Available online at www.sciencedirect.

com

ScienceDirect
Procedia CIRP 15 (2014) 212 – 217

21st CIRP Conference on Life Cycle Engineering

Remaining useful life assessment and its application in the decision for
remanufacturing
Yawei Hua,*, Shujie Liua, Huitian Luc, Hongchao Zhanga,b
a
School of Mechanical Engineering, Dalian University of Technology, Dalian, China
b
Dept.of Industrial Engineering, Texas Tech University, Texas, America
c
Dept. of Engineering Technology & Management, South Dakota StateUuniversity, South Dakota ,America
* Corresponding author. Tel.: 15998528442; fax:+0-000-000-0000.E-mail address:yaweihu@126.com

Abstract

Remaining useful life (RUL) of a mechanical device or equipment is the lifetime from current time to the time that the device fails. Remaining
useful life modelling and online assessment provide the timely message for system decision making in production management and condition-
based maintenance. RUL of a device is also essential in remanufacturing engineering in the decision for parts to be taken out service for
remanufacturing. To take full advantage of the life of a device before remanufacturing, the RUL modelling of the device through online
performance monitoring is critical to the remanufacturing plan. The inclusion of online performance monitoring will answer the basic questions
about the timing of the remanufacturing processes that will lead to the renewal of parts. This paper proposes an analytical decision method for a
remanufacturing program that is based on the remaining useful life online assessment. Support vector machine (SVM) model is used to
estimate the remaining useful life and the performance reliability based on the predefined threshold for failure, and the remanufacturing
characteristics. Furthermore, a corresponding remanufacturing strategy can be developed for the plan. This paper also illustrates an example of
on-line data monitoring for a double row cylindrical roller bearing, which showed the validity of the aforementioned method.
©2014
© 2014The
Elsevier B.V.
Authors. This is an
Published by open access
Elsevier B.V.article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Selection and peer-review under responsibility of the International Scientific Committee of the 21st CIRP Conference on Life Cycle
Selection and
Engineering in peer-review
the person of under responsibility
the Conference Chairof theTerje
Prof. K. Lien. Scientific Committee of the 21st CIRP Conference on Life Cycle
International
Engineering in the person of the Conference Chair Prof. Terje K. Lien
Keywords:Remaining useful life prediction; reliability assessment; remanufacturing program decisions;

1. Introduction control, etc. For remanufacturing product quality and cost

With the rapid development of Green Manufacturing
Technologies, remanufacturing engineering of mechanical
products has attracted more and more research on theory
development. In remanufacturing management, new research
has focused on the development of a remanufacturing scheme
for a device in service. Recently, extensive research has been
conducted on remanufacturing decision criteria for
remanufacturing of a mechanical device/components,
including setting up practical remanufacturing plans that
would lead to optimal remanufacturing objectives [1,2];
establishing a comprehensive assessment for remanufacturing
strategies with regards to economy, technique, and
environmental impact [3,4]; and analyzing different
remanufacturing processes for remanufactured product quality
optimization of the entire remanufacturing process, it is an
essential consideration at what RUL level time the device
should leave from service for remanufacturing. Some
researchers refer to this comprehensive evaluation and
assessment for remanufacturing as remanufacturability [5].
There is little research regarding analytical decision rules for
remanufacturing based on RUL assessment directly. In order
to make full use of the value of a device in its first service and
optimizing the remanufacturing processes, this search attempts
to model the link between RUL and remanufacturing time and
remanufacturability. This paper proposes a methodology to
make corresponding remanufacturing decisions, according to
variable RUL assessment results of the device. Additionally,
systematic research on decision making of remanufacturing
schemes by means related to RUL assessment will be a

