You are on page 1of 19

Mechanical Systems and Signal Processing 134 (2019) 106302

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing


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

Weibull accelerated failure time regression model for remaining


useful life prediction of bearing working under multiple
operating conditions
Pradeep Kundu a,⇑, Ashish K. Darpe a, Makarand S. Kulkarni b
a
Vibration Research Laboratory, Department of Mechanical Engineering, Indian Institute of Technology Delhi, 110016, India
b
Department of Mechanical Engineering, Indian Institute of Technology Bombay, 400076, India

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

Article history: The rolling element bearings in industry applications operate at different operating condi-
Received 6 June 2018 tions. The approaches for remaining useful life (RUL) prediction developed so far are lim-
Received in revised form 4 July 2019 ited to bearings operating under a single operating condition. Thus, separate models
Accepted 8 August 2019
need to be developed for each operating condition, which is a tedious and time-
consuming task. In this paper, a Weibull Accelerated Failure Time Regression (WAFTR)
model is presented that considers both operating condition parameters and condition
Keywords:
monitoring signal during model parameter estimation. Using the vibration signal, statisti-
Remaining useful life
Weibull regression
cal time domain features such as root mean square, kurtosis, peak and crest factor and fre-
Multiple operating conditions quency domain features such as variation of % signal energy in the different spectrum
Bearing bands (obtained after applying the bank of band pass filters) are extracted. Frequency
domain features are used for further study since better trendability are found with these
features compared to the statistical time domain features. For each bearing dataset, sixteen
features are available from sixteen different frequency bands. However, all do not have bet-
ter trendability and therefore principal component analysis (PCA) is used for dimensional-
ity reduction. The best PC value and operating conditions such as speed and load are used
in the WAFTR model for RUL prediction. The algorithm performance has been checked with
metrics such as bias, mean square error, updated score value, % unacceptable early predic-
tions, % unacceptable late predictions and % unacceptable predictions. The accuracy of the
RUL prediction is found superior with the model that includes the effect of operating con-
ditions and does not show significant bias as noticed in the RUL prediction when the oper-
ating conditions were not considered in the formulation of the model.
Ó 2019 Elsevier Ltd. All rights reserved.

1. Introduction

Strong market competition restricts the industries to sell their products at a competitive price. Therefore, to ensure higher
profit margin, the industries are focusing on servicing their traditional product-focused businesses. In such scenarios,
original equipment manufacturers (OEM) are made to share the risk of failures of products by engaging in the contractual
service agreement (CSA) with their customer. Under such agreements, the industry is completely responsible for machine
maintenance during the agreement period. The newer maintenance technology is oriented around Prognosis and Health

⇑ Corresponding author.
E-mail address: pkundu.pradeep@gmail.com (P. Kundu).

https://doi.org/10.1016/j.ymssp.2019.106302
0888-3270/Ó 2019 Elsevier Ltd. All rights reserved.
2 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Management (PHM). Effective implementation of the PHM requires an accurate model for predicting the remaining useful
life of the components.
In literature, various models are available from data driven approaches to physics based approaches for remaining useful
life (RUL) prediction. The physics based approaches use damage propagation models such as Paris law, stiffness based meth-
ods, etc. for RUL prediction. These models calculate the stress at each failure site as a function of loading conditions, product
geometry and the material properties and then predict the RUL [1–4]. Physics based models are defect specific, too stochastic
and complex for developing a reliable and robust mathematical model of RUL prediction [5]. On the other hand, data driven
approaches utilize the condition monitoring (CM) data and do not require any assumption or empirical estimation of phys-
ical parameter. In the data driven prognosis approaches, the RUL estimation of the bearing at different load and speed con-
ditions is the primary requirement of the industry [6]. Some work [7–11] is reported for bearing RUL estimation under
different operating conditions.
Liang et al. [7] used the recurrent neural network (RNN) model for bearing RUL prediction under multiple operating con-
ditions. Input layer, one or more hidden layer and output layer are the main elements of the neural network (NN). But the
type and optimum structure selection is still a major challenge with NN. Statistical approaches can be used to overcome this
problem. In this category, hidden semi-Markov model (HSMM) is proposed by Cartella et al. [8] for bearing RUL prediction
under multiple operating conditions. In the HSMM, the number of degradation states is often determined subjectively based
on experience. The large amount of historical data is also required for accurate state estimation and modelling [12]. In addi-
tion, in the case of bearing, lack of relation of the defined health-state change point to the actual defect progression since it is
often impractical to physically observe a defect in the bearing leads to interpretability issue with HSMM. Liao et al. [9] pre-
sented a framework for RUL estimation of the bearing under time-varying operating conditions. The root mean square (RMS)
value of the vibration signal is used as the degradation measure and assumed that this degradation phenomenon is governed
by a Brownian motion with linear drift and follows normal distribution. The model parameters are estimated for a particular
operating condition and if the operating condition changes the initial estimates of the model parameters were estimated off-
line based on acceleration degradation test experiment. Later the model parameters were updated using the Bayesian tech-
nique for further improvement in the RUL prediction. In this work, it was assumed that once a defect initiated in the bearing,
it will trend up linearly under a fixed operating condition, which is usually not linear specially for the bearings. Hence, Sin-
gleton et al. [10] used the extended Kalman filter (EKF) algorithm which can model nonlinear system dynamics for RUL pre-
diction for bearing working under multiple operating conditions. Huang et al. [11] proposed an adaptive skew-Wiener model
for RUL prediction. It was demonstrated that the closed skew s-normal distribution is more suitable than the normal distri-
bution to model the degradation drift in RMS-based health indicator extracted from the vibration signal. In this process, the
model parameters were updated based on the Kalman filter algorithm. This process only employs the information contained
in the current degradation rather than the information of entire sequence of observations. However, the Wiener process is a
time homogeneous process and all degradation processes do not have this property, specially bearings [13].
In all these models, initial estimates of the model parameters are not updated/changed if the operating conditions chan-
ged as it was done in [9]. In few works [7,11], the degradation data from each operating condition was considered for model
parameters estimation and then RUL was predicted for the bearing, assuming that all the bearings belong to a single popu-
lation. Obviously, the variability in the time to failure for bearings belonging to wide range of operating conditions is very
high and if the operating condition is not considered as a factor/parameter for RUL estimation then it leads to a high error in
the RUL prediction. Hence it can also be noticed that in these works the RUL was predicted accurately only in the end stage of
the bearing life. In some works [8,10], different models were trained for different operating conditions that is time consum-
ing and for each operating condition large amount of historical data is required.
In addition, the methods such as Brownian motion with linear drift, Kalman filter, Wiener process, Bayesian updating are
based on the sampling. Hence they don’t give repeatable (deterministic) results, i.e., not give same PDF of the RUL on every
repetition of the algorithm. This feature is very important criteria for existing verification, validation and certification pro-
tocols in the industry such as aerospace [14]. Hence these algorithms are not suitable for the online implementation. In addi-
tion, in these approaches, the type of initial distribution and their parameters need to be pre-defined to obtain meaningful
prediction results. Hence, for defining the component health state distribution, for each operating condition, large amount of
historical data is required [15]. In practical applications, availability of such prior knowledge about the distribution param-
eters is difficult.
To overcome the shortcomings of the existing RUL prediction models, a novel deterministic data driven Weibull Acceler-
ated Failure Time Regression (WAFTR) model is presented in this paper which considers the operating condition as a feature
in addition to the condition monitoring data. Consideration of operating conditions with the CM data will ensure that a single
model is required for RUL prediction instead of developing the different models for different operating conditions. Percent-
age signal energy calculated in sixteen different spectrum bands are used as features to represent the bearing degradation.
The % signal energy estimated in each spectrum bands are not found significant and hence feature dimensionality is reduced
using principal component analysis (PCA). The best PC value, the bearing applied load and operating speed are used as the
input variables for model parameter estimation. The result based on the performance metrics demonstrates that the pro-
posed method helps in systematically incorporating the operating conditions in the model and accurate RUL prediction
results within an error bound are observed.
The rest of the paper is organized as follows: the WAFTR model is presented in Section 2. Section 3 explains the procedure
used for bearing RUL prediction. Experimental setup, features used for RUL prediction, feature dimensionality reduction and
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 3

performance metrics used for validating the model performance are discussed in this section. Section 4 presents the results
and discusses the significance of operating condition consideration. Conclusions and scope for future work are discussed in
Section 5.

