sensors

Article
Naive Bayes Bearing Fault Diagnosis Based on
Enhanced Independence of Data
Nannan Zhang 1,2,3,4 , Lifeng Wu 1,2,3,4, *, Jing Yang 1,2,3,4 and Yong Guan 1,2,3,4
1 College of Information Engineering, Capital Normal University, Beijing 100048, China;
[email protected] (N.Z.); [email protected] (J.Y.); [email protected] (Y.G.)
2 Beijing Key Laboratory of Electronic System Reliability Technology, Capital Normal University,
Beijing 100048, China
3 Beijing Key Laboratory of Light Industrial Robot and Safety Verification, Capital Normal University,
Beijing 100048, China
4 Beijing Advanced Innovation Center for Imaging Technology, Capital Normal University,
Beijing 100048, China
* Correspondence: [email protected]; Tel.: +86-134-0110-8644

Received: 21 December 2017; Accepted: 1 February 2018; Published: 5 February 2018

Abstract: The bearing is the key component of rotating machinery, and its performance directly
determines the reliability and safety of the system. Data-based bearing fault diagnosis has become a
research hotspot. Naive Bayes (NB), which is based on independent presumption, is widely used
in fault diagnosis. However, the bearing data are not completely independent, which reduces the
performance of NB algorithms. In order to solve this problem, we propose a NB bearing fault
diagnosis method based on enhanced independence of data. The method deals with data vector
from two aspects: the attribute feature and the sample dimension. After processing, the classification
limitation of NB is reduced by the independence hypothesis. First, we extract the statistical
characteristics of the original signal of the bearings effectively. Then, the Decision Tree algorithm is
used to select the important features of the time domain signal, and the low correlation features is
selected. Next, the Selective Support Vector Machine (SSVM) is used to prune the dimension data
and remove redundant vectors. Finally, we use NB to diagnose the fault with the low correlation data.
The experimental results show that the independent enhancement of data is effective for bearing
fault diagnosis.

Keywords: Naive Bayes; decision tree; support vector machines; fault diagnosis

1. Introduction
The rolling bearing is the main component of rotating machinery. It carries the entire rotating
machinery and equipment operation, and a small fault may have a significant impact on the operation
of the entire device. Most of the problems with rotating machines are caused by bearing failure [1].
Therefore, the bearing fault diagnosis is of great significance. After the fault diagnosis of rotating
machinery, the machine can be repaired and handled in time, so as to avoid the catastrophic effect
caused by mechanical failure [2]. The related contents and techniques of fault diagnosis are introduced
in the literature [3–6]. Before the machine failure, the maintenance and treatment of the machine can
prevent the probability of failure and reduce the maintenance costs of the machine, as well as avoid
casualties caused by equipment failure.
The vibration analysis is the main tool for the diagnosis of rotating machinery [7], and vibration
signals analysis has been widely used in the field of fault diagnosis. In this field, the vibration spectrum
analysis technique has successful identified the faults [8–11]. Through the analysis of vibration signals,
the state of rotating machinery can be reflected. Sensors can be used to collect vibration signals of

Sensors 2018, 18, 463; doi:10.3390/s18020463 www.mdpi.com/journal/sensors

Sensors 2018, 18, 463 2 of 17

operating machinery, which contains rich information about the working state of machinery [12].
The mechanical health state is determined by analyzing the collected vibration signals. However,
the collected vibration signals are chaotic and irregular. Therefore, it is necessary to extract the most
representative, reliable and effective features from the acquired vibration signals.
The time domain signal feature of statistical analysis can be used to detect faults, which is mainly
to extract feature of data. However, it can only reflect whether rotating machinery and its state are
normal, and it can give diagnostic messages but not diagnose the fault, so further fault diagnosis is
needed. Nowadays, with the successful application of machine learning methods in various fields,
more and more machine learning methods are used in mechanical fault diagnosis. Neural networks,
as a typical method of machine learning, has been applied to the field of fault diagnosis. As the most
popular classifier, Support Vector Machines (SVM) has achieved some success in the field of fault
diagnosis. SVM is a powerful tool for classification, and it also plays a significant role in machine
fault diagnosis [13]. Samanta [14] proposed time-domain characteristics of the rotating machine that
can be used as the input of artificial neural network (ANN) to verify the effective application in
bearing fault diagnosis. Jack et al. [15] put forward SVM for bearing fault diagnosis. Yang et al. [16]
proposed that vibration signals can be decomposed to stationary intrinsic mode functions (IMFs),
and the input of ANN is the energy features extracted from IMF so as to identify the rolling bearing
failure. Al-Raheem et al. [17] proposed a new technique that used genetic algorithm to optimize the
application of Laplace wavelet shape and classifier parameters of ANN for bearing faults. In addition,
Shuang et al. [18] proposed a fault pattern recognition method based on the principal component
analysis (PCA) and SVM. However, the extracted multi-dimensional feature vector contains a large
amount of information, with high data redundancy, which results in higher computational costs.
Therefore, the high-dimensional characteristics need to be processed. Wu et al. [19] used the Manifold
Learning algorithm to reduce the dimension of the high-dimensional features and then the processed
are used as the input of wavelet neural network for bearing fault diagnosis. Sugumaran et al. [20]
applied Decision Tree to selecting feature, and then carried on the bearing fault diagnosis with the
kernel neighborhood fractional multiple support vector machine (MSVM). In another article [21], first,
the time domain statistical feature and histogram feature was extracted from time domain signals,
then the main feature was selected by the Decision Tree, last SVM and Proximal Support Vector
was used for fault diagnostics of roller bearing. In a recent study, Ran et al. [22] proposed a neural
network-based method to directly identify the original time series sensor data without feature selection
and signal processing for bearing fault diagnosis. In his other article [23], the network is a combination
of linear and nonlinear method, and also uses the depth network classifier of the original time series
sensor data to diagnose faults.
ANN and SVM have an extensive application in fault diagnosis. However, there are some
limitations. For example, the fitting problem and the local extremum can lead to slow operation speed
and inaccurancy in ANN training results, respectively [24]. Moreover, SVM has a problem with the
speed of testing and training. There are some limitations on multi-class, nonlinear and parameters
problem. The training of ANN and SVM is complex, and the cost on training space is high. NB is
used not only a small mount of training data, but also its simple structure, fast calculation speed and
high accuracy [25,26]. Due to the reliable theoretical basis, comprehensive prior knowledge and the
assumption of independent conditions among attributes, NB successfully applied in machine fault
diagnosis. Hemantha et al. used the Bayes classifier to diagnose the bearing fault, and verified that
NB on fault diagnosis has a good performance [27]. Girish et al. successfully applied NB classifier
to the welded joints fault diagnosis [28]. However, the independence assumption of vibration signal
of bearing fault is difficult to be realized in actual situations, which limits the algorithm. Therefore,
this paper mainly carries on the vector pruning from two aspects of the characteristic attributes and
the data dimension. First, Decision Trees are mainly used to select the main feature attributes [29].
Then, the redundancy of dimension vectors is removed by the proposed selective support vector
Sensors 2018, 18, 463 3 of 17

