7762 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO.

12, DECEMBER 2015

An Improved Exponential Model for

Predicting Remaining Useful Life
of Rolling Element Bearings
Naipeng Li, Yaguo Lei, Member, IEEE, Jing Lin , and Steven X. Ding

AbstractThe remaining useful life (RUL) prediction of the remaining useful life (RUL) of the bearing is predicted
rolling element bearings has attracted substantial attention in advance, the catastrophe could be avoided by predictive
recently due to its importance for the bearing health man- maintenance. Furthermore, the predictive maintenance is able
agement. The exponential model is one of the most widely
used methods for RUL prediction of rolling element bear- to make the machine have a maximum uptime with minimum
ings. However, two shortcomings exist in the exponential maintenance costs. Therefore, the RUL prediction of rolling
model: 1) the rst predicting time (FPT) is selected sub- element bearings has attracted increasingly more attention in
jectively; and 2) random errors of the stochastic process recent years [1][4].
decrease the prediction accuracy. To deal with these two The RUL prediction methods can be categorized into data-
shortcomings, an improved exponential model is proposed
in this paper. In the improved model, an adaptive FPT driven and model-based methods [5], [6]. Data-driven methods
selection approach is established based on the 3 interval, attempt to derive the degradation process of a machine from
and particle ltering is utilized to reduce random errors measured data using machine learning techniques. Gebraeel et al.
of the stochastic process. In order to demonstrate the developed an artificial neural network-based method to predict
effectiveness of the improved model, a simulation and four the RUL of bearings [7]. Maio et al. proposed a method based
tests of bearing degradation processes are utilized for the
RUL prediction. The results show that the improved model on relevance vector machine for estimating the bearing RUL
is able to select an appropriate FPT and reduce random [8]. In addition, several studies have been published on the
errors of the stochastic process. Consequently, it performs neuro-fuzzy system-based prediction methods [9][11]. Data-
better in the RUL prediction of rolling element bearings than driven methods could be beneficial, when mechanical prin-
the original exponential model. ciples are not straightforward or mechanical systems are so
Index TermsExponential model, rst predicting time complex that developing an accurate model is prohibitively
(FPT), particle ltering (PF), remaining useful life (RUL) expensive. The prediction accuracy of data-driven methods,
prediction, rolling element bearings. however, depends on not only the quantity but also the quality
of the measured data, which is full of challenges for real
I. I NTRODUCTION applications.
Model-based methods are to set up mathematical or physical

R OLLING element bearings are widely used in rotating

machinery. They generally work in tough environment;
hence, different kinds of faults may occur frequently. Any
models to describe the degradation process of a machine, and
estimate model parameters using measured data. Liao et al.
proposed a prediction method based on proportional hazard and
fault of a bearing probably causes breakdown of the entire logistic regression models to predict the RUL of rolling element
machine, which could lead to catastrophic consequences. If bearings [12]. Tian et al. proposed a proportional hazard model-
based method for the RUL prediction of the systems consisting
of bearings [13]. Liao employed the Paris model combined
Manuscript received January 8, 2015; revised April 24, 2015; with a genetic programming method to predict the RUL of
accepted May 30, 2015. Date of publication July 10, 2015; date of cur- bearings [14]. Model-based methods could incorporate both
rent version November 6, 2015. This work was supported in part by the expert knowledge and measured information. Consequently,
National Natural Science Foundation of China under Grant 51222503
and Grant 51475355, in part by the Provincial Natural Science Founda- they may work well in the RUL prediction of rolling element
tion Research Project of Shaanxi under Grant 2013JQ7011, and in part bearings.
by the Fundamental Research Funds for the Central Universities under Among all model-based studies, the exponential model is
Grant 2012jdgz01 and Grant CXTD2014001.
N. Li, Y. Lei, and J. Lin are with the State Key Laboratory for Manu- one of the most popular methods. It was first established
facturing Systems Engineering, Xian Jiaotong University, Xian 710049, by Gebraeel et al. in [15], where model parameters were
China (e-mail: [email protected]; [email protected]. updated using a Bayesian approach to incorporate measured
edu.cn; [email protected]).
S. X. Ding is with the Institute of Automatic Control and Complex information. After that, many variants and applications of the
Systems, University of Duisburg-Essen, 47057 Duisburg, Germany exponential model have been reported in the RUL prediction
(e-mail: [email protected]). and health management [16][19]. Gebraeel further improved
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. the model by updating the model parameters using multi-
Digital Object Identifier 10.1109/TIE.2015.2455055 ple historical degradation signals acquired through condition
0278-0046 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LI et al.: IMPROVED EXPONENTIAL MODEL FOR PREDICTING RUL OF ROLLING ELEMENT BEARINGS 7763

Fig. 1. Bearing degradation process.

