Measurement 95 (2017) 355–366

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Fault diagnosis of rolling element bearings based on Multiscale Dynamic

Time Warping
Tian Han a,⇑, Xueliang Liu a, Andy C.C. Tan b,c
a
School of Mechanical Engineering, University of Science and Technology Beijing, 30 Xueyuan Road, Haidian District, Beijing 100083, China
b
Faculty of Engineering and Science, Queensland University of Technology, George Street, Brisbane, Qld 4001, Australia
c
Faculty of Engineering, Universiti Tunku Abdul Rahman, Sungai Long Campus, Cheras 43000, Kajang, Selangor, Malaysia

a r t i c l e i n f o a b s t r a c t

Article history: A fault recognition method for rolling element bearings based on Multiscale Dynamic Time Warping
Received 12 November 2015 (MDTW) is proposed in this paper. After preprocessing using Empirical Mode Decomposition (EMD),
Received in revised form 8 July 2016 CWs are extracted from vibration signals. The normalization of CWs is necessary to eliminate the influ-
Accepted 12 October 2016
ence of amplitude variations before using MDTW. Following this process, one CW of the normal condition
Available online 13 October 2016
is selected as a template to calculate the distance of DTW (DDTW) with other CWs. The calculated DDTW
results can then be utilized to classify the bearing conditions since different conditions have different
Keywords:
DDTW bands. The proposed method is validated by the data from Bearing Data Centre of Case Western
Multiscale Dynamic Time Warping
Characteristic waveform
Reserve University. The analysis results indicate that the influence of variable speed and different defect
Rolling element bearings sizes can be effectively eliminated by DTW.
Fault recognition Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction an assumed constant rotating speed. In order to extract the fault

pattern of roller bearings, the measured frequency components
Rotating machines are prime movers of process industry. With are compared with the theoretical values. However, the vibration
challenging demands to extent their useful life and increase pro- signals of rotating machinery are non-stationary and have non-
duction capacity, modern rotating machineries are required to linear characteristics. Thus, it is difficult to directly extract fault
operate at high-speeds and are usually large in size. These machi- information from the original measured signals. With the progress
nes have many rotating parts and a fault in one of the components of computing capability, bearing fault diagnosis technology has
can lead to catastrophic failure of the whole system which will become an integrated discipline of computer, digital signal pro-
affect the safety of operation and subsequent production loss. Roll- cessing, artificial intelligence and other technologies.
ing element bearings are critical components of rotating machin- Compared with the conventional methods, current researches
ery and are confronted with vulnerability due to unsteady loads have focused on the following techniques: Continuous Wavelet
and speeds, and the corrosive effects of the working environment. Transform [3] which focused on the selection of real wavelet and
Conventional fault diagnostic techniques of rolling element complex wavelet according to the Maximum Energy to Shannon
bearings utilized vibration, acoustic, temperature and oil measure- Entropy ratio and Maximum Relative Wavelet Energy criterion is
ments to assess the condition of the bearings [1]. These techniques used to extract the statistical properties of the wavelet coefficients
are effective if the operating characteristics do not change signifi- of raw vibration signals and finally exploit these statistical proper-
cantly during operation. Vibration diagnostic technique, which is ties as inputs of machine learning techniques to achieve fault diag-
commonly used in bearings fault diagnosis, provides an effective nosis of rolling element bearings. Expert System [4] is established
and easy approach in analyzing the signals obtained from the sen- to verify the correctness of the algorithm utilizing the processed
sors [2]. The key step of fault recognition of rolling element bear- data and gathered a large number of fault features from the vibra-
ings based on vibration analysis is to process the measured tion signals to build an experienced database. It combines the algo-
signal correctly to extract the fault characteristics, which is directly rithm with the database to complete the fault diagnosis. Fuzzy
related to the theoretical calculations of the bearing dynamics with Diagnosis [5] deals with the change of inherent complexity of the
vibration waveform when a fault has occurred in a roller bearing.
Fuzzy entropy is used to measure the complexity and the self-
⇑ Corresponding author. similarity of the time series in fault diagnosis. Neural Network
E-mail address: [email protected] (T. Han).

http://dx.doi.org/10.1016/j.measurement.2016.10.038
0263-2241/Ó 2016 Elsevier Ltd. All rights reserved.
356 T. Han et al. / Measurement 95 (2017) 355–366