machine (SSVM). In this way, the redundant data is processed from two aspects, and the limitation of
the independence hypothesis on the NB is reduced. Finally, fault diagnosis model is established.
In this paper, NB, which is based on data independence improvement in fault diagnosis, is
proposed. The remainder of paper is organized as follows: in Section 2, there is a brief introduction
of the NB model. The fault diagnosis based on improved data independence is given by Section 3.
In Section 4, the fault diagnosis based on improved data independence is applied to roller bearing
diagnosis. Section 5 draws the conclusion of this paper.

2. NB Model
NB is a supervised learning classification method based on probability. NB has received much
attention due to its simple classification model and excellent classification performance. The training
model is shown in Figure 1.

Preparation stage Known sample L Unknown sample Y

Calculate the prior probabilities

For each category

Training stage

Calculate the conditional probability formula for under the condition

of

Application stage The category of the discriminant function is the category of

Figure 1. Naive Bayes training model.

(a) Preparatory stage

Suppose there are m categories and categories L = { L1 , L2 , · · · , Lm }. Each sample has
n attributes At = { At1 , At2 , · · · , Atn }, and each attribute set has d-dimensional feature vector
X = { X1 , X2 , · · · , X d } .
(b) Training stage
P( Li ) is the prior probability of each category, only related to the ratio of each category to the total
category, that is,
n
P( Li ) = i , 1 ≤ i ≤ m, (1)
n
where n is the number of known sample, and ni is the number of i-th categories.
Bayes is a classifier based on the maximum posterior probability. There is an unknown sample
classes Y = {y1 , y2 , · · · , yz }, and the idea is to calculate the probability of unknown samples in each
category. Finally, if the probability of the unknown sample Y is maximum in class Li , the unknown
sample is classified into category Li . NB is based on the Bayes theorem, and the NB classification
method is shown below:

P( Li /yh ) > P( L j /yh ), 1 ≤ i ≤ m, 1 ≤ j ≤ m, 1 ≤ h ≤ z, i 6= j. (2)

Sensors 2018, 18, 463 4 of 17

According to Bayes’s theorem, the probability formula of P( Li /yh ) can be obtained. The NB is a
Bayes theorem based on the independence of the characteristic conditions, so P( Li /yh ) can be defined
as follows:
P(yh /Li ) P( Li )
P( Li /yh ) = (3)
P ( y h ),
m
P(yh ) = ∑ P(yh /Li ) P( Li ), (4)
i =1

where P(yh ) is a constant, and it is only necessary to compute the formula P(yh /Li ) P( Li ) of
Equation (3).
According to the NB classification method, the value of the discriminant function P(yh /Li ) P( Li )
in each class is calculated for the unknown sample, where P( Li ) is a priori probability of each category,
as shown in Equation (1), and where P(yh /Li ) is the probability of yh under the condition of Li .
The attribute At gi is continuous property and independent of each other. In general, the attribute
2 ) [30]; then, P ( y /L ) is defined as follows:
variable obeys the Gaussian distribution At gi ∼ N (u gi , δgi h i

( )
1 (yh − u gi )2
P(yh /Li ) = √ exp − 2
, (5)
2πδgi 2δgi

2 are mean and variance of samples, respectively, and the formula is as follows:
where u gi and δgi

n
∑i=i 1 Xig
u gi = , (6)
ni
n
2 ∑i=i 1 ( Xig − u gi )
δgi = . (7)
ni − 1
From the above Equations (2) and (5)–(7), the posterior probability equation can be obtained:
( )
n
1 (yh − u gi )2
P( Li /yh ) = P( Li ) ∏ √ exp − 2
. (8)
g =1 2πδgi 2δgi

In the same way:

( )
n
1 (yh − u gj )2
P( L j /yh ) = P( L j ) ∏ √ exp − 2
. (9)
g =1 2πδgj 2δgj

(c) Application stage

According to the Equation (2), if P( Li ) P( Li /yh ) > P( L j ) P( L j /yh ), the unknown sample is judged
as class i; otherwise, it is judged as j.

