Mechanical Systems and Signal Processing 50-51 (2015) 414–426

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Fault diagnosis of rotating machinery with a novel statistical

feature extraction and evaluation method
Wei Li n, Zhencai Zhu, Fan Jiang, Gongbo Zhou, Guoan Chen
School of Mechanical and Electrical Engineering, China University of Mining and Technology, Xuzhou 221116, PR China

a r t i c l e in f o abstract

Article history: Fault diagnosis of rotating machinery is receiving more and more attentions. Vibration
Received 31 January 2013 signals of rotating machinery are commonly analyzed to extract features of faults, and the
Received in revised form features are identified with classifiers, e.g. artificial neural networks (ANNs) and support
20 April 2014
vector machines (SVMs). Due to nonlinear behaviors and unknown noises in machinery,
Accepted 27 May 2014
the extracted features are varying from sample to sample, which may result in false
Available online 16 June 2014
classifications. It is also difficult to analytically ensure the accuracy of fault diagnosis.
Keywords: In this paper, a feature extraction and evaluation method is proposed for fault diagnosis of
Fault diagnosis rotating machinery. Based on the central limit theory, an extraction procedure is given to
Feature extraction
obtain the statistical features with the help of existing signal processing tools. The
Fault classification
obtained statistical features approximately obey normal distributions. They can signifi-
Rotating machinery
cantly improve the performance of fault classification, and it is verified by taking ANN and
SVM classifiers as examples. Then the statistical features are evaluated with a decoupling
technique and compared with thresholds to make the decision on fault classification.
The proposed evaluation method only requires simple algebraic computation, and the
accuracy of fault classification can be analytically guaranteed in terms of the so-called
false classification rate (FCR). An experiment is carried out to verify the effectiveness of
the proposed method, where the unbalanced fault of rotor, inner race fault, outer race
fault and ball fault of bearings are considered.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction

Rotating machinery is widely used in many industrial fields. Usually vibration signals are used to detect the faults of the
machine components and reduce the catastrophic damage and the down-time of machinery by applying fault diagnosis
methods [1–4]. Generally speaking, there are two main steps in fault diagnosis: the first step is the feature extraction of
vibration signals with some signal processing tools, and the second step is the fault classification based on the extracted
features in the previous step.
How to extract the features from vibration signals is one of the key problems in the fault diagnosis of rotating machinery.
Since vibration signals usually have non-stationary and nonlinear behaviors due to the complexity of the structure and work
conditions of rotating machinery, it is difficult to achieve an effective fault diagnosis only with the time domain or the
frequency domain analysis. Hence features are extracted to indicate the characteristics of vibration signals. The conventional
feature extraction methods include time-domain methods, frequency-domain methods, and time-frequency methods [1].

n
Corresponding author. Tel.: þ86 51683590777; fax: þ 86 51683590708.
E-mail address: liwei_cmee@163.com (W. Li).

http://dx.doi.org/10.1016/j.ymssp.2014.05.034
0888-3270/& 2014 Elsevier Ltd. All rights reserved.
 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426 415