2. Weibull accelerated failure time regression model

In 1936, Waloddi Weibull discovered a probability distribution called Weibull distribution that reflects the fatigue life of a
mechanical component under random loads [16]. This distribution is widely used to analyse reliability and maintainability of
many mechanical systems. Mathematically, the cumulative density function (CDF), F ðt Þ of a simple Weibull model is
expressed as [17]
 b !
t
F ðt Þ ¼ 1  exp  ð1Þ
g
where, t is the time to failure, i.e., primary ageing parameter, g is the scale or characteristics life parameter, i.e., time at which
63.2% of failure occurs and b is the shape parameter. The shape parameter characterizes the dispersion of failures.

b < 1: Component risk of failure decreases as the ageing parameter value increases.
b ¼ 1: Component risk of failure is constant as the ageing parameter value increases.
b > 1: Component risk of failure increases as the ageing parameter value increases.

The probability of failure changes with time, but there are some additional factors that contribute to failure rate progres-
sion. These factors may be vibration, applied load, operating speed, temperature, humidity, pressure, component design con-
figuration, etc. These parameters are also called as secondary ageing parameters or covariates. The secondary ageing
parameters are divided into two categories: external ageing parameters and internal ageing parameters. The parameters
such as speed, load and pressure are called as external ageing parameters or external covariates. Although, the bearing
has same material, design configuration and operated under same operating conditions but usually they have different time
of failure due to different reasons such as manufacturing process, machining quality and end user of the product. Hence, the
bearing health is monitored using sensors such as vibration, acoustic emission, proximity, temperature, etc. The output of
these sensors are termed as CM parameters or internal ageing parameters or internal covariates. Two or more ageing param-
eters can be modelled simultaneously with WAFTR model. The WAFTR model CDF value can be expressed as
0 0 1b 1
t
F ðt; X Þ ¼ 1  exp@@ Pk A A ð2Þ
expao þ aX
i¼1 i i

th th
where, ao is the intercept, X i is the i covariate, i.e., the i secondary ageing parameter and ai is the corresponding regression
coefficient. The secondary ageing parameters (internal ageing parameters and external ageing parameters) are used to
update the g value as g is the function of chosen secondary ageing parameters. In an alternate expanded form, the CDF value
can be expressed as
 b !
t
F ðt; Im ; En Þ ¼ 1  exp  ð3Þ
exp½ao þ ða1 I1 þ a2 I2 :               am Im Þ þ ðb1 E1 þ b2 E2 :               bn En Þ

where, I1 , I2 . . ..Im are the m internal covariates and E1 , E2 . . ..En are the n external covariates, aj¼1;2;;m and bk¼1;2;;n represent
the regression coefficients for the internal and the external covariates respectively. In a generalized form, the CDF for WAFTR
model is given as
 b !
t
F ðt; I; EÞ ¼ 1  exp  ð4Þ
expðao þ aI þ bEÞ

The hazard rate function for the WAFTR model is


 b1
b t
hðt; I; EÞ ¼ ð5Þ
expðao þ aI þ bEÞ expðao þ aI þ bEÞ
The PDF value of the WAFTR is
f ðt; I; EÞ ¼ hðt; I; EÞ  ½1  F ðt; I; EÞ ¼ hðt; I; EÞ  Rðt; I; EÞ
 b1 "  b #
b t t
¼  exp  ð6Þ
expðao þ aI þ bEÞ expðao þ aI þ bEÞ expðao þ aI þ bEÞ
4 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

The unknown parameters in the WAFTR equation are calculated by maximum likelihood estimation (MLE) approach. The
likelihood function can be defined as

Y
l
  Yp
Lðb; ao ; a; bÞ ¼ f tj ; Ij ; Ej  Rðt k ; Ik ; Ek Þ ð7Þ
j¼1 k¼1

"  b # Y "  b #
Y
l
b  b1 tj
p
tk
Lðb; ao ; a; bÞ ¼  b jt  exp   exp 
j¼1 ðexp ao þ aIj þ bEj Þ expðao þ aIj þ bEj Þ k¼1
expðao þ aIk þ bEk Þ
ð8Þ

where, j indexes for the number of failure data points and k indexes for the number of suspension or degradation data points.
For example, if a bearing failed at 30 h of cycle time and if its health is monitored at every five hours, then 6 data points are
available to define the bearing health state. The intermediate five data points will be considered as degradation or censored
data points (k ¼ 5) and last data point will be considered as failure data point (j ¼ 1). In another scenario, if two bearings
failed at 30 and 40 h of cycle time, then 6 and 8 data points are available to define both the bearing health state respectively.
The intermediate data points of both the bearings will be considered as degradation or censored data points (k ¼ 5 þ 7 ¼ 12)
and last data point of both the bearings will be considered as failure data points (j ¼ 1 + 1 = 2). For computational conve-
nience, the log-likelihood function is used and can be expressed as

X
l
ln½Lðb; ao ; a; bÞ ¼ k lnðbÞ  b½ðao þ aI1 þ bE1 Þ  ðao þ aI2 þ bE2 Þ            ðao þ aIl þ bEl Þ þ ðb  1Þ ln tj
j¼1
" b  b  b #
t1 t2 tl
 þ þ  þ
ao þ aI1 þ bE1 ao þ aI2 þ bE2 ao þ aIl þ bEl
" b  b  b #
t1 t2 tp
 þ þ  þ ð9Þ
ao þ aI1 þ bE1 ao þ aI2 þ bE2 ao þ aIp þ bEp

Taking the partial derivative of Eq. (9) with respect to the parameters b; ao ; a; b and setting the derivative equal to zero,
optimal values of the parameters b; ao ; a; b can be obtained. The WAFTR model is a fully parameterised model. The historical
lifetime data, censored or degradation data, corresponding condition monitoring data and operating condition data can be
combined for estimating the RUL at any time ‘t’.
In the proposed approach, both the operating conditions and the degradation extent (through measured vibration param-
eter values) have been taken into consideration to improve the correctness of the model and possible improvement in the
accuracy of prediction. Earlier researchers did not include the degradation extent and operating conditions together in the
RUL prediction model. Consideration of operating conditions with the degradation extent presented in the current work
ensures that a single model is required for the RUL prediction instead of developing different models for different operating
conditions as has been done in the past.