2212-8271 © 2014 Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Selection and peer-review under responsibility of the International Scientific Committee of the 21st CIRP Conference on Life Cycle
Engineering in the person of the Conference Chair Prof. Terje K. Lien
doi:10.1016/j.procir.2014.06.052
 Yawei Hu et al. / Procedia CIRP 15 (2014) 212 – 217
significant development in remanufacturing theory. Therefore,
with further analytical development of the RUL model and
prognostics, the framework of a remanufacturing scheme
based on RUL online assessment in this paper can be practical
and significant.
The research on assessment of RUL has been developing
for many years. During the 1950s and 1960s, August Wöhler,
from Germany, introduced the concept of the S-Ncurve and
fatigue value in fatigue tests on the axle of railway vehicles. In
1963, Paris and Erdogan proposed an empirical model of
fatigue crack growth rate, which resulted in a breakthrough on
the research of fatigue problems using a stress-strength model
for dynamic failure assessment [6]. Zhang Feng adopted the
Hertz theory and fracture mechanics analysis to estimate the
damage tolerance by crack size and service life, and gave an
overall evaluation of fatigue damage [7]. Apart from the
physical and mechanical methods developed through
laboratory experiments, the mathematical model-based
methods are also used widely. Following R. Stribeck’s
research work, A. Palmgren developed a probabilistic
approach to prognosis of the life of mechanical components
[8]. In his research, he defined useful service life as the
statistical time during which 10 percent of the bearings failed
or 90 percent survived. After that, ARMA and ARIMA time
series models have been widely used for RUL modeling and
prediction. Liao [9] used SPR (statistical pattern recognition)
to evaluate the performance of equipment, and then the
ARMA model to predict the RUL of machining equipment.
With the development of information technology, artificial
intelligence technology (AI) has been also applied to RUL
assessment. For instance, Achtsevanos and Wang [10]
introduced a prediction method using a dynamic wavelet
neural network, and demonstrated its feasibility through a
rolling bearing example. Remaining useful life modeling and
prognostics using online monitoring data for a device in
operation is a relatively complex task. The assessment is
problematic as 1) RUL itself is a random variable; 2) RUL is
conditional to the device working under different
environmental conditions; 3) RUL assessment is based on the
system monotonic degradation and usually the degradation is
unmeasurable online; 4) RUL assessment is a probability
density distribution of the random variable of RUL; and 5) the
degradation dynamics are from a nonlinear, time-variant and
non-Gaussian system. In a general statement, online RUL
assessment is the recursive estimation of RUL of a device
based on physical performance observations. There are two
basic methods associated with RUL assessment in research
and practice. They are a physical model based method and a
data-driven method. This paper presents a framework for a
data-driven method for RUL using support vector machine
(SVM) method. Given the online monitoring degradation
signal (observations) of a device, this paper discusses the
prognosis of the future RUL characteristics of the device by
adapting the support vector machine regression method.
Assume the observation of degradation available online and
the device failure is defined in terms of the degradation the
paper presents the framework on RUL assessment based on
observations of degradation at an instant time. In accordance
with the residual life prognostic and reliability assessment,
213
this paper also suggests a strategic plan for remanufacturing
time, which depends upon the prediction results. These
comprehensive results include the changing operating
environment, the failure mechanism, and the nonlinear system
dynamics. A case study on double row cylindrical roller
bearing is also presented to demonstrate the data analysis and
feasibility of this method.
In the following section, a brief review of the support
vector machine method and prediction algorithm is
introduced. In Section III performance reliability calculations
and the RUL predictions are discussed. Section IV presents
the development method for a remanufacturing scheme based
on RUL assessment. Section V validates the study using an
industrial case with a detailed analysis of the experimental
results. Finally, Section VI gives the conclusions and
comments regarding the SVM method and suggestions for
future research of RUL and the relation to remanufacturing
engineering.
2. Support vector machine (SVM) model
2.1. Support vector regression (SVR)
Support vector machine can be applied to classification
problems, also to the case of regression. Support vector
regression (SVR) algorithm proposed in the middle of 1990s
is based on a statistical learning theory created by
Vapnik[11].Vapnik introduced an insensitive loss function
into the support vector machine (SVM) model, which can be
used to solve the problem of nonlinear regression. For the
nonlinear case, the basic idea of support vector machine
regression is that the sample points are mapped into a high
dimensional feature space by a nonlinear mapping
) : Rn o H , and linear regression is then applied in the high
dimensional feature space to get the nonlinear regression
estimation in the original space.
The estimate function is:
f ( x) ¢Z, )( x)²  b (1)
where Z is weight vector and b is the offset, Φ(x) is
nonlinear mapping of x into the H space.
To estimate the unknown coefficients Z and b , a
functional minimization problem can be constructed as
follows:
l
1 2
*(Z ) Z  C ¦] ( f ( xi )  yi , xi ) (2)
2 i 1
1 2
where the first term Z is the regularization item, and the
2
l
second term C ¦ ] ( f ( xi )  yi , xi ) is an empirical risk item,
i 1
C is a constant, which determines the balance of the
contributions to the function between the two terms.
Empirical risk metrics are determined by the H -insensitive
loss function, defined as
214 Yawei Hu et al. / Procedia CIRP 15 (2014) 212 – 217