[6,7] obtained the mode network signals through the study of The rest of the paper will be organized as follows. Section 2, the
known samples to accomplish the pattern recognition task of roll- theory of MDTW is briefly presented. Section 3 is dedicated to the
ing element bearing faults. This identification process is essentially resource of data collection, the methods of processing the vibration
to match the unidentified data with the prior data. The neural net- signals, and the proposed approach applied for fault diagnosis of
work classifier has the same problem as the traditional pattern rolling element bearings. Section 4, variable speed and different
recognition methods, which requires a significant amount of learn- severity are discussed to present the properties of MDTW. Section 5
ing samples. Unfortunately, in most practical situations, it is diffi- gives the conclusion.
cult to obtain typical fault samples, which is one of the limitations
of the techniques. Different types of rolling element bearings exhi-
bit different characteristics which can be extracted directly in the 2. Multiscale Dynamic Time Warping
time domain signals. Compared with characteristic statistics (such
RMS and impulse index), the time-domain Characteristic Wave- DTW is a distance measurement technique which is ideal for
form (CW) is intuitive, clear physical representation, and compre- comparing time series. It differs from Euclidean distance (ED) by
hensive reflection of the inherent characteristics of bearing allowing the vector components to compare with the ‘‘drift” from
vibration. Compared with transformed feature (characteristic fre- the corresponding exact positions. It is an algorithm for measuring
quency, wavelet coefficients, etc.), it is easy to access and does similarity between two sequences (e.g. time series) which may
not need any space transformation. Hence, fault diagnosis of rolling vary in time or speed. The sequences are ‘‘warped” non-linearly
element bearings based on time-domain CW signal matching [8,9] in the time dimension to determine a measure of their similarity
is an attractive area of research. which is independent of non-linear variations in the time domains
Recently, the study on time series has obtained widespread [25].
interest. One of the major applications of time series is in classifi- A time series is a sequence of observation which follows a cer-
cation field. For example, two signals in the time domain having tain order in time or space. For simplicity and without any loss of
similar overall component shapes may be out of synchronization generality, we assume that time is discrete within a certain period.
and generally are not of exactly the same length. In order to find The basic algorithm of DTW is as follows. Given two time series
the exact dissimilarity between these two signals and as a pre- of lengths N and M
processing step before comparing them, it is necessary to match
them to an appropriate alignment. However, many methods have X ¼ ðx1 ; x2 ; x3 ; . . . ; xN Þ ð2-1Þ
been applied using time domain waveform matching, which
required the length, dimension and phase of two waveforms to Y ¼ ðy1 ; y2 ; y3 ; . . . ; yM Þ ð2-2Þ
be the same. These requirements can be hard to achieve which
where xi and yj are represented by the sequences of values at point i
restrict its application and development. Dynamic Time Warping
and j in series X and Y, respectively.
(DTW) is a well-known technique for measuring distance between
To align the two time series for comparison, a N  M distance
two time sequences that are similar but locally out of phase
matrix D is constructed first. The element of matrix D is the dis-
through finding an optimal alignment. In the early stage, DTW
tance between the two points xi and yj which is represented by d
was mainly used in the recognition of isolated words and speeches.
(xi, yj). Typically, the Euclidean distance is used to calculated the
Since then, it has been employed for clustering and classification in
point-to-point distance by
many domains: database index [10,11], handwriting recognition
[12], in the field of biological engineering and biomedical informat-
dðxi ; yj Þ ¼ ðxi  yj Þ2 ð2-3Þ
ics [13], and gene regulatory network [14]. In the field of computer
vision, it was applied to gesture recognition and human activity Once the distance matrix is established, the path through the
recognition [15,16]. In addition, its application in signal processing matrix with minimal cumulative distance between the time series
[17,18], electrocardiogram analysis [19,20], and data mining [17] can be calculated. DTW distance corresponds to the path with min-
and process monitoring [21] have shown strong vitality. Recently, imal warping cost:
Zhen et al. [22] have explored its application in processing data
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
from motors for condition monitoring and shown promising XK
results. Dai et al. [23] have put forward a new fault diagnosis DTWðx; yÞ ¼ min k¼1 k
w ð2-4Þ
methodology for batch chemical processes, using a dynamic time
warping-based artificial immune system and experiments show where wk is the matrix element that also belongs to the element of a
that it has the capability of fault diagnosis of batch processes. Zhao warping path W.
et al. [24] have studied a new online fault diagnosis system for full
operating cycles of chemical processes and the dynamic artificial W ¼ ðw1 ; w2 ; w3 ; . . . ; wK Þ; wk ¼ ðnk ; mk Þ ð2-5Þ
immune system based on dynamic time warping is used for fault
identification in both the startup phase and the steady state, and wherenk ¼ 1; 2; 3; . . . ; N; mk ¼ 1; 2; 3; . . . ; M; k ¼ 1; 2; 3; . . . ; K;
shows it has high adaptation capability. maxðN; MÞ 6 K 6 N þ M
In this paper, DTW is applied for fault diagnosis of rolling ele- The warping path is subjected to three constraints:
ment bearings using the vibration signals. Based on time waveform
analysis of the vibration signals, CWs of different rolling element (1) Boundary condition: w1 = (1, 1) and wK = (N, M) the starting
bearing conditions are extracted. After normalization and Empiri- and ending points of the warping path must be the first
cal Mode Decomposition processing, different DDTW bands are and the last points of the aligned time series.
obtained through calculating DTW of the template and other (2) Monotonicity condition: Given wk = (nk, mk) then wk1 =
extracted CWs. Finally, fault recognition can be accomplished (nk1, mk1), where ðnk  nk1 Þ P 0 and ðmk  mk1 Þ P 0.
through matching the test CWs with the template. This technique This forces the points in w to be monotonically spaced in
has many advantages, among them, it ignores the diversity of time.
lengths and phases of two waveforms. Furthermore, the influence (3) Continuity condition: Given wk = (nk, mk) then
of speed fluctuations and load conditions can be overcome during ðnk  nk1 Þ 6 1, ðmk  mk1 Þ 6 1; so the continuity condition
fault classification. is formulated as wk  wk1 e {(1, 1), (1, 0), (0, 0)}
 T. Han et al. / Measurement 95 (2017) 355–366 357

11 11
Signal X Signal X
10 Signal Y 10 Signal Y
9 9
8 8
7 7
Amplitude

Amplitude
6 6
5 5
4 4
3 3
2 2
1 1
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Data points Data points
(a) Euclidean Distance (b) Distance of DTW

X: 1 1 1 1 10 3 3 3 3 3
9 9 9 9 36 1 1 1 1 1
4

9 9 9 9 36 1 1 1 1 1
4

9 9 9 9 36 1 1 1 1 1
4

9 9 9 9 36 1 1 1 1 1
4 10

81 81 81 81 0 49 49 49 1 1

1 1 1 1 64 1 1 1 1 1
2

1 1 1 1 64 1 1 1 1 1
2

1 1 1 1 64 1 1 1 1 1
2

1 1 1 1 64 1 1 1 1 1
2

1 1 1 1 64 1 1 1 1 1
2

(c) The best warping path

:Y

Fig. 1. The ED and DDTW.

