fault diagnosis of helical gear box using decision tree ... · pdf filefault diagnosis of...
TRANSCRIPT
22
Fault Diagnosis of Helical Gear Box Using Decision Tree
and Best-First Tree
1M. Amarnath,
2Deepak Jain,
3V. Sugumaran and
4Hemantha Kumar
1Department of Mechanical Engg., Indian Institute of Information Technology Design and
Manufacturing Jabalpur, Jabalpur, Madyapradesh, India,
2,3
SMBS, VIT University, Chennai Campus, Vandalur-Kelambakam Road, Chennai,
4Department of Mechanical Engineering, National Institute of Technology, Surathkal, Karnataka,
Mangalore, India,
ABSTRACT
Gear is one of critical transmission elements, found its wide applications in small, medium and large
machineries. It is well proved that vibration signals acquired from a rotating machine comprise of the
dynamic information about the health condition of the rotating machine. This paper uses vibration
signals acquired from machinery comprising of gears in good and simulated faulty conditions for the
purpose of fault diagnosis by machine learning approach. In this paper, a vibration based on machine
learning approach is presented for helical gear box as it plays critical role in the industries. This
approach has mainly three steps namely feature extraction, feature selection and classification. Here
statistical analysis was used for feature extraction and decision tree and best-first decision tree for
classification was taken and compared. The classification accuracies for different conditions were
calculated and compared to find the best classifier for the fault diagnosis of the helical gear box.
Keywords: Gear fault diagnosis, Fault diagnosis, Statistical analysis, Decision tree, Best-first
Decision tree
1. INTRODUCTION
Helical gear box condition monitoring has received significant attention from many years. The
distinctive failure mode of helical gear box is localized defects, which occur if a sizable piece of
material on the contact surface is shifted during operation, usually by fatigue cracking under cyclic
contact stressing. Hence failure alarm from helical gear box is frequently established on the detection
of the localized defects in early stage. If the helical gear box is operating under various speeds and
loads, it’s not possible to measure or estimate the inclemency of localized faults, hence certain
physical parameters such as vibration, sound, acoustic emission and wear debris have been considered
in detection and diagnosis of inchoate faults. Sound and vibrations generated by rotating machinery
often mask the features of fault-related signals generated by the machine elements like gears, bearings
and cams (Lai Wuxiang et al. [1], Zvokelj et al. [2], Li and Ma [3] and Tinta et al. [4]).Then again, it is
known that local faults in helical gear box induce impacts, as a result of which transient excitations
might be seen in vibration and sound signals. Villa et al. [5] presented findings which deals with the
detection of different mechanical faults (unbalance and misalignment) under a broad range of working
conditions of speed and load. The conditions tested in a test bench are alike to the ones that can be
International Journal of Research in Mechanical Engineering
Volume 1, Issue 1, July-September, 2013, pp. 22-33, © IASTER 2013
www.iaster.com, ISSN Online:2347-5188 Print: 2347-8772
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
23
found in dissimilar kinds of machines like for example wind turbines. The authors demonstrate how to
take advantage of the information on vibrations from the mechanical system under study in a broad
range of load and speed conditions. Utilizing such information the prognosis and detection of faults is
quicker and more reliable than the one obtained from an analysis over a limited range of working
conditions (e.g. nominal). Due to the presence of growing local faults, vibration and sound signals
from helical gear box have non-stationary characteristics; therefore analyses of the above signals in
faulty operating conditions become hard. Under such conditions, ceremonious application of
measurement of statistical parameters may not be practicable (Oguamanam et al. [6], Benko et al. [7]
and Benko et al. [8]). Loutas et al. [9] conducted analyses on lab-scale, single stage, gearbox using
different non-destructive inspection methodologies with advanced signal processing techniques.
