2015 International Conference on Soft-Computing and Network Security (ICSNS -2015), Feb.

25 – 27, 2015, Coimbatore, INDIA

FAULT LOCALIZATION IN MOTORCYCLES

USING WAVELET PACKET ENERGY
DISTRIBUTION
Veerappa B. Pagi Ramesh S. Wadawadagi
Dept. of Computer Science and Engg Dept. of Computer Applications
Basaveshwar Engineering College Basaveshwar Engineering College
Bagalkot, INDIA Bagalkot, INDIA
[email protected] [email protected]

Basavaraj S. Anami
KLE Institute of Technology
Hubli, INDIA
[email protected]

Abstract— Motorcycles produce sound signals with varying timing chain, cylinder kit, piston and crankshaft. During the
temporal and spectral properties under different working earliest stage of an engine's life, small irregularities in the
conditions. These sound patterns can be used as an important metals are encouraged to gradually form congruent grooves by
source of information for automatic diagnosis of faults in avoiding extreme operating conditions. Later in its life, after
motorcycles. Fault localization is a process of identifying the
mechanical wear has increased they start producing abnormal
exact source of failure from a set of observed fault indications.
The work proposed in this paper demonstrates the mechanism sound patterns. As a matter of fact, the valve train provides a
for identifying source of faults in motorcycles based on the passageway that allows air-fuel mixture into the cylinder and
distribution of energy in wavelet packet coefficients of the sound allows exhaust gas to pass out. Any deviation of crank angle
signal. The percentage of energy distribution among the nodes in valve opening or closing causes variation in peak
forms the feature vectors. The feature vectors thus obtained are combustion chamber leads to vary in sound. The timing gear
used to train the neural network classifier. Performance of the of engine is used to operate the intake and exhaust valves. The
proposed method is evaluated against six different types of faults. vibration of a loose chain can alter the valve setting resulting
The experimental results show 95% accuracy for fault source in change in sound. A cylinder is also important part of a
localization in motorcycles.
engine or pump, the space in which a piston travels. Wear in
Keywords-Fault source localization, Acoustic fault diagnosis, the cylinder and the piston creates a clearance between them,
Wavelet packet transform, artificial neural network which results into escaping of the burnt gases
.
I. INTRODUCTION
This causes drop in power output of the engine results in
Over the years the two wheeler sector is playing a key role change of sound. Crankshaft support bearings take all the
in India where majority of population lives in rural or semi- pressure generated by the engine power stroke. The reactive
urban areas without adequate transportation facilities. Society gases in the residual exhaust, which are at high temperature,
of Indian Automobile Manufacturers (SIAM) has forecast the mixed with water vapor, lead to corrosion reactions and
two-wheeler segment to register a growth of 6% to 8% by minute holes. These holes in the muffler change the firing
2013-14 [1]. The major factors that influence the growth in sound produced by an engine. In traditional motorcycle fault
two-wheeler segment includes the advent of economic and diagnosis, the technician depends on the experience and will
fuel-efficient engines, the maneuverability of motorcycles and introduce many subjective senses to diagnose the faults.
their endurance to critical road conditions. Over time, Unfortunately, to correctly classify when the engine is
motorcycles may turn faulty due to improper usage, poor operating the traditional procedure is not a precise approach.
maintenance, worn parts and for other inherent defects.

For a technology point of view, it is necessary to design an

Many a times these defects are reflected in terms of intelligent system to identify the faults in engine so that the
irregular sound patterns. Usually the sound generated by the worn parts can be replaced in time. So one of the most
motorcycles are originated from engine, rotary parts, tyre promising approach for automatic identification of faults cab
friction, exhaust and clutch plate [3]. The engine which is a be based on acoustic signals. This study build up a fault
central working part of a motorcycle consists of valve train, diagnosis system using discrete wavelet transform (DWT) and