This criterion limits the warping path from long jumps while the DTW calculation, which is widely used to quantify the com-
aligning sequences, and restricts the allowable steps in the warp- plexity of time series for a range of scales in multiscale entropy
ing path to adjacent cells. [26]. To obtain the coarse-grained time series at a scale factor of
Dynamic programming is used to find this warping path by s, the original time series x is divided into disjointed windows of
applying the following recurrence: length s, and the data points are averaged within each window.

cij ¼ dðxi ; yj Þ þ minfcði  1; j  1Þ; cði  1; jÞ; cði; j  1Þg ð2-6Þ

Scale factor τ =2
where d(xi, yj) the distance is found in the current cell, and c(i, j) is
the cumulative distance of d(xi, yj) and the minimum cumulative x1 x2 x3 x4 x5 x6 xi xi+1 xi+2
distances from the three adjacent cells.
As shown in Fig. 1, the template X [1, 1, 1, 1, 10, 3, 3, 3, 3, 3] and
the test signal Y [2, 2, 2, 2, 2, 10, 4, 4, 4, 4] have high similarity exclu-
sion of peak shift. According to Eqs. (2-3) and (2-6), the result of
y1(2) y2(2) y3(2) y(i+1)/2(2) = xi+xi+1 /2
the ED between X and Y is 121, while DDTW is 10. Obviously,
Scale factor τ =3
DDTW is much smaller than ED. The two signals are aligned by
DTW due to the warp of time and a best match is realized between
x1 x2 x3 x4 x5 x6 xi xi+1 xi+2
them through the best warping path. So, DDTW is effective to
express the similarity of two signals than ED.
Dynamic programming is used to calculate the similarity
between two time series, which is a technique of time complexity
O(n2). For two long time series, it is necessary to improve the cal- y1(3) y2(3) y(i+1)/2(3) = xi+xi+1+ xi+2 /3
culation efficiency. There are many methods to reduce or resample
the number of time series. In this paper, multiscale is introduced in Fig. 2. Schematic illustration of the coarse-graining procedure.
358 T. Han et al. / Measurement 95 (2017) 355–366

As shown in Fig. 2, two and three coarse-grained time series are 3.1. Vibration data
divided from the original time series using time scale factor and
are illustrated in the figure. The k-th coarse-grained time series All the bearing vibration data analyzed in this paper came from
is defined as follows. the Bearing Data Centre of Case Western Reserve University
(CWRU) (www.eecs.cwru.edu/laboratory/bearing). The deep
jsX
þk1   groove ball bearing (6205-2RS JEM SKF) was used in the tests. As
1 N
ysk;j ¼ xi ; 16j6 ; 16k6s ð2-7Þ shown in Fig. 4, the test rig consists of a 2 hp, three-phase induc-
s i¼ðj1Þsþk s
tion motor (left), a torque transducer (middle), and a dynamometer
(right). The torque transducers are used to collect speed and horse-
where s = 1, 2, . . . is the scale factor. With s equals to one, the power data of the motor. The dynamometer can be controlled to
coarse-grained time series is the original time series, and for a obtain the desired torque load levels. The test bearing supports
non-zero series, the original time series is divided into coarse the motor shaft at the drive end. Vibration data was collected using
grained vector series with length of N/s. accelerometers, which were attached on the housing with mag-
netic bases. Accelerometers were placed at 12 o’clock position at
both ends of the drive shaft. Single point faults with fault diame-
3. Fault recognition method based on MDTW ters of 7 mils, 14 mils and 21 mils (1 mil = 0.001 in.) were intro-
duced into the test bearing using electro-discharge machine. The
In this paper, fault recognition method based on MDTW is pro- vibration data were collected using a digital recorder with a sam-
posed for rolling element bearings fault detection. The flowchart of pling frequency of 12 kHz for different bearing conditions. The
the proposed method is illustrated in Fig. 3. First the vibration sig- tests were carried out under different loads varying from 0 to
nals of different fault conditions were obtained using accelerome- 3 hp with 1 hp increments. The corresponding speed varies from
ters. In order to improve the signal-noise ratio, EMD was applied to 1772 to 1730 rpm. Bearing faults were induced in the inner race,
reduce the effect of background noise. Then, data analysis tech- outer race and rolling element. This section describes the analysis
nique is used to extract the CWs from the de-noise signal, resulting of one working condition data as listed in Table 1.
in one normal CW being selected as the standard template. Then
normalization is preprocessed to erase the effect of amplitude dis- 3.2. Data analysis
crepancy. Finally, the bearing conditions are classified based on the
DDTW. The four types of bearing vibration signals and their corre-
sponding frequency spectra are illustrated in Fig. 5.
From Fig. 5(a), the time waveform or the normal bearing shows
periodic background noise, and impact waveforms generated by
Vibraon data other bearing and motor components. The impacts on the outer
Preprocessing
(Bearing Data CW extraon race and inner race bearing faults are more obvious as shown in
(EMD)
Centre of CWRU) the third and fourth plots of Fig. 5(a). However, from frequency
spectrum plots, it is difficult to identify the bearing fault character-
istic frequency components due to noise. For instance, the inner
race fault frequency is about 155.5 Hz based on theoretical calcula-
MDTW bands tions and is not visible in the frequency spectrum. The fault signals
Classificaon Normalizaon are overwhelmed by the background noise and the amplitude of
calculaon
155.6 Hz is too low to be identified since fault impact is in the form
of shock pulse it has a broad frequency spectrum. Furthermore the
Fig. 3. MDTW classification chain. inherent frequencies of the sensors, shaft and bearing parts are
modulated which lead to the generation of a large number of spec-
tral lines.
In order to comprehensively analyze the time waveform and
motor torque frequency spectrum, de-noising of the raw signals are necessary
transducer
before carrying out fault diagnosis. EMD used as a pretreatment
is used to decompose the non-stationary vibration signal into a
number of Intrinsic Mode Functions (IMF) [28].
fan end
3.3. EMD method

drive end EMD method is based on the assumption that any signal can be
composed into different simple intrinsic modes of oscillations.
Each mode should have the same number of zero-crossings and
Fig. 4. Rolling element bearings test rig [27]. extrema. There exists only one extremum between successive
zero-crossings. Each mode is independent of the others. Therefore,

Table 1
Working conditions of test bearing.