Time-domain methods are directly based on the time waveform, e.g. peak amplitude, root-mean-square amplitude,
variance, skewness, kurtosis, correlation dimension and fractal dimension [5–8]. Most of those features are statistics of the
time waveform. Frequency-domain methods are based on the transformed signal in frequency domain, i.e. Fourier spectrum,
cepstrum analysis, and envelope spectrum [9–11]. Wavelet analysis, short time Fourier transform, Wigner–Ville distribution,
Choi–Williams distribution and Hilbert–Huang transform are the time-frequency methods [12–17], which investigate
waveform signals in both time and frequency domains.
Many studies of fault classification were also carried out [1]. Fault classification based on artificial neural network (ANN)
techniques and support vector machine (SVM) techniques are popular in literatures for fault diagnosis of rotating
machinery. ANN was used as a fault classifier for both binary and multi-class fault classifications [18–21]. ANN requires
extensive training data and training time. It is difficult to be analytically interpreted, and it may not converge in the global
optima. An alternative to ANN is SVM, which is based on the statistical learning theory [22]. The solution to an SVM is the
classification hyperplane, which is well interpreted and unique. SVM has recently been applied to many practical problems
[23–27]. In order to obtain the training samples of features required by ANNs or SVMs, a number of vibration signals under
the same condition were processed to extract the features, or a given vibration signal was equally divided into many
partitions and then certain signal processing method was applied to extract the features of each partition. The length of
signals or partitions was determined by the applied signal processing method. The variation of extracted features can
significantly influence the accuracy of fault classification with ANN and SVM [2]. In [28], a decoupling technique has been
proposed to classify the faults based on simple algebraic computations without a training procedure, where feature vectors
were projected onto a set of orthogonal directions in order to indicate the fault information. There were also other fault
classification methods such as fuzzy inference method [29], genetic algorithm [30] and hidden Markov model method [31].
The effectiveness of a fault diagnosis method is indicated by the classification accuracy and the computation time. Hence
finding an accurate and fast method for fault diagnosis is an essential issue [23]. However ANN and SVM based fault
classification methods do not give any analytical guarantee on the accuracy. The classification accuracy of ANN and SVM is
usually estimated through experiments.
In this paper, a feature extraction and evaluation method is proposed for fault diagnosis of rotating machinery. Based on
the central limit theory, an extraction procedure is presented to compute statistical features with the help of existing signal
processing tools. Such statistical features are close to normal distributions. The raw features which are directly computed
with conventional feature extraction methods are compared with the statistical ones by using them as the inputs for ANN
and SVM based fault classifiers. Then the statistical features are evaluated with a decoupling technique, and a decision logic
is also given to indicate the classification results by making use of the means and variances of statistical features.
The proposed evaluation method only requires simple algebraic computation, and the accuracy of fault classification is
analytically guaranteed. The so-called false classification rate (FCR) can be ensured to be smaller than any given value. An
experiment is carried out to verify the effectiveness of the proposed method, where the unbalanced fault of a rotor, inner
race fault, outer race fault and ball fault of bearings are considered.
The rest of the paper is organized as follows. Section 2 introduces the experimental system. In Section 3, the statistical
feature extraction is proposed. The statistical feature evaluation is given in Section 4. In Section 5, the experimental results
are used to demonstrate the effectiveness of the proposed method. Finally, concluding remarks are given in Section 6.

2. Experimental system

The experiment is carried out with a machinery fault simulator as shown in Fig. 1, which contains a variable speed motor,
a variable speed motor controller, a flexible coupling, a shaft, a rotor disk with tapped holes in the edge and two rolling
element bearings. One of the bearings without defects is located in the bearing housing closer to the motor. The other
bearing is located in the bearing housing farther to the motor, and it could be replaced by the testing bearings with three
kinds of faults, i.e. defect in the inner race, defect in the outer race, and defect in a ball. An ICP accelerometer with a
bandwidth up to 10 kHz and a 0.1 V/g output is mounted on the right bearing housing, and the vibration signals from the
accelerometer are used to diagnose faults. The running speed of this simulator is set at 1198 RPM and driven by a 3-hp AC
motor. The sampling frequency is 2 kHz, and the sampling time is 10 s.
In order to develop the new fault diagnosis method, five sets of vibration data were obtained from the experimental
system under five health conditions: (i) under the normal condition; (ii) with the unbalanced fault of rotor; (iii) with the
inner race fault of a bearing; (iv) with the outer race fault of a bearing; and (v) with a ball fault of a bearing. The bearings
under different health conditions are replaced in the right bearing housing in order to simulate the faulty behaviors. The
rotor is unbalanced with bolts screwed in the tapped holes. Each set of data contained vibration signals of 10 s, and it was
divided into two parts: design data and test data. The design data is used to design the fault diagnosis system, while the test
data is used to testify the designed system. Fig. 2 shows parts of the vibration signals under different health conditions.
Given that a discrete time vibration signal xj is collected through the accelerometer under the j-th health condition of
rotating machinery. Due to the limits of computation capacity, only a partition of xj can be processed by computers at once.
It is well known that vibration signals may with inherent non-stationary characteristics and may also interfered by
unknown noises. Hence the raw features extracted from different partitions of xj could be varying. Actually it is also difficult
to determine the distributions of raw features. This is the reason why ANN and SVM techniques are employed to realize fault
classification in literatures, where a number of samples of features are used to train ANN or SVM. However such fault
416 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426

Fig. 1. Experimental setup.

normal condition
0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1.0

with unbalanced fault of rotor

0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1.0
Amplitude

with inner race fault of a bearing

0.1
0
−0.1
0 0.2 0.4 0.6 0.8 1.0

with outer race fault of a bearing

0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1.0

with a ball fault of a bearing

1
0
−1
0 0.2 0.4 0.6 0.8 1.0

Time(s)
Fig. 2. Vibration signals under different health conditions.

diagnosis methods based on ANN or SVM may still have some classification errors, and their accuracies can only be
estimated through experiments. Hence the so-called statistical features will be extracted from xj and then evaluated in order
to realize a fault diagnosis with an improved performance and a guaranteed accuracy.