] ( f ( x )  y, x ) y  f ( x) H ) : Rn o H based on the historical data to approximate the

implicit nonlinear mechanism f of the data sequence, using
­ 0 ,| y  f ( x ) |d H (3)
® support vector mapping as an ideal predictor.
¯| y  f ( x ) | H , otherwise For the given time series XN, XN={x1,x2,…,xN}, take the
first r (r  N ) data as training samples, and the rest of the
By introducing slack variables [i and [i* , the data set as the test sample. In order to make more effective
minimization problem of Eq. (2) can be transformed to the use of the data, phase space of the one-dimensional time
following optimization problem: series is reconstructed and the series is transformed into a
matrix form. The one-dimensional time series XN can be
1 l transformed into the following matrix form.
Z  C ¦ ([i  [i* )
2
min *(Z )
2 i 1
§ X1 · § x1 x2 xm · § xm 1 ·
s.t. Z, ) ( xi )  b  yi d [i  H (4) ¨ ¸ ¨ ¸ ¨ ¸
yi  Z, ) ( xi )  b d [i*  H X ¨ X2 ¸ ¨ x2 x3 xm 1 ¸
,Y ¨ xm  2 ¸ (8)
¨ ¸ ¨ ¸ ¨ ¸
[i , [ t 0, i 1, 2, , l
* ¨ ¸ ¨ ¸ ¨ ¸
i
© X r m ¹ © xr  m xr  m 1 xr 1 ¹ © xr ¹

The above constrained optimization problem is a typical where X i X i

m
( xi , xi 1 , , xi  m1 ) , and mis the forecasting
quadratic programming problem that can be solved by embedding dimension, and its value is determined by FPE
Lagrange multiplier method. The Lagrange function is (Final Prediction Error) criterion. The forecast regression
function is
1 2 l
L(Z, b, [i(*) ) Z  ¦Di (H  [i  yi  Z , ) ( xi )  b) l
2 i 1 f ( x) ¦ (D
i 1
*
i  Di )K ( xi , x )  b (9)
l l
C ¦ ([i  [ )  ¦ [D (H  [  yi  Z, ) ( xi )  b)
i
* *
i i
*
(5) Where
i 1 i 1
­ l
(Ki[i  Ki*[i* )] ° y j  ¦ (Di  Di )K ( xi , x j )  H D j  (0, C )
*

° i 1
b ® l
where Di(*) and Ki(*) are Lagrange multipliers. The extremum ° y  (D *  D )K ( x , x )  H
°̄ j ¦ i i i j D *j  (0, C )
condition of the function L is: i 1

In Eq. (9), t=m+1,m+2,…,r; K(.,.) is the kernel function.

wL l
If new information is added in the forecasting process, it is
Z  ¦ (D  Di )) ( xi ) 0
*

wZ i 1
i
called 1-step ahead forecasting. Data used for forecasting of
wL l the next moment are the actual observed value before m
wb
¦ (D *
i  Di ) 0 (6) moment, rather than the forecasting data.
i 1
According to Eq. (9) and X r  m1 ^xr  m1 , xr  m 2 , , xr ` ,
wL
C  Di(*)  Ki(*) 0 the 1-step ahead forecasting at time r is
w[i(*) r m
X r 1 ¦ (D
i 1
i  Di* ) K ( xi , xr m 1 )  b ,
Solving the dual problem of the above problem as
and then
r m