Bearing type Motor load (HP) Fault description (mils) Motor speed (rpm) Bearing conditions
SKF6205 3 Diameter: 21 1730 Ball fault
Inner race fault
Depth: 11 Outer race fault
(Centered @6:00)
 T. Han et al. / Measurement 95 (2017) 355–366 359

Normal FFT of Normal

0.5 0.04

0 0.02

-0.5 0
0 500 1000 1500 2000 0 200 400 600 800 1000
Ball FFT of Ball
0.5 0.02

0 0.01

-0.5 0
Amplitude

0 500 1000 1500 2000 0 200 400 600 800 1000

Inner race FFT of Inner race
5 0.1
X: 155.6
0 0.05 Y: 0.01265

-5 0
0 500 1000 1500 2000 0 200 400 600 800 1000
Outer race FFT of Outer race
5 0.1

0 0.05

-5 0
0 500 1000 1500 2000 0 200 400 600 800 1000
Data points Frequency(Hz)
(a) Time domain (b) Frequency domain

Fig. 5. Bearing vibration signals and their corresponding frequency spectrum.

each signal could be decomposed into a number of IMFs and each (6) Repeat the process and get the second IMF component c2(t).
IMF must satisfy the following two requirements: Repeat the process until the n-th IMF component cn(t) or its
allowance rn(t) is less than the default value, or when the
(1) On the whole signal, the number of extrema and the number residual component rn(t) is a monotonic function or a con-
of zero-crossings can differ at most by one. stant, the EMD decomposition is finished.
(2) At any point, the mean value of the enveloped local maxima (7) Finally, x(t) is composed by EMD as
and local minima is zero [29].
X
n
xðtÞ ¼ ci ðtÞ þ rn ðtÞ ð3-5Þ
An IMF represents a simple oscillatory mode compared with a i¼1
simple harmonic function. Based on the assumption, any signal
can be decomposed as follows:
The IMFs include different frequency bands ranging from low
to high. The frequency components contained in each frequency
(1) Find the maximum value points of the original signal, and fit
band are different and change with the variation of the original
the upper envelop e+(t) of local maxima using a cubic spline.
signal. Fig. 6 shows the result of a ball bearing’s vibration signal
In the same way, find all the minima of the original signal
after using EMD and it takes 5.963 s to complete the process.
and get the lower envelop e(t).
Though real-time de-noising always cause additional delays, as
(2) Make the mean value of the upper and lower envelop as the
long as only past measured values are used, the occurrence
mean envelop of the original signal, that is
and development of the fault need some time and lag the diag-
eþ ðtÞ þ e ðtÞ nosis by 5.963 s. This time has minimum effect in practical
m1 ðtÞ ¼ ð3-1Þ
2 situations.
1
(3) Get a new signal h1 ðtÞ with the low frequency removed by
subtracting m1(t) from the original signal, that is 3.4. Selection of CW
1
h1 ðtÞ ¼ xðtÞ  m1 ðtÞ ð3-2Þ After the original signal is decomposed by EMD, a number of
1
IMFs can be obtained. The Domain IMF can represent the charac-
(4) Generally, h1 ðtÞ
is not a stationary signal and it does not sat- teristic of the fault more precisely than the others IMFs. Thus, it
isfy the two conditions defined by IMF. The procedure is is important to find the Domain IMF based on the Fault Leading
k
repeated, if h1 ðtÞ satisfy the definition of IMF after k cycles, Rate Evaluation Algorithm before extracting the characteristic
the first order IMF component of the original signal is waveforms. This is carried out by the following steps:
k
c1 ðtÞ ¼ imf1 ðtÞ ¼ h1 ðtÞ ð3-3Þ
(1) Decompose the original fault signal x(t) which contains fault
(5) Get a new signal r1(t) with the high frequency removed by and noise to obtain the IMF components ci(t)(i = 1, 2, . . ., n)
subtracting c1(t) from the original signal, that is
X
n

r1 ðtÞ ¼ xðtÞ  c1 ðtÞ ð3-4Þ xðtÞ ¼ ci ðtÞ þ rn ðtÞ ð3-6Þ

i¼1
360 T. Han et al. / Measurement 95 (2017) 355–366

Outer race
6
4

Amplitude
2
0
-2
-4
-6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)
EMD
5
IMF1

0
-5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
1
IMF2

0
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
1
IMF3

0
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
1
IMF4

0
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.5
IMF5

0
-0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.1
IMF6

0
-0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.05
IMF7

0
-0.05
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.05
IMF8

0
-0.05
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.05
IMF9

0
-0.05
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.02
IMF10

0
-0.02
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.01
IMF11

0
-0.01
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
0.01
IMF12

0
-0.01
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
EMD
IMF13

0.02
0.01
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (s)

Fig. 6. EMDs of bearing outer race fault.

