SCIENCE CHINA

Information Sciences
. RESEARCH PAPER . June 2022, Vol. 65 162206:1–162206:17
https://doi.org/10.1007/s11432-020-2980-9

Prognostics based on the generalized diffusion

process with parameters updated by
a sequential Bayesian method
Hong PEI1 , Xiaosheng SI1,2* , Changhua HU1 , Jianxun ZHANG1 ,
Dangbo DU1 , Zhenan PANG1 & Shengfei ZHANG1
1
School of Missile Engineering, Rocket Force University of Engineering, Xi’an 710025, China;
2
School of Mechanical Engineering, Xi’an Jiaotong University, Xi’an 710049, China

Received 17 February 2020/Revised 6 May 2020/Accepted 29 May 2020/Published online 26 May 2022

Abstract The realistic degradation process for the engineering equipment is generally stochastic and com-
plicated owing to the uncertain operational condition and multiple functional loading, exhibiting the absolute
nonlinear distinction. Such a nonlinear degradation process is widely modeled as a generalized diffusion pro-
cess. When utilizing the generalized diffusion process-based model, certain model parameters are considered
as the random variables to characterize the unit-to-unit discrepancies. Hence, the estimation of these kinds
of parameters usually resorts to the Bayesian method. However, owing to the complex pattern of the model
parameters in the generalized diffusion process, computing the Bayesian updated parameters requires plenty
of repeated calculation and integration operations once the new degradation monitoring information is avail-
able. This will inevitably lower the computing efficiency and real-time performance. Toward this end, this
paper presents an adaptive prognostic method based on the generalized diffusion process to determine the
remaining useful life (RUL) of degraded equipment. First, a generalized diffusion process-based degradation
modeling framework is constructed to describe the health performance of stochastic degraded equipment
under complex conditions. Then, we utilize the maximum likelihood estimation (MLE) method to estimate
the initial model parameters by analyzing the historical degradation information. Furthermore, a sequential
Bayesian method is proposed to recursively update the stochastic model parameters in the generalized diffu-
sion process for particular equipment in service. Unlike the existing studies utilizing the Bayesian method,
the primary contrast in the presented method lies in that there is no need to implement the calculation
process with complicated integration repeatedly utilizing the whole degradation information obtained before
the current time. Particularly, the current measured information is incorporated into the estimates of the
stochastic parameters in the previous time to determine the corresponding posterior estimates at the current
time. This can avoid repeated calculation and raise the efficiency to a certain extent. Thereafter, the RUL
distribution is updated adaptively by incorporating the acquired posterior estimates. Finally, we provide
two practical case studies associated with the gyroscope and 2017-T4 aluminum alloy to demonstrate the
efficiency and advantage of the proposed sequential Bayesian method. The experimental results exhibit that
the proposed method can increase the RUL prediction accuracy compared with the existing methods in the
literature.
Keywords generalized diffusion process, stochastic model parameters, remaining useful life, maximum
likelihood estimation, sequential Bayesian methods

Citation Pei H, Si X S, Hu C H, et al. Prognostics based on the generalized diffusion process with parame-
ters updated by a sequential Bayesian method. Sci China Inf Sci, 2022, 65(6): 162206, https://doi.org/10.1007/
s11432-020-2980-9

1 Introduction
In engineering practice, dependable, secure, and economical functioning during the life cycle for safety-
critical equipment, including aviation control systems and nuclear power generators, is the basic and
necessary requirement from the task completion dimension. To obtain such requirement, massive experts
* Corresponding author (email: [email protected])

c Science China Press and Springer-Verlag GmbH Germany, part of Springer Nature 2022 info.scichina.com link.springer.com
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:2

and scholars are dedicated mainly in the research on prognostics and health management (PHM) tech-
nique [1–3]. Being a fundamental and organized tool for determining the health status and creating the
maintenance decision, the PHM can appraise the failure risk, predict the failure progression, and mod-
erate the functional cost via management actions, like maintenance, replacement, etc. Over the past few
decades, the PHM has attracted the major attention and been adopted in many industrial and military
fields [4, 5].
It is distinguished that the foundation of the PHM is the remaining useful life (RUL) prediction. In
general, the RUL of the particular equipment can be defined as the length of time from the present time
to the end of useful life [6, 7]. The responsibility of RUL prediction is to evaluate the degradation, and
to determine the RUL utilizing the in-situ condition monitoring (CM) data. Because a trouble that how
long a monitored equipment can survive is frequently encountered in the industry [8–12], the necessity
for executing the RUL prediction is transparent. In addition, RUL prediction can provide the vital
information supporting for creating maintenance plans. Particularly, on the basis of RUL prediction,
the time and way to take the management activities can be determined from both the cost-effective and
safety perspectives [13–15]. Accordingly, an enormous volume of theoretical achievements on the RUL
prediction have been created over continuous investigation and research.
With the rapid advancement of sensor technique, the mainstream methods for RUL prediction are
data-driven methods. This primarily consists of the machine learning-based methods and the statistical
data-driven methods [16]. The machine learning-based methods can establish the mapping relationship
from the monitoring data to the RUL value without any proceeding specialization, such as the neural
network (NN)-based methods [17] and the support vector machine (SVM)-based methods [18]. The
machine learning-based methods can be considered as a black box model and it is applied in various fields
owing to the strong universality. However, the deficiency for traditional machine learning-based methods
is that the prediction results are just point estimations [19, 20]. The statistical data-driven methods
have advantages in dealing with uncertainties for RUL prediction [21], whose typical representatives
are diffusion process-based methods [22], Gamma process-based methods [23], and inverse Gaussian
process-based methods [24]. As is well known, Gamma process-based methods and inverse Gaussian
process-based methods are only suitable for the strictly monotonic degradation process. Owing to the
complex operating environment and interaction of internal structure, the non-monotonic degradation is
frequently identified in existence. When facing with the non-monotonic circumstances, it is transparently
no longer suitable to utilize these two methods for degradation modeling and RUL prediction. In reverse,
the diffusion process-based methods are efficient for describing the non-monotonic degradation, signifying
such methods are closer to the actual situation and have the extensive application capacity. Moreover,
such methods can exhibit the distinguished mathematical characteristics and obtain the parameters
effortlessly. Additionally, temporal variability can be expressed by the diffusion process-based methods.
Thus, the research in this paper is on the basis of diffusion process-based methods. It is noted that the
calculation of the first hitting time (FHT) is not identical with that of non-first hitting time under the
diffusion process-based methods. As a result, it is necessary to distinguish from these two calculations
when implementing the RUL prediction.
As for diffusion process-based methods, Wiener process-based method is a special case. This method
attracts the major attention for degradation modeling in recent years. The first attempt to apply Wiener
process to degradation modeling and RUL prediction was implemented by Doksum and Hoyland [25].
Consequently, a large number of prognostic models based on Wiener process were developed to suit the
actual situation finer and enhance the accuracy of RUL [26–29]. Among these prognostic models, Bayesian
method is a typical method, which is usually utilized to update the stochastic model parameters, such as
the work in [28,29]. Unfortunately, Wiener process can only conform to the linear degradation process but
the various nonlinear degradation process may be identified in practice. When dealing with the nonlinear
degradation process, the activity, including time scale transformation and log-transformation can be
adopted to convert the nonlinear degradation process into the linear degradation process [30,31]. However,
these transformations have certain limitations and are only applicable for particular cases. Therefore, Si
et al. [32] presented the Brownian motion (BM) with a nonlinear drift to model the degradation process
and derived the FHT distribution of the diffusion process crossing a constant threshold by the time-
space transformation. Nevertheless, the model parameters are only estimated by maximum likelihood
estimation (MLE) method and cannot be updated once the new monitoring data are available. To
address this issue, the Bayesian method was adopted to update the stochastic parameters under a special
nonlinear form [33]. Tang et al. [34] researched a two-stage parameter estimation method and derived an
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:3