3. Extraction of statistical features

Assume that p 1 types of faults are going to be classified in rotating machinery. Besides, the normal condition should
also be classified. Therefore the considered health conditions of rotating machinery are of p types. The number of raw
features extracted from vibration signals is assumed to be q, and these raw features compose a raw feature vector
F j ¼ ½f j ð1Þ f j ð2Þ ⋯ f j ðqÞT

where fj(i) stands for the i-th raw features under the j-th health condition.
Without loss of generality, it is assumed that the length of xj under the j-th health condition is L which is sufficiently
large, i.e.
xj ¼ ½xj ð1Þ; xj ð2Þ; …; xj ðLÞ;
 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426 417

such that the distribution of raw features under the j-th health condition remain unchanged. Then a set of consecutive data
is arbitrarily selected from xj and denoted as
Dj ¼ ½xj ðnÞ; xj ðn þ1Þ; xj ðn þ 2Þ; …; xj ðn þ k  1Þ
where n is arbitrarily selected and 1 r n r L k þ1, k o L. With a fixed k, the number of possible selections of Dj is L k þ 1.
A sample of the raw feature vector could be obtained based on Dj by applying some existing feature extraction methods in
literatures. The parameter k determines the length of Dj, which could be chosen according to the applied raw feature
extraction method, the rotating speed of machinery and the computation power of computers. By arbitrarily selecting M sets
of Dj, a sequence of raw feature vectors under the j-th health condition can be obtained and the sequence is denoted as
fF gj g .
g ¼ 1;2;…;M

Since M sets of Dj are all arbitrarily selected in xj, fF gj g is a sequence of independent and identically distributed
g ¼ 1;2;…;M
random vectors with the same element-wise mean and variance, which are defined as
μ j ¼ EðF j Þ ¼ ½Eðf j ð1ÞÞ Eðf j ð2ÞÞ ⋯ Eðf j ðqÞÞT ð1Þ

and
s 2j ¼ VarðF j Þ ¼ ½Varðf j ð1ÞÞ Varðf j ð2ÞÞ ⋯ Varðf j ðqÞÞT : ð2Þ

According to the central limit theorem [33], when M approaches infinity, the random vector
1 M g
Sj ¼ ∑ F
Mg¼1 j

element-wisely converge in a normal distribution and

E Sj ¼ μ j ;

1
Var Sj ¼ s 2j :
M
In fact Sj is the sample average of fF gj g . Once xj is given, μ j and s 2j can be easily estimated, and then the mean and the
g ¼ 1;2;…;M
2
variance of Sj can be determined. For convenience the estimated mean and variance of Sj are denoted as μj and sj respectively.
With a sufficiently large M, Sj is defined as the statistical feature vector of vibration signals under the j-th health
condition. By repeating the procedure, a group of statistical feature vectors can be obtained, which approximately obey the
2
normal distribution with μj and sj .
In Figs. 3 and 4, 100 samples of the 1st raw features and the 1st statistical features under the normal condition, i.e. j¼1, are
shown in the same scale, respectively. The samples of raw features are computed based on 100 arbitrarily selected D1 from x1.
Each sample of the statistical features is computed based on 100 arbitrarily selected D1 from x1 and therefore totally 100  100
sets of D1 are employed. It is clear that the variance of statistical features is much smaller than that of raw features.
In Figs. 5 and 6, histograms of those samples using 15 bins are plotted and fitted normal density functions. It can be
observed that the statistical features are much closer to a normal distribution than the raw features. Based on those samples
the Lilliefors test [34] was also carried out, which confirmed that the statistical features come from a normal distribution
with a higher probability than the raw features (the p-value in case of the statistical features was 0.7610 at the 5%
significance level, while the p-value in case of the raw features was 0.3630).
The main procedure of the statistical feature extraction is summarized as follows:

 Step 1: arbitrarily select a set of consecutive data with the length k from a given vibration signal xj under the j-th health
condition;
 Step 2: compute the raw feature vector of the selected consecutive data with existing feature extraction methods;
 Step 3: repeat Step 1 and Step 2 for M times such that M samples of the raw feature vectors are obtained; and

0.25
Amplitude (× 10−6)

0.2

0.15

0.1
0 10 20 30 40 50 60 70 80 90 100
Sample No.

Fig. 3. The 1st raw features under the normal condition: energy levels around fr.
418 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426

0.25

Amplitude (×10−6)
0.2

0.15

0.1
0 10 20 30 40 50 60 70 80 90 100
Sample No.

Fig. 4. The 1st statistical features under the normal condition: sampled average of energy levels around fr, M ¼100, k ¼2048.