3. NB Fault Diagnosis Model Based on Enhanced Independence of Data

3.1. Fault Diagnosis Model

In order to improve the classification effect of NB, this paper enhances the independence between
data from two aspects of attribute characteristics and data dimension. The proposed fault diagnosis
model is shown in Figure 2. The fault diagnosis model includes three parts: signal acquisition, signal
processing and fault diagnosis.

• Signal acquisition: Acceleration sensor is used to obtain vibration signals of rolling bearings.
• Signal processing: The original vibration signal of the rolling bearing obtained from the sensor
contains a large amount of noise, so it is necessary to process the data to obtain valid data signals.
 Sensors 2018, 18, 463 5 of 17

Firstly, feature extraction is performed on the original signal acquired by using the time-domain
signal method. Then, the Decision Tree is used to select the main feature attributes from the
feature attributes. The data are processed from two directions of feature attribute and data
dimension, so that the data with strong independence can be obtained, which is beneficial to the
fault diagnosis of the bearing.
• Fault diagnosis: After the data is processed, we obtain data with low redundancy. Thus,
the impact of data independence assumption on NB model is reduced, and the fault diagnosis
can be made effectively.

Collect vibration signal

Extract Time-frequency domain features

Multi-dimensional feature

Main features by J48 algorithm

Low-correlationl feature

Training data set Testing data set

Adjust threshold
SSVM data pruning

Machine fault diagnosis

Figure 2. Fault diagnosis model based on the enhanced independence of data.

3.2. Feature Selection Using Decision Tree

The Decision Tree is a tree structure, which is mainly composed of nodes and branches, and the
nodes contain leaf nodes and intermediate nodes. The intermediate nodes are used to represent
a feature, and leaf nodes are used to represent a class label. The Decision Tree can be used for
feature selection [29]. The attributes that appear in the Decision Tree nodes provide important
information to promote classification. The J48 algorithm is mainly used to construct Decision Tree.
Therefore, we construct a Decision Tree using J48 algorithm. Then, we find the characteristic attribute
corresponding to the middle node of the decision tree, and remove the feature attribute that without
important information. The following describes the J48 algorithm for feature extraction:

(a) The acquired data is used as the input of the algorithm, and the output is the node of the
Decision Tree.
(b) The output Decision Tree nodes are divided into leaf nodes and intermediate nodes. The leaf
node represents the classification, the intermediate node represents the decision attribute,
and the branch represents the condition that the next decision data comes from the previous
decision attribute.
(c) The Decision Tree is used to find feature attributes from top to bottom until all nodes become
leaf nodes.
(d) Finding the criteria of decision attributes: the information gain of each feature is calculated
and the maximum information gain is chosen as the intermediate node of the Decision Tree.

Information gain is used to determine how to select the most appropriate features from a number
of attributes. Information gain is mainly determined by the information entropy. Information gain of
Sensors 2018, 18, 463 6 of 17

attribute At for the data set is: entropy of all attribute information minus the entropy of split attributes.
The At is a continuous attribute based on Gaussian distribution, so information entropy properties of
At is defined as follows:
Gain( At) = In f o ( L) − in f o At ( L), (10)
m
In f o ( L) = − ∑ P( j/L)logP( j/L), (11)
j =1

2
m L j Log2πeδij
In f o At ( L j ) = ∑ L 2
. (12)
j =1

Gain( At) is the information gain of the attribute At, In f o ( L) is the undivided information entropy,
and in f o At ( L) is the information entropy At after splitting. The variance x is given by the Formula (6),
and m is the number of classifications, and L j is a subset of data set L.

3.3. SSVM
SVM is a traditional classification method for two categories. In this paper, an optimal
classification hyperplane is constructed in the sample set, and two classes of samples are separated
from each other on the hyperplane. Generally, in the case of too much data, SVM can not completely
classify the two kinds of data into both sides of the hyperplane. Thus, we propose an SSVM algorithm
to remove the spatial redundancy problem of the vector.
SSVM data processing is divided into several steps, as shown in Figure 3.

Construct a Select data Reorganize

hyperplane pruning the data

Figure 3. Selective Support Vector Machine data processing flow chart.

Step 1: Constructing the optimal hyperplane of data.

In most cases, SVM is targeted at two types of problems [31]. The data set ( X, Y ) is divided into
training set and test set. The training set is ( X1 , Y1 ), ( X2 , Y2 ), · · · , ( Xn , Yn ). if Xi is the first class, Yi = 1.
if Xi is second class, Yi = −1. As shown in Figure 4, hyperplane H ( X ) separates the two-class data on
both sides.

Support
Vector

Support
Vector H(X)

Figure 4. Two categories of Support Vector Machine.