approximate analytical RUL distribution in a closed-form of a diffusion process-based degradation process

with measurement errors. From the iterative results by the Bayesian method, the posterior distribution
of stochastic parameters can be updated once the new degradation monitoring information is available.
In addition, Wang et al. [35] proposed an additive Wiener process model-based degradation modeling
for hybrid deteriorating systems, which consists of a linear degradation part and a nonlinear part. The
corresponding parameter estimation method similarly contained a two-stage procedure: offline estimating
by MLE and online updating via the Bayesian method. Moreover, the Bayesian method can assure the
accuracy of the stochastic parameters and RUL prediction.
Generally speaking, the estimation of stochastic parameters in diffusion process-based degradation
model usually resorts to the Bayesian method. Owing to the complicated form of the model parameters
in diffusion process, calculating the Bayesian updated parameters requires frequently repeated calculation
and integration operations once the new degradation monitoring information is available. This will
inevitably lower the computing efficiency and real-time performance. In particular, when the Wiener
process is selected to describe the linear evolving process of the degraded equipment, the above problem
may be non-existing owing to the special linear form of drift coefficient. Nevertheless, there will exists
another transparent limitation that the Bayesian updated parameters can be obtained by only utilizing the
current degradation measurement and cannot incorporate the whole degradation measurements. Thus,
Bayesian updated mechanism for Wiener process cannot assure the prognosis accuracy of RUL. As a
result, it is urgent to research a method to overcome the common problem for diffusion process and
special problem for Wiener process.
To do so, this paper presents an adaptive prognostic method based on the generalized diffusion process
to predict the RUL of degraded equipment. First, a generalized diffusion process-based degradation
modeling framework is constructed to describe the health performance of stochastic degraded equipment
under the complex conditions. Next, the MLE method is utilized to estimate the initial model parameters
by analyzing the historical degradation information. Furthermore, a sequential Bayesian method is
introduced to recursively update the stochastic model parameters in the generalized diffusion process for
particular equipment in service. Thereafter, the RUL distribution is updated adaptively by incorporating
the obtained posterior estimates. Overall, the main contribution of this study is that there is no need to
implement the calculation process with complicated integration repeatedly utilizing the whole degradation
information obtained before the current time compared with diffusion process-based model with the
Bayesian method. Specifically, the current measured information is incorporated into the estimates of
the stochastic parameters in the previous time to determine the corresponding posterior estimates at the
current time. This can overcome the above mentioned common problem for diffusion process and raise
the efficiency to some extent. In addition, the presented method can include the linear degradation model
as a special case. Because the estimates of the stochastic parameters in the previous time can reflect
the whole degradation information obtained before the current time, the presented method can avoid the
problem that the Bayesian estimate for the stochastic parameter in the current time only depends on the
current degradation measurement for Wiener process.
The remaining parts of the paper are organized as follows. Section 2 formulates the problems and
constructs the degradation models. We estimate the unknown parameters in our model and update the
stochastic parameters by a sequential Bayesian method in Section 3. The probability density function
(PDF) and expectation of the RUL are derived in Section 4. Section 5 provides two practical case studies
associated with the gyroscope and 2017-T4 aluminum alloy to illustrate the effectiveness and superiority
of the developed method. We conclude this paper and discuss the future direction in Section 6.

2 Problem formulation

2.1 A generalized diffusion-based degradation modeling

Since the modern engineering equipment is heavily affected by the internal structure wear and external
environmental loading, the practical degradation trajectory reflecting the health status of the concerned
equipment is changeable and stochastic. Generally speaking, there are certain disadvantages in using a
single linear or nonlinear degradation model, which is difficult to accurately describe the real situation
of degradation performance and guarantee the prediction accuracy. Therefore, it is natural to adopt a
generalized diffusion process that is the combination of multiple linear or nonlinear degradation processes
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:4

