International Journal of Mechanical and Production

Engineering Research and Development (IJMPERD)

ISSN(P): 2249-6890; ISSN(E): 2249-8001
Vol. 8, Issue 1 Feb 2018, 1135-1150
© TJPRC Pvt. Ltd.

EFFECTIVE TIME DOMAIN FEATURES FOR IDENTIFICATION OF

BEARING FAULT USING LDA AND NB CLASSIFIERS

B R NAYANA1 & P GEETHANJALI2

1
Department of Electrical & Electronics Engineering, Sir MVIT, Bangalore, India
2
School of Electrical Engineering, VIT University, Vellore, India
ABSTRACT

Recently, the mechanical fault detection of an induction motor (IM) from vibration signals using pattern
recognition has proven to be an effective method. This paper has studied for the first time statistical time domain features
mean absolute value (MAV), waveform length (WL), zero crossing(ZC), slope sign changes (SSC), simple sign
integral(SSI) and Willison amplitude (WAMP) for identification of the mechanical faults using linear discriminant
analysis (LDA) and naive Bayes (NB) classifiers. In this study, the effectiveness of the features is investigated using
parameters like accuracy, sensitivity and specificity individually and in groups for a total of 63 combinations. Each
feature set combination is investigated for 15datasets defined under 5 groups in different combinations of faulty and
normal working conditions. The results indicate that the feature set of SSI, WL,SSC and ZC features outperform the

Original Article
conventional features in the identification of faults and is found to be computationally effective. Further, NB classifier is
found to be better than LDA in identification of mechanical faults.

KEYWORDS: Fault Diagnosis, Statistical Features, Linear Discriminant Analysis Classifier, Naive Bayes Classifier,
Roller Bearings & Pattern Recognition

Received: Jan 05, 2018; Accepted: Jan 25, 2018; Published: Feb 07, 2018; Paper Id.: IJMPERDFEB2018135

1. INTRODUCTION

In recent years, there has been considerable evolution in the field of fault diagnosis of induction machines
with the aid of expert systems and artificial intelligence algorithms. Many condition monitoring techniques have
been successfully developed and implemented. Bearing faults are among the more prominently occurring faults [1]
and hence their diagnosis forms an essential part in condition monitoring of induction machines. Large number of
detection techniques have been developed based on signature analysis of either stator current or vibration signals.
Among this, vibration signals have been proven to be more reliable for diagnosing mechanical faults either
invasively or non-invasively. Condition monitoring of bearing faults with pattern recognition involves feature
extraction, feature selection, feature reduction and their classification. Typically, pattern recognition methods are
applied to diagnose the faults with time domain features like peak value, crest factor, kurtosis, etc.[2-3].Prior
researches in this area using time domain features like mean, standard deviation, shape factor, etc. have been found
to yield poor results [4]. Investigations using frequency domain features like power spectrum, power spectral
density, periodograms etc.[5-6] relies on the differences in frequency characteristics of fault conditions[7]. These
differences are non-significant and hence difficult to diagnose. As vibration signals are non-stationary in nature,
time–frequency domain analysis like spectrogram, wavelets transforms(WT) etc. have been used for extracting
features to identify the bearing faults[7-12]. This analysis using WT methodology [13] suffers a major setback due

www.tjprc.org [email protected]
1136 B R Nayana & P Geethanjali

to adjustable windowed Fourier transforms energy leakage occurring during signal processing. Another limitation of this
technique is that the success of this relies heavily on the choice of appropriate base function which determines the
frequency bands of the decomposed signals.

In the present work, the authors attempted on novel time domain features such as mean absolute value (MAV),
waveform length (WL), zero crossing (ZC), slope sign changes (SSC), simple sign integral(SSI) and Wilson
amplitude(WAMP) and found time domain features outperform frequency and time-frequency features. Though, frequency
and time-frequency features necessitate the dimensionality reduction methods prior to classifiers. These time domain
features do not require any feature reduction or selection schemes and hence found to be computationally effective.
Literature survey shows that various classification techniques such as k-nearest neighbor (KNN) [14,15] artificial neural
network (ANN)[3,16,17] support vector machine[18], linear discriminant analysis (LDA)[2] etc. have been employed to
study the performance of the extracted features. In this paper, the authors have used naive Bayes (NB) classifier and
compared the results thus obtained for variation in parameters like accuracy, sensitivity and specificity with that obtained
using LDA classifier. The results evince that time domain features identify the bearing faults with good accuracy compared
to other features considered in the literature [2], [13-14], [17-22]. Overall 63 feature set combinations from 6 features have
been employed for bearing fault diagnosis of 5 groups of data involving 15 datasets which has been drawn in combinations
of location of fault and load condition. Though condition monitoring schemes of bearing faults involve Feature extraction,
feature selection, features reduction and classification processes which will be handled by various methodologies like WT
for feature extraction and minimum-redundancy maximum-relevancy method for feature selection and differential
evolution algorithm for classification[21] and spectrum imaging has been implemented for feature extraction and
enhancement[22] in previous works. The present work focuses on the simplest scheme development to serve the purpose;
hence Feature extraction and classification alone are implemented. It should be noted that the approach is not limited to
these 6 features. Other features, such as MAV slope, log detection, peak factor, etc. can be also used. Detailed discussions
of which type of features are more useful than others for bearing fault diagnosis are beyond the scope of this paper.