3. Bearing RUL prediction procedure

Fig. 1 shows the flowchart of the proposed methodology used for RUL prediction. The methodology is divided in two
phases: offline and online. In the offline phase, the time domain (RMS, kurtosis, peak and crest factor) and frequency domain
(variation of % signal energy in different spectrum bands obtained after applying the bank of band pass filters) based features
are extracted from the raw vibration signal. Based on the trendability analysis, frequency domain based features are found
better than conventional time domain based features and hence used in further study. Percentage signal energy estimated in
different spectrum bands gives different trendability value with bearing degradation. A priori for a particular bearing it is
also difficult to know % signal energy estimated in which spectrum band is ideal to represent the degradation. In addition,
the feature (% signal energy) estimated in a particular spectrum band that is found to be ideal for a particular bearing may
have a poor trendability for another bearing in the same spectrum band. Hence, the size of these features estimated in dif-
ferent spectrum bands is reduced using PCA. The features are normalized while performing the fusion using principal com-
ponent. The normalization is done using the maximum and minimum values of the feature in the particular spectrum band
in the training dataset. On the extracted principal components, smoothing process is performed using the moving average
method as small areas of fluctuation in the feature may affect the prognosis model stability. First PC value is used as an inter-
nal covariate whereas, bearing operational load and speed are used as external covariates for calculating the unknown
parameters of the WAFTR model. The model performance is checked based on some performance metrics. In the online
phase, the RUL is estimated based on the obtained WAFTR model coefficients during model training. For new bearing, the
RUL is estimated at every 10 s based on the current time, operating condition data and first PC value extracted from fre-
quency domain based feature space. The overall procedure for implementing the proposed methodology is depicted in Fig. 1.
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 5

Training and Validation Operating


Condition

Offline Phase
Raw Vibration Feature Dimensionality Feature WAFTR Perform
Data from fleet Extraction and Reduction Creation Model ance
Smoothing
of failed Normalization Using PCA Parameter Metrics
Bearings Estimation

Optimum
Structure
Testing Operating
Condition

Online Phase
Raw Vibration Feature Dimensionality RUL
Feature
Data from a Extraction and Reduction Estimation
Smoothing
new bearing Normalization Using PCA Using
(real time) WAFTR

Fig. 1. Flowchart of the proposed bearing RUL estimation procedure.

3.1. Experimental setup

Fig. 2 is an experimental platform developed by FEMTO science & technology group. With this platform, run to failure
experiments on roller element bearings can be performed under specific operating conditions. Two accelerometers and
one thermocouple are mounted for measuring the horizontal, vertical vibration signal and temperature signal respectively.
The vibration signal was measured after every 10 s and at a sampling frequency of 25.6 kHz. The temperature signal was
measured after every 60 s and at a sampling frequency of 10 Hz [18]. The time to failure (TTF) values observed for bearings
with PRONOSTIA platform are shown in Table 1. The source [18] from which the data is taken (i.e. IEEE PHM 2012 challenge)
does not specify the threshold criteria and corresponding values and only the TTF data (Table 1) for each bearing is provided.

Fig. 2. Overview of PRONOSTIA Platform [18].


6 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Table 1
Observed TTF values for bearings (in seconds).

Operating Conditions
1. Speed = 1800 RPM 2. Speed = 1650 RPM 3. Speed = 1500 RPM
Load = 4000 N Load = 4200 N Load = 5000 N
1_1 (28030) 2_1 (9110) 3_1 (5150)
1_2 (8710) 2_2 (7970) 3_2 (16370)
1_3 (23750) 2_3 (19550) 3_3 (4340)
1_4 (14280) 2_4 (7510)
1_5 (24630) 2_5 (23110)
1_6 (24480) 2_6 (7010)
1_7 (22590 s) 2_7 (2300 s)

Hence, the vibration data till this TTF is considered as a threshold point for model training and testing. Some of the studies in
the literature [8,19–22] have followed the same approach.
During initial modelling, both temperature and vibration data were considered. No improvement in the RUL prediction
was observed with the temperature data and hence only vibration data is considered for further analysis. In Table 1, the
number before underscore represents the operating condition (as described by the speed and load values at the top of
the column) and the number after underscore represents the bearing number for that operating condition. The TTF values
observed for the given bearing at the given operating condition is indicated within the round brackets. For example, the bear-
ing number 1 which is operating at the speed 1800 rpm at 4000 N is found to fail after 28,030 s, while the TTF for bearing 3
operating at 1500 rpm at 5000 N load is recorded to be 4340 s.
The measured time domain vibration data for complete life cycle for all the bearings are shown in Fig. 3. The vibration
signal for some of the bearings (bearing test cases 1_2, 2_3, 2_5 and 3_2) show some sporadic and random high vibration
levels (shown encircled in Fig. 3), in the initial part of the experiment. One of the possible reasons for such data could be
improper assembly/installation of bearings before the experiment. Such datasets with abnormally high and erratic signal
levels may obscure the physics of the degradation process in the bearing and it is just and appropriate to leave out such data-
sets in the model building exercise. In addition, bearing 7 from operating condition (2) (OC2) exhibits an abnormal, irratic
and sporadic sharp rise in the vibration values (>16 g) in the middle of the experiment which cannot be related to the typical
failure mechanism. Based on overall vibration history, bearing 2 from operating condition (1) (OC1), bearing 3, 5 and 7 from
OC2 and bearing 2 from operating condition (3) (OC3) have not been considered in the present study.

3.2. Feature extraction and smoothing

Although the raw vibration data obtained from the sensors is very rich in content, most of the times it does not give
required usable information. For the purpose of diagnosis and prognosis, it is necessary to process the data for extracting
the feature that can be related to flaws in the bearing. Feature extraction step removes the unnecessary redundant and insen-
sitive information as well as content such as noise or sporadic impulses etc. that obscures the important bearing specific
information from the raw signal. The conventional time domain features such as RMS, kurtosis, peak and crest factor are
extracted from the raw vibration signal. Fig. 4 shows the trend of these feature value extracted from the vibration signal
for bearing 1 from the OC1 and bearing 6 from OC2. The variation in these features is observed only a few times the initial
level before the failure. Thus from the point of view of getting sufficient advance information using such features is ques-
tionable for prognosis. In the industry environment, most of the bearing failures appear suddenly and unfortunately, the pre-
sent conventional time domain features may not show a distinct monotonic degradation process. In the present work, the
variation of % signal energy in different spectrum bands is used to represent the bearing degradation. As degradation pro-
gresses in the bearing, the vibration characteristics continue to change as stiffness of the bearing changes due to fault pro-
gression. Hence as fault progresses, the signal energy is transfer from one frequency band to another frequency band. Hence
the signal energy in some frequency bands increases whereas in some frequency band decreases and this change in signal
energy can be used to represent the bearing degradation. Using the bank of band pass filters, sixteen different spectrum
bands were constructed to estimate the % signal energy. The vibration data used in the present work was recorded at a sam-
pling rate of 25.6 kHz and hence the vibration signal in the frequency range of 0–12.8 kHz is used. Sixteen different spectrum
band signals were obtained by applying the band pass filter in the range of 800 Hz. These sixteen different bands would be
(0 Hz–800 Hz), (801 Hz–1600 Hz), (1601 Hz–2400 Hz), . . .. . .. . .., (12001 Hz–12800 Hz). The % signal energy estimated using
vibration signal obtained in these 16 spectrum bands are considered as frequency domain based features to represent bear-
ing degradation. Fig. 5 shows the few of the extracted frequency domain based features for bearing 1 from OC1 and bearing 6
from OC2.
The extracted features should possess good correlation with the age and health state of the bearing. The level of corre-
lation can also influence the subsequent steps and even the final prognosis results [6,23]. The trendability parameter shows
the feature correlation to time and ranges between 0 and 1 and correlated to the machine performance. The trendability is
defined as [24]
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 7

Fig. 3. Raw vibration time domain plot for bearings working under different operating conditions.
8 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

(a) (b)
Fig. 4. Conventional features variation with time (a) for bearing 1 from OC1 (b) for bearing 6 from OC2.