 Sensors 2018, 18, 463 7 of 17

The hyperplane H (X) equation is given as in Equation (13) [32]:

w T K ( X ) + b = 0. (13)

The function K ( X ) is a kernel function, which maps the low-dimensional space to the
high-dimensional space, and avoids the fact that the data cannot be separated in the low-dimensional
space, where w is a vector, b is constant, and their values can be obtained by the optimization of the
following Equation [31]:
n
1
min : kwk + C ∑ ξ i , kwk = w T w, (14)
2 i =1

s.t. yi (w T K ( Xi ) + b) ≥ 1 − ξ i , ξ i ≥ 0. (15)

Parameter C is mainly used to adjust training error. ξ i is a slack variable [33]. After the solution
of the parameter, the optimal hyperplane H ( X ) is obtained [31]:

H ( X ) = sgn( ∑ yi ai K ( xi , x ) + b), (16)

iεSV

where S is the support vector for the DataSet ( X, Y ), where sgn is a symbolic function that mainly
returns the positive and negative of the parameter. K ( xi , x ) is a kernel function, and there are many
kinds of kernel functions. The Gaussian kernel function is better in the application, so the Gaussian
kernel function is used in this paper:

k x − yk
K ( x, y) = exp( ). (17)
2σ2
Step 2: Using the constructed hyperplane to select the data and remove the redundancy.
Firstly, a suitable threshold is selected, and the hyperplane K ( X ) is used to test the data. When the
test result does not reach the threshold, this data is chosen to be pruned.
Then, find the hyperplane boundary support vector.
Finally, find the point closest to each support vector, and judge if the closest distance is consistent
with the classification of the vector; then, keep it, or otherwise delete it.
This article uses the Euclidean distance to measure the distance between two points.
For high-dimensional data, the distance between two points is the distance of two vectors, for example,
X = ( x1 , x2 , · · · , xn ) and Y = (y1 , y2 , · · · , yn ), X and Y distanceD ( X, Y ) is written as:
s
n
D ( X, Y ) = ∑ (xi , yi )2 , 1 ≤ i ≤ n. (18)
i =1

Step 3: Reorganizing processing data, and obtaining new data.

SVM is mainly used for two types of data. This article mainly uses multiple categories of data.
First of all, the data in multiple categories were put into pairs, respectively. Then, two kinds of data are
pruned with SSVM. The data processing is divided into the following steps:

(1) Construct hyperplane for training.

(2) Test the data with a trained hyperplane.
(3) Set the appropriate threshold to find out the classification of the training data and training results
below the threshold.
(4) Finding the nearest neighbor of each support vector form data obtained in step (3), calculating
the distance between the support vectors and the data points, and setting the distance between
the points to itself be infinity.
(5) Find the nearest vector point of each support vector.
 Sensors 2018, 18, 463 8 of 17

(6) Determine whether the support vector is consistent with its corresponding nearest neighbor
vector classification result, and mark it as 0 if inconsistent.
(7) Remove the data marked as 0 in the data.
(8) Reorganize data to get new data.

According to the description of the SSVM, the SSVM pruning algorithm is the most important
part of the SSVM. The details of SSVM pruning algorithm are shown in Algorithm 1.

Algorithm 1 SSVM pruning algorithm.

Input:
The selected training sample <X,Y>,X = ( X1 , X2 , · · · , Xn );
Output:
Trimmed sample <X1,Y1>
1: Begin
2: Obtain support vector <Z,H> by SVM, Z = ( Z1 , Z2 , · · · , Zn )
3: for i:=1 to n do
4: for j:=1 to m do
5: Calculate the distance D ( Zi , X j ) between Z and X by Equation (18), When Zi is the
same as X j , define the distance D as infinite.
6: end
7: Find the nearest dimension vector X j between Zi and X
8: Judge whether Hi and Yj are the same, if not, let Yj = 0
9: end
10: Delete the sample data of the Y=0
11: return < X1, X2 >
12: end

4. Experiment and Analysis

4.1. Bearing Data Preprocessing

The data in this article is from bearing fault signals provided by the Case Western Reserve
University (CWRU) laboratories [34]. The experimental platform is shown in Figure 5. The
experimental platform consists of a torque tachometer, a 1.5 KW motor and a dynamometer. The
experimental data uses the acceleration signals collected by the acceleration sensors. The sensor is fixed
to the position of the driving end and the fan end of the motor shell at 12 o’clock with the magnetic
base, and the vibration signal is collected through the recorder. The type of bearing used in the test is
SKF6205-2RS deep groove ball bearing. The sampling frequency of the experiment is 12 KHz, the speed
is 1797 rpm, and the main data is collected from normal vibration signal and the fault vibration signal.

Figure 5. Experimental diagram of experimental platform for rolling bearing fault.

Sensors 2018, 18, 463 9 of 17

In this paper, the normal vibration signals and fault signals of bearings are analyzed, and the
samples of each type of signals are at least 12,100. The main samples of this paper are those samples
with no load and the 0.021 (inches) radius fault. Table 1 describes a normal bearing signal and five
kinds of fault bearing signals used in this paper. Six kinds of bearing data are described in Figure 6.
Acceleration(m/s 2 )

0.5
(a)

-0.5
0 0.002 0.004 0.006 0.008 0.01
time(s)
Acceleration(m/s 2 )

0.5
(b)

-0.5
0 0.002 0.004 0.006 0.008 0.01
time(s)
Acceleration(m/s 2 )

0.5
(c)

-0.5
0 0.002 0.004 0.006 0.008 0.01
time(s)
Acceleration(m/s 2 )

0.5
(d)

-0.5
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
time(s)
Acceleration(m/s 2 )

0.5
(e)

-0.5
0 0.002 0.004 0.006 0.008 0.01
time(s)
Acceleration(m/s 2 )

0.5
(f)

-0.5
0 0.002 0.004 0.006 0.008 0.01
times(s)

Figure 6. The time domain waveform of rolling bearings is shown in the figure. The x-axis is the
time unit of the second and y-axis is the driving end bearing accelerator data. (a) normal bearing
signal waveform; (b) inner fault signal waveform; (c) roller fault signal waveform; (d) outer fault
signal waveform at center @6:00; (e) outer ring fault signal at orthogonal @3:00; (f) outer fault signal
waveform at opposite @12:00.
Sensors 2018, 18, 463 10 of 17

Table 1. Description of CWRU dataset.