to construct the degradation model. Specifically, the generalized mathematical model can be formulated
as [36]
X (t) = x0 + λT f (t, θ) + σB B (t) , (1)
where X (t) denotes the degradation level of the concerned equipment at time t, x0 denotes the corre-
sponding initial degradation level. λT f (t, θ) represents the drift coefficient controlling the degradation
speed during the life cycle, where f (t, θ) with a parameter vector θ represents an n-dimension vector
composed of a series of fundamental functions, and λ is the parameter vectors composed of the coefficient
T
of the fundamental functions, i.e., λ ∈ Rn×1 . [·] denotes the transposition operation, σB represents the
diffusion coefficient, and B (t) is the standard BM. Note that σB and B (t) are constituted together to
describe the time-varying dynamics of the degraded equipment.
Remark 1. The degradation models defined in (1) can include most of the existing linear and nonlinear
degradation models as the special cases. That is to say, when f (t, θ) = t, the generalized degradation
model can be simplified as the simple linear Wiener process-based model in [29, 37]; when f (t, θ) =
Rt
0 µ (τ, θ) dτ , the generalized degradation model can be simplified
Rt
as the nonlinear diffusion process-
based model in [32, 34]. Additionally, when f (t, θ) = [t, 0 µ (τ, θ) dτ ]T , we can obtain the additive
hybrid degradation model in [36]. The specific forms of f (t, θ) should be determined according to the
practical performance of the concerned degraded equipment. Thus, the above equation can be used to
describe any degradation trajectory once selecting the proper fundamental functions.
To characterize the unit-to-unit heterogeneity among a batch of equipment, the parameter vectors
are generally treated as the stochastic vectors following a certain distribution. Note that λ and B (t)
need to be defined in the same probability space. It is assumed that λ is statistically independent
of B (t). The most commonly utilized distribution is the multidimensional normal distribution with
unknown parameters (µλ , Σλ ), that is, λ ∼ MVN (µλ , Σλ ). And the diffusion coefficient is conventionally
considered as the fix parameter for describing the similar properties. The basic idea on these parameter
estimation methods in most literature is to fully utilize the historical equipment measurement information
for offline estimation and the monitoring data of equipment in service for online updating λ.
Remark 2. The reasons that the parameter vectors λ are assumed to follow the multidimensional
normal distribution are to facilitate the derivation process and to obtain the analytical solutions. Addi-
tionally, such assumption has been adopted in the existing literature. When λ follows other multivariate
distributions, it may be difficult to derive the analytical expressions of RUL distribution. Thus, numer-
ical simulation algorithms are needed to obtain the numerical solutions of RUL distribution, which are
time-consuming and poorly operational.

2.2 RUL prediction for particular equipment conditional on the CM data

The main task is to predict the RUL adaptively for particular equipment conditional on the CM data. We
let X1:k = {X (t1 ) , X (t2 ) , . . . , X (tk )} represent the total monitoring data of the concerned equipment
up to time tk . According to most literature in the field of RUL prediction, the RUL at time tk can be
defined as the following equation from the point of FHT:

Lk = inf {lk : X (tk + lk ) > w |X1:k } , (2)

where X (tk + lk ) represents the degradation measurement at time tk + lk , lk represents the realization
value of Lk . w denotes the preset failure threshold specified by expert knowledge, engineering experiences
and industrial standard. Note that inf indicates the abbreviation of the infimum operation.
For the sake of deriving and updating the PDF fLk |X1:k (lk |X1:k ) of the RUL at time tk , the important
information in X1:k should be fully utilized. When considering the stochastic effect of the parameter
vectors λ, the above PDF can be formulated as follows based on the law of total probability:
Z Z Z
fLk |X1:k (lk |X1:k ) = · · · fLk |λ,X1:k (lk |λ, X1:k )f (λ |X1:k ) dλ, (3)

where fLk |λ,X1:k (lk |λ, X1:k ) represents the conditional PDF of the RUL at time tk , and f (λ |X1:k ) is the
PDF of λ conditional on X1:k . The unknown model parameters in (1) include the stochastic parameter
vectors λ, the fixed parameter vector θ of the fundamental functions and the fixed diffusion coefficient
σB . Usually, the Bayesian method is utilized to update λ and guarantee the accuracy of RUL prediction.
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:5

However, owing to the complicated form of the model parameters in generalized diffusion process, the
Bayesian updated parameters require plenty of repeated calculation and integration operations once the
new degradation monitoring information is available, resulting in the problems of the low computing
efficiency and poor real-time performance. Additionally, the PDF of the RUL defined in (3) contains
complex multiple integrals, which is difficult to directly solve (3). Thus, the following problems will be
investigated in the paper.
(1) How to determine the fixed parameters including θ and σB , and the hyper-parameters in the prior
distribution of λ?
(2) How to update the posterior distribution f (λ |X1:k ) of stochastic parameter vectors recursively
with the sequential Bayesian framework without plenty of repeated calculation?
(3) How to obtain the explicit form of the PDF of the RUL for the concerned equipment according to
the defined degradation model and update it utilizing the estimated parameters?

3 Parameter estimation
3.1 Estimating the fixed parameters and hyper-parameters in the prior distribution

As described previously, the parameters to be estimated offline in the generalized diffusion process-based
model can be classified as the fixed parameters and the hyper-parameters in the prior distribution,
which are denoted as Θ= [µλ , Σλ , θ, σB ]. Let {Xi (ti,j ) = xi,j , i = 1, . . . , N, j = 1, . . . , mi } represent the
monitored historical degradation data from other analogous tested equipment, where N denotes the
number of the tested equipment, mi denotes the number of degradation measurements for the group
of tested equipment, and ti,j represents the jth monitoring time for the ith group of tested equipment.
Based on the degradation model in (1), the degradation observation at the jth monitoring time for the
ith group of tested equipment can be expressed as
Xi (ti,j ) = x0,i + λT
i f (ti,j , θ) + σB B (ti,j ) . (4)
To simplify the calculation of the likelihood function, we define Ti = [Ti,1 , Ti,2 , . . . , Ti,mi ], Ti,j =
T
f (ti,j , θ), Ti ∈ Rn×mi , xi = [xi,1 , . . . , xi,mi ] . Without loss of generality, it is assumed that the degra-
dation observations for the different equipment are statistically independent with each other. According
to the independent increments property of diffusion process, it can be observed that xi follows the mul-
tivariate normal distribution, i.e., xi ∼ MVN (µi , Σi ). The corresponding mean and covariance can be
represented as
µi = x0,i I1 +TiT µλ , (5)
Σi = Ωi + TiT Σλ Ti , (6)
T mi ×1 2
where I1 = [1, 1, . . . , 1] , I ∈ R , Ωi = σB Qi ,
 
ti,1 ti,1 · · · ti,1
 
 ti,1 ti,2 · · · ti,2 
Qi =  . .
 