(a) (b)
Fig. 5. Percentage signal energy variation with time (a) for bearing 1 from OC1 (b) for bearing 6 from OC2.

P P P
jN X ðtN Þt N  X ðt N Þ t N j
TrendabilityðX; T Þ ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h P ih P i ð10Þ
P P
N X 2 ðt N Þ  ð X ðt N ÞÞ2 N t2N  ð t N Þ2

where, X ðtN Þ is the degradation feature value at time tN and N is total number of data points.
Table 2 shows the obtained trendability value for each time domain feature and best three frequency domain features.
From Fig. 4, non-monotonic and non-trendable behaviour was observed with the conventional time domain features. This
issue is confirmed from the poor trendability value for these features as shown in Table 2. Comparatively, frequency domain
based new set of features shows a high trendability value as shown in Table 2. From Table 2, Figs.4 and 5, the frequency
domain features have better trendability and monotonic behaviour and hence used for further study.

3.3. Dimensionality reduction

The feature space should have the better trendability with bearing life for accurate life prediction. The vibration signal
recorded from a typical bearing has different energy in different frequency bands and the energy associated with the fault
interaction in the bearing lies in particular frequency bands. However, as fault progresses, the signal energy is transfer from
one spectrum band to another spectrum band. Hence, % signal estimated in all sixteen spectrum bands have different trend-
ability. In addition, it is not mandatory that the energy extracted for a particular spectrum band which has the highest trend-
ability for one bearing will also have the highest trendability for another bearing. For example, as noticed from Table 2, %
energy estimated in second frequency band is the ideal feature for bearing 1_1 whereas for bearing 1_4 the ideal feature
is % energy estimated in the seventh frequency band. For a new bearing, it would be difficult to know, a-priori, % energy from
which frequency band will highest correlate with the bearing degradation. Hence, % signal energy estimated in all sixteen
spectrum bands need to be combined for estimating a reliable degradation index for bearing health. The unsupervised tech-
nique, the PCA is widely used for feature dimensionality reduction. The PCA converts a set of correlated features into a new
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 9

Table 2
Features trendability value comparison.

Bearing ID RMS Kurtosis Peak Crest factor Band 02 Band 08 Band 07


1_1 0.71 0.69 0.63 0.83 0.96 0.94 0.93
Bearing ID RMS Kurtosis Peak Crest Factor Band 02 Band 06 Band 07
1_3 0.58 0.69 0.63 0.71 0.96 0.94 0.93
Bearing ID RMS Kurtosis Peak Crest Factor Band 07 Band 14 Band 08
1_4 0.70 0.43 0.72 0.55 0.87 0.84 0.73
Bearing ID RMS Kurtosis Peak Crest Factor Band 15 Band 14 Band 13
1_5 0.22 0.27 0.08 0.38 0.94 0.92 0.89
Bearing ID RMS Kurtosis Peak Crest Factor Band 08 Band 07 Band 05
1_6 0.13 0.35 0.16 0.44 0.91 0.85 0.76
Bearing ID RMS Kurtosis Peak Crest Factor Band 07 Band 08 Band 12
1_7 0.35 0.45 0.39 0.71 0.94 0.87 0.80
Bearing ID RMS Kurtosis Peak Crest Factor Band 05 Band 13 Band 16
2_1 0.44 0.51 0.61 0.67 0.72 0.65 0.64
Bearing ID RMS Kurtosis Peak Crest Factor Band 09 Band 16 Band 07
2_2 0.78 0.46 0.66 0.46 0.88 0.79 0.78
Bearing ID RMS Kurtosis Peak Crest Factor Band 11 Band 16 Band 14
2_4 0.28 0.11 0.07 0.13 0.95 0.95 0.93
Bearing ID RMS Kurtosis Peak Crest Factor Band 16 Band 13 Band 14
2_6 0.12 0.52 0.006 0.51 0.96 0.93 0.93
Bearing ID RMS Kurtosis Peak Crest Factor Band 11 Band 10 Band 09
3_1 0.064 0.26 0.24 0.40 0.82 0.82 0.80
Bearing ID RMS Kurtosis Peak Crest Factor Band 11 Band 14 Band 05
3_3 0.67 0.31 0.64 0.17 0.87 0.84 0.84

set of linearly uncorrelated features called principal components (PC) using orthogonal transformation [23]. The singular
value decomposition (SVD) algorithm is used for extracting the PC. The frequency domain based features are normalized
while performing the fusion using principal component. Fig. 6 shows the variation in the first value of the principal compo-
nent. The reported trendability value after dimensionality reduction is shown in Table 3 which is almost the same as
obtained for highest trendable frequency domain based feature for the particular bearing. To reduce the variability of the
extracted principal components, the smoothing process is performed using moving average method as small areas of fluc-
tuation in the feature may affect the prognosis model stability.

3.4. RUL prediction procedure

The PC value obtained from frequency domain based features has better trendability with bearing age and is considered
as the secondary ageing parameter and an internal covariate for RUL prediction in WAFTR model. The bearing life depends on
the external operating conditions such as load and speed. Hence they are considered as secondary ageing parameters and
external covariates. As discussed earlier in Section 2, the reliability function at time ‘t’ for the current bearing problem is
defined as
"  b #
t
Rðt; PC; Load; SpeedÞ ¼ exp  ð11Þ
expðao þ a1  PC þ b1  Load þ b2  SpeedÞ

(a) (b)
Fig. 6. First Principal Component value variation with time (a) for bearing 1 from OC1 (b) for bearing 6 from OC2.
10 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Table 3
Frequency domain feature trendability after dimensionality reduction.

Bearing ID 1_1 1_3 1_4 1_5 1_6 1_7 2_1 2_2 2_4 2_6 3_1 3_3
Trendability 0.94 0.93 0.93 0.88 0.66 0.85 0.66 0.67 0.92 0.96 0.5 0.74

where, ao is the intercept, b is the distribution shape parameter and a1 , b1 and b2 are the regression coefficient for PC value,
load and speed respectively. The RUL of the bearing in present case is calculated based on the conditional reliability which is
defined as
Rðt þ t o Þ
Rðtjt o Þ ¼ ð12Þ
Rðt o Þ
where, to is the current time, t is the bearing time to failure and Rðtjto Þ is the conditional reliability value which is defined as
a component or system that will operate without failure for a mission time, t, given that it has already survived to a given
time, to . At current time ‘to ’, the bearing reliability can be calculated using Eq. (11) and the bearing TTF can be estimated by
extrapolating the current usage condition to critical conditional reliability value given in Eq. (12). The choice of critical con-
ditional reliability value depends on the user. The value of ‘t’ in the Eq. (12) is the predicted RUL value for bearing at time ‘to ’.