978-1-4799-1753-2/15/$31.00 ©2015 IEEE

 2015 International Conference on Soft-Computing and Network Security (ICSNS -2015), Feb. 25 – 27, 2015, Coimbatore, INDIA

neural network classifier for identifying different types of [13] PSD, STFT, DWT K-NN, SVM 95 - 97%, 70 - 93%
faults in motorcycles.
[3] CWT, DWT BPNN, SVM 90 -100%, 75 -100%
LITERATURE REVIEW
The literature survey is carried out to know the state-of-the- DB(1 - 10)
art applications of wavelets in fault detection and diagnosis Coif(1 - 5)
systems based on sound patterns. A range of applications from Bior(1.1 - 6.8) Rough sets,
automotive industry to other allied areas has been described in [14] Up to 99.84%
the literature. Table. 1 provides a comparative analysis of Rbio(1.1 - 6.8) Fuzzy logic
wavelet-based works for fault detection and diagnosis Sym(2 - 8)
systems. For study, we focus on issues such as features Meyer wavelet
extraction, learning methods used and performance evaluation
of each method as a criterion. However, the survey does not KNN, MPP,
explain the technical details of methods that can be applied [15] FFT, WT 95.5%
K- Means
directly on the sound signals.
In [2] a method to enhance the detection and diagnosis low- [16] WPT SVC 94 -100%
speed rolling-element bearing faults based on DWT is
addressed. The method involves the automatic extraction of
wavelet packets containing bearing fault-related features from
DWT analysis representation of machine vibration. Automated Sound intensity method is described in [7] for investigation of
selection of wavelet packets of interest is achived via adaptive mechanisms of engine front noise generation and the
network-based fuzzy inference system [ANFIS]. An adaptive corresponding countermeasures of a diesel engine using sound
wavelet filter based on Morlet wavelet is introduced by Lin intensity method. An approach for the classification of the
and Zuo [4] for detecting vibration signals of a gearbox with working condition of gear is proposed in [8]. In this approach,
early fatigue tooth crack. The parameters in the Morlet the vibration signals are decomposed into a finite number of
wavelet function are optimized based on the kurtosis intrinsic mode functions, and then the autoregressive (AR)
maximization principle. In [5] a multi-resolution wavelet model of each intrinsic mode functions (IMF) component is
analysis coupled with a neural network based approach is established. Finally, the corresponding autoregressive
applied for fault diagnostics of industrial robots. The analysis parameters and the variance of remnant are regarded as the
implements discrete wavelet transforms (DWT) with filters fault characteristic vectors and used as input parameters of
and decomposes the signal in various levels. The approximate SVM classifier to classify the working condition of gears.
and detailed coefficients of the decomposed signals are then Similarly in [9] an investigation of a fault diagnostic technique
used for training a feedforward neural network (FDNN) whose for internal combustion engines using DWT and neural
output determines the state of the robot. Wu et al. [10] used a network is proposed.
continuous wavelet transform (CWT) and artificial neural
network (ANN) to develop a mechanical fault diagnosis The features of the sound emission signal at different
system for a scooter engine platform. To extract information resolution levels are extracted by multi-resolution analysis and
for condition monitoring Zhi-yong and Jun [6] reviewed CWT Parseval’s theorem. Wavelet-based neural network is used for
characteristics represented by short duration transient signals. power disturbance recognition and classification. A fault
Wavelet selection and CWT are then implemented and diagnosis system for engines using an empirical mode
wavelet transform is applied to analyze the major sources of decomposition (EMD) and wavelet packet backpropagation
the engine front’s exterior radiation sound. neural network is presented in [11]. Anami et al. [12] designed
a fault detection system for motorcycles based on acoustic
signals. The approach employs the 1D central contour moment
TABLE 1. A COMPARATIVE STUDY OF WAVELET-BASED and their invariants of wavelet subbands and DTW classifier.
WORKS FOR FAULT DIAGNOSIS In [13], two feature extraction methods for acoustic signals
Learning from moving ground vehicles are investigated. The first one is
Reference Features Accuracy
Method based on spectrum distribution and the second one on WPT.
The performance of two methods is evaluated against KNN
[2] DB20 ANFIS 73 - 99% and SVM classifiers. The use of mechanical vibration and
sound signals for fault analysis in rotating machinery based on
[5] and [10] DWT, CWT ANN 90%, 95%
DWT and a SVM is studied in [2]. Vibration signals are used
[9] DB(4, 8, 20) ANN 95% for fault diagnosis of centrifugal pumps using wavelet analysis
[14]. Rough set theory is applied to generate the rules from the
[11] EMD, WPT BPNN 95 - 100% vibration signals. Based on the strength of the rules the faults
are identified. Wavelet features are computed using DWT
[12]
DB4 central and
DTW 81 - 100%
from the vibration signals and rules are generated using rough
contour sets and classified using fuzzy logic. For mechanical fault
 2015 International Conference on Soft-Computing and Network Security (ICSNS -2015), Feb. 25 – 27, 2015, Coimbatore, INDIA