 .. .. . . .
 . . .. 

ti,1 ti,2 · · · ti,mi
The degradation data for each equipment can together constitute all the monitored historical degra-
dation data, i.e., X = [x1 , x2 , . . . , xN ]T . Thus, from the characteristic of the multivariate normal
distribution, the log-likelihood function of all historical degradation data X over parameters vector Θ
can be written as [32]
N N N
ln (2π) X 1X 1 Xh i
ℓ (Θ |X ) = − mi − ln |Σi | − (xi − µi )T Σ−1
i (xi − µi ) . (7)
2 i=1
2 i=1 2 i=1

To maximize the log-likelihood function in (7), taking the first partial derivative of the log-likelihood
function with respect to µλ can yield
N
∂ℓ (Θ |X ) X 
Ti Σ−1 −1 T

= i (xi − x0,i I1 ) − Ti Σi Ti µλ . (8)
∂µλ i=1
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:6

For the specific values of Σλ , θ, and σB , let the derivative in (8) equal zero, and thus the estimated
results of µλ can be formulated as
"N #−1 N
X X
−1 T
Ti Σ−1


µ̂λ = T i Σi T i i (xi − x0,i I1 ) . (9)
i=1 i=1

After obtaining the results µ̂λ of MLE, we can further determine the profile log-likelihood function
over Σλ , θ, and σB by substituting (9) into (7), which can be formulated as
N N N
ln (2π) X 1X 1X T
ℓ (Θ |X ) = − mi − ln |Σi | − (xi − x0,i I1 ) Σ−1
i (xi − x0,i I1 )
2 i=1
2 i=1
2 i=1
"N # (10)
N h i
T −1 T
X X
T T −1


−2 (xi − x0,i I1 ) Σi Ti µ̂λ +µ̂λ Ti Σi Ti µ̂λ .
i=1 i=1

As a result, the results of MLE for Σλ , θ, and σB in Θ, can be obtained by maximizing the above
profile log-likelihood function. Subsequently, the corresponding estimated results can be recorded as
Σ̂λ , θ̂, and σ̂B . The commonly utilized method to maximize the profile log-likelihood function is multi-
dimensional search, which can be achieved by the “Fminsearch” function in MATLAB as described
in [32, 35]. Furthermore, the estimated result of (Σ̂λ , θ̂, σ̂B ) obtained by maximizing (10) is substituted
into (9). Until now, the fixed parameters and the hyper-parameters in the prior distribution of the
stochastic vector λ have been estimated utilizing the historical degradation measurement for all the tested
equipment. Owing to the fact that the stochastic vector λ describes the unit-to-unit heterogeneity, the
distribution parameters of λ are not identical for different equipment, and thus we will research how to
update the distribution parameters of λ based on the real-time degradation measurement for particular
equipment under the sequential Bayesian framework.

3.2 Stochastic parameter online updating using a sequential Bayesian method

The Bayesian estimation method can be served as the common method for parameters updating and
has been widely used in existing literature, such as [38, 39]. The Bayesian estimation method in [38]
belongs to the traditional Bayesian methods, indicating that the updated result in the last time is not
fully utilized during the updating process when the newly measured data are available. The work in [39]
offers a systematic introduction to the Bayesian state estimation framework. The main idea of such
framework is that the updated information in the last time is fully utilized to update the parameters
at the moment, which is consistent with the sequential Bayesian method. The differences between [39]
and the sequential Bayesian method lie in the complex recursive mechanism of stochastic parameters in
the state space equation. The complex recursive mechanism of stochastic parameters in [39] may not be
conducive to the derivation of analytical expressions for stochastic parameters. As a result, we adopt the
sequential Bayesian method to update the stochastic vector λ in the defined degradation model given the
monitored degradation information up to time tk for particular equipment in service.
As described before, the parameter vector λ follows the multidimensional normal distribution whose
initial mean and covariance can be determined by MLE method in Subsection 3.1. To facilitate the
illustration, the monitored degradation information up to time tk for particular equipment in service can
be denoted as X1:k . The main idea of sequential Bayesian method is that posterior estimates of λ in the
last time are regarded as the prior information in the current time. Hence, once the new degradation data
at time tk are available, we can fully utilize the current degradation information and posterior estimates
of λ at time tk−1 to compute the posterior estimates of λ at time tk . Based on this, the posterior
distribution of λ at time tk can be written as
p (X1:k |λ ) · p (λ) p (xk |X1:k−1 , λ ) · p (λ |X1:k−1 ) · p (X1:k−1 )
p (λ |X1:k ) = =
p (X1:k ) p (X1:k )
(11)
p (xk |X1:k−1 , λ ) · p (λ |X1:k−1 )
= ∝ p (xk |X1:k−1 , λ ) · p (λ |X1:k−1 ) ,
p (xk |X1:k−1 )
where p (xk |X1:k−1 , λ ) is the probability distribution of xk given X1:k−1 and λ, p (λ |X1:k−1 ) represents
the posterior distribution of λ at time tk−1 conditional on X1:k−1 . The above equation describes the recur-
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:7

rence relation of the posterior estimates at two consecutive moments in essence, which can effectively sim-
plify the repeated calculations. From the degradation model in (4), it can be observed that xk |X1:k−1 , λ
is a random variable following the normal distribution with the mean xk−1 +λT (f (tk , θ) − f (tk−1 , θ))
2
and variance σB (tk − tk−1 ), that is,
xk |X1:k−1 , λ ∼ N xk−1 +λT (f (tk , θ) − f (tk−1 , θ)) , σB
2