3.5. Performance evaluation method

Performance metrics are used for checking the accuracy of prognosis algorithm and are the function of prediction error.
Traditional performance metrics such as bias and mean square error (MSE) are defined as

1X N
Bias ¼ Ei ð13Þ
N i¼1

1 XN
MSE ¼ ðEi Þ2 ð14Þ
N i¼1

where, Ei is the difference between actual RUL and predicted RUL at time instant i.
The metric Bias may mislead the accuracy judgment as prediction error can be negative or positive whereas the metric
MSE gives the absolute value of prediction error [25].
The metric score was proposed by Zerhouni et al. [16]. This metric does not consider under-prediction and over-
prediction in RUL in the same manner. The score gives an exponential weightage to prediction error through an asymmetric
function, i.e., penalizes the over prediction more than under prediction. The score value lies between 0 and 1, and the model
is considered to be more accurate if score value is 1. The metric score is defined as
( )
exp ln ð0:5Þ:ðEi =5Þ if Ei  0
Score ¼ Av erage ð15Þ
expln ð0:5Þ:ðEi =20Þ if Ei > 0

In the metric score, the mathematical expression developed for over-prediction in RUL gives a very high exponential pen-
alty to the prediction error and may mislead the results. For example, from Fig. 7, the score value for over-prediction in RUL
curve is 0.24 while the score value for under-prediction in RUL curve is 0.61. However, both the curves are equidistant from

Fig. 7. Demonstration of RUL under-prediction and over-prediction phenomena.


P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 11

the actual RUL curve but lying on the opposite side. According to score metric, if the model is over predicting, then it should
be considered as a poor model while if the model is under predicting then should be considered as a good model. However,
both under-prediction and over-prediction in RUL are not considered good for the prognosis model. The under-prediction
may lead to economic burden while the over-prediction may cause unplanned outages and hence both should be avoided.
The score value is updated and called as update score value (USV). The mathematical expression in metric score developed
for under-prediction in RUL is considered as USV and defined as

Updated Score Value ðUSVÞ ¼ Av erageðexpln ð0:5Þ:ðjEi j=20Þ Þ ð16Þ


The metrics such as % unacceptable early predictions (PUEP), % unacceptable late predictions (PULP) and % unacceptable
predictions (PUP) can also be used for assessing the model performance. These metrics are specially designed based on the
error bound. The error bound is the area in which the prediction made by the prognosis model is acceptable. The metric PUEP
estimates the percentage of times a model under predict the RUL outside the acceptable area or error bound. The metric PULP
estimates the percentage of times a model over predict the RUL outside the acceptable area or error bound. Both under-
prediction and over-prediction in RUL value are not significant and hence model performance can also be checked with met-
ric PUP. This metric estimates the percentage of times the model prediction lies outside the acceptable region or error bound.

4. Results and discussion

The bearing vibration and operating condition data obtained from the test rig in [18] is used to test the performance of the
WAFTR model. The model accuracy is checked with performance metrics explained in Section 3.5.

4.1. RUL prediction including the tested bearing operating condition in training

Twelve bearing failure histories were available for model training and testing as discussed in Section 3.1. Leave one out
cross validation approach is used for checking the model performance. Eleven bearing failure histories are used for model
parameter estimation. Performance of the model was tested with the twelfth bearing that was not considered during model
parameter estimation. For example, bearing 3, 4, 5, 6 and 7 from OC1, bearing 1, 2, 4 and 6 from OC2 and bearing 1 and 3
from OC3 are used for model parameter estimation for the RUL prediction of the bearing 1 from OC1. Likewise, the process is
repeated for other bearings as shown in Table 4.
Table 5 shows the model coefficient values obtained with different training data sets. The bearing shape parameter value
is obtained in the range of 1.58 and 1.59. The intercept value lies in the range of 12.93 and 10.14, the coefficient for the PC
lies in the range of 0.33 and 0.16, for load the coefficient range is 4.89 and 6.27 and for the speed coefficient range is 15.86
and 17.53. With different training data sets, a low variation (0.69 for intercept, 0.05 for PC coefficient, 0.33 for load coeffi-
cient, 0.43 for speed coefficient and 0.004 for shape parameter) in the model coefficient values indicate the model consis-
tency. In addition, the coefficient value for the speed is highest compared to the load and the PC, indicating a strong
influence of speed parameter on the bearing life. The load has the second highest coefficient value and coefficient value
for PC is the lowest one. It may be noted that the external covariates load and speeds are the cause of the bearing failure
and internal covariate (PC extracted from the raw vibration signal) values are the effect due to external covariates.
The RUL prediction results with the WAFTR model are shown in Fig. 8. All the results shown in Fig. 8 correspond to the
90% critical failure probability, i.e., (1R(t+to)). The choice of critical failure probability value depends on the user and in the
present case, best RUL prediction results are obtained with the 90% failure probability. An area of 20%of the actual RUL is
considered as error bound. The model with Operating Condition Consideration (WOC) accurately predicts the RUL for all the
bearings within 20%error bound, except for the bearing 1_4.

Table 4
Training and testing model development when all operating conditions are considered.

OC1 OC2 OC3


1_1 1_3 1_4 1_5 1_6 1_7 2_1 2_2 2_4 2_6 3_1 3_3
Test Train Train Train Train Train Train Train Train Train Train Train
Train Test Train Train Train Train Train Train Train Train Train Train
Train Train Test Train Train Train Train Train Train Train Train Train
Train Train Train Test Train Train Train Train Train Train Train Train
Train Train Train Train Test Train Train Train Train Train Train Train
Train Train Train Train Train Test Train Train Train Train Train Train
Train Train Train Train Train Train Test Train Train Train Train Train
Train Train Train Train Train Train Train Test Train Train Train Train
Train Train Train Train Train Train Train Train Test Train Train Train
Train Train Train Train Train Train Train Train Train Test Train Train
Train Train Train Train Train Train Train Train Train Train Test Train
Train Train Train Train Train Train Train Train Train Train Train Test
12 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Table 5
Model parameters with different training data sets when tested bearing operating conditions are considered.

Tested Bearing ao (Intercept) a1 (PC Coefficient) b1 (Load Coefficient) b2 (Speed Coefficient) b (Shape Parameter)

1_1 10.68 0.26 5.25 16.11 1.59


1_3 11.41 0.25 5.53 16.64 1.58
1_4 10.68 0.26 5.25 16.11 1.59
1_5 11.46 0.20 5.52 16.68 1.59
1_6 11.67 0.29 5.65 16.84 1.58
1_7 11.62 0.21 5.59 16.80 1.59
2_1 12.93 0.16 6.27 17.53 1.59
2_2 11.37 0.19 5.50 16.59 1.59
2_4 10.90 0.26 5.29 16.32 1.59
2_6 10.14 0.21 4.89 15.86 1.59
3_1 11.05 0.21 5.30 16.45 1.59
3_3 11.08 0.33 5.35 16.48 1.58
Mean 11.25 0.24 5.44 16.53 1.58
Standard Deviation 0.69 0.05 0.33 0.43 0.004

Without the operating condition (WoOC) consideration, the model significantly underestimates the RUL value for the
bearings from OC1 and substantially overestimates for the bearings from OC2 and OC3. This happens because the model
assumes that all the bearings belong to a single population if operating conditions are not included in the RUL prediction
model. Hence, the large variation in the TTF values among different operating conditions causes bias in the model and results
in the underestimation or overestimation in the RUL prediction.
The performance metrics results obtained with the proposed methodology are listed in Table 6. For all the bearings, the
values of the bias and MSE metrics show that the incorporation of the operating conditions in the RUL prediction model sig-
nificantly reduces the overall prediction error. For example, for bearing 1_1, the bias reduces from 5.52e3 to 4.49e3 if the
operating conditions are included in the model. Similarly, for the same bearing, the MSE reduces from 3.9e7 to 2.15e7 if
the operating conditions are included in the model. In addition, for each case of the bearing, there is a significant improve-
ment in the value of updated score value (USV) for the case of WOC is observed compared to the case of WoOC. For example,
USV improves from 0.25 to 0.32 for bearing 1_1, while it improves from 0.38 to 0.59 for bearing 1_3, when the operating
condition parameters are considered. With operating condition (WOC) consideration, the PUP value is also very low for
all the bearings except for bearing 4 from OC1. This shows that the proposed methodology accurately predicts the RUL
within the error bound.
All these performance metrics indicate that the WAFTR model after considering operating conditions gives better results
compared to the model that ignores operating condition parameters and hence operating conditions should be considered
for accurate RUL prediction.