 T. Han et al. / Measurement 95 (2017) 355–366 361

(2) Compute the correlation coefficient ai(t)(i = 1, 2, . . ., n) Table 3

between the original fault signal x(t) and the IMF compo- Geometry details of SKF6205 ball bearing.

nents ci(t)(i = 1, 2, . . ., n). Outer race Inner race Number of Contact Balls
R þ1 diameter diameter balls angle diameter
xðtÞci ðtÞdt
ai ¼ R þ1 1
R þ1 2 1=2 i ¼ 1; 2; . . . ; n ð3-7Þ 52 (mm) 25 (mm) 9 0° 6.75 (mm)

1
x2 ðtÞdt  1 i
c ðtÞdt
(3) Mark the normal signal corresponding to the original fault
signal x(t) as xnor(t) and compute the correlation coefficient Table 4
The characteristics frequencies of faulty bearing at 1730 rpm.
bi(t)(i = 1, 2, . . ., n) between the normal signal xnor(t) and
the IMF components ci(t)(i = 1, 2, . . ., n). Bearing fault Normal (rotating Inner Outer Rolling
R þ1 frequency) race race fault element
1
xnor ðtÞci ðtÞdt fault fault
bi ¼ R R þ1 2 1=2 i ¼ 1; 2; . . . ; n ð3-8Þ
þ1 Characteristics 28.8 156 103.2 67.9
1
x2nor ðtÞdt  1 i
c ðtÞdt
frequency (Hz)
(4) Define the fault factor as ni
jai  bi j
ni ¼ Pn  100% ð3-9Þ malfunction, d is the rolling element diameter, D is pitch diameter,
i¼1 jai  bi j
n is the number of rolling elements and a is contact angle.
(5) According to the values of the fault factor ni (i = 1, 2, . . ., n), According to the bearing parameters as listed in Table 3, the
the largest value is the Domain IMF. theoretical characteristics frequencies of faulty bearing are listed
in Table 4.
From the Table 2, the n with IMF1(65%) is the largest of all, as a As mentioned above, the vibration data were collected with a
result, IMF1 is regarded as the Domain IMF used in this work and sampling frequency of 12 kHz. The data points of one period
the other IMFs are ignored. The rest of the original fault signals can (L = 1 T) of CW for each condition is calculated as follows:
be dealt with using the same approach.
The selection of CW is crucial in the proposed method, which Sample Frequency 12; 000
NN ¼ ¼  417 ð3-13Þ
can directly affect the accuracy of fault diagnosis. The selection fr 28:8
of CWs needs to satisfy the follows requirements: (1) The CWs
should preserve the main information for decision-making for at
Sample Frequency 12; 000
least one period; (2) The CWs appear repeatedly; (3) For same type NI ¼ ¼  77 ð3-14Þ
fI 156
of conditions, the CWs are similar, and the length of CW has to be
the same.
It is well known that discrete frequencies of the outer race, Sample Frequency 12; 000
NO ¼ ¼  116 ð3-15Þ
inner race or rolling element malfunction can be used for identifi- fO 103:2
cation of bearing faults. Regarding the characteristics frequencies
of inner race, outer race, or rolling element, there are formulas
Sample Frequency 12; 000
for calculating the discrete frequencies. NB ¼ ¼  177 ð3-16Þ
fB 67:9
 
N d
fI ¼ f 1 þ cos a ð3-10Þ where NN, NB, NO, NI are sample number of CW for normal, ball,
2 r D
outer race and inner race, respectively.
  Based on these requirements, the calculations and bearing char-
N d
fO ¼ f r 1  cos a ð3-11Þ acteristic, one of CW is first determined through data analysis in
2 D
the time domain. The other CWs are then obtained by searching
"  2 # for similar of CWs at different period of the signal using correlation
D d
fB ¼ fr 1  cos2 a ð3-12Þ function.
d D After the above process, four CWs of the four bearing conditions
are extracted from the Domain IMF are shown in Fig. 7, respec-
where fr is rotation frequency, fI and fO are frequencies for inner and tively. Normal bearing contains envelops of sine waves of low fre-
outer race malfunctions, respectively, fB is the frequency for ball quency range. The signal of the rolling element fault is complex,
and its fault component also appears in the low frequency region.
Table 2 As a result, for normal bearing and rolling element fault they
Values of a, b and n of each IMF of the outer race. required more points to express the waveform features in time
domain. On the contrary, the signals of outer race and inner race
IMF a b n (%)
faults show impulsive impacts with short time period, which rep-
1 0.9515 7.1433e04 65.00
resent their characteristics. According to the calculated results, a
2 0.1889 0.0014 12.82
3 0.1527 0.0064 10.00 small sample of data is used to express the waveform features in
4 0.1265 0.0017 08.53 the time domain. Based on CWs analysis and considering speed
5 0.0290 0.0056 01.60 fluctuation, the length of normal bearing is 417 points, the inner
6 0.0078 0.0059 00.13 race fault samples have 77 points, and the outer race fault samples
7 0.0018 0.0089 00.49
8 0.0019 0.0083 00.44
have 116 points while the rolling element fault samples have 177
9 0.000789 0.0031 00.16 points.
10 0.000123 0.0026 00.17 After the CWs selection process, 30 CWs for each bearing condi-
11 0.0000426 0.0015 00.10 tion are selected. A CW of the normal bearing is selected as the
12 0.000105 0.0036 0.24
template. The CWs for the four bearing conditions are calculated
13 0.000261 0.0051 0.33
by DTW.
362 T. Han et al. / Measurement 95 (2017) 355–366

1 1 1 1

0.5 0.5 0.5 0.5

0 0 0 0

-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 100 200 300 400 0 100 200 300 400 0 50 100 150 0 50 100 150
Sequence1 Sequence2 Sequence1 Sequence2
1 1 1 1

0.5 0.5 0.5 0.5

0 0 0 0

-0.5 -0.5
-0.5 -0.5
-1 -1
-1 -1 0 50 100 150 0 50 100 150
0 100 200 300 400 0 100 200 300 400
Sequence3 Sequence4
Sequence3 Sequence4
(a) (b)
1 1 1 1

0.5 0.5 0.5 0.5

0 0 0 0

-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 0 20 40 60
Sequence1 Sequence2 Sequence1 Sequence2
1 1 1 1

0.5 0.5 0.5 0.5

0 0 0 0

-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 0 20 40 60
Sequence3 Sequence4 Sequence3 Sequence4
(c) (d)
Fig. 7. CWs of four bearing conditions, (a) normal condition, (b) ball defect, (c) outer race defect and (d) inner race defect.