2. METHODOLOGY
2.1. Dataset Description

The vibration recordings from experiments conducted using a 2 HP Reliance Electric motor by CWRU (Bearing
vibration dataset, 0000) has been used to derive five groups (A-E). The drive end (DE) bearing (6205-2RS JEM SKF
make) and fan end (FE) bearing (6203-2RS JEM SKF bearing) are selected as the test bearings. Motor bearings were
seeded with faults using electric discharge machining (EDM). Fault depths of 0.007 inch, 0.014 inch and 0.021inch with
0.040 inch of diameter were artificially created at the inner raceway (IR), rolling element (RE) and outer raceway (OR).
Faulted bearings with respect to all 3 faults (3F) were reinstalled into the test motor and vibration data was recorded for
motor loads of 0 to 3 HP (motor speeds of 1797 to 1730 RPM) individualistically. Vibration data have been collected using
accelerometers, which were placed at 12 O’ clock position at DE and FE of the motor housing at a sampling rate of
12kHzfor a duration of 10 seconds. The data recorder had been equipped with a low - pass filter at the input stage for
antialiasing. The benchmark study made by [23] indicates that the central load zone is at 6 O’ clock position. In addition,
the study revealed the fact that some of the vibration signals recorded on DE and FE positions were non useable due to lack
of clarity, clipping of the data, and contamination of data due to the presence of significant electrical noise signals. The
five groups (A-E) derived are as shown in Table1 with varying combinations of bearing fault depth in milli inches (FD)

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1137
Bearing Fault using LDA and NB Classifiers

and load conditions to explore the effectiveness of time domain features and classifiers considered in this study.

Overall 8 normal and 60 faulty working conditions are used for the analysis. The faulty conditions considered are
obtained from DE data for 3F with 3 FDs and FE data for 3F with 2 FDs, each under 4 different loading conditions.
Correspondingly, 8 normal working conditions for DE and FE are considered each of 4 different loading conditions. In
Group A, the datasets A-i, iii and v are derived to study a four class classification of N and defects with IR, RE and OR for
identical load and identical FD conditions. Dataset A-i is drawn for DE N and all 3F with FD of 7, thus includes 4 working
conditions for each load respectively.

Table 1: Basic Information of 5 Groups

No of Working No of
Group Dataset Description
Conditions Classes
i. DE-N and 3F of FD 7 for each load 4
ii. DE-N and 3F of FD 7 for all 4 loads together 16
iii. DE-N and 3F of FD 14 for each load 4
A 4
iv. DE-N and 3F of FD 14 for all 4 loads together 16
v. DE-N and 3F of FD 21 for each load 4
vi. DE-N and 3F of FD 21 for all 4 loads together 16
i. DE-N and 3F of 3FDs (7,14, 21) for all 4 loads together 40
ii. DE- N and 3F of 3 FDs (7,14, 21) for each load 10
B 4
iii. DE- N and 3F of 2 FDs(7 & 21) for all 4 loads together 28
iv. DE- N and 3F of 2 FDs (7 & 21) for each load 7
i. DE-3F of 2 FDs(7 & 21) for each load
C 6 6
ii. FE-3F of 2 FDs(7 & 21) for each load
iii. DE & FE-3F of FD 7 for each load
D 6 6
iv. DE & FE 3F of FD 21for each load
E DE- N and 3F of 3 FDs (7,14, 21) for each load 10 10

Datasets A-iii and A-v are derived in a similar manner for FDs of 14 and 21 respectively. A four class
classification is studied in datasets A-ii, A-iv and A-vi for identical FD conditions, but being independent of load with FDs
of 7,14 and 21 respectively. Thus, datasets A-ii, A-iv and A-vi includes 16 working conditions. Group B, is also derived
for the same four class classification, with different combination of FDs and load conditions. The datasets B-i and B-ii
deals with FDs of7, 14 and 21; and FDs of 7 and 21 are considered for datasets B-iii and B-iv. However, dataset B-i and B-
iii are implemented irrespective of load condition, thus 40 and 28 working conditions are employed correspondingly.
Whereas, datasets B-ii and B-iv are implemented with 10 and 7 working conditions respectively for each load condition.
Group C considers the working conditions of same bearings with all 3Fof FD7 and 21to perform a 6 class classification for
every load condition. Hence, group C includes 2 datasets, one for DE and one for FE respectively. A similar analysis is
performed with group D wherein, the datasets are for all 3Fwithsame FD over DE and FE. Therefore one dataset is for FD
of 7 and other for FD of 21 respectively. Group E, includes N and all 3F for FDs of 7,14 and21 of DE for every load
condition. Therefore, 10 working conditions are employed of for a 10 class classification.