monitoring [16]. Si et al. combined Bayesian updating with

the expectation maximization (EM) algorithm for parameter
estimation and offered a closed-form RUL distribution [19]. It
can be seen from these studies that the exponential model works
well for the RUL prediction of systems with exponential-like
degradation processes. However, there are still two shortcom-
ings in this model. Fig. 2. (a) Simulated degradation process of a bearing and (b) RUL
prediction result.
One shortcoming is that the first predicting time (FPT) is
selected subjectively. To explain the concept of FPT, a simple
example for describing the degradation process of a rolling tion accuracy. This can be explained using a simulation shown
element bearing is presented in Fig. 1. The root mean square in Fig. 2, where a stochastic process of a bearing degradation is
(RMS) of vibration signals is used as the degradation index simulated and the RUL of the bearing is predicted using the
in this example. It is seen that the degradation process of the exponential model [19]. The error-free process describes the
bearing generally consists of two stages, i.e., I: the normal op- actual degradation process of the bearing, without the interfer-
eration stage and II: the failure stage. The degradation index in ence of random errors. It is noticed that there is an obvious
Stage I is stable, which means that the bearing is healthy. Once negative correlation between random errors of the stochastic
a fault occurs, the degradation process changes from Stage I process and deviations of the prediction result. For example, the
to Stage II, where the degradation index generally increases positive random error at time t1 leads to the negative deviation
exponentially as the fault gets severer. The major task in Stage between the predicted RUL B1 and the actual RUL B1 . Vice
I is to monitor the health state and detect incipient faults. When versa, the negative random error at time t2 causes the positive
a fault is detected, the predicting process is triggered, and the deviation between B2 and B2 . This negative correlation can be
bearing RUL is predicted during this process. The time to start explained as follows. Taking time t1 for example, the value
predicting is defined as FPT, as shown in Fig. 1. It is important of the degradation process increases from A1 to A1 because
to select an appropriate FPT value for the exponential model. of the positive random error. As a result, the fault severity
If an inappropriate FPT value is selected, either interference is overestimated compared with the actual severity. A severer
noises could be included in the predicting process or critical fault generally indicates a shorter RUL. Therefore, the overes-
information may be excluded from the predicting process. timated severity leads to a shorter predicted RUL than the ac-
Both cases could decrease the RUL prediction accuracy. The tual one.
appropriate FPT should be the time of fault occurrence, which To overcome these two shortcomings in the exponential
is difficult to be detected because incipient faults of a bearing model, this paper proposes an improved exponential model for
often occur randomly and the fault characteristics are quite predicting the RUL of rolling element bearings. The major
weak. Therefore, the FPT in the exponential model is generally contributions of this work are: 1) an adaptive FPT selection
selected subjectively, which restricts the applications of the approach is established based on the 3 interval; and 2) random
exponential model. Several approaches for selecting FPT have errors of the stochastic process are reduced utilizing particle
been reported in literature, such as the engineering norm ISO filtering (PF), and therefore, the RUL prediction accuracy is
10816 and the approach based on the longest time constant raised.
of a machine and statistical properties of a candidate baseline The remainder of the paper is organized as follows: In
[20], [21]. In these approaches, the FPT is selected based on an Section II, we briefly describe the basic theory of the ex-
alarm, which is set according to the statistical properties of large ponential model and PF, respectively. Section III shows the
numbers of systems. Therefore, the approaches fail to adjust improved exponential model for predicting the RUL of rolling
the alarm adaptively according to the variation of an individual element bearings. In Section IV, the model is evaluated using a
system. simulation of a degradation process of bearings. In Section V,
The other shortcoming of the exponential model is that ran- the model is demonstrated using accelerated degradation tests
dom errors of the stochastic process decrease the RUL predic- of bearings. Conclusions are drawn in Section VI.
7764 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 12, DECEMBER 2015

II. B ASIC T HEORY conditional joint probability density function (PDF) of S1:k is
calculated as follows:
A. Exponential Models
As mentioned earlier, many variants of the exponential model  k
1
have been reported in literature [16][19]. The model consid- p(S1:k |, ) =
2 2 t
ered in this work furthers the idea in [19]. In this exponential
model, the parameters are estimated using the Bayesian updat- (s t )2 k
(s s t)2
exp (3)
1 1 j j1
ing and EM algorithm, and the RUL is offered as a closed-form
2 2 t1 2 2 t
distribution. j=2
1) Development of Degradation Models: A degradation
process is stochastic in nature due to inherent randomness in where t = tj tj1 is the constant time interval. Then, the
manufacturing and operating. Therefore, it is natural to model a joint posterior PDF of and conditioned on S1:k is still
degradation process as a stochastic one {xk = x(tk ), tk 0}, normal resulted from the normal distribution of and , i.e.,
where xk is the system state at time tk . Generally, a degrada- , |S1:k N (,k , ,k
2 2
, ,k , ,k , k ), with [15]
tion model consists of deterministic and stochastic parts. The
deterministic part represents a constant physical phenomenon, ,k

which is common to all systems of a given population, whereas s1 02 +0 2 t1 2 +12 tk 02 t1 sk 12 +1 2 0.5 4

the stochastic part captures the variation of the degradation =
(02 + 2 t1 ) (12 tk + 2 )02 12 t1
process of an individual system [16], [22]. In the exponential 2
model, it is assumed that the system state can be represented by ,k

measurements, and the state xk at time tk is given by 02 2 t1 2 +12 tk

= 2
  (0 + 2 t1 ) (12 tk + 2 )02 12 t1
  2
xk = + exp tk + B(tk ) tk (1) ,k
2

sk 12 +1 2 0.5 4 02 + 2 t1 12 s1 02 +0 2 t1
where is a known constant;  and  are random variables =
(02 + 2 t1 ) (12 tk + 2 )02 12 t1
characterizing the stochastic part; is a constant representing 2
the deterministic part; and B(tk ) is a Brownian motion (BM) ,k

following a normal distribution of N (0, 2 tk ), which represents 12 2 t1 02 + 2 t1

random errors in the stochastic process. = 2
(0 + 2 t1 ) (12 tk + 2 )02 12 t1
For convenience, the exponential model is transformed into
the logged format and is defined as k

  0 1 t1
2 =  2 (4)
sk = ln[xi ] = ln( ) +  tk + B(tk ) (0 + 2 t1 ) (12 tk + 2 )
2
where we know that the posterior estimations of and can be
= + tk + B(tk ) (2) easily updated once new measurements are available. However,
there are unknown parameters needed to be estimated before
where = ln( ) follows N (0 , 02 ) and =  2 /2 fol- updating, which can be expressed as = [ 2 , 0 , 02 , 1 , 12 ].
lows N (1 , 12 ). It is assumed that = 0, to simplify the The unknown parameters will be estimated using the EM
analysis. (i) 2(i) (i) 2(i) (i) 2(i)
algorithm. Given k = [k , 0,k , 0,k , 1,k , 1,k ] as the
2) Parameter Estimation: Suppose that we have observed (i)
a sequence of measurements S1:k = {s1 , s2 , . . . , sk }. Since estimation in the ith step, let [ (k |k )]/k = 0, then the
(i+1)
random errors B(tk ) are independent identically distributed parameter estimation in the next step k is obtained as (5)
(i.i.d.) normal random variables, if and are known, the [19], shown at the bottom of the page.