3.5. Normalization of CW and fault classification range of DDTW. In order to eliminate the influence of amplitude
variations, it is necessary to normalize the amplitudes of CWs.
It is expected that the amplitudes of bearing data can change As shown in Fig. 8(b), the normalized DDTW of each bearing
with different load and speed conditions, such as due to the vibra- fault condition shows good consistency. The four bearing condi-
tion transfer path, impact strength and others. Faulty rolling ele- tions are well separated and the results show distinctive DDTW
ment bearings cause obvious impacts and generate non bands for each bearing condition. Another difference when com-
stationary large vibration amplitudes due to the fluctuation in paring with the result in Fig. 7(a) is that the DDTW increases from
speed and the characteristic of bearing components. For the DTW the 30–100 unit. This is due to the normalization process which
calculation, the amplitude of CW greatly affects the calculation of increases the amplitude of normal condition (original range is from
distance. The results in Fig. 8(a) shows a large dispersion of DDTW 0.2 to 0.2), for a ball fault the range is (0.5 to 0.5), and the range
for an outer race defect. Based on the DTW theory, the high vibra- reduces for those of inner fault (2.0 to 2.0) and outer race fault
tion amplitude of the outer race fault results in the largest DDTW. (5.0 to 5.0).
Furthermore, the variable CW for the outer race due to path trans-
fer or impact results in extreme fluctuation of DDTW. On the con- 3.6. Mutliscale Dynamic Time Warping calculation
trary, DDTW for a normal bearing has the smallest values since the
template is selected from CWs of normal condition. Small DDTW New time series are produced after multiscale processing and
indicates that the template and the normal CWs are very similar. the new series not only keep the morphology of the original CW
The four conditions of DDTW described above cannot be used to but also reduce the length of data points. The new series are used
classify with each other. as CW and identify faults based on DTW. However, different scale
The changes of speeds and other external effects cause the gen- factors may affect fault recognition of the bearing condition and a
eration of vibration amplitudes of each CW. The DDTW of each large scale factor may lead to waveform distortion. Therefore, it is
bearing condition is not consistence and can severely affect the necessary to select optimum scale factor to ensure the accuracy of
 T. Han et al. / Measurement 95 (2017) 355–366 363

Normal malization processing. For the original sequences (s = 1), four bear-
120 Outer race ing states can be divided into different bands. With s = 5, optimum
Inner race classification results can be obtained since the distances among the
100
Ball four states are at the maximum. Meanwhile, the length of the orig-
inal time series is shortened as shown in Eq. (2-7). All the other
scale factors lead to unsatisfactory classification results.
80 According to the above analysis, the recognition of different
DDTW

bearing fault conditions can be carried out through calculating

60 the test CW with the template by DTW. The bearing conditions
can be identified by comparing the calculated DDTW. In order to
validate the consistency of the technique, six groups of CWs were
40 selected from four bearing conditions which were randomly
selected to calculate the DDTW. The results are shown in
20 Tables 5 and 6.
From Table 5, it is obvious that the DDTW of the CWs of the var-
ious bearing fault conditions are in the range of corresponding
0 interval shown in the Fig. 7(b). The results indicate that DTW is
5 10 15 20 25 30
suitable for identifying the bearing fault conditions with high accu-
Samples
racy. The time of DTW is calculated and listed in Table 5. The
(a) Without normalization length of CW affects the processing time. The results indicate that
DTW is fast and an efficient way for bearing fault diagnosis.
140 From Table 6, the DTW available for bearing fault recognition is
Normal
further validated by Multiscale-DTW and obviously shorten the
Outer race
120 processing time. Hence, Multiscale-DTW is an efficient data pro-
Inner race
Ball
cessing method for fault diagnosis. Though the length of data in
100 this calculation is insufficient and its effect on time is not particu-
larly evident. It is therefore suitable for using in other fields dealing
with complex time series.
80
DDTW

In order to illustrate the effectiveness of the proposed methods,

the classification accuracies are compared with other methods
60
[30,31] using the same experiment data source (the Bearing Data
Centre of Case Western Reserve University). As shown in Fig. 10,
40 the results indicate that MDTW has the most high classification
accuracy 100%.
20

0 4. Discussion on efficiency under different working conditions

5 10 15 20 25 30
Samples Considering the working conditions of a rolling element bear-
(b) Normalization ings are not uniform due to variation of load, speed and fault sever-
ity. This can lead to the change of CWs and has an influence on the
Fig. 8. Comparison of DDTW of CWs (a) without normalization and (b) with DDTW results. In this section, the efficiency of the proposed system
normalization. is analyzed by considering variable speed situations and different
fault severity, sizes as listed in Table 7. In case 1, the fault size
(21 mils) is fixed and three speed conditions (1730 rpm,
1
1750 rpm and 1772 rpm) are analyzed In case 2, the speed
(1730 rpm) is kept constant with three different fault sizes (7, 14
0.8 and 21 mils) are calculated.
The same process as described in the previous section is used
DTW distance

0.6 for case 1 and case 2. The DDTW results with a scale factor of 5
outer race
inner race
for the various bearing conditions are shown in Fig. 11 (variation
ball with speed) and Fig. 12 (variation with fault size). Four bearing
0.4
normal conditions similar to those describe above can be clearly classified
under different speeds and fault sizes.
0.2 From Fig. 11, the results show that the four bearing conditions
are well separated under different running speeds based on DDTW.
0 The results show that the change of speed does not affect CW and
0 1 2 3 4 5 6 7 8 9 10 the vibration signals of the four conditions still have high similarity
Scale Factor under different speeds. Although the change in speed affects the
amplitude of the vibration, normalization processing eliminates
Fig. 9. Fault classification with different scale factor.
the influence. Based on the above analysis and results, it indicates
that the proposed method is suitable for fault recognition of rolling
the calculated results. Fig. 9 shows the classification results using element bearings operating in non-uniform speed machinery.
different scale factors. Thirty (30) CWs were selected for each bear- In condition monitoring and fault diagnosis, finding incipient
ing state. A fixed scale factor is used to perform calculations of fault is critical and can lead to prevention of unexpected catas-
mean value and standard deviation of DTW distance after the nor- trophic machine failure. From the Fig. 12, it indicates that DDTW
364 T. Han et al. / Measurement 95 (2017) 355–366

Table 5
Results of DTW algorithm of CW under test.