2.2. Feature Extraction

The temporal characteristics, hidden in the vibration signals are extracted using newly attempted time domain
features such as mean absolute value, simple sign integral, waveform length, Willison amplitude, zero crossing, and slope
sign for bearing fault diagnosis. The mathematical description of proposed features is presented in this section.

Mean Absolute Value (MAV): Mean absolute value is the average of absolute value of data for a segment of

www.tjprc.org [email protected]
1138 B R Nayana & P Geethanjali

length L and is defined in equation (1). MAV is similar to average rectified value and can be calculated using the moving
average of full-wave rectified vibration signal.

(1)

Simple Sign Integral (SSI): Simple sign integral is the integral of square of data samples. It determines the energy of the
data segment and is computed using equation (2).

(2)

Waveform Length (WL): Waveform length is the cumulative length of the waveform over the time segment. It is related
to amplitude, time and frequency information of the data segment and is calculated using equation (3).

(3)

Willison Amplitude (WAMP): Willison amplitude is the number of times the difference between amplitude of adjacent
samples exceeds a predefined threshold value. It is calculated using equation (4) and (5).

(4)

where (5)

and is the threshold value and chosen as 0.5.

Zero Crossing (ZC): Zero crossing is the number of times the signal crosses zero. This is a feature, provides information
about frequency of the signal and is calculated from (6) which satisfy equation (7).

(6)

(7)

To abstain from the background noise a small threshold of =0.5 is chosen.

Slope Sign Change (SSC): Slope sign change is another feature that characterizes the frequency and is computed using
equation (8) and satisfying equation (9).

(8)

(9)

Slope sign change indicates the number of changes between positive and negative slope among three consecutive

segments. A threshold =0.5 is used for avoiding the interference in vibration signal.

These 6 features extracted from the bearing vibration signals are given as input for LDA and NB classifiers. The
effectiveness of the features is studied in 63 feature set (FS) combinations.

FS={FS1,FS2,FS3,FS4,FS5,FS6,FS7,FS8………………..,FS61,FS62,FS63} (10)

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1139
Bearing Fault using LDA and NB Classifiers

where FS1-FS6 are FSs of individual features, FS7 is set of MAV,SSI and FS8 is set of MAV,WL similarly further FSs
are derived with combinations of 2,3,4,5 and 6 features. The performance of these time domain features has been
investigated using vibration data recorded by Case Western Reserve University (CWRU) [24]. Each working condition of
the dataset contains 10 seconds of vibration signal form DE and FE from which, 65536 samples (5.46 seconds) are
considered for processing and are segmented into windows of length 1024 with 50% overlapping. Features are extracted
for each segment thus resulting in a feature length of 128 for every feature pertaining to respective working condition.
Based on the experimental motor bearing data discussed in section 2.1, the analysis results are drawn in section 3 to verify
the effectiveness of these 6 features in 63 combinations for bearing fault diagnosis, with respect to accuracy, sensitivity and
specificity.

2.1 Classification

LDA classifier

Linear discriminant analysis is the most common technique used for data classification and dimensionality reduction.
Linear discriminant analysis easily handles the case where the within-class frequencies are unequal and their performances
have been examined on randomly generated test data. LDA approach for classification considers posterior probability,
prior probability and cost of classifying an observation to a particular class. Thus the objective is to minimize the
classification cost and the minimization function is defined as

(11)

where

is the predicted class.

N is the number of classes.

is the posterior probability of class k for observation X.

cost(c|k) is the cost of classifying an observation as c when its true class is k.

The posterior probability that an observation X belongs to class k given as

(12)

Where P(k) represents the prior probability of class k.

P(X) is a normalization constant, that is, the sum over k of P(X|k)P(k)

P(X|k) is the multivariate normal density function and is given as

(13)

Where is the covariance matrix of kth class.

(14)

and is the mean of kth class.

www.tjprc.org [email protected]
1140 B R Nayana & P Geethanjali

The LDA classifier steps can be summarized as to estimate the prior probabilities, mean and covariance matrix for
each class. Further, for a new observation X estimate the class using equation (11).

NB classifier

Naive Bayes is based on Bayes theorem suited to solve the high dimensional problems. Parameter estimation for
naive Bayes models uses the method of maximum likelihood and performs better in many complex real world situations

The advantage of NB classifier is it requires a small amount of training data to estimate the parameters. The algorithm
for implementation of NB classifier is as follows:

If there are ‘m’ classes: C1,C2,C3…Cm , and the feature vector X : [x1,x2,x3,…. xn], for n number of features, the naive
Assumption of class conditional independence computed using equations (15) and (16).

(15)

(16)

NB classifier predicts that X belongs to Class Ci iff

(17)

The maximum posteriori hypothesis can be stated as

(18)

Maximize P(X/Ci)P(Ci) as P(X) is constant. (19)

where P(Ci) is class prior probability.

P(X) is the prior probability of X.

P(Ci/X) is the posterior probability.

P(X/Ci) is the posterior probability of X conditioned on Ci.