2(i+1) 1 s21 2s1 (,k + ,k t1 ) + ,k
2 2
+ ,k

k = + 2(k ,k ,k + ,k ,k ) + t1 2,k + ,k
2
k t1

k (s s 2 2
 j j1 ) (s j s j1 )t ,k + (t) 2
,k + 2
,k
+
j=2
t

(i+1) 2(i+1) 2
0,k = ,k , 0,k = ,k
(i+1) 2(i+1) 2
1,k = ,k , 1,k = ,k (5)
LI et al.: IMPROVED EXPONENTIAL MODEL FOR PREDICTING RUL OF ROLLING ELEMENT BEARINGS 7765

3) RUL Prediction: The RUL Lk at tk is defined as follows: 2) SIS Algorithm: The recursive propagation of the
a posteriori PDF is only a conceptual solution, which can-
Lk = inf {lk : s(lk + tk ) |S1:k } (6) not be calculated analytically, because it requires the evalua-
tion of complicated high-dimensional integrals. Therefore, the
where is the prespecified failure threshold, and SIS algorithm is employed to approximate the a posteriori
s(lk + tk ) = sk + lk + W (lk ), for lk 0 (7) PDF, numerically. SIS approximates the state PDF using a set
of particles {i0:k }i=1:Ns sampled from an importance PDF
with q(k |z1:k ), where Ns is the number of particles. Using SIS,
the state of the system can be approximated using the following
W (lk ) = B(lk + tk ) B(tk ) (8) discrete density:

which is a standard BM. Given the estimated parameters from 

Ns

(5), the PDF of the RUL can be expressed by the following [19]: p(k |z1:k ) wki 0:k i0:k (14)
i=1
sk
fLk |S1:k (lk |S1:k ) =   where wki is the weight of particle i0:k .
2lk3 ,k
2 l + 2
k k Once a new measurement is available, the particle weights
can be updated and normalized as follows:
( sk ,k lk )
2

exp  , lk 0. (9) p zk |ik p ik |ik1 wi

2 l + 2
2lk ,k
i i
wk = wk1 i i
, wki = N k . (15)
k k q k |k1 , z1:k s i
i=1 wk

In general, the importance PDF q(k |z1:k ) is chosen as

B. PF Algorithm p(k |k1 ), and then, (15) is simplified as follows:
PF is derived from traditional filtering algorithms, such as
wi
Kalman filtering and extended Kalman filtering [23]. Bayesian wki = wk1
i
p zk |ik , wki = Nsk . (16)
theory is used in PF, which attempts to evaluate the state of i=1 wki
a system based on measurements. A sequential importance 3) Resampling Step: One problem in the SIS algorithm
sampling (SIS) algorithm is developed and also applied in PF is the particle degeneracy, i.e., after a few iterations, all but
[24]. Based on the Bayesian theory and SIS, PF is particularly one particle have negligible weights. To deal with this problem,
effective in estimating states of nonlinear and/or non-Gaussian a resampling step is utilized. The basic idea of resampling is
systems, and it has been widely used in the fields of diagnosis to eliminate particles with small weights and to concentrate
and prognosis [25][31]. upon particles with large weights. A suitable measure of the
1) Bayesian Theory: It is assumed that the state of a sys- degeneracy of particles is the effective sample size introduced
tem k at tk can be described in the following state-space model: in [32] and defined as follows:
k = fk (k1 , k ) (10) Ns
Ne =
(17)
1 + var wki
where fk : Rn Rn Rn is the state transition function,

and k is the i.i.d. state noise. where wki = p(ik |z1:k )/q(ik |ik1 , z1:k ) is referred to the
Generally, we cannot obtain the system state but uncertain true weight. The effective sample size cannot be calculated
measurements. Therefore, the system state should be evaluated exactly, but we can estimate it by
based on the measurements. The relationship between them is
1
denoted as follows: Ne = N
2 . (18)
s i
i=1 wk
zk = hk (k , k ) (11)
When Ne falls below a threshold NT , a new particle set
where hk : Rn Rn Rnz is the measurement function,
{ik }i=1:Ns will be generated by resampling Ns times, follow-
and k is the i.i.d. measurement noise.
ing the principle of P (ik = ik ) = wki , and the weights are
Based on the knowledge of the transition function and the i
reset to wk = 1/Ns .
previous state estimation, the a priori PDF of the present state
is calculated as follows:
 III. I MPROVED E XPONENTIAL M ODEL
p(k |z1:k1 ) = p(k |k1 )p(k1 |z1:k1 )d k1 . (12) The flowchart of the improved exponential model for predict-
ing RUL of rolling element bearings is shown in Fig. 3.
Once a new measurement is available, the a priori PDF is up- Vibration signals are acquired from rolling element bearings,
dated according to (13) to generate the a posteriori PDF. That is, and then, kurtosis and RMS are extracted from the vibration
signals, one of which is used for health monitoring and the
p(zk |k )p(k |z1:k1 ) other is used for RUL prediction. It is well known that kurtosis
p(k |z1:k ) = . (13)
p(zk |z1:k1 ) is sensitive to incipient faults but is not informative for the
7766 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 12, DECEMBER 2015

new trigger mechanism is applied to restrict the interference

of random noises, i.e., the predicting process is triggered when
l + 1 consecutive kurtosis values exceed the 3 interval. The
process of the trigger mechanism is described as follows:
First, let l = 0, find the first time when the kurtosis value
exceeds the 3 interval, and select the time as FPT0 .
Second, let l = l+1, find the time tf when the consecutive l+
1 kurtosis values {mf +k }k=0:l satisfy {|mf +k | > 3}k=0:l ,
and select the time tf as FPTl .
Third, let l increases from 1 until l satisfies FPTl = FPTl1 ,
and output FPTl as the final FPT value.
The theory behind this trigger mechanism is that abnormal
states caused by random noises are almost impossible to appear
l + 1 times consecutively during the normal operation stage.
When the selected FPT starts to keep stable with the increase of
l, it may imply that a fault has occurred.
After the FPT is decided, RMS is input into the degradation
model for the RUL prediction of the bearing. In order to reduce
random errors of the stochastic process, PF is employed in this
study to estimate the health state. The RUL prediction process
of the improved exponential model is explained as follows.
First, the model parameters are initialized as 2 = 02 , 0 =
,0 , 02 = ,0
2
, 1 = ,0 and 12 = ,0 2
using the approach
proposed in [15]. Then, initial particles {s0 }i=1:Ns are sampled
i

from p(s0 |s0 , 0 ) N (s0 , 02 t). Once the measurement sk

at tk is available, model parameters are updated using (4) and
further estimated with (5). With the estimated parameters and
the following variant of the degradation model:

sk = sk1 + t + W (t) (19)

the PDF of the state is described as follows:

1
sk |sk1 , k ) = 
p( 
2t ,k2 t + 2
k

sk sk1 ,k t)
( 2
exp  (20)
2t ,k2 t + 2
Fig. 3. Flowchart of the improved exponential model. k

development of faults, while RMS is able to reflect the increase where sk is the actual state at tk ; k = [k2 , ,k , ,k
2
, ,k ,
2
of the vibration energy with the development of faults [33]. ,k ] are the estimated parameters; sk1 is the estimated state
That is why in this study, kurtosis and RMS are adopted as the at tk1 , which is different from the measurement sk1 . It should
monitoring and predicting indexes, respectively. In the normal be noticed that, if sk1 is replaced by sk1 , (20) will change to
operation stage, kurtosis is used for health monitoring and FPT p(sk |sk1 , k ), which is the PDF of the state acquired using
selection. Once an FPT is decided, the predicting process is the original exponential model.
triggered, and RMS is utilized as the predicting index for the Now, we need to focus on how to estimate the actual
RUL prediction of the bearing. state sk at tk based on the measurement sk . Here, we take
In order to select the appropriate FPT, an adaptive approach is p(sk |sk1 , k ) as the importance PDF, which contains the
established based on the 3 interval. At first, the history data of information of the state estimation at tk1 . Then, particle sets
the bearing in the normal operation stage are utilized to decide {sik }i=1,2,...,Ns are sampled from (20). These particles can
the 3 interval [ 3, + 3], by calculating the mean only approximate the a priori PDF of the state. In order to
and standard deviation of kurtosis. Then, the 3 interval is approximate the a posteriori PDF, particle weights should be
utilized to identify normal and abnormal states of the bearing. updated according to the measurement sk as follows:
When new kurtosis mf at tf is available, it is compared with 
2 
the 3 interval. If mf is beyond the 3 interval, the health 1 sk sik wki
wk = wk1 
i i
exp , wk
i
=  .
state is considered to be abnormal. The abnormal state may be 2k2 tk 2k2 tk Ns
i=1 wk
i

caused by either a fault or random noises. In this approach, a (21)

LI et al.: IMPROVED EXPONENTIAL MODEL FOR PREDICTING RUL OF ROLLING ELEMENT BEARINGS 7767

With the particles and their corresponding weights, the TABLE I

RUL P REDICTION P ROCESS U SING THE
a posteriori PDF of the system state can be approximated using I MPROVED E XPONENTIAL M ODEL
the following discrete density:

Ns

sk |sk , k )
p( wki sk sik . (22)
i=1

Through consecutive iteration and updating, the particle whose

motion curve is closer to the actual degradation process ac-
quires a higher weight. The health state of the bearing is
estimated by

Ns

sk = wki sik . (23)

i=1

With the weighted average of updated particles, the estimated

state sk is closer to the actual state than the measurement sk .
After the particle weights have been updated, Ne in (18) is
calculated to measure the degeneracy of particles. When Ne

falls below a threshold NT , a new set of particles {sik }i=1:Ns
are resampled, and the particle weights are reset to wki = 1/Ns .
Finally, the PDF of the RUL is calculated using

sk
fLk |S1:k (lk |S1:k ) =  
2lk3 ,k
2 l + 2
k k

( sk ,k lk ) 2
exp  , lk 0. (24)
2 l + 2
2lk ,k k k

The RUL prediction process using the improved exponential

model is summarized in Table I.

IV. S IMULATION
Here, a simulation of a degradation process of rolling ele-
ment bearings is utilized to evaluate the effectiveness of the
improved exponential model in the bearing RUL prediction.

A. Simulations of the Bearing Degradation Process

We simulate the vibration signals that represent the degrada-
tion process of rolling element bearings [34]. The parameters of
the bearing in the simulation are as follows: out race diameter of
29.1 mm, inner race diameter of 22.1 mm, 13 rollers, and con-
tact angle of 0 . The bearing has an outer race fault, and its ro-
tating speed is 1800 r/min. The sampling frequency is 25.6 kHz,
and each sample contains 2560 data points, i.e., 0.1 s, and the
sampling is repeated every 10 s.
A sample in the normal operation stage and a sample in Fig. 4. Simulation of the bearing: (a) a sample in the normal operation
stage, (b) a simple in the failure stage, and (c) vibration signals of the
the failure stage are shown in Fig. 4(a) and (b), respectively. whole lifetime.
There exist obvious impacts with a period of = 0.0059 s in
the sample of the failure stage, which is corresponding to the
fault characteristic frequency of 169 Hz. To simulate the whole
B. RUL Prediction Based on the Simulation
degradation process of the bearing, the signals are repeated
200 times with the increasing fault severity, and Fig. 4(c) shows The improved exponential model is used to predict the bear-
the vibration signals of the whole lifetime. ing RUL in this simulation. As described in Section III, kurtosis
7768 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 12, DECEMBER 2015

Fig. 7. Comparison between p(sk |sk1 , k ) and p(sk |sk , k ) at

tk = 830 s.

Fig. 5. FPT selection in the simulation: (a) Kurtosis and (b) RMS.