(tk − tk−1 ) . (12)
Hence, the corresponding PDF of xk |X1:k−1 , λ can be formulated as
"
2 #
1 xk − xk−1 − λT (f (tk , θ) − f (tk−1 , θ))
p (xk |X1:k−1 , λ ) = p · exp − 2 (t − t . (13)
2 (t − t
2πσB k k−1 ) 2σB k k−1 )

Because of the conjugate prior distribution in Bayesian framework, the posterior distribution of λ
at any time can be considered as the multidimensional normal distribution. That is to say, the mean
and covariance of λ |X1:k−1 at time tk−1 conditional on X1:k−1 can be denoted as µλ,k−1 and Σλ,k−1
respectively. Accordingly, the mean and covariance of λ |X1:k at time tk conditional on X1:k can be also
denoted as µλ,k and Σλ,k respectively. The posterior distribution of λ at time tk can be provided by the
following theorem.
Theorem 1. For the specific equipment in service, the mean and covariance of λ |X1:k at time tk
conditional on degradation data X1:k can be updated as
" #−1
T
(f (tk , θ) − f (tk−1 , θ)) (f (tk , θ) − f (tk−1 , θ)) −1
Σλ,k = 2 (t − t + Σλ,k−1 , (14)
σB k k−1 )
" #−1
T
(f (tk , θ) − f (tk−1 , θ)) (f (tk , θ) − f (tk−1 , θ))
µλ,k = 2 (t − t + Σ−1
λ,k−1
σB k k−1 )
  (15)
(f (tk , θ) − f (tk−1 , θ)) (xk − xk−1 ) −1
· 2 (t − t + Σλ,k−1 µλ,k−1 .
σB k k−1 )

Proof. As illustrated previously, it can be found that λ |X1:k−1 ∼ MVN (µλ,k−1 , Σλ,k−1 ), and thus its
PDF can be formulated as
 
1 1 T −1
p (λ |X1:k−1 ) = n 1 · exp − (λ − µ λ,k−1 ) Σ λ,k−1 (λ − µ λ,k−1 ) . (16)
(2π) 2 |Σλ,k−1 | 2 2

Substituting p (xk |X1:k−1 , λ ) in (13) and p (λ |X1:k−1 ) in (16) into (11), we can further obtain the
posterior distribution of λ at time tk , which can be expressed as
p (λ |X1:k ) ∝ p (xk |X1:k−1 , λ ) · p (λ |X1:k−1 )
"
2 #
xk − xk−1 − λT (f (tk , θ) − f (tk−1 , θ))
 
1 T −1
∝ exp − 2 (t − t · exp − (λ − µλ,k−1 ) Σλ,k−1 (λ − µλ,k−1 )
2σB k k−1 ) 2
"
T
λT (f (tk , θ) − f (tk−1 , θ)) (f (tk , θ) − f (tk−1 , θ)) λ
= exp − 2 (t − t
2σB k k−1 )

λT (f (tk , θ) − f (tk−1 , θ)) (xk − xk−1 )

+ 2 (t − t
σB k k−1 )
2 
(xk − xk−1 ) 1 1 T
− 2 (t − t − λT Σ−1 T −1 −1
λ,k−1 λ + λ Σλ,k−1 µλ,k−1 − µλ,k−1 Σλ,k−1 µλ,k−1
2σB k k−1 ) 2 2
" !
(xk − xk−1 )2 1 T −1
= exp − 2 (t − t + µλ,k−1 Σλ,k−1 µλ,k−1
2σB k k−1 ) 2
 
T (f (tk , θ) − f (tk−1 , θ)) (xk − xk−1 ) −1
+λ 2 (t − t +Σλ,k−1 µλ,k−1
σB k k−1 )
! #
T
T (f (tk , θ) − f (tk−1 , θ)) (f (tk , θ) − f (tk−1 , θ)) 1 −1
−λ 2 (t − t + Σλ,k−1 λ .
2σB k k−1 ) 2
(17)
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:8

Moreover, from the multidimensional normal property of λ |X1:k at time tk conditional on X1:k , we have
 
1 1 T −1
p (λ |X1:k ) = n 1 exp − (λ − µλ,k ) Σλ,k (λ − µλ,k )
(2π) 2 |Σλ,k | 2 2
  (18)
1 T −1 T −1 1 T −1
∝ exp − λ Σλ,k λ + λ Σλ,k µλ,k − µλ,k Σλ,k µλ,k .
2 2

As a result, the posterior parameters can be determined by comparing (17) and (18). Specifically, the
third term of the exponential in (17) corresponds to the first term of the exponential in (18), and the
second term of the exponential in (17) corresponds to the second term of the exponential in (18). Under
these two corresponding relationships, the constant terms of the exponential in (17) and (18) seem to be
not important. Thus, µλ,k and Σλ,k can be formulated as
" #−1
T
(f (tk , θ) − f (tk−1 , θ)) (f (tk , θ) − f (tk−1 , θ))
Σλ,k = 2 (t − t +Σ−1
λ,k−1 ,
σB k k−1 )
" #−1
T
(f (tk , θ) − f (tk−1 , θ)) (f (tk , θ) − f (tk−1 , θ))
µλ,k = 2 (t − t +Σ−1
λ,k−1
σB k k−1 )
 
(f (tk , θ) − f (tk−1 , θ)) (xk − xk−1 ) −1
· 2 (t − t +Σλ,k−1 µλ,k−1 .
σB k k−1 )

This completes the proof of Theorem 1.

On the basis of Theorem 1, the stochastic parameter vectors λ can be updated for the equipment
once new degradation data are obtained. The posterior estimators of λ at time tk are associated with
the current degradation data and posterior estimators of λ at time tk−1 . Such update mechanism can
avoid plenty of repeated calculation and integration operations for traditional Bayesian method during
the update process, which is beneficial for estimating model parameters quickly and guaranteeing the
accuracy of RUL prediction effectively.