With many attributes, it is computationally expensive to evaluate P(X/Ci). Being conditionally independent and
computationally expensive are the only drawbacks of this classifier.

2.2 Performance Metrics

Classification of bearing conditions for groups A to E are employed with 63 feature set combinations and the
performance is assessed for 50% training and 50% testing sizes respectively. The performance is evaluated for all FS
combinations based on accuracy, sensitivity, specificity of LDA and NB classifiers. Sensitivity, SE is defined as the rate of
overall number of true positives (TP) (correctly classified patterns) to the total number of actual positive patterns (TAp)

(20)

Specificity, SP is defined as the rate of total number of true negative (TN) to the total number of actual negative
patterns(TAn) :

(21)

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1141
Bearing Fault using LDA and NB Classifiers

The overall accuracy AC, is estimated as the percentage of rate of TP and TN to total number of patterns, N under
consideration for classification.

(22)

However, the overall accuracy contributed by every feature depends even on positive prediction value and negative
prediction value, that is, even if sensitivity is 1 or specificity is 1 accuracy is not necessary to be achieved as 100%.
3. RESULTS

The classification performance of a classifier is investigated with 3 parameters in this work as discussed in section
2.4. Hence,Figure.1 and figure.2 shows a plot of Accuracy, Sensitivity and specificity as a function of the FS number for a
four class classification of N, IR, OR, RE with a fault depth of 7,14 and21 using LDA and NB classifiers respectively for
group A. It is observed that the accuracy is 100% of the data sets I and II for all individual features except ZC and SSC.
The performance of these features does not improve even in combined form. In data set iii the accuracy is 40% to 80%
with the same features which had excelled in performance for data set I and ii whereas the features ZC and SSC have given
accuracy of 95-100%. However in data set v performances of all features are better than dataset iii, and it is noticed that the
feature WL, which had given the least accuracy in dataset iii exhibits maximum accuracy in this dataset. It is seen that as
the number of features grouped increases the performance also improves and the accuracy reaches 100%. It can be
witnessed that ZC and SSC are the features resulting in maximum accuracy when implemented in combination with any
other feature for datasets iii to vi. Table 2, presents the number of features required for attaining maximum accuracy for
every dataset of group A. It is interesting to note that both in LDA and NB classifications as seen in figure 1 and 2, the
sensitivity and specificity dips for the features ZC and SSC, both for individual and combined cases thus resulting in poor
accuracy levels in the classification. But ZC and SSC when combined with any other feature will excel in performance and
exhibit maximum accuracy. Therefore, it is important to analyze all parameters of a feature before either considering or
rejecting it for classification. The classification with NB as shown in figure 2,shows better results when compared to LDA
in figure 1, as the patterns do not get scattered.

Table 2: Number of Features Required to Attain Maximum Efficiency

for the Datasets of Group A.
Group A
Dataset DE-N&3F-7D DE-N&3F-14D DE-N&3F-21D
Load 4L L-0 L-1 L-2 L-3 4L L-0 L-1 L-2 L-3 4L L-0 L-1 L-2 L-3
Max AC 100 100 100 100 100 100 100 100 100 98 100 100 100 100 100
No of Features 1 1 1 1 1 3 3 1 2 3 3 2 2 1 1

Table 3: Number of Features Required to Attain Maximum Efficiency

for the Datasets of Group B
Group B
Dataset DE-N, 3F, 3 FD(7,14,21) DE-N, 3F, 2 FD(7,21)
Load 4L L-0 L-1 L-2 L-3 4L L-0 L-1 L-2 L-3
Max AC 91.4 97.2 98.3 90 92.3 100 100 100 100 100
No of Features 4 5 4 5 3 3 2 2 4 3

Figure 3, (i) presents the accuracy of classification for 4 datasets using LDA and NB. It is observed that for LDA
though many FSs exhibit 100% accuracy the average AC of NB is greater than LDA. The reason is, when compared to NB

www.tjprc.org [email protected]
1142 B R Nayana & P Geethanjali

sensitivity of LDA is less and specificity remains to be almost same with respect to each FS, thus the numbers of false
positive patterns are more with LDA classifier. Further, it is observed that datasets B-iii and B-iv exhibits 100% and
requires less number of features when compared to datasets B-i and B-ii as shown in table 3. This is for the known fact of
resonant vibration signals interfering with the vibration signals of FD 14, as discussed earlier for the observations made by
group A.

Figure 1: Accuracy, Sensitivity and Specificity as a Function of Feature

Set Number for a Four Class Classification of N, IR, OR and RE with
(a) FD-7, (b) FD-14 and (c) FD-21 Using LDA for Group A.

Figure 2: Accuracy, Sensitivity and Specificity as a Function of Feature

Set Number for a Four Class Classification of N, IRD, ORD, BD with
(a) FD-7, (b) FD-14 and (c) FD-21 Using NB for Group A.
In General, NB Is Better than LDA.