Ramroop et al. [10] reported a comparative study of conventional vibration and acoustic monitoring
techniques to diagnose defects in industrial multi stage gearbox, operating under healthy and faulty
conditions. Tekiner and Yesilyurt [11] carried out experiments to find best suitable machining
conditions using statistical values of sound signals. Their work highlighted the influence of sound
signal to analyze the machining process parameters viz. flank wear, surface roughness, chip
morphology and built-up edge formation. Byrtus and Zeman [13] presented an original method of the
mathematical modeling of gear drive nonlinear vibrations. Their proposed model includes
nonlinearities caused by gear mesh interruption, parametric gearing excitation caused by time-varying
meshing stiffness and nonlinear contact forces acting between journals of the rolling-element bearings
and the outer housing. The nonlinear model is then used for investigation of gear drive vibration,
especially for detection of nonlinear phenomena like impact motions, bifurcation of solution and
chaotic motions in case of small static load and in resonant states. The theoretical method is used for
investigation of two-stage gearbox nonlinear vibration. Amarnath et al. [14] presented the results of
experimental investigations carried out to assess wear in spur gears of back-to-back gearbox under
accelerated test conditions. The studies considered the estimation of operating conditions such as film
thickness and their effects on the fault growth on teeth surface. Modal testing experiments have been
carried out on the same gear starting from healthy to worn out conditions to quantify wear damage.
The results provide a good understanding of dependent roles of gearbox operating conditions and
vibration parameters as measures for effective assessment of wear in spur gears. Saravanan et al. [15]
presented the effectiveness of wavelet-based features for fault diagnosis of a helical gear box using
artificial neural network (ANN) and proximal support vector machines (PSVM). The statistical feature
vectors from Morlet wavelet coefficients are classified using J48 algorithm and the predominant
features were given as input for training and testing ANN and PSVM and their relative accuracy in
classifying the faults in the bevel gear box was compared. The Naïve Bayesian classifier is very
simple and efficient and highly sensitive to feature selection, so the research of feature selection
especially for it is significant. Naïve Bayes and Bayes net algorithms were effectively used for tool
condition monitoring as well [16].
From the literature one can understand that many classification algorithms have been used for
classifying the faults in helical gear box and other rotating members. In order to say strongly that a
particular algorithm is better compared to other algorithms a detailed comparative study needs to be
done. Therefore, this paper mainly deals with the performances of Decision Tree and Best-first
Decision Tree algorithms in classifying gear faults.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
24
2. EXPERIMENTAL SETUP AND PROCEDURE
Fig.1 shows the experimental setup. The setup consists of a 5 HP two stage helical gearbox. The gear
box is driven by a 5.5 HP, 3-phase induction motor with a rated speed of 1440 rpm. The speed of the
motor is controlled by an inverter drive and for the present study the motor is operated at 80 rpm. In
other words the speed of the first stage of the gearbox is 80 rpm. With a step-up ratio of 1:15, the
speed of the pinion shaft in the second stage of the gear box is 1200 rpm. Table.1 summarizes the
specifications of the test rig.
Fig.1 Experimental setup of two stage helical gearbox
Table 1- Specifications of helical gearbox
First stage Second stage
Number of teeth 44/13 73/16
Pitch circle diameter (mm) 198 /65 202 /48
Pressure angle (°) 20 20
Helix angle 20 15
Modules 4.5/ 5 2.75 / 3
Speed of shafts 80 rpm (input) 1200 rpm (output)
Mesh frequency 59 Hz 320 Hz
Step - up ratio 1:15
Rated power 5 HP
Power Transmitted 2.6 HP
The pinion is connected to a D.C motor (which is used as generator) to generate 2 kW power, which is
dissipated in a resistor bank. Therefore, the actual load on the gearbox is only 2.6 HP which is 52% of
its rated power 5 HP. In industrial environment utilization of load varies from 50% to 100%. In the
case of traditional dynamometer, additional torsional vibrations can occur due to torque fluctuations.
This is avoided in this case by using D.C motor and resistor bank.
Tyre couplings are fitted between the electrical machines and gear box so that backlash in the system
can be restricted to the gears. The motor, gear box and generator are mounted on I-beams, which are
anchored to a massive foundation. Vibration signals are measured using a B&K accelerometer which
is installed close to the test bearing. Signals are sampled at a sampling frequency of 8.2 kHz. The
experimental setup with equipment and sensors is shown in Fig. 2.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
25
Fig. 2 Photograph of experimental set up with sensors and equipment
Overhaul time of a new gear box is more than one year. It is very difficult to study the fault detection
procedures without seeded fault trials. Local faults in a gear box can be classified into three categories.