Data Type Motor Load (HP) Fault Diameter (Inches) Label

Normal 0 0 1
Inner race 0 0.021 2
Ball 0 0.021 3
Out race fault at center @6:00 0 0.021 4
Out race fault at orthogonal @3:00 0 0.021 5
Out race fault at opposite @12:00 0 0.021 6

4.2. Application of Improved Algorithm in Bearing Fault

Fault diagnosis model is constructed according to Figure 2. This paper chooses West University of
rolling bearing samples and the numbers of each state are at least 121,200. Test data and training data
account for half of the total data. The detailed description of various bearing States is shown in Table 2.

Table 2. Description of the data sets.

Data Type The Number of Training The Number of Testing Label

Normal 121 121 1
Inner race 121 121 2
Ball 121 121 3
Out race fault at center @6:00 121 121 4
Out race fault at orthogonal @3:00 121 121 5
Out race fault at opposite @12:00 121 121 6

In this paper, the vibration signal is mainly processed from three aspects.
First, the feature extraction is performed by the time domain method.
The statistical characteristics of signal vibration amplitude will change with the location and the
size of the fault. The time domain waveform is dynamically transformed over time. The amplitude of
the vibration signal can reflect the characteristic information of the signal intuitively. The time domain
waveform information can be used to diagnose the state of the bearing by analyzing the amplitude,
shape and other characteristics of the waveform. The time domain characteristic parameters are
different due to different fault types and different fault degree. Generally speaking, the time domain
feature provides the global characteristics of bearing state, and can effectively extract the bearing
fault feature.
In the actual situation, there is various information of bearing fault, and a faults are often
accompanied with other faults, such as bearing deformation, corrosion and so on. In order to diagnose
the fault more effectively, we need to extract the feature of bearing fault data. In this paper, 17-time
domain extraction methods are used to extract the features of the signal.
In Table 3, X (n) is the representative of the signal sample n = 1, 2, ..., m, and m represents the
number of samples. Seventeen time domain feature attributes is: T1 the average value, T2 absolute
mean, T3 effective value, T4 average power, T5 square amplitude, T6 peak, T7 peak-to-peak, T8 variance,
T9 standard deviation, T10 skewness, T11 kurtosis, T12 waveform, T13 Crest index, T14 impluse index,
T15 margin index, T16 skewness index and T17 kurtosis index.
Second, the main feature selection of feature extraction data is made by the Decision Tree.
The main description of the J48 algorithm is given in Chapter 3, and the output tree structure
shown in Figure 7. It can be seen from the diagram that the main characteristics of bearing data are T1 ,
T5 , T12 and T17 .
Sensors 2018, 18, 463 11 of 17

Table 3. Time domain analysis of bearing fault data.

Number Characteristic Equation Number Characteristic Equation

d d
∑ n =1 X (n) ∑n=1 | X (n)|
1 T1 =
q d d 2 T2 = d
d
T3 = ∑n=1 d ∑n=1 ( X (n))2
( X (n))2
3 4 T4 =
d
√ d
∑ | X (n)| 2
5 T5 = ( n=1 d ) 6 T6 = max | X (n)|
d
∑n=1 ( X (n)− T1 )2
7 T7 = max ( X (n)) − min( X (n)) 8 T8 = d −1
q d d
∑n=1 ( X (n))3
T8 = ∑n=1 d−1 1
( X (n)− T )2
9 10 T10 = d
d
( X (n))4
11 T11 = ∑n=1 d 12 T12 = TT32
13 T13 = TT36 14 T14 = TT62
15 T15 = TT56 16 T16 = TT103
9
17 T17 = TT114
9

3 -

2 3

2 2 5

4 2 2 6 2 5 6 4 3

4 3

Figure 7. A part of the Decision Tree.

The 17 characteristic attributes obtained by feature extraction are interrelated with each other,
which leads to data redundancy. The attributes with low correlation are obtained by extracting the
main features with J48 so that the independence of data can be enhanced.
The description and significance of these four main time-domain features are as follows:

• average value ( T1 ): T1 is mainly used to reflect the trend of the bearing fault signal,
• square amplitude ( T5 ): T5 is mainly used to describe the energy of signals,
• waveform index ( T12 ): T12 is sensitive to fault signals with stable waveform,
• kurtosis index T17 : kurtosis is sensitive to bearing defects and can reliably reflect the state of
rolling bearings. It is not easy to be affected by temperature, speed, etc. and comprehensive
analysis of kurtosis, peak factor, and effective value.

In Figure 7, the intermediate node represents the attribute of the decision with an ellipse, and the
leaf node represents the classification result with a rectangle. The data between nodes are the
classification condition. The graph is a part of the Decision Tree. Class label is a class with the
highest probability in classification result when it has little effect on feature selection.
Sensors 2018, 18, 463 12 of 17

Third, the main feature of extraction is pruned with SSVM.