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1143
Bearing Fault using LDA and NB Classifiers

Figure 3: (i)Accuracy, (ii)Sensitivity and (Iii)Specificity as a Function of

Feature Set Number for a Four Class Classification of N, IR, OR, RE is
Dataseted Irrespective of Load, and with Respect to Each Load for
(a) FD-7,FD-14 and FD-21, and (b) FD-7 and FD-21 Using LDA,
(c) and (d) Using NB for Group B. In General, NB is Better than LDA.

In figure 3, (a) and (b) are for LDA and; (c) and (d) are for NB, which clearly shows LDA has scattered accuracy
patterns along with less sensitivity, whereas average specificity remains same for LDA and NB respectively. Figure 4,
shows accuracy, sensitivity and specificity for group C and it is observed that LDA performance is slightly better than NB.
In addition to this, it can be observed that for NB sensitivity does not show considerable improvement as number of
features combined increases like in other groups. However, its value is less in FSs where feature WL, WAMP, ZC exist
individually or in combination. Mainly when feature WL, WAMP and WL, ZC combination exists with any other feature
poor sensitivity and is attained. Conversely, specificity is good for all FSs when compared to LDA, as it improves with the
number of features combined and reaches maximum at the earliest. Therefore NB classifier performs better than LDA
except for load 0HP condition. Table 4 presents the number of features required to attain maximum accuracy for datasets
of group C. It is clear from the table that good accuracy ranges are obtained for dataset C-i, as it deals with DE data.
However, in datasetC-ii the accuracy ranges have not reached 100% with any FS, as it deals with FE data. In which, certain
data files are either non diagnosable or having electrical noises or the signal is clipped off as stated in the benchmark study
made by [23]. Therefore, it can be observed from figure 4 that the accuracy of classification is less for dataset C-ii of load
2HP condition and the same can be seen in figure 5, for dataset D-i of load 2HP condition. As C-ii and D-i of load 2HP
condition employs the same file which has more noise, the accuracy of classification is less. However, when the signals
were processed for complete 10 seconds instead of 5.4 seconds as considered in this work earlier, the accuracy levels are
considerably improved as seen in table 4 and 5 for the respective datasets. Table 5 presents the number of features required
to attain maximum accuracy for datasets of groups D and E. For the datasets of group E accuracy, sensitivity and
specificity are plotted figure 6, and the figure illustrates that among the existing 63 FS, the dips are formed due to the FSs

www.tjprc.org [email protected]
1144 B R Nayana & P Geethanjali

WL and WAMP. It can also be observed that though sensitivity and specificity for the feature SSC is 1 for all load
conditions with both classifiers the accuracy is not 100%. For the reason, that positive prediction and negative prediction
values are not sufficiently enough. Overall, for all datasets the FS of SSI, WL, SSC and ZC can give the maximum
accuracy. However, the FS of WL, SSC and ZC features is sufficient for most of the dataset to attain maximum accuracy
and for some datasets single feature is sufficient as seen in figure 7. Therefore the present approach is simple and
computationally cost effective.

Table 4: Number of Features Required to Attain Maximum

Efficiency for the Datasets of Group C
Group C
Dataset i-DE-3F-7D&21D ii-FE-3F-7D&21D
Load L-0 L-1 L-2 L-3 L-0 L-1 L-2 L-3
Max AC 97 100 100 100 95.05 94.01 90.1 97.4
No of Features 3 3 2 2 4 2 4 4

Table 5: Number of Features Required to attain Maximum

Efficiency for the Datasets of Group D & E
Group D E
Dataset i-DE&FE-3F-7D ii-DE&FE-3F-21D DE-N, 3F, 3 FD(7,14,21)
Load L-0 L-1 L-2 L-3 L-0 L-1 L-2 L-3 L-0 L-1 L-2 L-3
Max AC 99.5 97.14 93 100 99.74 98.7 98 99.5 97.3 91.95 91.83 98.3
No of Features 3 2 4 3 3 4 3 3 3 4 4 4

Figure 4: Average (i)Accuracy,(ii)Sensitivity, (iii)Specificity of Group C,

for a 6 Class Classification of all 3 Faults for DE with FD of 7 and 21
(a) and for FE (b) Using LDA, Similarly (c) and (d) Using NB.
In General, NB Is Better than LDA

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1145
Bearing Fault using LDA and NB Classifiers

Figure 5: Average Accuracy, Sensitivity and Specificity of Group D,

for a 6 Class Classification of all 3 Faults for DE and FE with FD of 7
(a) and 21(b) Using LDA, Correspondingly (c) and (d) Using NB.
Overall, LDA Performed Better than NB