max W (Di(*) )
1 l
¦ (Di*  Di )(D *j  D j ) )( xi ), )( x j ) X r 2 ¦ (D i  Di* ) K ( xi , xr m  2 )  b for 1-step ahead forecast
2 i, j 1 i 1
l l at time r+1. For multiple-step ahead forecasting, the model
 H ¦ (D  Di )  ¦ yi (D  Di )
*
i
*
i (7) forecasts will be used instead of unavailable observation
i 1 i 1
l
inputs at the time.
s.t. ¦ (D i  Di* ) 0
i 1
3. Remaining useful life assessment
0 d Di(*) d C , i 1, 2, ,l
In this work, a simple case with one failure mode and one
the solution is D (D1 , D2 , , Dn ) ˈ D * (D1* , D2* , , Dn* ) . The single performance variable for degradation is demonstrated
solution for the weights is based on the Karush-Kuhn-Tucker for conditional performance reliability and RUL assessment.
conditions that result in the optimal solution satisfies DiDi* 0 Assume that a single unit’s performance variable yt is
. That is one of D i ˈ D i* is zero at least, and their increasing with performance time and the system failure is
corresponding sample points’ vector is the support vector. defined as when performance yt reaches and exceeds the
critical value Ycl. Assuming yt follows a Gaussian distribution,
2.2. The forecasting model of SVM the probability of system failure in the process can be
expressed as a conditional probability [12], so the failure
The forecasting model of SVM is searching for mapping probability at time t  l 't is:
 Yawei Hu et al. / Procedia CIRP 15 (2014) 212 – 217 215

f R( t  l ˜ 't ) at the time t  l ˜ 't is less than the predefined

F (t  l ˜ 't | t ) P(Y (l ) t Ycl ) ³ yCL
f ( y(t  l ˜ 't | t ))dy (10)
where Y(l) is the predictive performance variable at time
t  l 't . Fig.1 shows the reliability assessment in the
degrading process. The shaded area is the probability of
failure for each 't .
The Conditional reliability at time t  l 't , the system has
been survived to current time t is:
R(t  l ˜ 't | t ) R(t f ) / R(tc )
where R(tc ) is the reliability at current time tc, and R(t f ) is
the reliability at future time t  l 't .
The online sensors taking real-time device performance
measurements can update the performance variables to obtain
performance reliability and reliability assessment for current
and future time points.
reliability R*, that is R(t ) ! R * and R(t  l ˜ 't )  R * , the
probability of the products failure is relatively large at the
time. Therefore to improve operational reliability the product
remanufacturing is recommended, called the remanufacturing
phase, as in Fig. 2(b).
c) When the current time reliability R(t) is less than the
predefined threshold reliability R*, and the future reliability
R( t  l ˜ 't ) at the time t  l ˜ 't is also less than the
predefined threshold reliability R*, that is
(11) R(t )  R * and R(t  l ˜ 't )  R * , remanufacturing is not
recommended, instead recycling for available material reuse,
called the recycling phase for over used products, as in
Fig.2(c).
The predicted performance reliability of the device or
component is established by SVM model based on on-line
performance monitoring date, and remanufacturing decisions
can be developed according to the reliabilities and the
relationships. The following case study demonstrates the
procedure and the feasibility of the method.

Fig. 1. Reliability assessment in the degradation process. Fig. 2. (a). Remanufacturing decision criteria: re-use.

4. Remanufacturing scheme

A remanufacturing scheme for a part can be developed by

taking the optimum utilization of the part life time before
remanufacturing and deciding the best time to pull the part
from service for remanufacturing based upon total cost
optimization and circular economics. Based on this analysis, a
more effective remanufacturing decision method is proposed
to determine the time to remanufacture and the
remanufacturing plan to perform based on the online Fig. 2. (b). Remanufacturing decision criteria: remanufacturing.
monitoring data and the forecasts information of the device
performance, named a part condition-based
remanufacturability. For remanufacturing decision criteria,
calculating the current time reliability R(t), predicting
reliability R( t  l ˜ 't ) of the time at t  l ˜ 't , and according
to the predefined threshold reliability R*, the remanufacturing
scheme can be categorized according to the relationship of
R(t), R( t  l ˜ 't ) and R*. There are three different solutions
for the remanufacturing scheme [13], as shown in the
Fig. 2. (c). Remanufacturing decision criteria: recycling.
following:
a) When the current time reliability R(t) is greater than the 5. A case study
predefined threshold reliability R*, and the future reliability
R( t  l ˜ 't ) at the time t  l ˜ 't is also greater than the This case experiment is set up on the study of reliability
predefined threshold reliability R*, that is prediction and RUL on Rexnord ZA-2115 double row
R(t ) ! R * and R(t  l ˜ 't ) ! R * , products can be used cylindrical roller bearing. The vibration data on the bearing
without remanufacturing, called the re-use phase, as in Fig. are taken for the performance variable and the bearing failure
2(a). is pre-set by a threshold. The online QAC (data acquisition)
b) When the current time reliability R(t) is greater than the system consists of an NI DAQCard-6062E data acquisition
predefined threshold reliability R*, and the future reliability board, a PCB 353B33 acceleration sensor, and data
216 Yawei Hu et al. / Procedia CIRP 15 (2014) 212 – 217