(a) Surface wear spalling (b) cracked tooth and (c) loss of a part of tooth due to breakage of tooth at
root or at a point on working tip (broken tooth or chipped tooth). There are different methods to
simulate faults in gears viz. electric discharge machining (EDM), grinding, adding iron particles in
gearbox lubricant and over loading the gear box i.e. accelerated test condition. The simplest approach
is partial tooth removal. This simulates the partial tooth break, which is common in many industrial
applications (Staszewski et al. [17], Yesilyurt [18], Yesilyurt [19] and Loutridis [20]).
3. FEATURE EXTRACTION
Statistical analysis of vibration signals concedes different parameters. The statistical parameters used
for this study are mean, standard error, median, standard deviation, sample variance, kurtosis,
skewness, range, minimum, maximum and sum. These features are extracted from the vibration
signals. They are named as ‘statistical features’. Brief descriptions of the extracted features are given
in the following Table 2.
Table 2- Statistical Features
Sl.
No.
Name of the
Feature Feature Description
1 Standard error
2
2
2
1Standard error of the predicted,
2
x x y yY y y
n x x
2 Standard
deviation
22
Standard Deviation ( 1)
x x
n n
3 Sample variance
22
Sample Variance ( 1)
x x
n n
3-phase
Induction
Motor
Two stage gearbox
FFT analyzer
Resistorbank
D.C. motor
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
26
4 Kurtosis
4 2( 1) 3( 1)Kurtosis
( 1)( 2)( 3) ( 2)( 3)
ix xn n n
n n n s n n
5 Skewness
3
Skewness 1
ix xn
n s
6 Range It refers to the difference in maximum and minimum signal point values for
a given signal.
7 Minimum value It refers to the minimum signal point value in a given signal.
8 Maximum value It refers to the maximum signal point value in a given signal.
9 Sum It is the sum of all feature values for each sample.
4. CLASSIFIER
Decision tree is a tree established on knowledge methodology applied to interpret classification rules
[21]. A standard tree got from J48 comprises of a number of branches, one root, a number of nodes
and a number of leaves (See Fig.4). One branch of which is a chain of nodes from root to a leaf and
each node have one attribute. An attribute in a tree gives the information of the importance of the
associated attribute. The procedure of making the Decision Tree and using the same for feature
selection is explained in detail by Sugumaran et al [21].
4.1 Gini Index
Gini index is another criterion to evaluate the impurity of a node, which is applied in the CART
system (Breiman et al [22]). If pi is probability that the instance is in class i and pj stand for the
probability that an instance is in class j, the Gini index is given in form (Breiman et al [22]).
(1)
The Gini index has an interesting interpretation in terms of sample variances (Breiman et al [22]).
When all the instances at a node that are in class j are allotted the value 1and the other instances are
allotted the value 0, the sample variance for these values is pj (1 − pj), i.e. pj – pj2. Restating this for all
classes and adding the variances as the sum of the pj over all classes is 1, it can be written that the Gini
index can be presented as
(2)
If only two classes are used, the Gini index can be modified to 2p1p2. Figure 3 tells the relationship
between the Gini index and the class probability of the first class in a two-class problem, accompanied
by the relationship between the information value and the same class probability. From the figure, we
can view that in both cases, the impurity is maximum if the two classes has equal probability (i.e. 0.5
for each class). And if any of class probability is 1, the impurity is 0. This means the node is
absolutely pure.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
27
Fig. 3 The information value and the Gini index in a two-class problem.
5. RESULTS AND DISCUSSION
The vibration signals were recorded for normal and abnormal conditions of helical gear box. Totally
420 samples were collected; out of which 60 sample were from no load condition. For 100% load with
10%, 20%, 30%, 40%, 80% and 100% fault, 60 samples from each condition were collected. The
statistical parameters explained in Section 3 are treated as features and used as input to the algorithm.
The corresponding status or condition (10%, 20%, 30%, 40%, 80%, 100%, expt no load) of the
classified data will be the required output of the algorithm. This input and corresponding output
together forms the dataset.
5.1 Effect of number of features and Feature selection
Totally fourteen descriptive statistical features were extracted from vibration signals. All of them may
not be significant for the classification purpose. One cannot say before in hand which are the features
will be helpful for the purpose of classification while using machine learning algorithms. As more
number of irrelevant features actually may reduce the performance of the classification algorithm.