4 RUL prediction
In Section 2, a generalized diffusion-based degradation model is established and the concept of the RUL
at time tk is defined. Nevertheless, the life distribution should be derived first as the basis of RUL
prediction. Thus, the life of the equipment can be defined as the time when the performance indicator
firstly crosses the failure threshold, which is formulated as

T = inf {t : X (t) > w |X (0) < w } , (19)

where t represents the realization value of T . The PDFs of life and RUL should be derived from (19) and
(2). To implement the corresponding derivation, the following two lemmas should be given at first.
Lemma 1. When ignoring the stochastic nature of λ, the conditional PDF of the life for the equipment
with the defined degradation process can be approximated as [32]
"
2 #
w − λT f (t, θ) + tλT f ′ (t, θ) w − λT f (t, θ)
fT |λ,X1:k (t |λ, X1:k ) = p · exp − 2 , (20)
2
2πt3 σB 2tσB

where f ′ (t, θ) represents the partial derivation of f (t, θ) with respect to t. The work in [32] provides
the detailed proof process, and thus the proof is omitted in the paper.
Lemma 2. Ignoring the unit-to-unit variability, the conditional PDF of the RUL at the time tk for the
equipment with the defined degradation process can be approximately expressed as
"
2 #
wk − λT f ∗ (lk , θ) + lk λT f ′ (tk + lk , θ) wk − λT f ∗ (lk , θ)
fLk |λ,X1:k (lk |λ, X1:k ) = p · exp − 2 , (21)
2πlk3 σB
2 2lk σB

where f ∗ (lk , θ) = f (tk + lk , θ) − f (tk , θ), wk = w − X (tk ).

 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:9

Proof. According to the general diffusion process-based model, the degradation level at time tk can be
formulated as
X (tk ) = x0 + λT f (tk , θ) + σB B (tk ) . (22)
And the degradation level at any time t (t > tk ) can be also formulated as

X (t) = x0 + λT f (t, θ) + σB B (t) . (23)

Combining these two equations, we have

X (t) − X (tk ) = λT (f (t, θ) − f (tk , θ)) + σB (B (t) − B (tk )) . (24)

Since the time t and tk satisfy the relationship t = tk + lk , thus the above equation can be further
represented as
U (lk ) = λT (f (tk + lk , θ) − f (tk , θ)) + σB B (lk )
(25)
= λT f ∗ (lk , θ) + σB B (lk ) ,
where U (lk ) = X (t) − X (tk ). It can be observed that the RUL at time tk can be considered as the time
when the new stochastic process {U (lk ) , lk > 0} crosses the threshold wk = w − X (tk ).
For the transformed stochastic process {U (lk ) , lk > 0}, according to the theorem in [32], we can easily
obtain the PDF of time when {U (lk ) , lk > 0} crosses the threshold wk . Therefore, when ignoring the
unit-to-unit variability, the conditional RUL at time tk for the concerned equipment can be represented
as
"
2 #
wk − λT f ∗ (lk , θ) + lk λT f ′ (tk + lk , θ) wk − λT f ∗ (lk , θ)
fLk |λ,X1:k (lk |λ, X1:k ) = p · exp − 2 . (26)
2πlk3 σB
2 2lk σB

This completes the proof of Lemma 2.

Note that the conditional PDFs of the life and RUL ignoring the stochastic nature of λ are provided in
(20) and (21), and it is natural to consider that the law of total probability should be adopted to derive
the PDFs of the life and RUL. The specific expressions can be formulated as
Z Z Z
fT |X1:k (t |X1:k ) = · · · fT |λ,X1:k (t |λ, X1:k )f (λ |X1:k ) dλ
  (27)
= Eλ|X1:k fT |λ,X1:k (t |λ, X1:k ) ,
Z Z Z
fLk |X1:k (lk |X1:k ) = ··· fLk |λ,X1:k (lk |λ, X1:k )f (λ |X1:k ) dλ
(28)
 
= Eλ|X1:k fLk |λ,X1:k (lk |λ, X1:k ) .
From the previous description, it can be found that λ ∼ MVN (µλ , Σλ ). However, the complex multiple
integrals exist in the process of implementing the law of total probability. It may be especially difficult to
derive the corresponding theoretical expressions. To address this problem, we give the following lemma
for facilitating the derivation of the PDFs of the life and RUL.
Lemma 3. It is assumed that the stochastic vector ρ follows an n-dimensional normal distribution with
corresponding mean µρ and covariance Σρ , where the covariance matrix is a positive definite matrix. υ1 ,
υ2 and c are the given constants, and a, b are n-dimensional vectors. Then we can obtain the following
formulation:
"
2 !#
T

υ2 − bT ρ
Eρ υ1 − a ρ exp −
2c
s "
2 #
cn υ1 aT Σρ b + caT µρ υ2 − bT µρ
 
= υ1 − exp − , (29)
|bbT Σρ + cI| c + bT Σρ b 2 (c + bT Σρ b)

where I is an identity matrix and I ∈ Rn×n . The proof of Lemma 3 was provided in [36]. This lemma is
so critical to derive the theoretical expressions of the PDFs of life and RUL. On the basis of Lemma 3,
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:10

the PDF of the life in (27) and the PDF of the RUL in (28) can be further expressed as
" #
2
+ f T (t, θ) Σλ f (t, θ) − aT Σλ f (t, θ) + tσB 2 T


w tσB a µλ
fT |X1:k (t |X1:k ) = 2 + f T (t, θ) Σ f (t, θ)
tσB λ
s "
2 # (30)
(tσB2 )n−1 w − f T (t, θ) µλ
× 2 I| × exp − 2 (tσ 2 + f T (t, θ) Σ f (t, θ)) ,
2πt2 |f (t, θ) f T (t, θ) Σλ + tσB B λ

T
where aT = [f (t, θ) − tf ′ (t, θ)] .