Figure 6: Accuracy of a 10 Class Classification, of DE with all

3 Faults and all 3 FDs Including Normal Working
Condition of Group E, Using LDA and NB

Figure 7: Number of Features Required for each Dataset to Attain

Maximum Accuracy with Respect to Load Conditions

www.tjprc.org [email protected]
1146 B R Nayana & P Geethanjali

4. DISCUSSIONS

The statistical time domain features MAV, SSI, WL, WAMP, ZC, SSC is introduced for bearing fault diagnosis
using LDA and NB classifiers. Even though some studies have been performed [3], where time domain features are applied
but the feasibility of above discussed features for bearing fault diagnosis has not been investigated so far. Our findings with
the datasets derived with a new set of features are in agreement with the previous studies. The investigations performed
disclose the fact that features considered will perform well for the combination of features associated with time and
frequency. To be precise, the feature SSC is associated with time and the feature ZC is associated with the frequency of a
signal, when these 2 features are employed together then it leads to maximum accuracy. However, the parameter
sensitivity, specificity, positive prediction and negative prediction values and others to be considered before choosing the
combination of features to develop the scheme for automated bearing fault diagnosis, which can provide best
discrimination with less computation time. It is perceived that features ZC and SSC exist as main features in determining
the maximum accuracy for almost all the data sets discussed above. Either feature ZC and SSC together or ZC and SSC
along with WL will give maximum accuracy. But, overall4 features are sufficient for the authors to get maximum accuracy
in all datasets of groups A to E. Though in dataset B for certain load conditions 5 features are providing maximum
accuracy, it is not exhibiting considerable improvement. Therefore the FS of SSI, WL, SSC and ZC features will provide
maximum accuracy for all datasets of group A-E. The application of these features for bearing fault diagnosis has a number
of benefits over other methods proposed so far. Such as, the features selected are one dimensional simple and fast to
estimate and also they are less in number. This avoids the need for feature selection and reduction processes. Secondly,
classifications can be implemented by simple classifiers like LDA and NB classifiers; hence diagnosis can be implemented
with fewer computations. Altogether these factors constitutes that the proposed method is highly appropriate for real-time
analysis. Another important advantage commonly believe is, these features provide same information as time, frequency
and time-frequency analysis of the signals as implemented by[2-4][8-9].

For the comparison between results obtained from the proposed method and the existing methods in literature,
only the works which have used identical dataset are considered. However, the differences exist with respect to the
vibration data considered is 12kHz or 48kHz and the number of channel inputs taken into consideration. Few authors have
chosen single channel input, either DE or FE data like in present work for diagnosis and some authors consider both DE
and FE data for every working condition. The accuracy obtained from the proposed method gives the best accuracy for
group A, and it is equivalent to the best presented. This is in corroboration with earlier reporting’s[18, 21]. Specifically
datasets A-II, which is a four class classification for FD of 7 being, irrespective of load is implemented by authors in [21]
excluding load 0HP condition. Further WPD for feature extraction and mRMR for feature selection and DE-EAM for
classification are employed and average accuracy of 96.1% is obtained. Similarly spectrum imaging and feature
enhancement is applied for feature extraction, and classification is realized using ANN by the authors in [22] for dataset A-
i with 2HP load and obtains an accuracy of 96.9%. Data setA-iii is assessed for load 0, 1 and 2HP using 20 features by
authors in [18] and attained an accuracy of 99.56%,100%,99.89% respectively. But100% accuracy is obtained in the
present work for all datasets of group A by using a maximum of 3 features. For group B, identical implementation of
dataset B-i is performed by authors in [17] and in [19]. In the former work the authors would generate feature vectors using
characteristics of the feature ZC and classify the fault using Feed forward NN and results indicates for 20 ZC intervals
accuracy obtained is 92.45% using a window length of 1024. In the later work, the authors have made use of NPE, SOM
along with many classifiers including LDA. Further, they have extracted10 time domain and 10 frequency domain features

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1147
Bearing Fault using LDA and NB Classifiers

to realize using classifiers and have attained an accuracy of 96% without denoising and 98% after denoising the data for
LDA classifier respectively. For B-iii dataset authors of [20] have implemented by fuzzy inference methodology and by
excluding normal working condition besides obtains a maximum accuracy of 73%. But in the present work, 100%
accuracy is obtained either when implemented irrespective of load or with respect to each load, as shown in table 2 for B-
iii and B-iv datasets. However, though accuracy does not excel for all cases, they are in substantiation with earlier works.

The comparison of datasets C and D group of present work and identical implementation in[13]can be discussed
for every load condition. The authors have extracted 5 IMFs by EEMD and classified with SVM classifier It is seen from
table 6that the present work has performed better for datasets C-i and D-ii. But for datasets of C-ii and D-i are in par with
earlier work with a difference of 0% to 2% except for the cases in which electrical noise is assumed to be present as
discussed earlier. However, in the present work the window length is 1024 and 5.4 seconds of data are processed for every
working condition, whereas in [13] the window length is 3000 and 10seconds of complete data is processed.