Also, they increases the computational resources required. Researchers have to extract all descriptive
statistical features and then select the good ones. Here, decision tree was used for feature selection. In
decision tree, the feature that occurs first will be the root node and the same will be the best feature for
classification. The other features in the tree are in the order of importance. In decision tree, only seven
features were present namely, standard error, maximum, minimum, range, mean, median and kurtosis
in the order of importance. This situation in general means only seven features are enough to classify
the gear box conditions. However, one needs to confirm the same to be on the safer side. For this
purpose, the effect of features on classification accuracy study was carried out (Sugumaran et al [21]).
Initially, only first best feature alone was used with decision tree and found the classification accuracy.
This happens to be mean as it is in the top node of the decision tree. Then, top two best features were
used with decision tree and found the classification accuracy (mean and standard error).
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
28
5.2 Decision Tree
The dataset is used with decision tree algorithm for the purpose of feature selection and classification.
The generated decision tree is shown in Fig.4. The rectangles represent classes (condition of the
helical gear box). In rectangle the information about the condition is given using abbreviations e.g.
‘expt no load’. Then within parenthesis, there are two numbers separated by a slash or one number.
The first number (in case of two numbers) or the only number represents the number of data points
that support the decision. Meaning, if one follows the rule (as described above ‘if-then’ rules) how
many data points will be correctly classified is given as first number. The second number (after slash)
is optional and it represents the number of data points that is against the rule followed. Meaning, if one
follows a rule, how many data points will be incorrectly classified is given as second number.
FiFig.4 Decision tree
The detailed class-wise accuracy of the J48 algorithm is presented in Fig.5. Out of the terms used in
Fig.8, ‘TP rate’ and ‘FP rate’ are very important. The ‘TP rate’ stands for true positive and its value
should be close to ‘1’ for better classification accuracy. The ‘FP rate’ stands for false positive and its
value should be close to ‘0’ for better classification accuracy. In the study, one can appreciate the
closeness of ‘TP rate’ to ‘1’ and ‘FP rate’ to ‘0’. The both values confirm that the built model is good
one.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
29
Fig.5 Detailed accuracy by class
The classification accuracy of the decision tree algorithm is presented in Fig.9. The interpretation of
the confusion matrix is as follows:
The diagonal elements in the confusion matrix (Refer Fig.6) show the number of correctly
classified instances.
In the first row, the first element shows number of data points that belong to ‘10.0% fault’
class and classified by the classifier as ‘10.0 fault’.
In the first row, the fifth element shows the number of data points belonging to ‘10.0%
fault’ class but misclassified as ‘80% fault’.
In the first row, the sixth element shows the number of ‘10% fault’ data points
misclassified as ‘100.0% fault’.
In the first row, the seventh element shows the number of ‘10% fault’ data points
misclassified as expt no load and so on.
In the present study, the faulty conditions are clearly distinguishable from 10% fault condition by the
algorithm. Therefore, one can note ‘0’ in first column third and fourth elements which correspond to
misclassification of faulty conditions. However, there are misclassifications in other conditions. They
are given in non-diagonal elements. Here, out of 420 data points, 65 data points were misclassified by
the algorithm. Actually, this is an error of about 15.5%, which is acceptable for many practical
applications.
a b c d e f g Classified as
51 0 0 0 4 3 2 a =10.0
8 48 0 2 0 2 0 b = 20.0
0 5 46 4 1 1 3 c=30.0
0 1 1 48 0 7 3 d = 40.0
3 0 0 0 57 0 0 e =80.0
2 0 0 5 0 52 1 f = 100.0
1 0 1 2 3 0 53 g = expt no load Fig. 6 Confusion matrix for decision tree
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
30
Summary of stratified cross validation is given below.
Total Number of Instances 420
Correctly Classified Instances 355 84.52%
Incorrectly Classified Instances 65 15.47%
5.3 Best First Tree Algorithm
From the decision tree, seven features that are contributed for classification were only selected for
training and testing. The selected features were trained and tested by using Best first tree algorithm
with post pruning processes.