Fig. 8. (a) Bearing state estimation and (b) RUL prediction in the
simulation.

number is set to 1000. The parameter estimation process is

shown in Fig. 6. It is observed that all parameters are not
estimated accurately at the beginning of the predicting process.
Fig. 6. Parameter estimation in the simulation.
When enough RMS values are used in the parameter estima-
tion, all of these parameters converge to the actual values. To
and RMS are extracted from the vibration signals, which are visualize the state estimation process of the improved expo-
used for FPT selection and RUL prediction, respectively. nential model, the state PDF p(sk |sk1 , k ) of the original
The FPT selection result is shown in Fig. 5. It is seen exponential model [19] and the state PDF p(sk |sk , k ) of the
that kurtosis is sensitive to the incipient fault of the bearing improved model in the middle of the process, i.e., tk = 830 s,
but not informative for the development of the fault. RMS are presented in Fig. 7. It is seen that, with the help of the
has an obvious exponential-like growth during the degradation state estimation process of PF, the improved exponential model
process but reflects nothing characterizing the fault occurrence. actually changes the result of the original model. For further
Therefore, it is reasonable to adopt kurtosis as the monitoring identifying the effectiveness of the improved exponential model
index and use RMS as the predicting index. In Fig. 5(b), the in state estimation and RUL prediction, the results of the
RMS is divided into two stages by the FPT. Before the FPT, improved exponential model are compared with those of the
the values of RMS are stable, while they increase exponentially original exponential model and a commonly used PF-based
after the FPT. Therefore, tf = 330 s indicates the initial time prediction model, i.e., the Paris model [27]. In this process,
of the exponential-like degradation process and is suitable to be the same FPT and particle number are used for them. The
the FPT in this simulation. comparison results are shown in Fig. 8. The actual RUL is
The values of RMS after tf = 330 s are input into the im- calculated with RULk = tEoL tk , where RULk is the actual
proved exponential model for RUL prediction, and the particle RUL at tk , and tEoL is the time index at the end of lifetime [35].
LI et al.: IMPROVED EXPONENTIAL MODEL FOR PREDICTING RUL OF ROLLING ELEMENT BEARINGS 7769

TABLE II
S CORES OF T WO P ERFORMANCE M ETRICS IN THE S IMULATION

Fig. 10. Rolling element bearings before and after a test.

Fig. 9. Overview of the experimental system.

In Fig. 8(a), the state estimation result of the original model

almost coincides with the measurement; thus, the estimation
result still includes random errors of the stochastic process. The
Paris model performs slightly better than the original exponen-
tial model, but it still has random fluctuation in the estimation
result. The result of the improved model, however, reflects
a smoother degradation process and succeeds in restraining
random errors of the stochastic process. In Fig. 8(b), the RUL
prediction results of all three models produce large errors at the
beginning. The prediction results of the original model and the
Paris model have larger fluctuation than that of the improved
model. As time goes on, all of them converge to the actual RUL.
In addition, the improved model converges fastest and provides
the most accurate result among them. This corresponds to our
requirements for RUL prediction that more accurate prediction
results are expected when the machine is closer to the final
failure. The results could be explained as follows. At the
beginning, RMS values are not enough to accurately estimate
the model parameters. Once the model parameters are estimated
appropriately, random errors of the stochastic process become
the major causes of the prediction errors. In fact, the improved Fig. 11. Vibration signals of four bearings: (a) bearing 1, (b) bearing 2,
(c) bearing 3, and (d) bearing 4.
exponential model is able to reduce random errors by using PF.
Therefore, it obtains a best prediction result in this simulation.
In order to compare the performance of the three models summarized in Table II. It is noticed that the CRA score of the
quantitatively, two commonly used performance metrics, i.e., improved model is the highest among them, which means that
cumulative relative accuracy (CRA) and convergence proposed the improved model has the most accurate prediction result. The
in [35], are calculated. CRA quantifies how accurately a pre- convergence score of the improved model is the lowest among
diction method performs at specific time indexes. The range of them, which implies that the improved model approaches the
scores for CRA is [0, 1], where the perfect score is 1. Con- actual RUL fastest.
vergence quantifies how fast a method approaches the actual
RUL. A lower score means a faster convergence. CRA and
V. E XPERIMENTAL D EMONSTRATIONS
convergence are calculated at the time indexes from the half
to the end of lifetime, since these time indexes are closer to Here, vibration signals of the whole lifetime acquired from
the final failure and, therefore, more meaningful than the early accelerated degradation tests of rolling element bearings are
stages for RUL prediction. The scores of the two metrics are used to verify the effectiveness of the improved model.
7770 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 12, DECEMBER 2015

Fig. 12. FPT selection results: (a) bearing 1, (b) bearing 2, (c) bearing 3, and (d) bearing 4.