diagnosis, a hybrid method that combined wavelet packet samples are directly segmented into a number of samples of
decomposition and support vector classification (SVC) is duration one second each for uniformity in processing. The
proposed [16]. The wavelet packet is employed to decompose segment arbitrarily starts at the local maxima of the first 50
the vibration signal to obtain the energy ratio in each millisecond (ms) duration and continues till one second
frequency band. Taking energy ratios as feature vectors, the duration to make one segment. Finally, the feature vector is
pattern recognition results are obtained by the SVC. constructed over the segments of sound signals.

The aforementioned literature mainly focuses on fault

diagnosis systems for engines, gearbox, machinery, Sound Signal
monoblock centrifugal pump, industrial robots etc. Wavelet Acquisition
packets combined with soft computing tools such as fuzzy
logic, SVM and variants of ANNs are adequately used for
Segmentation
analysis of sound patterns. However, it is observed that these
techniques are sensitive to Doppler Effect, noise produced by
moving parts and atmospheric variations. Further, these
methods are expensive and require extra hardware. In order to Feature Extraction
overcome the shortcomings of previous works, we address the Using WPED
issue of fault diagnosis of motorcycles based on the
distribution of energy in the wavelet packet coefficients of the
sound signatures.
Training Data Testing Data
The obtained feature vector is used to train the neural network.
It is very hard to make a practical comparison between the
proposed method and other works in the literature. This is
mainly due to the use of specific databases, recorded in an No
ANN is Yes
isolated sound-proof environment However, the proposed
trained?
method shows good results for the sound signals recorded in Trained
real-time environment and hence it is more suitable for real-
time implementation.
Fault
Localization
By
II. PROPOSED METHODOLOGY Classification
Fig. 1 illustrates the overview of the proposed
methodology. The acquired sound signals are segmented into Fig. 1. Block diagram of the proposed methodology
smaller units of one second duration for feature extraction. The
feature vectors comprising the percentage distributions of the
wavelet packet energy are input to the ANN classifier. The C. Feature extraction using wavelet packets analysis
performance is analyzed using the energy distributions among
approximation coefficients, detail coefficients, and the
Wavelet packet decomposition [18] is now considered as a
combination of nodal energy distributions of the both.
major tool for signal processing and analysis. A wavelet
packet is a square integrable modulated waveform with zero
A. Acquisition of sound samples
mean, well localized in both position and frequency. Wavelet
packet transform is applied to both low pass results
The sound signals are acquired using Sony ICD-PX720
(approximations) and high pass results (details). It leaves the
digital voice recorder under the supervision of automobile
experts in authorized service stations. The automotive high-frequency part alone and keeps splitting the low-
guidelines suggest the ideal sampling rate for recording is 9 to frequency part. In wavelet packet decomposition, we can
30 kHz, but this is suitable only for recording in anechoic choose to split the high-frequency part also into a low-
chamber. Since the objective is to process the real-world frequency part and a high-frequency part. Thus, in general,
signals a higher sampling frequency of 44.1 kHz with 16 bits wavelet packet decomposition divides the frequency space
quantization is used. This helps us to capture the sound signals into various parts and allows better frequency localization of
effectively in the noisy environments. In [17], a detailed signals.
discussion on arrangement of recording environment is
provided. Further, motorcycles of the same age are considered From Fig. 2, it is evident that the separability among different
for recording to maintain uniformity in the processing. faults is clear. But for some samples, the energy distribution is
interleaved, leaving scope for confusion. Hence, ANN
B. Segmentation
classifier is felt suitable for this work, since it can be
To resemble a real-time system, no preprocessing is carried employed in problems where the class boundaries are not
out over the acquired sound samples. The acquired sound clear. The features are extracted from the samples recorded
 2015 International Conference on Soft-Computing and Network Security (ICSNS -2015), Feb. 25 – 27, 2015, Coimbatore, INDIA