Table 6: Comparison of Present Work and X. Zhang et el [13]

2013 for Groups C and D
load/ X. Zhang et al[13] Present Work
exp C-i C-ii D-i D-ii C-i C-ii D-i D-ii
L-0 96.81 96.85 100 98.17 96.88 95.05 99.48 99.74
L-1 97.04 95.37 100 98.89 99.74 93.75 97.14 98.89
L-2 99.33 98.81 100 98.81 100 90.88 92.96 98.8
L-3 99.7 99.83 100 98.65 100 96.88 100 99.48

Group E is for a 10 class classification of all 3F of 3 FDs and N for each load condition. Although the authors in
[2] have implemented a 10 class classification by LDA same as group E for a load of 3HP, using 9 time domain features
and 5 time-frequency features for the vibration data of 48kHZ and have developed TR-LDA1 and TR-LDA2 algorithms to
achieve 100% classification accuracy in classification and also has presented a comparison by implementing with different
classifiers. It is observed LDA exhibits 98% of accuracy, and the same is achieved in present work using 4 features by
LDA and NB classifiers. A similar implementation is performed by former authors for load conditions of 1 and 2 HP in
[14] with the aid of K-means clustering and the implementations are identical to dataset E L-2 and E L-3 in present work.
Correspondingly the authors have obtained 98.5% and 98.9% It is interesting to note that in the present work combination
of four features are sufficient to obtain an AC of 92% as given in table 5, which once again illustrates that a combination of
few good features are sufficient rather than a complicated algorithm for diagnosing the bearing faults, with minimum
computational cost.

5. CONCLUSIONS

Fault diagnosis is a crucial part of condition monitoring of bearings to avoid unprepared repairs and cost-effective
damages caused by failures. The condition monitoring scheme involves suitable feature extraction, feature reduction,
feature selection and classification processes, among which feature extraction and classification play important role in the
scheme. Once the feature extracted are effective enough to reveal all the characteristics of fault condition then feature
reduction and feature selection processes can be evaded.

In this paper, for the first time statistical time domain features MAV, SSI, WL, WAMP, ZC, SSC is employed for
the identification of the mechanical faults using LDA and NB classifiers. In this study the effectiveness of each feature is
investigated pertaining to accuracy, sensitivity and specificity in 63 combinations for 15 datasets which are drawn from 5

www.tjprc.org [email protected]
1148 B R Nayana & P Geethanjali

groups. The FS of SSI, WL, SSC and ZC will contribute for maximum accuracy in all the cases. However, for many
datasets FS of one, two and three features are giving better accuracy in which the features are WL, SSC and ZC. It is
studied from the results that increasing the number of features will not contribute to improve the classification accuracy;
instead if a feature represent time characteristics of a fault condition and another feature for frequency then their
combinations are giving best accuracy results, conversely combining many features which are redundant in characteristics
will not contribute much to improve classification accuracy therefore investigations are to be conducted to find the best
feature combination and then employ classification with minimal number of features which reduces the overheads of
dimensionality reduction schemes like feature selection and feature reduction. In addition, to this the present approach
avoids the computational burden on classifiers. The low computational complexity of these features constitutes it a highly
favorable feature to be employed as part of a system for real-time automated fault diagnosis schemes. The success of the
present approach is verified through comparing the performance of classification problems from other researchers. It can be
concluded that the features employed newly with NB classifier achieves more satisfactory results to discriminate the fault
condition from vibration signal than the other methods do. While our system can achieve promising results for handling the
fault diagnosis of roller bearings, our future work might focus on the following issues to improve the present approach (1)
diagnosing the fault severities of each fault individually and even when present in combinations. (2) Diagnosing and
investigating OR fault for different load zone conditions and to localize the OR faults. (3) Finally, to indicate the severity
of fault and by defining certain fault level indicators, this aids in bearing performance prognostics in future.

REFERENCES

1. Zhang P, Du Y, Habetler TG, Lu B (2011) A Survey of Condition Monitoring and Protection Methods for Medium-Voltage
Induction Motors. IEEE Transactions on Industry Applications 47:34–46. doi: 10.1109/TIA.2010.2090839

2. Jin X, Zhao M, Chow TWS, Pecht M (2014) Motor Bearing Fault Diagnosis Using Trace Ratio Linear Discriminant Analysis.
IEEE Transactions on Industrial Electronics 61:2441–2451. doi: 10.1109/TIE.2013.2273471

3. Prieto MD, Cirrincione G, Espinosa AG, et al (2013) Bearing Fault Detection by a Novel Condition-Monitoring Scheme
Based on Statistical-Time Features and Neural Networks. IEEE Transactions on Industrial Electronics 60:3398–3407. doi:
10.1109/TIE.2012.2219838

4. Chebil J, Hrairi M, Abushikhah N (2011) Signal analysis of vibration measurements for condition monitoring of bearings.
Australian Journal of Basic and Applied Sciences 5:70–78.