The J48 algorithm is mainly used to extract attribute vector so that the connection between data is
reduced and the independence between data is enhanced. This paper mainly uses SSVM as mentioned
above to reduce the similar attributes on the data dimension. The more similar the attribute is, the more
redundant it would be. The data redundancy between the pruned data will be reduced so that the
independence of the data dimension can be enhanced.
SSVM is used to select the appropriate data for pruning. When the data is removed excessively or
removed too little, the classification result will be affected. Therefore, it is very important to choose
the appropriate threshold. The threshold in this article is the accuracy rate of test data tested by
SVM. When the accuracy is greater than a certain value, we think that these kinds of data are not
redundant, so we do not prune it. Therefore, the classification data, which is below the threshold,
is selected, and then remove the nearest neighbor inconsistent data. Table 4 shows the selected data
corresponding to the pruning data and the pruned training data set, and Figure 8 is the test accuracy
of the bearing data corresponding to the selection threshold. From Table 4 and Figure 8, it can be
concluded that the data trimming is too small to make the classification effect not obvious, and too
much data pruning will result in important data loss. It can be seen from Figure 8 that, when the
threshold is 0.9, the corresponding accuracy is the highest than others. Therefore, the training data
with a threshold below 0.9 is selected for SSVM pruning. Only in this way can the fault diagnosis be
performed effectively.

0.9917

0.991
(0.9 , 0.9917)
0.99

0.989
accurancy

0.988

0.987

0.986

0.985

0.984
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
Threshold

Figure 8. The accuracy of the data corresponding to the threshold.

Table 4. The corresponding threshold data.

Threshold (Training Accuracy) 1 0.95 0.90 0.85 0.80 0.60

The number of pruning 190 187 179 155 102 0
The number of training 536 539 547 571 624 726

After processing, the vibration data from the three aspects above, the redundant data is
removed from the feature vector and the dimension vector, respectively. Figure 9 shows the
three-dimensional data of time domain feature extractiont, three-dimensional data after J48 select
feature, and three-dimensional data after J48 and SSVM trimming. The axes x, y, and z in Figure 9 are
dimensional features. Among them, Figure 9a selects three dimensions of mean, absolute mean and
effective value, Figure 9b,c select three dimensions of mean, waveform index and kurtosis index. It can
be seen from Figure 9 that each class of data has obvious overlap in Figure 9a, the overlap ratio of each
Sensors 2018, 18, 463 13 of 17

kind of data in Figure 9b is obviously lower than of Figure 9a, and Figure 9c obviously separates each
type of category data. Therefore, it is shown from Figure 9 that the redundancy between the processed
data is greatly reduced, so that the correlation between the data is reduced, and the influence of NB
independence assumption on the fault diagnosis is finally reduced.

Normal

Inner race fault

0.5 Ball fault

Out race fault at center@6:00

0.4
Effective value

Out race fault at center@3:00

0.3
Out race fault at center@12:00
(a)

0.2

0.1

0
1
0.04
0.5 0.02
0
absolute mean 0 -0.02 mean

Normal
# 105
Inner race fault
2
Ball fault

Out race fault at center@6:00

1.5
kurtosis index

Out race fault at center@3:00

(b)

1 Out race fault at center@12:00

0.5

0
2.5
2 0.08
0.06
0.04
1.5 0.02
0
waveform 1 -0.02 mean

Normal
# 105
Inner race fault
2
Ball fault

Out race fault at center@6:00

1.5
kurtosis index

Out race fault at center@3:00

1 Out race fault at center@12:00

(c)

0.5

0
2.5
2 0.1
0.05
1.5
0
waveform 1 -0.05 mean

Figure 9. Bearing data description. (a) the original signal time domain feature extraction fault
three-dimensional; (b) J48 select the characteristics of the three-dimensional fault data diagram; (c) the
three-dimensional fault data diagram after J48 and SSVM pruning.
Sensors 2018, 18, 463 14 of 17

The processing bearing fault data correlation is low, which reduces the limitation of the
independence assumption on NB fault diagnosis. Table 5 is the confusion matrix of NB fault diagnosis
for the processed data, and Table 6 is a confusion matrix for bearing fault diagnosis using an NB model
without redundant vibration data. As can be seen from the table, the model has been improved for
each category after redundancy removal.

Table 5. Confusion matrix of the processing bearing fault data on test sets.

Predicted Classes
Actual Classes
1 2 3 4 5 6
1 121 0 0 0 0 0
2 0 121 0 0 0 0
3 0 0 121 0 0 0
4 0 0 0 121 0 0
5 0 0 0 0 121 0
6 0 0 0 0 0 121

Table 6. Confusion matrix of NB on test sets.

Predicted Classes
Actual Classes
1 2 3 4 5 6
1 120 0 1 0 0 0
2 0 121 0 0 0 0
3 0 0 117 0 0 4
4 0 1 0 117 2 1
5 0 1 0 3 117 0
6 0 0 0 0 0 121

testing accuracy NB
0.98
Clafficies Accuracy

0.96 testing accuracy NB+J48

0.94
testing accuracy NB+J48+SVM

0.92
testing accuracy NB+J48+SSVM

0.9
class 1 class 2 class 3 class 4 class 5 class 6

Figure 10. Testing accuracy comparison of each condition in the experiment.

