 0.8500 0.9667 0.7834 0.8834
Mis-set
Valve
Faulty
crank
Faulty
cylinde
r kit
Muffler
leakage
Silence
r
leakage
Faulty
timing
chain
Classifi
cation
accurac
y
Mis-set
Valve
49 0 3 2 6 0 0.8167
Faulty
crank
0 60 0 0 0 0 1.0000
Faulty
cylinder kit
1 0 51 8 0 0 0.8500
Muffler
leakage
0 0 2 58 0 0 0.9667
Silencer
leakage
6 0 1 4 47 2 0.7834
Faulty
timing
chain
3 0 0 0 4 53 0.8834
Overall classification accuracy 0.8834

88
amount of time for spectral computations and ANN training. The fault localization in presence of
multiple faults will be attempted in the future work.
ACKNOWLEDGEMENTS
We the authors thank Shrinivas M. S. and Basavaraj Kumbar for recording and testing help,. We
thank Prof. Vishwanath Kagawade and Prof. Sangamesh Magi for language corrections. The
service stations in Bagalkot city need special mention for sharing the fault diagnostic expertise.
REFERENCES
[1] http://www.moneylife.in/article/car-sales-in-india-likely-to-expand-3-5-in-fy14-siam/32142.html
[SIAM : Society of Indian Automobile Manufacturers, Predictions for two-wheeler sales in Financial
year 2013-14].
[2] Basavaraj S. Anami, Veerappa B. Pagi & Sangamesh M. Magi, (2011) “Wavelet based acoustic
analysis for determining health condition of two-wheelers”, Elsevier J. Applied Acoustics, Vol. 72,
No. 7, pp464-469
[3] Zhang JunHong & Han Bing, (2005) “Analysis of engine front noise using sound intensity
techniques”, J. Mechanical Systems and Signal Processing, Vol. 19, pp. 213-221
[4] J.D. Wu, E.C. Chang, S.Y. Liao, J.M. Kuo & C.K. Huang, (2009) “Fault classification of a scooter
engine platform using wavelet transform and artificial neural network”, Proceedings of the
International MultiConference of Engineers and Computer Scientists, IMECS 2009, Hong Kong,.
March 18–20 2009, Vol. 1, pp58-63
[5] Wei, Liao, Pu, Han, & Xu, Liu, (2009) “Fault diagnosis for engine based on EMD and wavelet packet
BP neural network”, Proceedings of the Third International Symposium on Intelligent Information
Technology Application, pp672-676
[6] K.Heidarbeigi, Hojat Ahmadi, M. Omid & A. Tabatabaeefar, (2009) “Fault diagnosis of Massey
Ferguson gearbox using power spectral density”, J. Agricultural Technology, Vol. 5, No. 1, pp1-6
[7] J. Lin & M. J. Zuo, (2003) “Gearbox fault diagnosis using adaptive wavelet filter”, J. Mechanical
Systems and Signal Processing, Vol. 17, No. 6, pp1259–1269
[8] Junsheng, Cheng, Dejie, Yu, & Yu, Yang, (2008) “A Fault Diagnosis Approach for Gears Based on
IMF AR model and SVM”, EURASIP J. Advances in Signal Processing ,Vol. 2008, Article ID
647135, 7 pages doi:10.1155/2008/647135
[9] http://www.numbertheory.org/book/cha6.pdf
[10] http://www.mathworks.in/help/toolbox/signal/pmusic.html
[11] Kishan, Mehrotra, Chilukuri K. Mohan & Sanjay, Ranka, (1996) “Elements of Artificial Neural
Networks”, The MIT Press
[12] Shuxiang, Xu & Ling, Chen, (2008) “A novel approach for determining the optimal number of hidden
layer neurons for FNN’s and its application in data mining”, Proceedings of the 5th International
Conference on Information Technology and Applications, Queensland, pp683-686

89
Authors
Basavaraj S. Anami B. S. Anami, received the Ph.D. from the University of Mysore,
and the M. Tech. from Indian Institute of Technology, Chennai, both in Computer
Science, in 2003 and 1986, respectively. From 1983 to 2008, he was with the
Basaveshwar Engineering College, Bagalkot, India, where he worked on several research
projects in image processing. Presently he is with KLE Institute of Technology, Hubli,
India, as Principal. He visited Saginaw Valley State University, USA, as Visiting
Professor, from August to December 2006. He has authored two books and over 50
refereed papers. He is a fellow of Institution of Engineers, India, and member of Indian Society of
Technical Association and Computer Society of India. He has guided five students for Ph. D.
Veerappa B. Pagi V. B. Pagi, received the M. E. degree from the Bangalore University,
Bangalore, and B. E. Degree from Karnatak University, Dharwad, India, both in
Computer Science, in 2007 and 1994, respectively. From 1994 to 1999, he was with Sir
M. Visvesvaraya Institute of Technology, Bangalore, where he worked as Lecturer in the
Department of Computer Science. In 1999, he joined Basaveshwar Engineering College,
Bagalkot, where he is currently working as Assistant Professor, in the Department of
Computer Applications. He has published 9 papers in International journals and 3 papers in conferences.
He has guided over 10 projects in the area of Image processing and 3 projects in the area of sound signal
processing.
Ramesh S. Wadawadagi R. S. Wadawadagi received the B.Sc., M.Sc. degrees in
Computer Science from Karnatak University Dharwad, and the M. Tech.. degree in
Computer Science from University of Mysore, in 2001, 2004, and 2009, respectively.
From 2009 he is working as Assistant Professor in the department of Computer
Applications, Basaveshwar Engineering College, Bagalkot, Karnataka State, India. His
areas of interest include pattern recognition, data mining and system simulation. He has
guided over 10 projects for the postgraduate students. His current research activities
include vehicle engine diagnostics, neural networks, and web content mining.

A coustic pseudo spectrum based fault Localization in motorcycles

More Related Content

What's hot (9)

Similar to A coustic pseudo spectrum based fault Localization in motorcycles (20)

Recently uploaded (20)

A coustic pseudo spectrum based fault Localization in motorcycles