14

12
The number of samples

10

0
0.12 0.14 0.16 0.18 0.2 0.22 0.24
Amplitude (×10−6)

Fig. 5. The distribution of the 1st raw feature under the normal condition, which is compared with a fitted normal distribution.

16

14

12
The number of samples

10

0
0.166 0.168 0.17 0.172 0.174 0.176 0.178 0.18
Amplitude (×10−6)

Fig. 6. The distribution of the 1st statistical feature under the normal condition, which is compared with a fitted normal distribution.

 Step 4: compute Sj, i.e. the sampled average of the raw feature vectors, and define it as the statistical feature vector under
the j-th health condition.

In the next section, the evaluation method of statistical features is proposed to realized fault classification by using the
2
knowledge of μj and sj , where the accuracy of fault classification is ensured.

4. Evaluation of statistical features

The following statistical feature matrix of rotating machinery is defined for the purpose of fault classification:
μF ¼ ½μ1 μ2 ⋯ μp  A Rqp ð3Þ

where μj ; j ¼ 1; 2; …; p, is the estimated mean of Sj under the j-th health condition.

Define
μout ðjÞ ¼ ½μ1 ⋯ μj  1 μj þ 1 ⋯ μp :
 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426 419

A left null space Zj of μout ðjÞ is defined and computed by

Z j μout ðjÞ ¼ 0: ð4Þ
A row vector Hj from Zj is selected for the j-th health condition, which satisfies
H j μj a 0 ð5Þ

and
H j ¼ ½H j ð1Þ H j ð2Þ ⋯ H j ðqÞ:
Hence the vector Hj can distinguish μj from other ones. Hj is also called the decoupling vector for the j-th health condition
[28]. Such a decoupling vector exists if and only if
rankðμout ðjÞÞ orankðμF Þ:
Besides, decoupling vectors exist for all health conditions if and only if
rankðμF Þ Z p:
Once the decoupling vectors are obtained for each health condition, they can be used to classify the health conditions.
In the following Stest is used to denote the statistical feature vector of vibration signals under an unknown health condition,
which are going to be classified. Then the following indexes are used to indicate the health condition:
r j ¼ H j Stest ; j ¼ 1; 2; …; p: ð6Þ
In fact the number of indexes is equal to that of health conditions.
2
If the rotating machinery is under the i-th health condition, Stest should be close to the normal distribution with μi and si .
Then Eq. (6) can be rewritten as
r j ¼ H j ðμi þ Δi Þ ð7Þ
2
where Δi is unknown but Δi approximately obeys the normal distribution with zero mean and Hence rj is a summation of si .
random variables approximately obeying normal distributions [33]. When i¼j, rj approximates a normal distribution whose
mean is H j μj as
r j ¼ H j μj þH j Δj :

When ia j, rj approximates a normal distribution whose mean is 0 as

r j ¼ H j μi þH j Δi ¼ H j Δi :
Since rj is a summation of random variables approximately obeying normal distributions, the variance of rj under the i-th
health condition is also the summation of the variances of those random variables [33]. The variance of rj under the i-th
health condition is denoted as s2rji .
Then the following logic rule is applied to classify the health conditions of rotating machinery:
if jr j j A ½0; J j  then C j ¼ 0
if jr j j2
= ½0; J j  then C j ¼ 1 ð8Þ

where
J j ¼ βsj;max ð9Þ

is the threshold with β being a positive constant, and Cj ¼0 means the machinery is not under the j-th health condition,
while Cj ¼1 means the machinery is under the j-th health condition. The parameter sj;max is defined as
sj;max ¼ maxfsrji g;
iaj

which is the largest variance of rj with i aj.

The main procedure of the statistical feature evaluation is summarized as follows:

 Step 1: compute the decoupling vector Hj for the j-th health condition according to (3)–(5);
 Step 2: compute the index rj according to (7); and
 Step 3: classify the health condition according to (8), where Jj is determined with (9).

The false classification rate (FCR) is defined to evaluate the performance of fault classification.

Definition. The FCR is the probability that the rotating machinery is not under the j-th health condition, but jr j j=
2 ½0; J j .

Since rj is close to a normal distribution and sj;max is the largest possible variance of rj (except under the j-th health
condition), according to the properties of normal distribution [33], the FCR can be ensured to be lower than a given value
with (8) and (9) by selecting a proper β. Hence an analytical bound of FCR is established with the knowledge of the
distributions of the statistical feature vectors.
420 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426

For example, set β¼3. When the machinery is under the i-th health condition and i aj, the probability that
rj 2
= ½0; 3sj;max 
is less than 0.3%. Therefore
FCR r 0:3%:
The proposed method can guarantee the accuracy of fault diagnosis in terms of FCR by using the information about
means and variances of the statistical feature vectors. It is also of practical interests to have the knowledge of the diagnosis
accuracy. Besides, the proposed evaluation method only requires simple algebraic computations, and no training procedure
is involved.