In order to verify the validity of this algorithm in bearing data, the data simulation is carried
out by MATLAB (Version 8.6, The MathWorks, MA, USA). Figure 10 and Table 7 are bearing fault
diagnosis results. In Figure 10 and Table 7,the meaning of NB+J48+SVM is that first data is selected
by J48,then the data after feature selection is pruned by SVM and the fault diagnosis of NB is finally
carried out. Compared with other experimental results, the bearing fault diagnosis experimental
results on JSSVM-NB is better than removing the data redundancy by feature vector and data vector.
Compared with other experiments, the accuracy of the fault diagnosis model is 99.17%. Table 8 shows
the comparison of results of about JSSVM-NB and reference [35], which have the same data for bearing
Sensors 2018, 18, 463 15 of 17

fault diagnosis. It can be seen from Tables 7 and 8 that the JSSVM-NB model is effective for rolling
bearing fault diagnosis.

Table 7. The corresponding threshold data.

Methods Accuracies
NB 98.21%
NB + J48 + SVM 98.48%
NB + J48 + SSVM (JSSVM-NB) 99.17%

Table 8. The comparison results in bearing fault diagnosis.

State JSSVM-NB Reference [35]

Normal 100% 98.31%
Inner race 100% 97.73%
Ball 97.5% 95.04%
Out race fault at center (@6:00, @3:00 and @12:00) 99.17% 98.02%

5. Conclusions
In this paper, in order to improve the independence assumption, the bearing data processing is
carried out from two aspects of the attribute vector and the dimension vector, and the bearing data
with higher data independence is obtained for the bearing fault diagnosis of the NB. NB is based on
the conditional independence hypothesis of Bayes. However, in the actual case, it is difficult for the
bearing data vector to achieve independence. Therefore, the redundancy is removed from the feature
attribute vector and dimension of bearing data in this paper, so that the connection between data is
reduced and the bearing condition monitoring on NB be enhanced. It be seen from the simulation
results. The NB improved the data independence has realized the fault diagnosis of the different parts
of rolling bearing, and can be applied to the other fault diagnosis of the industrial.

Acknowledgments: This work was supported by the National Natural Science Foundation of China (61702348,
61772351, 61602326), the National Key Technology Research and Development Program (2015BAF13B01, the
National Key R&D Plan (2017YFB1303000, 2017YFB1302800), and the Project of the Beijing Municipal Science
& Technology Commission (LJ201607). The work is also supported by the Youth Innovative Research Team of
Capital Normal University.
Author Contributions: Nannan Zhang and Lifeng Wu conceived and designed the experiments; Jing Yang
performed the experiments; Yong Guan analyzed the data; Lifeng Wu contributed analysis tools; Nannan Zhang
wrote the paper. All authors contributed to discussing and revising the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Jacobs, W.; Malag, M.; Boonen, R.; Moens, D.; Sas, P. Analysis of bearing damage using a multibody model
and a test rig for validation purposes. Struct. Health Monit. 2011, 14, 971–978.
2. Li, W.; Mechefske, C.K. Detection of induction motor faults: A comparison of stator current, vibration and
acoustic methods. J. Vib. Control 2006, 12, 165–188.
3. Ding, S.X.; Yang, Y.; Zhang, Y.; Li, L. Data-driven realizations of kernel and image representations and their
application to fault detection and control system design. Automatica 2014, 50, 2615–2623.
4. Chadli, M.; Abdo, A.; Ding, S.X. H-/H∞ fault detection filter design for discrete-time Takagi-Sugeno fuzzy
system. Automatica 2013, 49, 1996–2005.
5. Chibani, A.; Chadli, M.; Peng, S.; Braiek, N.B. Fuzzy Fault Detection Filter Design for T-S Fuzzy Systems in
Finite Frequency Domain. IEEE Trans. Fuzzy Syst. 2016, 25, 1051–1061.
6. Youssef, T.; Chadli, M.; Karimi, H.R.; Wang, R. Actuator and sensor faults estimation based on proportional
integral observer for TS fuzzy model. J. Frankl. Inst. 2016, 354, 2524–2542.
Sensors 2018, 18, 463 16 of 17