from the cluttered environment because of which the

developed methodology has potential of adaptation for real-
world applications. III. RESULTS AND DISCUSSION
Motorcycles of Indian two wheelers market, namely Hero
Honda (Now Hero Motocorp), Honda motors, TVS motors,
D. Artificial neural network architecture
Bajaj motors are considered for experimentation. A dataset
containing 866 sound samples, including healthy and faulty
Despite advancements in soft and evolutionary computing motorcycles, is constructed using the procedure explained in
techniques, the ANN and its descendants are still popular for subsection 3.1. The sound samples with one of the faults,
classification problems with scope for approximation. Fig.3. namely, Valve Setting problem (VS), Crank Fault (CF),
shows the overview of the architecture of the ANN. The Cylinder Kit problem (CK), Timing Chain problem (TC)
feature vectors containing the energy distribution values are Muffler Leakage (ML) and Silencer Leakage (SL) are
input to the neural network. The six output nodes correspond considered. The recording is carried out in authorized service
to the six-bit output vector indicating the type of the fault in stations under the supervision of automobile experts.
the motorcycle. The number of nodes in the hidden layer is The recorder is held 500 mm from the center line of the
varied for different combinations based on the number of input exhaust end, at the angle of 45o measured from the center line
nodes and an evaluation criterion. The neural network is of the exhaust end and at the height of the exhaust pipe. The
trained using backpropagation learning algorithm. The motorcycle is held in idling state. The recording is carried out
stabilized weights are reloaded and test vectors are input in real-time environment with disturbances around, with an
during testing. intention of applicability in the real-time systems.
The subband energies of both healthy and faulty sound
signals are analyzed for classification. For the first subband,
no significant difference is observed for the percentage energy
in the approximation coefficients of the wavelet packets. But
for subbands two thru six, the percentage of energy in
approximation coefficients differs for healthy and faulty
motorcycles.
This can be attributed to the uneven cycles of operation of
the engine in presence of a fault. Further, the faults related to
exhaust subsystem, like silencer or muffler leakage, contribute
Fig. 2.a. Percentage energy in approximation coefficients to drastic change in sound. This fact leaves scope for future
research in terms of analysis of energies of wavelet packets of
different faults for fault source localization in subsystems.
In order to evaluate the performance of the proposed
method, we consider four different schemes. For each scheme,
we focus on issues such as features used, number of inputs and
nodes in the hidden layer for ANN, ratio of samples used for
training, and validation, testing number of iterations, mean
squared errors.

A. Scheme 1: Using energy distribution in approximation

coefficients
In this scheme, the energy distribution in approximation
Fig. 2.b. Percentage energy in detailed coefficients coefficients is used as features. The first five subbands of the
wavelet packet distribution resulting in five input feature
vectors are given as input to neural classifier. The approach
I H examines for six different faults and hence ANN generates six
different outputs each indicating one fault. A dataset
O consisting of 20 samples of each type of fault are used. Table.
I H 2 shows the classification accuracy for different combinations
Output
Input of training, validation and testing.
O

I H

Fig.3. ANN architecture

 2015 International Conference on Soft-Computing and Network Security (ICSNS -2015), Feb. 25 – 27, 2015, Coimbatore, INDIA