5. Experimental results

In order to demonstrate the effectiveness of the proposed method, the typical radial basis function (RBF) ANN and SVM
classifiers are also employed to classify the health conditions of the experimental system. Firstly, the raw features and the
statistical features are extracted and used to train and test the RBF ANN and SVM separately, and the performances of ANN
and SVM with different features are compared. Secondly, the statistical features are evaluated according to the procedure in
Section 4.
Many feature extraction methods can be directly applied to obtain some raw features [1], i.e. extracting features in time-
domain, extracting features in frequency-domain and extracting features in time–frequency-domain. The raw features could
be energy level, entropy, etc. In the experiment, firstly the discrete time Fourier transformation is used to calculate the
energy levels around harmonic frequencies, i.e. f r ; 2f r ; 3f r ; …; 10f r with fr being the rotating frequency. These energy levels
are simply treated as the raw features of vibration signals under different health conditions. In the experimental system, the
number of considered health conditions is 5 and the number of raw features is 10, i.e. p¼ 5 and q ¼10.

5.1. Fault diagnosis with ANN and SVM

The RBF ANN is used for classification, function approximation, time series prediction and system control [1]. The RBF
ANN has three layers: an input layer, a hidden layer of neurons with a non-linear RBF activation function and a linear output

1st Output
2
1
0
0 30 60 90 120 150

2nd Output
2

0
Outputs of RBF network

0 30 60 90 120 150
3rd Output
2

0
0 30 60 90 120 150

2 4th Output

0
0 30 60 90 120 150
2 5th Output

0
0 30 60 90 120 150
Sample No.
Fig. 7. Fault diagnosis result with raw features and RBF ANN. Samples from No. 1 to No. 30 are under the normal condition. Samples from No. 31 to No. 60
are with the unbalanced fault of rotor. Samples from No. 61 to No. 90 are with the inner race fault. Samples from No. 91 to No. 120 are with the outer race
fault. Samples from No. 121 to No. 150 are with a ball fault. The circled samples are misclassified.
 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426 421

layer. The output of the network is a linear combination of radial basis functions of the inputs and neuron parameters. In the
experiment, the RBF ANN is designed for classifying the health conditions. Its inputs are the samples of feature vectors.
The number of the outputs of the RBF ANN is five. Each output indicates the occurrence of the corresponding health
condition, i.e. the system is classified as the j-th condition when the j-th output of the ANN is around 1; and the system is
not classified as the j-th condition when the j-th output of the ANN is around 0.
The SVM constructs a hyperplane or a set of hyperplanes in a high dimensional space, which can be used for
classification, regression, or other tasks [2]. A good separation of different classes is achieved by the hyperplane that has the

4
Output Values

0
0 30 60 90 120 150
Sample No.

Fig. 8. Fault diagnosis result with raw features and SVM. Samples from No. 1 to No. 30 are under the normal condition. Samples from No. 31 to No. 60 are
with the unbalanced fault of rotor. Samples from No. 61 to No. 90 are with the inner race fault. Samples from No. 91 to No. 120 are with the outer race fault.
Samples from No. 121 to No. 150 are with a ball fault. The circled samples are misclassified.

1st Output
2

0
0 30 60 90 120 150
2nd Output
2

1
Outputs of RBF network

0
0 30 60 90 120 150

3rd Output
2

0
0 30 60 90 120 150
4th Output
2
1
0
0 30 60 90 120 150
5th Output
2

0
0 30 60 90 120 150
Sample No.
Fig. 9. Fault diagnosis result with statistical features and RBF ANN. Samples from No. 1 to No. 30 are under the normal condition. Samples from No. 31 to
No. 60 are with unbalanced fault of rotor. Samples from No. 61 to No. 90 are with inner race fault. Samples from No. 91 to No. 120 are with outer race fault.
Samples from No. 121 to No. 150 are with a ball fault.
422 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426

largest distance to the nearest training data point of any class. The SVM classifier for the experimental system is designed
with the LIBSVM programm [35] to classify the health conditions of the experimental system. Its inputs are the samples of
feature vectors, and its output directly indicates the index of the estimated health condition, i.e. the system is classified as
the j-th condition when the output of the SVM is j.
The RBF ANN and the SVM classifiers are both trained and tested with the raw features and statistical features separately.
One hundred samples of feature vectors under each health condition are used, where 70 samples under each health
condition are used for training and the rest are for testing. The classification results of RBF ANN and SVM with the raw
features are shown in Figs. 7 and 8 respectively. The classification results of RBF ANN and SVM with the statistical features
are shown in Figs. 9 and 10 respectively. It can be observed that the ANN and the SVM with the raw features have

4
Output Values

0
0 30 60 90 120 150
Sample No.