A. Introduction to the Tests and Vibration Data TABLE III

S ELECTED FPT OF F OUR B EARINGS
An experimental system named PRONOSTIA [36] is shown
in Fig. 9. This system is designed to test methods for bearing
fault detection, diagnosis, and RUL prediction. In order to con-
duct accelerated degradation tests of bearings in a few hours, a
radial force, which is equal to the bearings maximum dynamic
load of 4 kN, is applied on the tested bearings. The force is the FPT selection approach proposed in this paper is used to
generated by a cylinder pressure, and the pressure is delivered decide the FPT of the four bearings. To make a comparison,
through a pressure regulator. During the tests, the rotating speed another approach proposed by Ginart et al. [20] is also used to
of the bearing keeps 1800 r/min. Accelerometers are fixed on select the FPT. The results are shown in Fig. 12, and all FPT
the outer race of the bearing, and vibration signals are captured. values are given in Table III, where N/A means that no FPT
The sampling frequency is 25.6 kHz. Each sample contains is selected during the whole lifetime. It is shown in Fig. 12 that
2560 data points, i.e., 0.1 s, and the sampling is repeated every the alarm set by Ginarts approach is much higher than the 3
10 s. The vibration signals are transmitted into a PC for data interval and fails to distinguish the normal and the abnormal
visualization and storage through a National Instruments (NI) states. On the contrary, the 3 interval is able to detect the
data acquisition (DAQ) card. The bearing useful-life ends at abnormal states. However, during the normal operation stage,
the time when the amplitude of the vibration signal exceeds there are some abnormal states caused by random noises instead
20 g [37]. of faults, which are marked by in Fig. 12. Due to the anti-
Each bearing is naturally degraded during the tests without interference ability of our approach, the time when faults occur
seeding a fault in advance. Depending on the diversity of is appropriately selected as the FPT.
different bearings and the randomness of degradation processes, Fig. 13 plots all RMS values of the four bearings and the
the fault modes may be slightly different for distinct bearings. corresponding FPT decided by the two approaches. It is seen
As a result, the vibration signal pattern is different for each that RMS has a more obvious degradation trend than kurtosis
tested bearing. Fig. 10 shows pictures of a tested bearing before in Fig. 12, with the development of faults. Since the faults of
and after a test. It is seen that different kinds of faults occur on bearings 1 and 2 get severe gradually, it is slightly difficult to
the balls and inner race of the bearing. The vibration signals exactly give the time of the fault occurrence. Consequently,
during the whole lifetime of four tested bearings are shown the FPT values selected by our approach have a slight time
in Fig. 11. The signal amplitudes of bearings 1 and 2 have delay, as shown in Fig. 13. The delayed time, however, is much
gradually increasing trends, which indicates that the faults get smaller than that of Ginarts approach, particularly in the case of
severe gradually, whereas the signal amplitudes of bearings 3 bearing 1. In addition, the small delay is acceptable for the RUL
and 4 show rapid increases at the end of lifetime, thus repre- prediction of bearings. For bearings 3 and 4, they behave abrupt
senting abrupt degradation processes. Taking bearing 4 as an degradation processes. Therefore, it is not as difficult as bear-
example, more detailed information about the vibration signals ings 1 and 2 to select the FPT. Our approach still performs better
in the normal operation stage and the failure stage is presented than Ginarts approach in selecting the FPT of bearings 3 and 4.
in Fig. 11. It is shown that there exist obvious impacts in the In order to further identify the effectiveness of the 3 inter-
vibration signals of the failure stage than those of the normal val, 14 intervals are utilized in our approach to select the
operation stage. FPT of four tested bearings. The selection results are shown
in Fig. 14. It is seen that the 2 and 4 intervals select much
too late FPT for bearing 1. For bearing 2, the 1 and 2
intervals offer much too early FPT. For bearing 3, all of them
B. FPT Selection
have the same FPT. And for bearing 4, the 1 interval presents
As described in the proposed method, kurtosis and RMS are much too early FPT. In conclusion, the 3 interval has the best
first extracted from the vibration signals. Based on kurtosis, performance among these intervals.
LI et al.: IMPROVED EXPONENTIAL MODEL FOR PREDICTING RUL OF ROLLING ELEMENT BEARINGS 7771

Fig. 13. RMS division results: (a) bearing 1, (b) bearing 2, (c) bearing 3, and (d) bearing 4.

Fig. 14. FPT selection results with different intervals: (a) bearing 1, (b) bearing 2, (c) bearing 3, and (d) bearing 4.

Fig. 15. RMS estimation results: (a) bearing 1, (b) bearing 2, (c) bearing 3, and (d) bearing 4.

Fig. 16. RUL prediction results: (a) bearing 1, (b) bearing 2, (c) bearing 3, and (d) bearing 4.

C. RUL Prediction
results are presented in Fig. 16. For comparison, the original
The RUL of four bearings is predicted using the improved exponential model [19] and the Paris model [27] are used to
exponential model with a particle number of 1000. The RMS predict the RUL of the four bearings as well. The same FPT
estimation results are shown in Fig. 15, and the RUL prediction and particle number are used for them, and the estimation
7772 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 12, DECEMBER 2015