The detailed class-wise accuracy of the Best first tree is presented in Fig.7. Misclassification details
for both the conditions are presented as confusion matrix in Fig.8.Out of the terms used in Fig.8, ‘TP
rate’ and ‘FP rate’ are very important. The ‘TP rate’ stands for true positive and its value should be
close to ‘1’ for better classification accuracy. The ‘FP rate’ stands for false positive and its value
should be close to ‘0’ for better classification accuracy. In the study, from Fig.8, one can appreciate
the closeness of ‘TP rate’ to ‘1’ and ‘FP rate’ to ‘0’. The both values confirm that the built model is
good one. The classification accuracy of the best first tree algorithm is presented in Fig.7.
Fig.7 Detailed accuracy by class - Best first tree
In the present study, the faulty conditions are clearly distinguishable from 10% fault condition by the
algorithm. Therefore, one can note ‘0’ in first column third and fourth elements which correspond to
misclassification of faulty conditions. However, there are misclassifications in other conditions. They
are given in non-diagonal elements. Here, out of 420 data points, 62 data points were misclassified by
the algorithm. Actually, this is an error of about 14.7 %, which is acceptable for many practical
applications.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
31
a b c d e f g Classified as
51 0 0 0 3 4 2 a =10.0
8 48 0 2 0 2 0 b = 20.0
0 6 46 2 2 0 4 c=30.0
0 0 1 51 0 5 3 d = 40.0
1 0 0 0 59 0 0 e =80.0
1 0 0 6 0 52 1 f = 100.0
1 0 0 7 1 0 51 g = expt no load
Fig. 8 Confusion matrix with best first tree
The Summary of stratified cross validation is given below.
Total Number of Instances 420
Correctly Classified Instances 358 85.23 %
Incorrectly Classified Instances 62 14.76 %
Fig. 9 illustrates the comparison of classification efficiency of J48 Decision tree and Best first decision
tree algorithm with variation in number of objects.
Fig. 9 Number of features Vs Classification accuracy
6. CONCLUSION
Condition monitoring of a helical gear box was carried out using vibration signals and the statistical
features were extracted. They were classified using decision tree and best first tree classifiers and the
superior feature – classifier combination was found. Dimensionality reduction was carried using
decision tree. The classification accuracy of statistical feature using decision tree was found to be
84.52% while with best first tree it was found to be 85.23%. The statistical feature with best first tree
combination yields the highest accuracy as per the conditions in this study. Although the difference in
classification accuracy between the best first tree and decision tree may be marginal, the kind of
features makes a huge impact on the study results.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
32
7. REFERENCES
[1] L. Wuxing, P. W. Tse, Z. Guiicai and Shitielin, Classification of gear faults using cumulant and
the radial basis function, Mechanical Systems and Signal Processing, 2004;18: 381-389.
[2] M. Zvokelj, S. Zupan, I. Prebil, Multivariate and multiscale monitoring of large-size low-speed
bearings using Ensemble Empirical Mode Decomposition method combined with Principal
Component Analysis, Mechanical Systems and Signal Processing,2010; 24: 1049-1067.
[3] C. J. Li and J. Ma, Wavelet decomposition of vibrations for detection of bearing localized
defects, NDT&E International, 1997; 30 (3):143-149.
[4] D. Tinta, J. Petrov, U. Benko, j. Juric, A. Rakar, M. Zele, J. Tavar, J. Rejec and A. Stefanovska,
Fault diagnosis of vacuum cleaner motors, Control Engineering Practice,2005; 13 :177–187.
[5] L. F. Villa, A. Renones, J. R. Peran and L. J. de Miguel,
Statistical fault diagnosis based on
vibration analysis for gear test-bench under non-stationary conditions of speed and
loadMechanical Systems and signal Processing, Mechanical Systems and Signal Processing,
2012; 29: 436-446.
[6] D.C.D. Oguamanam, H.R.Martin and J.P. Huissoon, On the Application of the Beta Distribution
to Gear Damage Analysis, Applied Acoustics, 1995; 45 (3):247-261.
[7] U. Benko, J. Petrov, D. Juricic, J. Tavcar and J. Rejec, An approach to fault diagnosis of vacuum
cleaner motor based on sound analysis, Mechanical Systems and Signal Processing, 2005;19:
427-445.
[8] U. Benko, J. Petrov, D. Juricic, J. Tavcar, J. Rejec and A. Stefanovska , Fault diagnosis of a
vacuum cleaner motor by means of sound analysis, Journal of Sound and Vibration,2004; 276:
781-806.