Fig. 10. Fault diagnosis result with statistical features and SVM. Samples from No. 1 to No. 30 are under the normal condition. Samples from No. 31 to No.
60 are with the unbalanced fault of rotor. Samples from No. 61 to No. 90 are with the inner race fault. Samples from No. 91 to No. 120 are with the outer
race fault. Samples from No. 121 to No. 150 are with a ball fault.

Table 1
The thresholds with β ¼ 3.

J1 J2 J3 J4 J5

3:67E  06 6:85E  07 2:54E  06 4:96E  06 3:59E  6

−5
x 10
2
J1
1
r1
0
0 10 20 30 40 50 60 70 80 90 100

−6
x 10
1
J2
0.5
r2
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
Amplitude

4
J3
2
r3
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
5 J4
r4
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
4
J5
2
r5
0
0 10 20 30 40 50 60 70 80 90 100
Sample No.
Fig. 11. The values of rj with the 100 samples of statistical feature vectors from the test data under the normal condition.
 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426 423

misclassifications. The circles in Fig. 7 and the rectangular in Fig. 8 indicate the wrong classified samples. The accuracy of
ANN with raw features is around 97% and with statistical features is 100%; and the accuracy of SVM with raw features is
around 82% and with statistical features is 100%. By using the statistical features, the accuracies of ANN and SVM based fault
classification methods are significantly improved. Although the calculation of statistical features requires higher computa-
tion effort, it is not a problem in off-line fault diagnosis.
It is worth mentioning that the accuracies of these fault classification methods can only be estimated through
experiments, and no bounds of the accuracies can be established mathematically. In the next subsection, the evaluation
2
method of statistical features is employed to realized fault classification by using the knowledge of μj and sj , where the
accuracy of fault classification is ensured.

5.2. Fault diagnosis with the proposed method

This subsection presents the fault diagnosis results with the proposed statistical feature extraction and evaluation
method. At first, μ j and s 2j in (4) and (5) are estimated based on 500 samples of the raw feature vectors under the j-th health
condition, where the design data is used. After that the decoupling vector Hj is computed according to (4) and (5). Finally rj
is calculated according to (6) and then the health condition is classified through the decision logic (8) with β¼3, such that
FCR is ensured to be smaller than 0.3%. The thresholds are shown in Table 1.
With M¼100 and k¼ 2048, 100 test samples of statistical feature vectors are extracted from the test data under each
health condition in order to testify the proposed method. Hence totally 500 samples are used for the testing purpose.
The values of rj computed with test samples under different health conditions are shown in Figs. 11 (normal condition),
12 (unbalanced fault of rotor), 13 (inner race fault of a bearing), 14 (outer race fault of a bearing) and 15 (a ball fault of a
bearing). For each test sample, rj are computed and depicted. According to (8), when rj is larger than the threshold Jj, the
corresponding health condition is classified, i.e. Cj ¼1. In the experiment, only one false classification is observed in Fig. 12,
which is the circled test sample misclassified by r1 and J1. In Figs. 11 and 13–15, no false classification is observed. Hence the
false classification rate estimated from the experiment is around 0.2%, which is smaller than the analytical bound of FCR. It is
also possible to select a larger β, such that the thresholds are larger and consequently the FCR is smaller. Nevertheless some
faults may not be detected in this case.

−6
x 10
5
J1
r1
0
0 10 20 30 40 50 60 70 80 90 100
−4
x 10
4
2 J2
r2
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
Amplitude

4
J3
2
r3
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
5
J4
r4
0
0 10 20 30 40 50 60 70 80 90 100

−6
x 10
4
J5
2
r5
0
0 10 20 30 40 50 60 70 80 90 100
Sample No.
Fig. 12. The values of rj with the 100 samples of statistical feature vectors from the test data with unbalanced fault of rotor. The circled sample is
misclassified.
424 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426

x 10
4
J1
2
r1
0
0 10 20 30 40 50 60 70 80 90 100

x 10
1
J2
0.5
r2
0
0 10 20 30 40 50 60 70 80 90 100
Amplitude

x 10
1
J3
0.5
r3
0
0 10 20 30 40 50 60 70 80 90 100

x 10
5 J4
r4
0
0 10 20 30 40 50 60 70 80 90 100

x 10
4
J5
2
r5
0
0 10 20 30 40 50 60 70 80 90 100
Sample No.
Fig. 13. The values of rj with the 100 samples of statistical feature vectors from the test data with inner race fault of a bearing.