4.2. RUL prediction without including any of the data set belonging to tested bearing operating condition during training

In the previous section, the bearing test data at various operating conditions were used in the model development/train-
ing. Thereafter, the test data of another bearing which belongs to the same operating conditions were used to check the per-
formance of the model (see Table 4). It is important to check the results for the bearing data obtained from different
operating condition than those used for the model development. Hence, for life estimation of bearings obtained from
OC1, the model parameters were estimated using bearing failure histories from OC2 and OC3. Likewise, the process is
repeated for other operating condition bearings as shown in Table 7. Table 8 shows the values of model coefficient obtained
with different training data sets. The predicted RUL results without considering any of the data set belonging to tested bear-
ing operating conditions are shown in Fig. 9 and performance metrics results are listed in Table 9.
From Fig. 9, it is clear that if the model incorporates the operating conditions (WOC model), the results are better com-
pared to when they are ignored (WoOC model), substantiating what was concluded at the end of Section 4.1. In addition,
from Table 9, bearings of all the operating conditions exhibit superior performance metrics for WOC model compared to
the WoOC model, indicating the need of inclusion of operating parameters in the model parameter estimation. However,
with the WOC model, for operating conditions (1) and (3), the accuracy of prediction in the RUL is not high when none of
the data sets belonging to the test bearing operating condition is included during the training (Fig. 9) as compared to the
case when data sets belonging to the test bearing operating condition is included during the training (Fig. 8). For the OC2
(bearing test cases 2_1, 2_2, 2_4 and 2_6), the prediction results are almost similar to the corresponding RUL results in
Fig. 8 and the WOC model estimates the RUL almost close to the actual RUL in each of the cases.
This happens as the test operating condition (either OC1 or OC3) is out of the range of the training operating conditions
(either OC2 and OC3 or OC1 and OC2) for estimating the RUL results shown in Fig. 9. However, for the four bearings from
OC2, the model estimates the RUL quite close to the actual RUL, as the test data from OC2 lies within the range of the training
operating conditions defined by OC1 and OC3. It may be noted from Table 1 that the TTF values for the bearings with OC2 lie
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 13

Fig. 8. RUL prediction results when data set belonging to tested bearing operating condition is included during training.

between OC1 and OC3. Hence for the life estimation of a bearing at OC1 and OC3, the model is biased and gives higher error
in RUL prediction. The model bias can be minimised by training the model with run-to-failure data corresponding to the two
extreme operating conditions (maximum and minimum operating conditions). If the operating conditions of the test bearing
lie within the range of the operating conditions with which the model has been trained, the model is expected to predict the
RUL accurately. In the present work, this is substantiated for the dataset belonging to the OC2. It may be emphasized that
14 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Fig. 8 (continued)

inclusion of even one set of data corresponding to the test bearing operating condition improves the accuracy of prediction.
This may be clarified by comparing the RUL prediction results (for the model WOC) for bearing test case 3_1 in Fig. 8 with
that obtained for the corresponding case in Fig. 9. In Fig. 8, in the model WOC, the data set for bearing case 3_3 was included
in the model parameter estimation and hence at least one dataset with OC3 was included. However, Fig. 9 belongs to the case
wherein none of the bearing datasets belonging to the OC3 is considered in the model parameter estimation. Thus the oper-
ating conditions for both bearings 3_1 and 3_3 are not represented in the model.
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 15

Table 6
Performance Evaluation Results when all operating conditions are considered during model parameter estimation.

Bearing ID 1_1 1_3 1_4 1_5 1_6 1_7 2_1 2_2 2_4 2_6 3_1 3_3
Bias (WOC) 4.49e3 1.18e3 1.04e4 1.3e3 266.53 1.14e3 1.1e3 547.9 649.6 1.2e3 484.01 119.3
Bias (WoOC) 5.92e3 2.96e3 9.95e3 2.48e3 1.47e3 534.78 8.2e3 1.3e4 1.6e4 1.65e4 1.8e4 1.6e4
MSE (WOC) 2.15e7 2.36e6 1.09e8 1.88e6 5.02e5 1.44e6 1.28e6 4.03e5 4.53e5 1.48e6 2.6e5 1.7e5
MSE (WoOC) 3.9e7 1.38e7 1e8 6.8e6 3.11e6 1.08e6 6.8e7 1.6e8 2.6e8 2.7e8 3.5e8 2.7e8
USV (WOC) 0.32 0.59 0.02 0.6 0.83 0.62 0.38 0.57 0.46 0.27 0.45 0.53
USV (WoOC) 0.25 0.38 0.02 0.44 0.62 0.74 0.002 0.001 8.2e-5 4.6e-5 1.1e-5 2.8e-8
PUEP (WOC) 13.66 0 0 0 0 0 0 0 0 0 0 0
PUEP (WoOC) 61.46 20.38 0 0 1.59 0 0 0 0 0 0 0
PULP (WOC) 0 0 100 0 0 0 0 0 0 0 0 0
PULP (WoOC) 0 0 100 0 0 0 100 100 100 100 100 100
PUP (WOC) 13.66 0 100 0 0 0 0 0 0 0 0 0
PUP (WoOC) 61.46 20.38 100 0 1.59 0 100 100 100 100 100 100

Table 7
Training and testing model development when any of data set belonging to tested bearing operating condition is not included during training.

OC1 OC2 OC3


1_1 1_3 1_4 1_5 1_6 1_7 2_1 2_2 2_4 2_6 3_1 3_3
Test Test Test Test Test Test Train Train Train Train Train Train
Train Train Train Train Train Train Test Test Test Test Train Train
Train Train Train Train Train Train Train Train Train Train Test Test

Table 8
Model parameters with different training data sets when any of data set belonging to tested bearing operating condition is not included during training.