VS TC ML SL
TABLE. 2. APPROXIMATION COEFFICIENTS WITH UNIFORM
SIZED SAMPLE SETS VS 49 0 0 0

Combination Trainin Validatio TC 0 50 0 0

Testing All
Tr Vn Ts g n
ML 0 0 50 0
70 15 15 0.9882 1.0000 0.9474 0.9917
SL 1 0 0 50
60 20 20 0.9861 1.0000 0.9600 0.9917
Accuracy 0.9800 1.0000 1.0000 1.0000
50 25 25 0.9833 1.0000 0.9677 0.9917
40 30 30 1.0000 1.0000 1.0000 1.0000
D. Discussion
The experimental results show that the classification accuracy
B. Scheme 2. Energy distribution in approximation, detail and of the proposed methodology varies from 0.89 to 1.00. The
combined coefficients higher classification accuracy indicates the better separability
of the sound signals from engine parts to exhaust subsystems.
The energy distribution in approximation coefficients, Here, the faults ML and SL are related to exhaust subsystem
detailcoefficients and combination of approximation and detail while the others are related to engine. Further, the testing is
coefficients are used as features. A dataset containing fault carried out on the sample sets of uniform sizes and sample sets
types VS, CF, CK, TC, ML and SL were randomly selected to of varying sizes for different faults. When all the six fault
form a statistically distributed data. In this dataset 30% of types are considered, testing is possible with combinations of
samples are used for training, while 35% for validation and 10, 20 and 30 samples of each type of fault. When four fault
remaining 35% for testing. Table. 3 shows the results of fault types are considered the testing is carried out with the
classification employed with this scheme. In this paper Ea combinations of 50 samples. The selection of features also has
refers to energy in approximation coefficients and Ed refers to significant impact on the classification rate. When four faults
energy in detail coefficients. The Ea-Ed is used to indicate the VS, TC, ML and SL are chosen, with 50 samples each,
energy in the combined approximation and detail coefficients. minimum accuracy of 0.98 is observed. When Ea alone is
used, with varying sizes of partitions for training, validation
TABLE. 3. CLASSIFICATION ACCURACY BASED ON and test sets, minimum classification accuracy observed is
EA, ED AND COMBINED EA-ED 0.95. When all the fault samples are input, with Ea, TC yields
Feature set Training Validation Testing the minimum accuracy of 0.7027. Use of Ed features resulted
in minimum accuracy of 0.8649 for CK. The features Ea and
Approx Coefficients (Ea) 0.9323 0.9184 0.8968 Ed, when used together, yield minimum classification
accuracy of 0.8269. For various combinations of training,
Detail Coefficients (Ed) 0.9850 0.9677 0.9548
validation and test sample sets, experimented with Ea features,
Combined Ea and Ed 0.9474 0.9032 0.8239 yields the same performance for all the faults, except for VS.
In majority of these scenarios VS is misclassified as CK, since
Overall accuracy 0.9549 0.9299 0.8913 both the fault types belong to the same subsystem.
V. CONCLUSION
C. Scheme 3: Approximation coefficient energy distribution
(Ea) on 4 faults T The work proposes classification of sound samples into one of
the six faults based on the acoustic signals produced by the
The percentages of energy distribution in both the motorcycles. It employs the distribution of the energy in
approximation and detail coefficients are used as features. In subbands of wavelet packets with ANN classifier. The ANN
this case four types of faults, namely, VS, TC, ML and SL are classifier has given the minimum performance of 0.8913 with
considered. A dataset of 50 samples for each of the four types test samples. Energy distribution among the detail coefficients
of faults is considered. Table. 4 shows the results of yields the best results compared to other features
classification for distribution of 30%, 35% and 35% of the combinations. Hence, the approach is proved to be robust to
total number of samples for training, validation and testing real-time implementation. The scope of the work is limited to
respectively. For this scheme, training, validation and testing fault source localization of motorcycles. It can be extended to
sample set yields classification accuracies of 1.0000, 1.0000 fault source localization of vehicles and other fault diagnosis
and 0.9857 respectively and average classification accuracy is applications based on sounds.
0.9950. Hence, with smaller number of fault types under
consideration, the classification accuracies are more ACKNOWLEDGMENT
appreciable than with more number of fault types. We the authors thank the service stations in Bagalkot for
the recording support and fault diagnosis expertise.
TABLE. 4. CONFUSION MATRIX FOR ALL SAMPLES WITH FOUR
FAULTS
 2015 International Conference on Soft-Computing and Network Security (ICSNS -2015), Feb. 25 – 27, 2015, Coimbatore, INDIA