−6
x 10
4
J1
2
r1
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
1
J2
0.5
r2
0
0 10 20 30 40 50 60 70 80 90 100
Amplitude

−6
x 10
4
J3
2
r3
0
0 10 20 30 40 50 60 70 80 90 100
−4
x 10
1
J4
0.5
r4
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
4
J5
2
r5
0
0 10 20 30 40 50 60 70 80 90 100
Sample No.
Fig. 14. The values of rj with the 100 samples of statistical feature vectors from the test data with outer race fault of a bearing.

6. Conclusions

In this paper, a fault diagnosis approach has been proposed for rotating machinery based on a novel statistical feature
extraction and evaluation method. The statistical features are indeed the sampled average of some conventional features,
and these conventional features can be obtained by analyzing arbitrarily selected partitions of a given vibration signal with
existing signal processing tools. According to the central limit theory, the obtained statistical feature vectors are close to
normal distributions, and their means and variances could be estimated. With the statistical features, the performance of
ANN and SVM based fault classifiers could be significantly improved. However ANN and SVM still cannot provide any
analytical guarantee on the accuracy of fault classification. Hence an evaluation method has also been proposed for
statistical features, where the means and the variances of statistical features have been used to calculate the decoupling
vectors and the thresholds based on simple algebraic computations. The health condition is classified when the
corresponding threshold is exceeded. With the proposed method the accuracy of fault classification can be analytically
 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426 425

−6
x 10
4
J1
2
r1
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
1
J2
0.5
r2
0
0 10 20 30 40 50 60 70 80 90 100
Amplitude

−6
x 10
4
J3
2
r3
0
0 10 20 30 40 50 60 70 80 90 100
−6
x 10
5 J4
r4
0
0 10 20 30 40 50 60 70 80 90 100
−4
x 10
1
J5
0.5
r5
0
0 10 20 30 40 50 60 70 80 90 100
Sample No.
Fig. 15. The values of rj with the 100 samples of statistical feature vectors from the test data with a ball fault of a bearing.

guaranteed in terms of FCR, which is of practical interests. The experiment has also been carried out to demonstrate the
effectiveness of the proposed method.

Acknowledgments

The research was supported by Program for New Century Excellent Talents in University (NCET-13-1018), National
Natural Science Foundation of China (51275513) and the Project Funded by the Priority Academic Program Development of
Jiangsu Higher Education Institutions (PAPD).

References

[1] A.K.S. Jardine, D. Lin, D. Banjevic, A review on machinery diagnostics and prognostics implementing condition-based maintenance, Mech. Syst. Signal
Process. 20 (7) (2006) 1483–1510.
[2] A. Widodo, B.S. Yang, Support vector machine in machine condition monitoring and fault diagnosis, Mech. Syst. Signal Process. 21 (6) (2007)
2560–2574.
[3] 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 Syst. Appl. 38 (4) (2011) 3819–3826.
[4] E. Zio, P. Baraldi, G. Gola, Feature-based classifier ensembles for diagnosing multiple faults in rotating machinery, Appl. Comput. 8 (4) (2008)
1365–1380.
[5] H.R. Martin, Statistical moment analysis as a means of surface damage detection, in: Proceedings of the International Modal Analysis Conference, 1989,
pp. 1016–1021.
[6] E. Volker, H.R. Matin, Application of kurtosis to damage mapping, in: Proceedings of the International Modal Analysis Conference, 1986, pp. 629–633.
[7] D. Logan, J. Mathew, Using the correlation dimension for vibration fault diagnosis of rolling element bearings I: basic concepts, Mech. Syst. Signal
Process. 10 (1996) 241–250.
[8] J. Yang, Y. Zhang, Y. Zhu, Intelligent fault diagnosis of rolling element bearing based on SVMs and fractal dimension, Mech. Syst. Signal Process. 21
(2007) 2012–2024.
[9] V.K. Rai, A.R. Mohanty, Bearing fault diagnosis using FFT of intrinsic mode functions in Hilbert–Huang transform, Mech. Syst. Signal Process. 21 (8)
(2007) 3030–3041.
[10] Y. Yang, D.J. Yu, J.S. Cheng, A fault diagnosis approach for roller bearing based on IMF envelope spectrum and SVM, Measurement 40 (9–10) (2007)
943–950.
[11] C.M. Harris, A.G. Piersol, Harris Shock and Vibration Handbook, McGraw-Hill, New York, 2002.
[12] W.J. Wang, P.D. McFadden, Early detection of gear failure by vibration analysis I: calculation of the time–frequency distribution, Mech. Syst. Signal
Process. 7 (1993) 193–203.
[13] S.K. Lee, P.R. White, Higher-order time–frequency analysis and its application to fault detection in rotating machinery, Mech. Syst. Signal Process. 11
(1997) 637–650.
[14] N. Baydar, A. Ball, A comparative study of acoustic and vibration signals in detection of gear failures using Wigner–Ville distribution, Mech. Syst. Signal
Process. 15 (2001) 1091–1107.
[15] R. Rubini, U. Meneghetti, Application of the envelope and wavelet transform analyses for the diagnosis of incipient faults in ball bearings, Mech. Syst.
Signal Process. 15 (2001) 287–302.
[16] N. Li, R. Zhou, Q. Hu, X. Liu, Mechanical fault diagnosis based on redundant second generation wavelet packet transform, neighborhood rough set and
support vector machine, Mech. Syst. Signal Process. 28 (2012) 608–621.
[17] Y. Lei, J. Lin, Z. He, M.J. Zuo, A review on empirical mode decomposition in fault diagnosis of rotating machinery, Mech. Syst. Signal Process. 35 (2013)
108–126.
426 W. Li et al. / Mechanical Systems and Signal Processing 50-51 (2015) 414–426