7. Paya, B.; Esat, I.I.; Badi, M.N.M. Artificial Neural Network Based Fault Diagnostics of Rotating Machinery
Using Wavelet Transforms as a Preprocessor. Mech. Syst. Signal Process. 1997, 11, 751–765.
8. Lynagh, N.; Rahnejat, H.; Ebrahimi, M.; Aini, R. Bearing induced vibration in precision high speed routing
spindles. Int. J. Mach. Tools Manuf. 2000, 40, 561–577.
9. Wardle, F.P. Vibration Forces Produced by Waviness of the Rolling Surfaces of Thrust Loaded Ball Bearings
Part 1: Theory. Arch. Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 1988, 202, 305–312.
10. Mevel, B.; Guyader, J.L. Routes to Chaos in Ball Bearings. J. Sound Vib. 2007, 162, 471–487.
11. Vafaei, S.; Rahnejat, H. Indicated repeatable runout with wavelet decomposition (IRR-WD) for effective
determination of bearing-induced vibration. J. Sound Vib. 2003, 260, 67–82.
12. Lei, Y.; Lin, J.; He, Z.; Zi, Y. Application of an improved kurtogram method for fault diagnosis of rolling
element bearings. Mech. Syst. Signal Process. 2011, 25, 1738–1749.
13. Wang, T.; Qi, J.; Xu, H.; Wang, Y.; Liu, L.; Gao, D. Fault diagnosis method based on FFT-RPCA-SVM for
Cascaded-Multilevel Inverter. ISA Trans. 2016, 60, 156–163.
14. Samanta, B.; Al-Balushi, K.R. Artificial Neural Network Based Fault Diagnostics of Rolling Element Bearings
Using Time-Domain Features. Mech. Syst. Signal Process. 2003, 17, 317–328.
15. Jack, L.B.; Nandi, A.K. Support vector machines for detection and characterization of rolling element bearing
faults. Arch. Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 2001, 215, 1065–1074.
16. Yang, Y.; Yu, D.; Cheng, J. A roller bearing fault diagnosis method based on EMD energy entropy and ANN.
J. Sound Vib. 2006, 294, 269–277.
17. Al-Raheem, K.F.; Roy, A.; Ramachandran, K.P.; Harrison, D.K.; Grainger, S. Application of the
Laplace-Wavelet Combined with ANN for Rolling Bearing Fault Diagnosis. J. Vib. Acoust. 2008,
130, 3077–3100.
18. Shuang, L.; Meng, L. Bearing Fault Diagnosis Based on PCA and SVM. In Proceedings of the International
Conference on Mechatronics and Automation, Harbin, China, 5–8 August 2007; pp. 3503–3507.
19. Wu, L.; Yao, B.; Peng, Z.; Guan, Y. Fault Diagnosis of Roller Bearings Based on a Wavelet Neural Network
and Manifold Learning. Appl. Sci. 2017, 7, 158, doi:10.3390/app7020158.
20. Sugumaran, V.; Sabareesh, G.R.; Ramachandran, K.I. Fault diagnostics of roller bearing using kernel based
neighborhood score multi-class support vector machine. Expert Syst. Appl. 2008, 34, 3090–3098.
21. Sugumaran, V.; Ramachandran, K.I. Effect of number of features on classification of roller bearing faults
using SVM and PSVM. Expert Syst. Appl. 2011, 38, 4088–4096.
22. Zhang, R.; Peng, Z.; Wu, L.; Yao, B.; Guan, Y. Fault Diagnosis from Raw Sensor Data Using Deep Neural
Networks Considering Temporal Coherence. Sensors 2017, 17, 549.
23. Zhang, R.; Wu, L.; Fu, X.; Yao, B. Classification of bearing data based on deep belief networks. In Proceedings
of the Prognostics and System Health Management Conference, Chengdu, China, 19–21 October 2017;
pp. 1–6.
24. Mohamed, E.A.; Abdelaziz, A.Y.; Mostafa, A.S. A neural network-based scheme for fault diagnosis of power
transformers. Electr. Power Syst. Res. 2005, 75, 29–39.
25. Sharma, R.K.; Sugumaran, V.; Kumar, H.; Amarnath, M. A comparative study of NB classifier and Bayes net
classifier for fault diagnosis of roller bearing using sound signal. Decis. Support Syst. 2015, 1, 115.
26. Mccallum, A.; Nigam, K. A Comparison of Event Models for NB Text Classification. In Proceedings of the
AAAI-98 Workshop on Learning for Text Categorization, Madison, WI, USA, 26–27 July 1998; Volume 62,
pp. 41–48.
27. Kumar, H.; Ranjit Kumar, T.A.; Amarnath, M.; Sugumaran, V. Fault Diagnosis of Bearings through Vibration
Signal Using Bayes Classifiers. Int. J. Comput. Aided Eng. Technol. 2014, 6, 14–28.
28. Krishna, H. Fault Diagnosis of welded joint through vibration signals using NB Algorithm. In Proceedings
of the International Conference on Advances in Manufacturing and Material Engineering, Mangalore, India,
27–29 March 2014.
29. Sugumaran, V.; Muralidharan, V.; Ramachandran, K.I. Feature selection using decision tree and classification
through Proximal Support Vector Machine for fault diagnostics of roller bearing. Mech. Syst. Signal Process.
2007, 21, 930–942.
30. Quinlan, J.R. Improved Use of Continuous Attributes in C4.5. J. Artif. Intell. Res. 1996, 4, 77–90.
31. Brereton, R.G.; Lloyd, G.R. Support vector machines for classification and regression. Analyst 2010,
135, 230–267.
Sensors 2018, 18, 463 17 of 17

32. Cristianini, N.; Shawe-Taylor, J. An Introduction to Support Vector Machines: And Other Kernel-Based Learning
Methods; Cambridge University Press: Cambridge, UK, 2000; pp. 1–28.
33. Zhang, X.; Liang, Y.; Zhou, J.; Zang, Y. A novel bearing fault diagnosis model integrated permutation entropy,
ensemble empirical mode decomposition and optimized SVM. Measurement 2015, 69, 164–179.
34. Loparo, K. Bearings Vibration Data Set; Case Western Reserve University: Cleveland, OH, USA. Available
online: http://www.eecs.case.edu/laboratory/bearing/welcome-overview.htm (accessed on 20 July 2012).
35. Wu, S.D.; Wu, C.W.; Wu, T.Y.; Wang, C.C. Multi-Scale Analysis Based Ball Bearing Defect Diagnostics Using
Mahalanobis Distance and Support Vector Machine. Entropy 2013, 15, 416–433.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).