5. Bouchikhi E, Houssin E, Choqueuse V, et al (2012) Induction machine fault detection enhancement using a stator current high
resolution spectrum. In: IECON 2012-38th Annual Conference on IEEE Industrial Electronics Society. IEEE, pp 3913–3918

6. Garcia-Perez A, Romero-Troncoso R de J, Cabal-Yepez E, Osornio-Rios RA (2011) The Application of High-Resolution

Spectral Analysis for Identifying Multiple Combined Faults in Induction Motors. IEEE Transactions on Industrial Electronics
58:2002–2010. doi: 10.1109/TIE.2010.2051398

7. Blodt M, Chabert M, Regnier J, Faucher J (2006) Mechanical Load Fault Detection in Induction Motors by Stator Current
Time-Frequency Analysis. IEEE Transactions on Industry Applications 42:1454–1463. doi: 10.1109/TIA.2006.882631

8. Peng ZK, Chu FL (2004) Application of the wavelet transform in machine condition monitoring and fault diagnostics: a
review with bibliography. Mechanical Systems and Signal Processing 18:199–221. doi: 10.1016/S0888-3270(03)00075-X

9. Jawadekar A, Paraskar S, Jadhav S, Dhole G (2014) Artificial neural network-based induction motor fault classifier using
continuous wavelet transform. Systems Science & Control Engineering 2:684–690. doi: 10.1080/21642583.2014.956266

Impact Factor (JCC): 6.8765 NAAS Rating: 3.11

Effective Time Domain Features for Identification of 1149
Bearing Fault using LDA and NB Classifiers

10. Lou X, Loparo KA (2004) Bearing fault diagnosis based on wavelet transform and fuzzy inference. Mechanical Systems and
Signal Processing 18:1077–1095. doi: 10.1016/S0888-3270(03)00077-3

11. Sugato Ghosh, Inertia Effect Under Couple Stress Fluid in Laminar Flow on Porous Journal Bearing, International Journal of
Mechanical and Production Engineering Research and Development (IJMPERD), Volume 5, Issue 5, September - October
2015, pp. 1-12

12. Schmitt HL, Silva LRB, Scalassara PR, Goedtel A CLASSIFICATION OF SIMULATED MOTOR SIGNALS WITH BEARING
FAULTS USING NEURAL NETWORKS, WAVELETS, AND PREDICTABILITY MEASURES.

13. Seshadrinath J, Singh B, Panigrahi BK (2014) Investigation of Vibration Signatures for Multiple Fault Diagnosis in Variable
Frequency Drives Using Complex Wavelets. IEEE Transactions on Power Electronics 29:936–945. doi:
10.1109/TPEL.2013.2257869

14. Zhang X, Zhou J (2013) Multi-fault diagnosis for rolling element bearings based on ensemble empirical mode decomposition
and optimized support vector machines. Mechanical Systems and Signal Processing 41:127–140. doi:
10.1016/j.ymssp.2013.07.006

15. Zhao M, Jin X, Zhang Z, Li B (2014) Fault diagnosis of rolling element bearings via discriminative subspace learning:
Visualization and classification. Expert Systems with Applications 41:3391–3401. doi: 10.1016/j.eswa.2013.11.026

16. Zhang L, Xiong G, Liu H, et al (2010) Bearing fault diagnosis using multi-scale entropy and adaptive neuro-fuzzy inference.
Expert Systems with Applications 37:6077–6085. doi: 10.1016/j.eswa.2010.02.118

17. Samanta B, Al-Balushi KR, Al-Araimi SA (2003) Artificial neural networks and support vector machines with genetic
algorithm for bearing fault detection. Engineering Applications of Artificial Intelligence 16:657–665. doi:
10.1016/j.engappai.2003.09.006

18. William PE, Hoffman MW (2011) Identification of bearing faults using time domain zero-crossings. Mechanical Systems and
Signal Processing 25:3078–3088. doi: 10.1016/j.ymssp.2011.06.001

19. Wu S-D, Wu P-H, Wu C-W, et al (2012) Bearing Fault Diagnosis Based on Multiscale Permutation Entropy and Support
Vector Machine. Entropy 14:1343–1356. doi: 10.3390/e14081343

20. Zhang S, Li W (2014) Bearing Condition Recognition and Degradation Assessment under Varying Running Conditions Using
NPE and SOM. Mathematical Problems in Engineering 2014:1–10. doi: 10.1155/2014/781583

21. Raj AS, Murali N (2013) Early Classification of Bearing Faults Using Morphological Operators and Fuzzy Inference. IEEE
Transactions on Industrial Electronics 60:567–574. doi: 10.1109/TIE.2012.2188259

22. Liu C, Wang G, Xie Q, Zhang Y (2014) Vibration Sensor-Based Bearing Fault Diagnosis Using Ellipsoid-ARTMAP and
Differential Evolution Algorithms. Sensors 14:10598–10618. doi: 10.3390/s140610598

23. Amar M, Gondal I, Wilson C (2015) Vibration Spectrum Imaging: A Novel Bearing Fault Classification Approach. IEEE
Transactions on Industrial Electronics 62:494–502. doi: 10.1109/TIE.2014.2327555

24. Smith WA, Randall RB (2015) Rolling element bearing diagnostics using the Case Western Reserve University data: A
benchmark study. Mechanical Systems and Signal Processing 64-65:100–131. doi: 10.1016/j.ymssp.2015.04.021.

25. Case Western Reserve University Bearing Data Center Website. Available online:
http://csegroups.case.edu/bearingdatacenter/pages/download-data-file/ (accessed on 23rd March 2015.

www.tjprc.org [email protected]