" #
2 T
wk (lk σB + (f ∗ (lk , θ)) Σλ f ∗ (lk , θ) − aT ∗ 2 T
L Σλ f (lk , θ)) + lk σB aL µλ
fLk |X1:k (lk |X1:k ) = T
2 + (f ∗ (l , θ)) Σ f ∗ (l , θ)
lk σB k λ k
v " 2 # (31)
2 )n−1 (w − (f ∗ (lk , θ))T µλ )
u
u (lk σB
×t T
× exp − ,
2πlk2 |f ∗ (lk , θ) (f ∗ (lk , θ)) Σλ + lk σB2 I| 2 + (f ∗ (l , θ))T Σ f ∗ (l , θ))
2(lk σB k λ k

T
where aT ∗ ′
L = [f (lk , θ) − lk f (tk + lk , θ)] . When we obtain the offline estimated results for θ and σB ,
and the online updated results for µλ and Σλ at tk by the methods in Section 3, the PDFs of the life
and RUL can be adaptively determined once new CM data are gained, which is beneficial to make the
scheduled maintenance activities scientifically. Additionally, the expectations of the life and RUL can be
further derived on the basis of the relevant content in the probability statistics. The specific expressions
can be formulated as Z ∞
E (T ) = tfT |X1:k (t |X1:k ) dt, (32)
0
Z ∞
E (Lk ) = lk fLk |X1:k (lk |X1:k ) dlk . (33)
0
Owing to the complicated form of fT |X1:k (t |X1:k ) and fLk |X1:k (lk |X1:k ), it is especially difficult to
directly acquire the theoretical explicit solutions by the integral operation in (32) and (33). Nevertheless,
the numerical solution can be easily obtained by converting integral operation into a large number of
summations with the help of MATLAB.
To facilitate the understanding and application, the main procedure of the sequential Bayesian updated
diffusion process model for adaptive RUL prediction is summarized in Algorithm 1.

Algorithm 1 Sequential Bayesian updated diffusion process model for adaptive RUL prediction
Input: Historical degradation data X, and online degradation data X1:k ;
Output: PDFs and expectations of life and RUL;
Step1: Construct the generalized diffusion process in (1) to describe the degradation model for the engineering equipment;
Step2: Estimate the fixed parameters and the hyper-parameters in the prior distribution based on X by MLE, namely Θ̂;
Step3: Initialize the hyper-parameters of stochastic vector λ utilizing the offline estimated results;
Step4: Update the hyper-parameters of stochastic vector λ at next time by (14) and (15);
Step5: k = k + 1, and go to Step4. When stochastic vector λ at any time is updated, go to Step6;
Step6: Predict the RUL and quantify uncertainty based on (30)–(33).

5 Experimental studies
For the verification purpose, this section provides two experimental studies related with the gyropic drift
data of inertial navigation systems and the fatigue-crack growth data of 2017-T4 aluminum alloy to
implement the research on RUL prediction. The state-of-the-art prognostic methods in [32, 35], which
are the nonlinear Wiener process-based model without the updating mechanism, and additive Wiener
process-based model updated by traditional Bayesian method respectively, will be introduced to compare
the prediction performance. To better illustrate the comparison results, the nonlinear Wiener process-
based model without the updating mechanism, and additive Wiener process-based model updated by
traditional Bayesian method can be recorded as Methods 1 and 2. Note that f (t, θ) of the proposed
method can be selected as a relatively simple form [t, exp (θt)]T , and the nonlinear drift coefficient of
Methods 1 and 2 are considered as the exponential function form.
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:11

3.0
#1
#2
2.5 #3
#4
#5
2.0

Gyropic drift (°/h)

1.5

1.0

0.5

0
0 5 10 15 20 25
Condition monitoring time (h)

Figure 1 (Color online) The drift data of five gyroscopes.

Table 1 The estimated values of all the parameters for the three methods
2 2
µλ1 µλ2 σλ 1
σλ 2
ς θ σB
Our model 1.13E−11 6.35E−9 8.53E−11 8.12E−11 8.35E−11 0.820 0.063
Method 1 9.34E−9 – 8.66E−9 – – 0.815 0.066
Method 2 1.37E−11 6.63E−9 8.57E−11 8.14E−11 – 0.823 0.061

5.1 The RUL prediction for gyroscope

As a class of typical mechanical and electrical products, the gyroscope is mainly employed to measure
the relevant information related with the angular velocity and plays an irreplaceable role in various
fields including missiles, rockets, aircraft, and warships. Owing to the complicated structure and diverse
environment, the gyroscope exhibits the degradation characteristic of the combination of multiple linear
and nonlinear degradation processes instead of a single degradation process. In reality, such degradation
may accumulate as time passes by, which can lead to bigger and bigger drift coefficient. In other words,
the observed drift coefficient through the precision testing procedure can reflect the health condition
of the gyroscope. Furthermore, we can implement the research on RUL based on the observed drift
coefficient and the given technical specifications (such as the preset failure threshold). Note that once
the drift coefficient exceeds the failure threshold, the gyroscope requires to carry out maintenance or
replacement to guarantee the precision of the inertial navigation system.
Five groups of historical drift data of gyroscopes provided in [32] are utilized for degradation modeling
and RUL prediction, which are depicted in Figure 1. The drift data of all the gyroscopes show an upward
trend in the life cycle as a whole, indicating the performance degradation of gyroscopes to some extent. In
this experiment, the monitoring interval for the concerned gyroscopes is 2.5 h. According to Figure 1, it
can be observed that the mean time to failure (MTTF) is around 21.5 h when experiencing the continuous
operation. As for every gyroscope, 9 suits of drift coefficient data are collected by CM technique. As
described previously, we can estimate the fixed parameters and hyper-parameters in our model by the
proposed MLE method utilizing these drift data of five gyroscopes, and the corresponding parameters
can be summarized in Table 1. It is worth mentioning that because f (t, θ) equals [t, exp (θt)]T , the mean
T T σ2 ς
and covariance of λ = [λ1 , λ2 ] can be represented as µλ = [µλ1 , µλ2 ] and Σλ = [ λς 1 σ2 ], where ς
λ2
represents the covariance of λ1 and λ2 . Analogously, the estimated parameters for Methods 1 and 2 can
also obtained by analyzing the drift data of five gyroscopes, which are listed in the same table.
After acquiring the offline estimated values of all the parameters in the proposed model, it is necessary
to select a specific gyroscope for parameters updating and RUL prediction. Hence, gyroscope #4 can be
selected as a candidates and the corresponding failure threshold is defined as the last drift data, which is
different from the work in [32], i.e., w = 0.5042. That is to say, its true lifetime is 22.5 h. Consequently, we
can update the parameters of the stochastic vector via the sequential Bayesian method for the monitored
gyroscope at each CM time and substitute the updated results into the PDF and expectation of the RUL
in (31) and (33). To investigate the prediction capability of these three methods in the RUL prediction,
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:12