References
[1] [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] [2] j. Altmann & j. Mathew, “Multiple band-pass autoregressive
demodulation for rolling-element bearing fault diagnosis”, Elsevier J.
Mechanical Systems and Signal Processing Vol. 15, No.5, September
2001, Pages 963–977.
[3] [3] Jian-Da Wu, Jian-Bin Chain, Chen-Wei Chung, & Hao Yu, “Fault
Analysis of Engine Timing Gear and Valve Clearance Using Discrete
Wavelet and a Support Vector Machine”, Int. Journal of Computer
Theory and Engineering, Vol.4, No.3, June 2012, pp386-390.
[4] [4] J. Lin, M. & J. Zuo, “Gearbox fault diagnosis using adaptive wavelet
filter”, Mechanical Systems and Signal Processing”, Vol.17, No.6, 2003,
pp1259-1269.
[5] [5] Datta Aveek, Mavroidis Constantinos, Krishnasamy Jay, Hosek
Martin. Neural network based fault diagnostics of industrial robots using
wavelet multi- resolution analysis. In: Proceedings of the American
control conference Marriott Marquis Hotel at Times Square New York
City, USA, 2007.
[6] [6] Hao Zhi-yong & Han Jun, “Identification of diesel front sound
source based on continuous wavelet transform”, Follow Journal of
Zhejiang University - Science A, Springer Journals, Vol.5, No.9, 2004,
pp1069-1075.
[7] [7] Zhang Junhong & Han Bing, “Analysis of engine front noise using
sound intensity techniques”, Mechanical Systems and Signal Processing,
Vol.19, 2005, pp213-221.
[8] [8] Junsheng Cheng, “A fault diagnosis approach for gears based on
IMF-AR model and SVM”, EURASIP J. Advances in Signal
Processing, 2008:doi:10.1155/2008/647135.
[9] [9] Jian-Da Wu & Chiu-Hong Liu, “Investigation of engine fault
diagnosis using discrete wavelet transform and neural network”, Elsevier
J. Expert Systems with Applications, 35, 2008, pp1200–1213.
[10] [10] J. D. Wu, E.C. Chang, S.Y. Liao, J.M. Kuo & C.K. Huang, “Fault
classification of a scooter engine platform using wavelet transform and
artificial neural network”, Proc. Int. Conf. Engineers and Computer
Scientists, IMECS 2009, Hong Kong, March 18-20, Vol.1, 2009, pp58-
63.
[11] [11] Wei Liao, Pu Han & Xu Liu, “Fault diagnosis for engine based on
EMD and wavelet packet BP neural network”, 3rd Int. Symp. Intelligent
Information Technology Application, 2009, pp672-676.
[12] [12] Basavaraj Anami, Veerappa Pagi & Sangamesh Magi, “Wavelet
based acoustic analysis for determining health condition of two-
wheelers”, Elsevier J. Applied Acoustics, Vol.72, No.7, 2011, pp464-
469.
[13] [13] A. Aljaafreh, L. Dong, “An evaluation of feature extraction
methods for vehicle classification based on acoustic signals.” Int. Conf.
in Networking, Sensing and Control (ICNSC), 2010.
[14] [14] V Muralidharan, V Sugumaran “Rough set based rule learning and
fuzzy classification of wavelet features for fault diagnosis of monoblock
centrifugal pump” Elsevier J. Measurement 46 (9), 2014, pp3057-3063.
[15] [15] Y. Sun & H. Qi., “Dynamic target classification in wireless sensor
networks”. In the Pattern Recognition, ICPR ,19 Int. Conf. 2008.
[16] [16] Pengfei Li, Yongying Jiang & Jiawei Xiang, “Experimental
Investigation for Fault Diagnosis Based on a Hybrid Approach Using
Wavelet Packet and Support Vector Classification” The Scientific World
Journal Vol. 2014, Article ID 145807,
http://dx.doi.org/10.1155/2014/145807.
[17] [17]
http://www.mqld.org.au/fileadmin/user_upload/state_files/qld/Document
s/Forms/Pg120_Sound_Basics.pdf (For recording environment setup).
[18] [18] A. N. Akansu, W. A. Serdijn & I. W. Selesnick, “Emerging
applications of wavelets: A review”, Physical Communication, Vol.3,
No.1, 2010, pp1-18.