Tested Bearing ao (Intercept) a1 (PC Coefficient) b1 (Load Coefficient) b2 (Speed Coefficient) b (Shape Parameter)

1_1 3.12 0.35 0 5.91 1.59


1_3 3.12 0.35 0 5.91 1.59
1_4 3.12 0.35 0 5.91 1.59
1_5 3.12 0.35 0 5.91 1.59
1_6 3.12 0.35 0 5.91 1.59
1_7 3.12 0.35 0 5.91 1.59
2_1 0.09 0.11 0 9.61 1.54
2_2 0.09 0.11 0 9.61 1.54
2_4 0.09 0.11 0 9.61 1.54
2_6 0.09 0.11 0 9.61 1.54
3_1 3.47 0.30 0 13.07 1.54
3_3 3.47 0.30 0 13.07 1.54

In the proposed approach, both the operating conditions and the degradation extent (through measured vibration param-
eter values) have been taken into consideration to improve the correctness of the model and possible improvement in the
accuracy of prediction. It is also shown that if only degradation extent is considered in the model, the RUL prediction results
are significantly poor. Earlier studies did not include the degradation extent and operating conditions together in the RUL
prediction model. Consideration of operating conditions with the degradation extent presented in the current work ensures
that a single model is required for the RUL prediction instead of developing different models for different operating condi-
tions as has been done in the past. In addition, even when none of the data set belonging to the tested bearing operating
condition is considered, the results are still better with the WOC model compared to when the model is built without con-
sidering operating conditions (WoOC). This shows the importance of considering operating conditions during model param-
eter estimation. In many other models like the PHM, the indicator variables are considered as uncorrelated. This may not give
accurate RUL predictions if the correlations are significant. The proposed method addresses this issue by taking the first prin-
cipal component, after conducting a PCA analysis on the indicator variables (% signal energy estimated in different spectrum
bands), to capture the effect of correlation among the indicator variables. The effect of these variables is incorporated in the
WAFTR model through a multiplier to the Weibull characteristic life.
In summary, from Figs. 8 and 9 and Tables 6 and 9 it is observed that if the operating conditions during model parameter
estimation are not considered, than the prediction accuracy of the model is significantly affected. Table 10 shows the signif-
icance of the proposed methodologies compared to the methodology used by other authors for mechanical component
prognosis.
16 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Fig. 9. RUL prediction results when none of the data sets belonging to the tested bearing operating condition is included during training.

5. Conclusions and scope for future work

The bearings in practice may run at different operating speed, load, and in diverse environmental conditions. It is noticed
that all the available work in literature for rotating machinery prognosis has considered only internal covariates, i.e., vibra-
tion, temperature, humidity, etc. during model development. The effect of external covariates, i.e., load and speed on the fail-
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 17

Fig. 9 (continued)

ure rate progression has not been explored. Consideration of operating conditions with the CM data will ensure that a single
model is enough for RUL prediction instead of developing different models for different operating conditions. A Weibull
Accelerated Failure Time Regression (WAFTR) model is presented here that considers the effect of the operating condition
with the condition monitoring data during model development. Using raw vibration signal, frequency domain based features
as the internal covariates are found better compared to the conventional time domain statistical features. PCA is applied on
the frequency domain based features for dimensionality reduction. Best PC value along with operating load and operating
18 P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302

Table 9
Performance Evaluation Results when only two operating conditions (not belongs to tested bearing operating condition) are considered at a time during model
parameter estimation.

Bearing ID 1_1 1_3 1_4 1_5 1_6 1_7 2_1 2_2 2_4 2_6 3_1 3_3
Bias (WOC) 1.80e4 1.4e4 3.61e3 1.4e4 1.4e4 1.2e4 167.5 43.41 694.3 1.2e3 2.98e3 2.34e3
Bias (WoOC) 2.23e4 1.8e4 8.05e3 1.84e4 1.8e4 1.6e4 1.1e4 9.04e3 1.06e4 1.12e4 1.3e4 1.16e4
MSE (WOC) 3.25e8 1.9e8 1.32e7 1.95e8 1.9e8 1.4e8 1.1e6 4.3e4 4.8e5 1.44e6 8.9e6 5.53e6
MSE (WoOC) 4.98e8 3.3e8 6.48e7 3.38e8 3.3e8 2.7e8 1.3e8 8.24e7 1.13e8 1.24e8 1.65e8 1.38e8
USV (WOC) 0.03 0.04 0.18 0.04 0.04 0.05 0.69 0.78 0.45 0.28 0.04 0.05
USV (WoOC) 0.014 0.02 0.04 0.02 0.02 0.02 0.002 0.0051 0.0012 6.5e-4 1.8e-4 2.5e-6
PUEP (WOC) 100 100 100 100 100 100 0 0 0 0 100 100
PUEP (WoOC) 100 100 100 100 100 100 0 0 0 0 0 0
PULP (WOC) 0 0 0 0 0 0 12.18 0 0 0 0 0
PULP (WoOC) 0 0 0 0 0 0 100 100 100 100 100 100
PUP (WOC) 100 100 100 100 100 100 12.18 0 0 0 100 100
PUP (WoOC) 100 100 100 100 100 100 100 100 100 100 100 100

Table 10
Benchmarking of the current approach vis-à-vis past methods.

Reference Methodology Advantages Disadvantages


[5,16,26] Neural - Model accuracy is very high - Computation time is high
Network - Longer prediction horizon - Requires large CM data for accurate prediction
- Over fitting
- Difficult to determine the optimum structure of the network
- The neural network is unstable and gives different results for differ-
ent number of run
[27] Support - Gives good prediction in RUL with large - No standard method to choose the kernel function
Vector data sets
Machine
[28,29] State Space - Good for non-stationary signals and - A large amount of data set requires for estimating the failure state
Modeling dynamic systems - Not suitable for some of the practical situations such as offshore well
- Good for identifying the multiple failure drilling and wind turbines in which online measurement of the crack,
modes wear or pit depth is not possible to obtain and such approaches
require the current state of the system for model parameter
estimation.
[12,30] Kalman - Can predict the current and future fail- - Require the knowledge of system dynamics to build the explicit
Filter ure state model equations which require domain knowledge
- A large amount of data set requires for estimating the failure state
- Not suitable for some of practical situations such as offshore well
drilling and wind turbines in which online measurement of the crack,
wear or pit depth is not possible to obtain and such approaches
require the current state of the system for model parameter
estimation.
Proposed WAFTR - With less CM histories from each operat- - At least three different operating condition data is required for the
Model ing condition, accurate RUL results can parameter estimation.
be obtained - Give accurate results for all the operating conditions that lie within
- Computation time is low the range used for model training.
- The model has inherent feature selection
capability
- Censored data also can be used to
achieve RUL prediction

speed is modelled in the WAFTR model. To show the importance of considering the operating conditions in the RUL predic-
tion model, the performance of the proposed approach is checked with and without consideration of the operating condition
data. A high underestimation or overestimation in the RUL prediction is observed when the operating condition data is
ignored in the model. It is found that when the operating condition data is included in the model the prediction accuracy
improves significantly.
The study reveals that in case if the dataset belonging to the test bearing operating conditions is not included in the model
training, the accuracy of the RUL prediction is found to be still reasonably high if the test bearing operating conditions lie
within the range of operating conditions of the training bearing datasets. These results are significant as the model built
can be useful to predict a reasonably accurate RUL for the unforeseen operating conditions that were not used in the model
development. Thus, the proposed model serves to cover a range of operating conditions in a more effective way and improves
the RUL prediction accuracy without having to train different models for different operating conditions. Although the pro-
posed methodology is validated with bearings tested under a constant operating condition, its suitability and/or improve-
ment under variable operating conditions will be explored in future.
P. Kundu et al. / Mechanical Systems and Signal Processing 134 (2019) 106302 19