[9] T.H. Loutas, G. Sotiriades, I. Kalaitzoglou and V. Kostopoulos, Condition monitoring of a
single-stage gearbox with artificially induced gear cracks utilizing on-line vibration and acoustic
emission measurements, Applied Acoustics, 2009; 9 (70): 1148-1159
[10] G. Ramroop, K. Liu, F. Gu, B. S. Paya and A.D. Ball, Airborne Acoustic Condition Monitoring,
www.maintenenceengineering.com (Accessed Dec. 2003).
[11] Z. Tekıner and S. Yesilyurt, Investigation of the cutting parameters depending on process sound
during turning of AISI 304 austenitic stainless steel, Materials and Design, 2004; 25: 507–513.
[12] N.R. Sakthivel, B.B. Nair, V. Sugumaran, R.S. Rai, Decision support system using artificial
immune recognition system for fault classification of centrifugal pump. International Journal of
Data Analysis Techniques and Strategies, 3 (1): 66-84.
[13] M. Byrtus and V. Zeman, On modeling and vibration of gear drives influenced by nonlinear
couplings, Mechanism and Machine Theory, 2011; 46 (3): 375-397.
International Journal of Research in Mechanical Engineering
Volume-1, Issue-1, July-September, 2013, www.iaster.com ISSN
(O) 2347-5188
(P) 2347-8772
33
[14] M. Amarnath, C. Sujatha and S. Swarnamani, Experimental studies on the effects of reduction in
gear tooth stiffness and lubricant film thickness in a spur geared system, Tribology
International, 2009;42 (2): 340-352.
[15] N. Saravanan , V.N.S. Kumar, Siddabattuni and K.I. Ramachandran, Fault diagnosis of spur
bevel gear box using artificial neural network (ANN)and proximal support vector machine
(PSVM), Applied Soft Computing, 2010;10 (1): 344-360.
[16] M. Elangovan, K.I. Ramachandran, V. Sugumaran, Studies on Bayes classifier for condition
monitoring of single point carbide tipped toolbased on statistical and histogram features, Expert
Systems with Applications,2010;37: 2059–2065.
[17] Staszewski, W. J., K. Worden., and G. R. Tomlinson Time-frequency analysis in gearbox fault
detection using the Wigner-ville distribution and pattern recognition, Mechanical Systems and
Signal Processing,1997; 11(5): 673-692.
[18] Yesilyurt, I., Fengshou and A.D. Ball Gear tooth stiffness measurement using modal analysis
and its use in wear fault severity assessment of spur gears, NDT&E International, 2003; 36:
357-372.
[19] Yesilyurt, I., Gearbox Fault Detection and Severity Assessment Using Vibration Analysis, Ph.D
Thesis, University of Manchester, UK, 1997.
[20] Loutridis S.J., Damage detection in gear system using empirical mode decomposition.
Engineering structures, 2004; 26(12): 1833- 1841.
[21] V. Sugumaran, V. Muralidharan and K.I. Ramachandran, Feature selection using Decision Tree
and classification through Proximal Support Vector Machine for fault diagnostics of roller
bearing, Mechanical Systems and Signal Processing, 2007; 21(2): 930-942.
[22] Breiman, L., J. H. Friedman, R. A. Olshen, and C. G. Stone, Classification and Regression
Trees. Wadsworth International Group, Belmont, California, USA, 1984.
[23] M Saimurugan, K.I. Ramachandran, V. Sugumaran, N.R. Sakthivel, Multi component fault
diagnosis of rotational mechanical system based on decision tree and support vector machine,
Expert Systems with Applications, 38 (4):3819-3826.
[24] R.A. Kumar, V. Sugumaran, BHL Gowda, CH Sohn,Decision tree: A very useful tool in
analysing flow-induced vibration data, Mechanical Systems and Signal Processing,22 (1):202-
216.
[25] S. Ravikumar, K.I. Ramachandran, V. Sugumaran, Machine learning approach for automated
visual inspection of machine components,Expert Systems with Applications,38 (4): 3260-3266.
[26] V. Indira, R. Vasanthakumari, V. Sugumaran, Minimum sample size determination of vibration
signals in machine learning approach to fault diagnosis using power analysis,Expert Systems
with Applications, 37 (12): 8650-8658.