[18] M. Demetgul, I.N. Tansel, S. Taskin, Fault diagnosis of pneumatic systems with artificial neural network algorithms, Expert Syst. Appl. 36 (7) (2009)
10512–10519.
[19] J.D. Wu, C.H. Liu, An expert system for fault diagnosis in internal combustion engines using wavelet packet transform and neural network, Expert Syst.
Appl. 36 (3) (2009) 4278–4286.
[20] C. Castejn, O. Lara, J.C. Garca-Prada, Automated diagnosis of rolling bearings using MRA and neural networks, Mech. Syst. Signal Process. 24 (1) (2010)
289–299.
[21] G.F. Bin, J.J. Gao, X.J. Li, B.S. Dhillon, Early fault diagnosis of rotating machinery based on wavelet packets-empirical mode decomposition feature
extraction and neural network, Mech. Syst. Signal Process. 27 (2012) 696–711.
[22] V. Vapnik, The Nature of Statistical Learning, Springer, New York, 1995.
[23] Q. Hu, Z. He, Z. Zhang, Y. Zi, Fault diagnosis of rotating machinery based on improved wavelet package transform and SVMs ensemble, Mech. Syst.
Signal Process. 21 (2007) 688–705.
[24] S. Abbasion, A. Rafsanjani, A. Farshidianfar, N. Irani, Rolling element bearings multi-fault classification based on the wavelet denoising and support
vector machine, Mech. Syst. Signal Process. 21 (7) (2007) 2933–2945.
[25] F. Camci, R.B. Chinnam, General support vector representation machine for one-class classification of non-stationary classes, Pattern Recognit. 41 (10)
(2008) 3021–3034.
[26] Q. Wu, Fuzzy fault diagnosis based on fuzzy robust v-support vector classifier and modified genetic algorithm, Expert Syst. Appl. 38 (5) (2011)
4882–4888.
[27] Z. Shen, X. Chen, X. Zhang, Z. He, A novel intelligent gear fault diagnosis model based on EMD and multi-class TSVM, Measurement 45 (1) (2012)
30–40.
[28] W. Li, F. Jiang, Z. Zhu, G. Zhou, G. Chen, Fault diagnosis of bearings based on a sensitive feature decoupling technique, Measur. Sci. Technol. 24 (2013)
035602.
[29] X. Lou, K.A. Loparo, Bearing fault diagnosis based on wavelet transform and fuzzy inference, Mech. Syst. Signal Process. 18 (5) (2004) 1077–1095.
[30] Y. Lei, Z. He, Y. Zi, Q. Hu, Fault diagnosis of rotating machinery based on multiple ANFIS combination with GAs, Mech. Syst. Signal Process. 21 (5) (2007)
2280–2294.
[31] Y. Xu, M. Ge, Hidden Markov model-based process monitoring system, J. Intell. Manuf. 15 (2004) 337–350.
[33] T.P. Ryan, Modern Engineering Statistics, John Wiley & Sons, New Jersey, 2007.
[34] W.J. Conover, Practical Nonparametric Statistics, John Wiley & Sons, New York, 1980.
[35] C. Chang, C. Lin, LIBSVM: a library for support vector machines, ACM Trans. Intell. Syst. Technol. 2 (2011) 27.