References

[1] C.J. Li, H. Lee, Gear fatigue crack prognosis using embedded model, gear dynamic model and fracture mechanics, Mech. Syst. Signal Process. 19 (2005)
836–846, https://doi.org/10.1016/j.ymssp.2004.06.007.
[2] S. Marble, B.P. Morton, Predicting the remaining life of propulsion system bearings, IEEE Aerosp. Conf. (2006) 1–8, https://doi.org/10.1109/
AERO.2006.1656121.
[3] D. An, J.H. Choi, N.H. Kim, A tutorial for model-based prognostics algorithms based on matlab code, Annu. Conf. Progn. Heal. Manag. Soc. 3 (2012) 1–9.
http://www.phmsociety.org/sites/phmsociety.org/files/phm_submission/2012/phmc_12_122.pdf.
[4] D. An, J.H. Choi, N.H. Kim, Identification of correlated damage parameters under noise and bias using Bayesian inference, Struct. Heal. Monit. 11 (2012)
293–303, https://doi.org/10.1177/1475921711424520.
[5] A.K. Mahamad, S. Saon, T. Hiyama, Predicting remaining useful life of rotating machinery based artificial neural network, Comput. Math. Appl. 60
(2010) 1078–1087, https://doi.org/10.1016/j.camwa.2010.03.065.
[6] M. Zhao, B. Tang, Q. Tan, Bearing remaining useful life estimation based on time-frequency representation and supervised dimensionality reduction,
Meas. 86 (2016) 41–55, https://doi.org/10.1016/j.measurement.2015.11.047.
[7] L. Guo, N. Li, F. Jia, Y. Lei, J. Lin, A recurrent neural network based health indicator for remaining useful life prediction of bearings, Neurocomputing 240
(2017) 98–109, https://doi.org/10.1016/j.neucom.2017.02.045.
[8] F. Cartella, J. Lemeire, L. Dimiccoli, H. Sahli, Hidden semi-markov models for predictive maintenance, Math. Probl. Eng. (2015), https://doi.org/10.1155/
2015/278120.
[9] H. Liao, Z. Tian, A framework for predicting the remaining useful life of a single unit under time-varying operating conditions, IIE Trans. 45 (2013) 964–
980, https://doi.org/10.1080/0740817X.2012.705451.
[10] R.K. Singleton, E.G. Strangas, S. Aviyente, Extended kalman filtering for remaining-useful-life estimation of bearings, IEEE Trans. Ind. Electron. 62
(2014) 1781–1790, https://doi.org/10.1109/TIE.2014.2336616.
[11] Z. Huang, Z. Xu, X. Ke, W. Wang, Y. Sun, Remaining useful life prediction for an adaptive skew-Wiener process model, Mech. Syst. Signal Process. 87
(2017) 294–306, https://doi.org/10.1016/J.YMSSP.2016.10.027.
[12] J. Lee, F. Wu, W. Zhao, M. Ghaffari, L. Liao, D. Siegel, Prognostics and health management design for rotary machinery systems – reviews, methodology
and applications, Mech. Syst. Signal Process. 42 (2014) 314–334, https://doi.org/10.1016/j.ymssp.2013.06.004.
[13] T. Xia, Y. Dong, L. Xiao, S. Du, E. Pan, L. Xi, Recent advances in prognostics and health management for advanced manufacturing paradigms, Reliab. Eng.
Syst. Saf. 178 (2018) 255–268, https://doi.org/10.1016/J.RESS.2018.06.021.
[14] S. Sankararaman, M.J. Daigle, K. Goebel, Uncertainty quantification in remaining useful life prediction using first-order reliability methods, IEEE Trans.
Reliab. 63 (2014) 603–619, https://doi.org/10.1109/TR.2014.2313801.
[15] M.S. Kan, A.C.C. Tan, J. Mathew, A review on prognostic techniques for non-stationary and non-linear rotating systems, Mech. Syst. Signal Process. 62
(2015) 1–20, https://doi.org/10.1016/j.ymssp.2015.02.016.
[16] J.B. Ali, B. Chebel-Morello, L. Saidi, S. Malinowski, F. Fnaiech, Accurate bearing remaining useful life prediction based on Weibull distribution and
artificial neural network, Mech. Syst. Signal Process. 56 (2015) 150–172, https://doi.org/10.1016/j.ymssp.2014.10.014.
[17] C.E. Ebeling, An Introduction to Reliability and Maintainability Engineering, McGraw-Hill, 2004.
[18] P. Nectoux, R. Gouriveau, K. Medjaher, E. Ramasso, B. Chebel-Morello, N. Zerhouni, C. Varnier, PRONOSTIA: An experimental platform for bearings
accelerated degradation tests, IEEE Int. Conf. Progn. Heal. Manag. (2012) 1–8.
[19] S. Hong, Z. Zhou, E. Zio, K. Hong, Condition assessment for the performance degradation of bearing based on a combinatorial feature extraction method,
Digit. Signal Process. 27 (2014) 159–166, https://doi.org/10.1016/j.dsp.2013.12.010.
[20] Z. Liu, M.J. Zuo, Y. Qin, Remaining useful life prediction of rolling element bearings based on health state assessment, Proc. Inst. Mech. Eng. Part C J.
Mech. Eng. Sci. 230 (2016) 314–330, https://doi.org/10.1177/0954406215590167.
[21] R.K. Singleton, E.G. Strangas, S. Aviyente, Time-frequency complexity based remaining useful life (RUL) estimation for bearing faults, 9th IEEE Int.
Symp. Diagnostics Electr. Mach. Power Electron. Drives. (2013) 600–606, https://doi.org/10.1109/DEMPED.2013.6645776.
[22] L. Ren, W. Lv, S. Jiang, Machine prognostics based on sparse representation model, J. Intell. Manuf. 29 (2018) 277–285, https://doi.org/10.1007/s10845-
015-1107-8.
[23] G. Niu, F. Qian, B.K. Choi, Bearing life prognosis based on monotonic feature selection and similarity modeling, Proc. Inst. Mech. Eng. Part C J. Mech. Eng.
Sci. 230 (2015) 3183–3193, https://doi.org/10.1177/0954406215608892.
[24] B. Zhang, L. Zhang, J. Xu, Degradation feature selection for remaining useful life prediction of rolling element bearings, Qual. Reliab. Eng. Int. 32 (2015)
547–554, https://doi.org/10.1002/qre.1771.
[25] Q. Li, Z. Gao, D. Tang, B. Li, Remaining useful life estimation for deteriorating systems with time-varying operational conditions and condition-specific
failure zones, Chinese J. Aeronaut. 29 (2016) 662–674, https://doi.org/10.1016/j.cja.2016.04.007.
[26] Z. Tian, An artificial neural network method for remaining useful life prediction of equipment subject to condition monitoring, J. Intell. Manuf. 23
(2012) 227–237, https://doi.org/10.1007/s10845-009-0356-9.
[27] V.T. Tran, H. Thom Pham, B.S. Yang, T. Tien Nguyen, Machine performance degradation assessment and remaining useful life prediction using
proportional hazard model and support vector machine, Mech. Syst. Signal Process. 32 (2012) 320–330, https://doi.org/10.1016/j.ymssp.2012.02.015.
[28] D.A. Tobon-Mejia, K. Medjaher, N. Zerhouni, G. Tripot, A data-driven failure prognostics method based on mixture of Gaussian hidden Markov models,
IEEE Trans. Reliab. 61 (2012) 491–503, https://doi.org/10.1109/TR.2012.2194177.
[29] A. Heng, S. Zhang, A.C.C. Tan, J. Mathew, Rotating machinery prognostics: state of the art, challenges and opportunities, Mech. Syst. Signal Process. 23
(2009) 724–739, https://doi.org/10.1016/j.ymssp.2008.06.009.
[30] C.K.R. Lim, D. Mba, Switching Kalman filter for failure prognostic, Mech. Syst. Signal Process. 52 (2015) 426–435, https://doi.org/10.1016/j.
ymssp.2014.08.006.

You might also like