TABLE IV reduce random errors of the stochastic process. A simulation

S CORES OF T WO P ERFORMANCE M ETRICS FOR T ESTED B EARINGS
and four tests of bearing degradation processes are used to
demonstrate the effectiveness of the improved model. In order
to identify the benefits of our adaptive FPT selection approach,
it is compared with Ginarts approach. The results show that our
approach performs better in restricting the interference caused
by random noises and selects a more appropriate FPT. To show
the superiority of state estimation using PF, the performance
of the improved exponential model is compared with those of
the original exponential model and the Paris model. The results
clearly demonstrate the effectiveness of the improved model in
reducing random errors of the stochastic process and increasing
and prediction results are also presented in Figs. 15 and 16, the accuracy of RUL prediction for bearings.
respectively. Fig. 15 indicates that the estimation results of the Although this study has improved the prediction accuracy
original model almost coincide with the actual RMS, while of the exponential model by selecting an appropriate FPT and
two PF-based methods, i.e., the Pairs model and the improved reducing random errors, the failure threshold still influences
exponential model, may reduce random errors of the stochastic the prediction accuracy of the exponential model. The fact
processes and therefore reflect the global degradation processes is that the failure threshold is generally set subjectively and
of the bearings. In Fig. 16, all of them generate inaccurate limited research has been carried out on adaptively setting
results at the beginning because of the lack of RMS values. failure thresholds in RUL prediction. Therefore, we will in-
When enough RMS values are available and used, the original vestigate adaptive threshold setting approaches in our future
exponential model and the Paris model still produce large research.
prediction errors, particularly in the case of bearing 2. The
improved exponential model, however, converges to the actual
RUL fastest and presents the most accurate prediction results R EFERENCES
among them. This should be explained as follows. The original [1] Z. G. Tian, An artificial neural network method for remaining useful
exponential model is unable to accurately estimate the health life prediction of equipment subject to condition monitoring, J. Intell.
Manuf., vol. 23, no. 2, pp. 227237, Apr. 2012.
state of the bearings; thus, the prediction results are influenced [2] A. Malhi, R. Q. Yan, and R. X. Gao, Prognosis of defect propaga-
by random errors of the stochastic process. The Paris model tion based on recurrent neural networks, IEEE Trans. Instrum. Meas.,
performs better than the original exponential model in the vol. 60, no. 3, pp. 703711, Mar. 2011.
[3] Y. N. Qian, R. Q. Yan, and S. J. Hu, Bearing degradation evaluation
state estimation with the help of PF, but it may not estimate using recurrence quantification analysis and Kalman filter, IEEE Trans.
the model parameters effectively. Therefore, the results of the Instrum. Meas., vol. 63, no. 11, pp. 25992610, Nov. 2014.
Paris model still have large random errors. The improved ex- [4] Z. G. Tian, L. Wong, and N. Safaei, A neural network approach
for remaining useful life prediction utilizing both failure and suspension
ponential model incorporates the superiority of PF in the state histories, Mech. Syst. Signal Process., vol. 24, no. 5, pp. 15421555,
estimation and the advantage of the Bayesian updating and EM Jul. 2010.
algorithm in the parameter estimation. Consequently, it presents [5] J. Liu, W. Wang, F. Ma, Y. B. Yang, and C. S. Yang, A data-model-fusion
prognostic framework for dynamic system state forecasting, Eng. Appl.
the best prediction results among the three models. Artif. Intell., vol. 25, no. 4, pp. 814823, Jun. 2012.
To quantify the prediction performance of the three models, [6] D. T. Liu et al., Lithium-ion battery remaining useful life estimation
CRA and convergence [35] are calculated at the time indexes based on fusion nonlinear degradation AR model and RPF algorithm,
Neural Comput. Appl., vol. 25, no. 3/4, pp. 557572, Dec. 2014.
from the half to the end of the lifetime. The scores of the [7] N. Z. Gebraeel, M. A. Lawley, R. Liu, and V. Parmeshwaran, Residual
two metrics are summarized in Table IV. It is noticed that the life predictions from vibration-based degradation signals: A neural net-
exponential model has the highest CRA scores and the lowest work approach, IEEE Trans. Ind. Electron., vol. 51, no. 3, pp. 694700,
Jun. 2004.
convergence scores among the three models for all of the tested [8] F. D. Maio, K. L. Tsui, and E. Zio, Combining relevance vector machines
bearings. Therefore, the improved exponential model has the and exponential regression for bearing residual life estimation, Mech.
most accurate prediction results and the fastest convergence Syst. Signal Process., vol. 31, pp. 405427, Aug. 2012.
[9] J. Liu, W. Wang, and F. Golnaraghi, A multi-step predictor with a
speed among them. In conclusion, the improved exponential variable input pattern for system state forecasting, Mech. Syst. Signal
model performs best in the RUL prediction of the bearings. Process., vol. 23, no. 5, pp. 15861599, Jul. 2009.
[10] C. C. Chen, B. Zhang, G. Vachtsevanos, and M. E. Orchard, Machine
condition prediction based on adaptive neuro-fuzzy and high-order parti-
cle filtering, IEEE Trans. Ind. Electron., vol. 58, no. 9, pp. 43534364,
VI. C ONCLUSION Sep. 2011.
[11] A. Soualhi, H. Razik, G. Clerc, and D. D. Dinh, Prognosis of bearing
To enhance the performance of the original exponential failures using hidden Markov models and the adaptive neuro-fuzzy infer-
model in RUL prediction of rolling element bearings, an im- ence system, IEEE Trans. Ind. Electron., vol. 61, no. 6, pp. 28642874,
proved exponential model has been proposed in this paper. Jun. 2014.
[12] H. T. Liao, W. B. Zhao, and H. R. Guo, Predicting remaining useful
There are two major contributions in this work. First, an life of an individual unit using proportional hazards model and logistic
adaptive FPT selection approach is established based on the regression model, in Proc. Annu. Rel. Maintain. Symp., Newport Beach,
3 interval, in which a new trigger mechanism is applied CA, USA, Jan. 2326, 2006, pp. 127132.
[13] Z. G. Tian and H. T. Liao, Condition based maintenance optimization for
to restrict the interference of random noises. Second, PF is multi-component systems using proportional hazards model, Rel. Eng.
utilized to estimate the health state of the bearings and therefore Syst. Safety, vol. 96, no. 5, pp. 581589, May 2011.
LI et al.: IMPROVED EXPONENTIAL MODEL FOR PREDICTING RUL OF ROLLING ELEMENT BEARINGS 7773