DDTW Data1 Data2 Data3 Data4 Data5 Data6 Interval Accuracy (%) Time (s)
Normal 9.08 8.75 8.43 9.30 9.76 9.25 [5, 15] 100 0.10
Outer race 86.07 88.34 87.50 84.19 85.78 86.71 [75, 89] 100 0.04
Inner race 93.58 94.42 96.50 95.09 93.28 93.65 [90, 100] 100 0.02
Ball 57.74 56.89 49.16 57.83 52.44 49.67 [40, 70] 100 0.04

Table 6
Results of Mutiscale-DTW with a scale factor of 5.

DDTW Data1 Data2 Data3 Data4 Data5 Data6 Interval Accuracy (%) Time (s)
Normal 1.56 1.44 1.82 1.54 2.20 1.73 [1, 3] 100 0.003
Outer race 15.13 15.86 14.51 13.85 15.28 15.05 [12, 17] 100 0.002
Inner race 18.48 17.80 17.87 18.53 18.82 17.59 [17, 21] 100 0.001
Ball 9.19 7.14 8.91 9.46 11.34 10.78 [5, 11] 100 0.002

can well classify normal and abnormal conditions for different

fault size. In case 2, although different fault size can change the
CW, the four conditions are still well separated since the CW con-
tains most of the condition information.
Based on the above analysis of the two cases, the efficiency and
outstanding capability of the proposed method is validated under
different working speeds and fault size, and is validated with real
bearing data to confirm its suitability for the real application in
industry.

5. Conclusions

A fault recognition method based on DTW is proposed for roll-

ing element bearings fault classification. The results confirm its
outstanding classification performance of DTW in time series.
Fig. 10. Comparison of different methods’ classification accuracies. The application of the method to rolling element bearings faults
classification indicates that the proposed method has a great
potential in real practical application with non-uniform speeds
Table 7
Information of the samples of the bearing balls. and alternating load conditions. It further confirms that:

Case 1 (0.021 mils) Different speed (rpm)

(1) The proposed method is simple and easy to process since the
1730 1750 1772
Case 2 (1730 rpm) Fault size (mils)
DDTW is the only feature for fault recognition without the
7 14 21 need for other classifiers.

30
Normal
Outer race
Inner race
25 Ball
1730rpm 1750rpm 1772rpm

20
DDTW

15

10

0
10 20 30 40 50 60 70 80 90 100
Samples

Fig. 11. DDTW of four bearing conditions with variable speed.

 T. Han et al. / Measurement 95 (2017) 355–366 365

30
Normal
Outer race
Inner race
25 Ball
0.007'' 0.014'' 0.021''

20

DDTW
15

10

0
10 20 30 40 50 60 70 80 90 100
Samples

Fig. 12. DDTW of four bearing conditions under different the fault sizes.