acquisition software developed in LabVIEW. The sampling

frequency is 20 kHz and the sampling interval is 20480
pitches. It is designed to calculate the vibration intensity of
the signal through every 10 min length of sampling. In this
experiment, a data set of the vibration signal with a total of
150 data points is collected. The study data set is plotted in
Fig. 3.
According to the modeling and the calculation method
described above, the first 140 experimental data points were
regarded as the training set to establish the SVM regression
model using the training program written in MATLAB. The
last 10 data points are used for model validation by comparing Fig. 5. The relative error of prediction.
with predicted data points, which are the data points based on
the forecasting model of SVM. One step ahead forecast for
the last 10 data points and a graphic comparison between
model forecasting and the actual measured data are shown in
Fig.4. It should be noted that the prediction error will increase
severely with increasing prediction step number. In this study
the one step sampling interval for the bearing in service is a
10 min time span and one-step ahead forecasting with less
error is recommended. In order to further improve the model
prediction accuracy, the wavelet decomposition is employed
for data signal pre-process. The data were decomposed by the
db5 wavelet function, and a significant low frequency
component and two high-frequency components were
selected. SVM models were then established for the different Fig. 6. Conditional reliability prediction curve.
sub-sequences on the result of the wavelet analysis to predict
the corresponding component. Conditional reliabilities for the forecasts of the
Finally, an overall prediction was obtained through sequences performance variable were then estimated by Equations (11)
that were reconstructed by the components of the three and (12). In accordance with the practical application, the
frequencies. The percentage errors of the predictions for the vibration intensity critical threshold for the failure of the
one step ahead forecasting were shown in Fig. 5. The results double row cylindrical roller bearing is set at 0.22 A/g.
showed that the prediction accuracy of SVM model, around Assuming that the vibration intensity is a normally distributed
±2% to 10%, is acceptable in industrial applications. random variable at every given time, the reliabilities of the
last 10 predicted data points were calculated, and the results
shown in Table 1. In order to make decisions for the
remanufacturing scheme, a conditional reliability prediction
curve using the predicted results is showed in Fig. 6. In this
case study, the reliability threshold R* for decision is pre-set
at 0.6. Prediction results show that the bearing can be “re-
used” at the first 5 predication steps. If the bearing survives in
service until the sixth step, the bearing will be recommended
for remanufacturing, and after that time point, the bearing
needs to be recycled, without remanufacturing based on the
pre-set decision rules (refer to the figure 2 for the scheme).
The sampling interval is 10 min of the bearing run time.
Fig. 3. Vibration intensity data. The case study shows the framework of conditional
reliability and the RUL online assessment with the
remanufacturing scheme decision. It can be seen that
connecting performance reliability and RUL with a
remanufacturing decision scheme or plan is natural and
conceptually straightforward. However, there are still some
aspects which need to be intensively studied: 1) Degradation
cannot usually be observed directly. The degradation process
must be assessed from another stochastic process that is
measurable in application. 2) In conditional performance
reliability assessment, the performance variable is assumed to
follow a Gaussian distribution, and the analytical closed
formulation is available. However it is not always the case in
Fig. 4. Thecomparison between model forecasting data and the actual data. real world applications. 3) In the proposed method for a
 Yawei Hu et al. / Procedia CIRP 15 (2014) 212 – 217 217