[14] L. X. Liao, Discovering prognostic features using genetic programming Naipeng Li received the B.S. degree in me-
in remaining useful life prediction, IEEE Trans. Ind. Electron., vol. 61, chanical engineering from Shandong Agricul-
no. 5, pp. 24642472, May 2014. tural University, Shandong, China, in 2012. He
[15] N. Z. Gebraeel, M. A. Lawley, R. Li, and K. R. Jennifer, Residual-life is currently working toward the Ph.D. degree in
distributions from component degradation signals: A Bayesian approach, mechanical engineering in the State Key Lab-
IIE Trans., vol. 37, no. 6, pp. 543557, 2005. oratory for Manufacturing Systems Engineering,
[16] N. Z. Gebraeel, Sensory-updated residual life distributions for compo- Xian Jiaotong University, Xian, China.
nents with exponential degradation patterns, IEEE Trans. Autom. Sci. His research interests include machinery
Eng., vol. 3, no. 4, pp. 382393, Oct. 2006. condition monitoring, intelligent fault diagnos-
[17] A. H. Elwany and N. Z. Gebraeel, Sensor-driven prognostic models for tics, and remaining useful life prediction of ro-
equipment replacement and spare parts inventory, IIE Trans., vol. 40, tating machinery.
no. 7, pp. 629639, Apr. 2008.
[18] M. Y. You, F. Liu, W. Wang, and G. Meng, Statistically planned and
individually improved predictive maintenance management for continu-
Yaguo Lei (M15) received the B.S. and Ph.D.
ously monitored degrading systems, IEEE Trans. Rel., vol. 59, no. 4,
pp. 744753, Dec. 2010. degrees in mechanical engineering from Xian
[19] X. S. Si, W. B. Wang, M. Y. Chen, H. H. Chang, and H. Z. Dong, A Jiaotong University, Xian, China, in 2002 and
degradation path-dependent approach for remaining useful life estimation 2007, respectively.
He is currently a Full Professor of mechanical
with an exact and closed-form solution, Eur. J. Oper. Res., vol. 226,
no. 1, pp. 5366, Apr. 2013. engineering at Xian Jiaotong University. Prior to
[20] A. Ginart, I. Barlas, J. Goldin, and J. L. Dorrity, Automated feature joining Xian Jiaotong University in 2010, he was
a Postdoctoral Research Fellow with the Univer-
selection for embeddable prognostic and health monitoring (PHM) archi-
tectures, in Proc. IEEE Autotestcon, Anaheim, CA, USA, Sep. 1821, sity of Alberta, Edmonton, AB, Canada. He was
2006, pp. 195201. also an Alexander von Humboldt Fellow with
[21] V. T. Tran and B. S. Yang, An intelligent condition-based maintenance the University of Duisburg-Essen, Duisburg,
Germany. His research interests focus on machinery condition moni-
platform for rotating machinery, Expert Syst. Appl., vol. 39, no. 3,
pp. 29772988, Feb. 2012. toring and fault diagnosis, mechanical signal processing, intelligent fault
[22] X. S. Si, C. H. Hu, X. Y. Kong, and D. H. Zhou, A residual storage diagnostics, and remaining useful life prediction.
life prediction approach for systems with operation state switches, IEEE Dr. Lei is a member of the American Society of Mechanical En-
Trans. Ind. Electron., vol. 61, no. 11, pp. 63046315, Nov. 2014. gineers; a Senior Member of the Chinese Mechanical Engineering
[23] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, A tutorial Society; and an Editorial Board Member of Neural Computing and
on particle filters for online nonlinear/non-Gaussian Bayesian tracking, Applications, Advances in Mechanical Engineering, The Scientic World
Journal, International Journal of Mechanical Systems Engineering, Chi-
IEEE Trans. Signal Process., vol. 50, no. 2, pp. 174188, Feb. 2002.
[24] A. Doucet, N. D. Freitas, and N. Gordon, An introduction to sequential nese Journal of Engineering, International Journal of Applied Science
Monte Carlo methods, in Sequential Monte Carlo Methods in Practice. and Engineering Research, etc.
New York, NY, USA: Springer-Verlag, 2001, pp. 314, ch. 1.
[25] M. E. Orchard and G. Vachtsevanos, A particle-filtering approach for
on-line fault diagnosis and failure prognosis, Trans. Inst. Meas. Control, Jing Lin received the B.S., M.S., and Ph.D.
vol. 31, pp. 221246, Jun. 2009. degrees from Xian Jiaotong University, Xian,
[26] B. Zhang et al., A probabilistic fault detection approach: Application China, in 1993, 1996 and 1999, respectively, all
to bearing fault detection, IEEE Trans. Ind. Electron., vol. 58, no. 5, in mechanical engineering.
pp. 20112018, May 2011. He is currently a Professor with the State
[27] D. An, J. H. Choi, and N. H. Kim, Prognostics 101: A tutorial for particle Key Laboratory for Manufacturing Systems En-
filter-based prognostics algorithm using Matlab, Rel. Eng. Syst. Safety, gineering, Xian Jiaotong University. From July
vol. 115, no. 1, pp. 161169, Jul. 2013. 2001 to August 2003, he was a Postdoctoral
[28] E. Zio and G. Peloni, Particle filtering prognostic estimation of the Fellow with the University of Alberta, Edmonton,
remaining useful life of nonlinear components, Rel. Eng. Syst. Safety, AB, Canada, and a Research Associate with the
vol. 96, no. 3, pp. 403409, Mar. 2011. University of WisconsinMilwaukee, Milwaukee,
[29] M. E. Orchard, P. Hevia-Koch, B. Zhang, and L. Tang, Risk measures for WI, USA. From September 2003 to December 2008, he was a Research
particle-filtering-based state-of-charge prognosis in lithium-ion batteries, Scientist with the Institute of Acoustics, Chinese Academy of Sciences,
IEEE Trans. Ind. Electron., vol. 60, no. 11, pp. 52605269, Nov. 2013. Beijing, China, under the sponsorship of the Hundred Talents Program.
[30] M. Abbas, A. Ferri, M. E. Orchard, and G. Vachtsevanos, An intelli- His current research directions are in mechanical system reliability, fault
gent diagnostic/prognostic framework for automotive electrical systems, diagnosis, and wavelet analysis.
in Proc. IEEE Intell. Veh. Symp., Istanbul, Turkey, Jun. 1315, 2007, Dr. Lin was a recipient of the National Science Fund for Distinguished
pp. 352357. Young Scholars in 2011.
[31] B. Zhang et al., Rolling element bearing feature extraction and anomaly
detection based on vibration monitoring, in Proc. Mediterranean Conf.
Control Autom., Ajaccio, France, Jun. 2527, 2008, pp. 17921797.
[32] J. S. Liu and R. Chen, Sequential Monte Carlo methods for dynamic sys- Steven X. Ding received the Ph.D. degree in elec-
tems, J. Amer. Statist. Assoc., vol. 93, no. 443, pp. 10321044, 1998. trical engineering from the Gerhard Mercator
[33] Y. G. Lei, Z. J. He, Y. Y. Zi, and Q. Hu, Fault diagnosis of rotating University of Duisburg, Duisburg, Germany,
machinery based on multiple ANFIS combination with GAs, Mech. Syst. in 1992.
Signal Process., vol. 21, no. 5, pp. 22802294, Jul. 2007. From 1992 to 1994, he was a Research
[34] Y. F. Wang and P. J. Kootsookos, Modeling of low shaft speed bearing and Development Engineer with Rheinmetall
faults for condition monitoring, Mech. Syst. Signal Process., vol. 12, GmbH, Dusseldorf, Germany. From 1995 to
no. 3, pp. 415426, May 1998. 2001, he was a Professor of control engineering
[35] A. Saxena, J. Celaya, B. Saha, S. Saha, and K. Goebel, Metrics for offline with the University of Applied Science Lausitz,
evaluation of prognostic performance, Int. J. Prognost. Health Manage., Senftenberg, Germany, where he served as
vol. 1, no. 1, pp. 120, Jan. 2010. Vice President during 19982000. He is cur-
[36] P. Nectoux et al., PRONOSTIA: An experimental platform for bear- rently a Full Professor of control engineering and the Head of the
ings accelerated degradation tests, in Proc. Int. Conf. Prognost. Health Institute for Automatic Control and Complex Systems (AKS) at the
Manage., Denver, CO, USA, Jun. 1821, 2012, pp. 18. University of Duisburg-Essen, Duisburg. His research interests include
[37] IEEE PHM 2012 Prognostic Challenge. Outline, Experiments, Scor- model-based and data-driven fault diagnosis, fault-tolerant systems,
ing of Results, Winners. [Online]. Available: http://www.femto-st.fr/f/d/ real-time control, and their applications in industry, with a focus on
IEEEPHM2012-Challenge-Details.pdf automotive systems and chemical processes.