(2) In practical application of condition monitoring and fault [10] H. Bing, G. Trajcevski, P. Scheuermann, X. Wang, E. Keogh, Querying and
mining of time series data: experimental comparison of representations and
classification, it is difficult to collect sufficient samples rep-
distance measures, Proc. VLDB Endowment 1 (2) (2008) 1542–1552.
resenting the various conditions, which limit the application [11] E. Keogh, M. Pazzani, Scaling up dynamic time warping for data mining
of some of the published methods to a certain extent. The applications. In Proceedings of the sixth ACM SIGKDD international conference
proposed method can overcome this problem since DTW on Knowledge discovery and data mining, ACM, 2000, pp. 285–289.
[12] V. Vuori, J. Laaksonen, E. Oja, J. Kanges, Experiments with adaptation strategies
does not required large amount of samples to train the algo- for a prototype-based recognition system for isolated handwritten characters,
rithm. The critical step is to extract the representative CW. Int. J. Doc. Anal. Recogn. 3 (2001) 150–159.
(3) Multiscale-DTW is an effective way to shorten the length of [13] Germain Forestier, Florent Lalys, Laurent Riffaud, Brivael Trelhu, Pierre Jannin,
Classification of surgical processing using time warping, J. Biomed. Inform. 45
time series. Its outstanding performance can be used in (2) (2012) 255–264.
many fields. [14] Cien-Pang Lee, Yungho Leu, Wei-Ning Yang, Constructing gene regulatory
(4) The contribution of the proposed method can overcome the networks from microarray data using GA/PSO with DTW, Appl. Soft Comput.
12 (3) (2012) 1115–1124.
problems of rotating machines with non-uniform speeds [15] Antonio Hernández-Vela, Miguel Ángel Bautista, Xavier Perez-Sala, Víctor
and fluctuating loading conditions while in operation since Ponce-López, Sergio Escalera, Xavier Baró, Oriol Pujol, Cecilio Angulo,
DTW can align the time series to eliminate the effect of Probability-based dynamic time warping and bag-of-visual-and-depth-words
for human gesture recognition in RGB-D, Pattern Recog. Lett. 50 (2014) 112–
speed variations and normalization can overcome the varia- 121.
tion of alternating loads. [16] Chuan-Jun Su, Chang-Yu Chiang, Jing-Yan Huang, Kinect-enabled home-based
rehabilitation system using Dynamic Time Warping and fuzzy logic, Appl. Soft
Comput. 22 (2014) 652–666.
Acknowledgment [17] Jingwei Liu, Qiansheng Cheng, Zhongguo Zheng, Mingping Qian, A DTW-based
probability model for speaker feature analysis and data mining, Pattern
Recogn. Lett. 23 (11) (2002) 1271–1276.
This research project is supported by ‘‘Beijing Higher Education [18] Rossana Muscillo, Maurizio Schmid, Silvia Conforto, Tommaso D’Alessio, Early
Young Elite Teacher Project” (Grant No. YETP0366). recognition of upper limb motor tasks through accelerometers: real-time
implementation of a DTW-based algorithm, Comput. Biol. Med. 41 (3) (2011)
164–172.
References [19] B. Huang, W. Kinsner, ECG frame classification using dynamic time warping,
in: W. Kinsner, A. Sebak, K. Ferens (Eds.), Proceedings of the Canadian
[1] J.L.H. Silva, A.J. Marques Cardoso, Bearing faults diagnosis in three-phase Conference on Electrical and Computer Engineering—IEEE CCECE, vol. 2, IEEE
induction motors by extended park’s vector approach. Industrial Electronics Computer Society, Los Alamitos, CA, USA, 2002, pp. 1105–1110.
Society, 32nd Annual Conference of IEEE 2005, pp. 2585–2590. [20] T. Syeda-Mahmood, D. Beymer, F. Wang, Shape-based matching of ECG
[2] J. Shiroishi, Y. Li, S. liang, T. Kurfess, S. danyluk, Bearing condition diagnostics recordings, in: A. Dittmar, J. Clark, E. McAdams, N. Lovell (Eds.), Engineering in
via vibration and acoustics emission measurement, Mech. Syst. Signal Process. Medicine and Biology Society, EMBS 2007, 29th Annual International
11 (1997) 693–705. Conference of the IEEE, IEEE Computer Society, Los Alamitos, CA, USA, 2007,
[3] P.K. Kankar, Satish.C. Sharma, S.P. Harsha, Fault diagnosis of ball bearings using pp. 2012–2018.
continuous wavelet transform, Appl. Soft Comput. 11 (2) (2011) 2300–2312. [21] K. Gollmer, C. Posten, Supervision of bioprocesses using a dynamic time
[4] Mingbo Zhao, Xiaohang Jin, Zhao Zhang, Bing Li, Fault diagnosis of rolling warping algorithm, Control Eng. Pract. 4 (9) (1996) 1287–1295.
element bearing via discriminative subspace learning: visualization and [22] D. Zhen, T. Wang, F. Gu, A.D. Ball, Fault diagnosis of motor drives using stator
classification, Expert Syst. Appl. 41 (7) (2014) 3391–3401. current signal analysis based on dynamic time warping, Mech. Syst. Signal
[5] Jinde Zheng, Junsheng Cheng, Yu Yang, A rolling bearing fault diagnosis Process. 34 (1–2) (2013) 191–202.
approach based on LCD and fuzzy entropy, Mech. Mach. Theor. 70 (2013) 441– [23] Y. Dai, J. Zhao, Fault diagnosis of batch chemical processes using a dynamic
453. time warping (DTW)-based artificial immune system, Ind. Eng. Chem. Res. 50
[6] M. Barakat, M. EI Badaoui, F. Guillet, Hard competitive growing neural network (8) (2011) 4534–4544.
for the diagnosis of small bearing faults, Mech. Syst. Signal Process. 37 (1–2) [24] J. Zhao, Y. Shu, J. Zhu, et al., An online fault diagnosis strategy for full operating
(2013) 276–292. cycles of chemical peocesses, Ind. Eng. Chem. Res. 53 (13) (2013) 5015–5027.
[7] Huang Wang, Peng Chen, Intelligent diagnosis method for rolling element [25] Tomasz Gorecki, Using derivatives in a longest common subsequence
bearing faults using possibility theory and neural network, Comput. Ind. Eng. dissimilarity measure for time series classification, Pattern Recogn. Lett. 45
60 (4) (2011) 511–518. (2014) 99–105.
[8] Z.W. Pan, J.W. Xu, A supervised fuzzy ART neural network for pattern [26] S. Pan, T. Han, A.C.C. Tan, et al., Fault diagnosis system of induction motors
classification, J. Univ. Sci. Technol. Beijing 22 (3) (2000) 262. based on multiscale entropy and support vector machine with mutual
[9] B. Samanta, K.R. Al-Balushi, S.A. Al-Araimi, Artificial neural networks and information algorithm, Shock Vib. 2016 (7) (2016) 1–12.
support vector machines with genetic algorithm for bearing fault detection, [27] Bearing Data Center Website, Case Western Reserve University <http://www.
Eng. Appl. Artif. Intell. 16 (2003) 65. eecs.cwru.edu/laboratory/bearing>.
366 T. Han et al. / Measurement 95 (2017) 355–366

[28] Huanhuan Liu, Minghong Han, A fault diagnosis method based on local mean [30] J. Yuan, X. Liu, Semi-supervised learning and condition fusion for fault
decomposition and multi-scale entropy for roller bearings, Mech. Mach. Theor. diagnosis, Mech. Syst. Signal Process. 38 (2) (2013) 615–627.
75 (2014) 67–68. [31] X. Zhang, J. Zhou, Multi-fault diagnosis for rolling element bearings based on
[29] Wei Liao, P. Han, X. Liu, Fault diagnosis for engine based on EMD and wavelet ensemble empirical mode decomposition and optimized support vector
packet BP neural network, in: Proceedings of the 3rd International Conference machines, Mech. Syst. Signal Process. 41 (1) (2013) 127–140.
on Intelligent Information Technology Application IEEE Press, 2009, pp. 672–
676.