remanufacturing scheme, the reliability threshold is necessary. There is much research ongoing for developing a
predefined. This may be selected by experience and the recursive online estimation of RUL for a nonlinear, non-
experts’ knowledge without mathematical support. 4) The Gaussian, time variant degradation dynamics.
degradation process may be very complex in practice with
high nonlinearity, time-variant parameters, and non-Gaussian Acknowledgements
noise.
This work is jointly supported by the 973 Basic Research
Table 1.Conditional reliability results and Development Plan of China (No. 211CB013402) and the
T n X Â E(%) R Special Fundamental Research Funds for Central Universities
140 1 0.1804 0,1701 -1.6798 0.9694 of China(No.DUT11RC(3)60).
141 2 0.1665 0.1716 -3.0881 0.9355
References
142 3 0.1672 0.1715 -2.5335 0.9030
143 4 0.1976 0.1765 10.7196 0.8565 [1] Z. Ghazalli, A. Murata. Development of an AHP–
CBR evaluation system forremanufacturing: end-of-
145 5 0.1970 0.1923 2.3932 0.7280 life selection strategy [J]. International Journal of Sustainable
146 6 0.2171 0.2053 5.4510 0.5162 Engineering, 2011; 4(1):2-15.
[2] Xugang Zhang, Hua Zhang, Zhigang Jiang, Yanhong Wang. A decision-
147 7 0.1968 0.1902 3.3853 0.4478 making approach for end-of-life strategies selection of used parts [J].
The International Journal of Advanced Manufacturing Technology,
148 8 0.2041 0.2088 -2.3256 0.2964
online,2013.
149 9 0.2294 0.2251 1.8587 0. 1259 [3] Yanbin Du, Huajun Cao , Fei Liu, Congbo Li, Xiang Chen. An
integrated method for evaluating the remanufacturability of used
150 10 0.2629 0.2617 0.4426 0.0074 machine tool [J]. Journal of Cleaner Production, 2012;20(1): 82-91.
[4] Yanbin DU, Huajun CAO, Fei LIU, et al. Evaluation of machine tool
remanufacturing scheme based on entropy weight and AHP [J].
6. Concluding remarks Computer Integrsted Manufacturing System, 2011;17(1):84-88.
[5] Shiyun DONG, Binshi XU, Lihong DONG, et al. Experiment research
This paper presented a framework for the performance on metal magnetic memory methods for life prediction of
reliability and RUL online assessment using support vector remanufacturing old parts [J]. China Surface Engineering, 2006;
machine (SVM) method, and the conditional performance 19(5):71-75.
[6] Paris PC. The Growth of Cracks due to Variations in Load[D].
reliability and RUL are connected to a remanufacturing DissertationˈLehigh University, 1962;1-237.
decision scheme. The support vector regression model was [7] Feng ZHANG. Damage limitation and life prediction of wheel
established for performance degradation modelling and steel[J].Journal of Mechanical strength, 1997;19(4):52-55,74.
forecasting, and the assessment of RUL. Decisions regarding [8] Gu J, Barker D, Pecht M. Prognostics implementation of electronics
whether a mechanical device of interest should be removed under vibration loading [J].Microelectronics Reliability, 2007; 47(12):
1849-1856.
from service for remanufacturing or not is based on the RUL [9] Wenzhu LIAO, Ershun PAN, Ying WANG, Lifeng XI. Research of
index and performance reliability estimates. The concept predicting machine's remaining useful life based on statistical pattern
connects the remanufacturing scheme with the performance recognition and auto-regressive and moving average model [J]. Journal
reliability and RUL is an idea beneficial to remanufacturing of ShanhHai JiaoTong university, 2011;45(7): 1000-1005.
theory development. This paper also illustrated a case [10] Vachtsevanos G, Wang P. Fault prognosis using dynamic wavelet neural
networks [J].Autotestcon Proceedings, 2001;857-870.
example on double row cylindrical roller bearings using the [11] Jinfang YANG, Yongjie ZHAI, Dongfeng WANG, et al. Time series
vibration intensity signal as the performance variable for prediction based on support vector regression [J]. Proceedings of the
reliability assessment. Wavelet decomposition for the signal CSEE, 2005;25(17)˖110-114.
process was also employed to combine with the support [12] Huitian Lu, William J.Kolarik, Susan S.Lu. Real-time performance
vector machine method, showing a better prediction result. reliability prediction[J].IEEE transactions reliability,2001;50(4):353-
357.
The conceptual work shows that the method is worth while in [13] Xugang ZHANG, Hua ZHANG, Zhigang JIANG. Remanufacturing
preventative maintenance, environment protection and energy scheme decision model and application based on remaining useful
sustainability through remanufacturing engineering. However, estimation [J]. Journal of mechanical engineering, 2013;49(7):51-57.
the SVM method is primarily a data-driven method. Data
collection, preparation, and training for the model building are