1.0 2.0
Predicted RUL Predicted RUL
True RUL True RUL
0.8
1.5
0.6
PDF

PDF
1.0
0.4

0.5
0.2

0 0
20 20
17.5 17.5
15 15
M 30 Mo 12.5 30
on 12.5 10 nit
ori 10
ito 20 7.5 20
rin 7.5 ng 5 )
gt 5 10 ) tim 10 L (h
im 2.5 L (h e( 2.5 R U
e(
h) 0 0 RU h) 0 0

Figure 2 (Color online) The PDFs of the RULs under the Figure 3 (Color online) The PDFs of the RULs under the
proposed method with the drift data of the gyroscopes. Method 1 with the drift data of the gyroscopes.

we depict the PDF curves and the corresponding predicted values from the first monitoring point to the
eighth monitoring point for these three methods in Figures 2–4. It can be seen from these three figures
that all the PDF curves can cover the predicted results of the RUL well over all the CM times. Moreover,
as more and more drift data are acquired, the PDF curves for these three methods will become higher
and tighter, which implies that the uncertainties of RUL prediction can be decreased since more useful
information is incorporated into the results of the RUL prediction. However, it can be easily observed
from Figures 2–4 that the predicted RULs by Method 1 deviate greatly from the true RUL, especially
at the last stage of the life cycle. This is because the parameter estimation of Method 1 is just MLE
and cannot update the model parameters. Besides, the degradation model for Method 1 only contains
the nonlinear component and can be regarded as the special case of the proposed method and Method 2,
and thus the fitting capacity by Method 1 is inferior to that by other two methods. As for the proposed
method and Method 2, there is little difference in prediction effect of RUL intuitively.
For the quantitative comparison, two generally utilized indicators including relative error (RE) and
mean-squared error (MSE) are selected to evaluate the prediction performance of RUL. The former
mainly reflects the prediction accuracy, while the latter takes both prediction accuracy and uncertainty
into consideration. Specifically, RE and MSE of RUL prediction can be formulated as

|E (Lk ) − l̃k |
REk = × 100%, (34)
˜lk
Z ∞
2
MSEk = (lk − l̃k ) fLk |X1:k (lk |X1:k ) dlk , (35)
0

where E (Lk ) and l˜k represent the predicted RUL and actual RUL at the observation time respectively.
fLk |X1:k (lk |X1:k ) denotes the PDF of the RUL utilizing the available degradation data up to time tk
based on (31). Based on the above illustration, Table 2 provides the REs of the predicted RULs for these
three methods. From Table 2, it can be found that all the REs of the predicted RULs for the proposed
model are below 10%, which indicates the effectiveness of our model in RUL prediction of the gyroscope.
At the first and second monitoring points, the REs of the predicted RULs by Method 1 are less than
those by other two methods. This is due to the fact the offline parameter estimation results by MLE are
accurate at this time. However, there is no updating mechanism for Method 1, and thus the parameter
estimation results will not change with new data acquisition. As a result, the REs of the predicted RULs
for the proposed model are less than those by Method 1 after the second monitoring points. In addition,
since our model and Method 2 utilize the updating mechanism, the prediction effects of these two methods
are satisfied as a whole. Specifically, the prediction effects by our model are slightly better than those
by Method 2. The average REs of the predicted RULs for these three methods are 6.5825%, 10.6337%,
and 7.9225%. Moreover, the calculating times for our model and Method 2 are respectively 16.39 s and
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:13

20
Proposed method
1.0 Predicted RUL Method 1
True RUL Method 2
15
0.8

0.6
PDF

MSE
10
0.4

0.2
5
0
20
17.5
Mo 1512.5 30
nit 10 0
ori 7.5 20
ng 0 2.5 5 7.5 10 12.5 15 17.5 20
tim 5 10 (h)
e( 2.5 RUL
h) 0 0 Monitoring time (h)

Figure 4 (Color online) The PDFs of the RULs under the Figure 5 (Color online) The MSE comparison of the pre-
Method 2 with the drift data of the gyroscopes. dicted RULs for three methods.

Table 2 The comparison of RE (%) for these three methods

Time (h)
2.5 5 7.5 10 12.5 15 17.5 20
Our model 7.61 5.39 4.84 4.17 5.96 8.64 8.38 7.67
Method 1 3.64 4.42 5.48 6.96 9.06 12.3 17.85 25.36
Method 2 7.76 6.91 6.64 6.01 6.37 8.00 10.3 11.39

41.57 s, which implies that the proposed sequential Bayesian method can overcome the problem of plenty
of repeated calculation and integration operations.
Furthermore, we provide the MSEs of the predicted RULs for these three methods in Figure 5. It can
be observed that MSEs of the predicted RULs in the life cycle for these three methods have the downward
trend as a whole. That is to say, as more drift data are available, MSEs of the predicted RULs can be
decreased. Nevertheless, all the MSEs of the predicted RULs by our model are smaller than those by
Methods 1 and 2. Note that at the late period in the life cycle (12.5 h, 15 h, and 20 h), the prediction
effects by our model are slighter than those by other methods, which is inconsistent with the results of
RE. Because the MSE takes not only the prediction accuracy, but also the uncertainty into consideration.
Although the REs of the predicted RULs for Method 1 are so large at the late period, less uncertainties
can lead to small MSE. The MSE values at 20 h for the proposed method, Method 1, and Method 2 are
0.2696, 0.4770, and 0.3022.

5.2 The RUL prediction for 2017-T4 aluminum alloy

As a vital metal material, 2017-T4 aluminum alloy is frequently applied in the military equipment and
spacecraft. Engineering practice shows that the performance of such material is evaluated by the length
of fatigue cracks. In general, when the crack length reaches or exceeds the preset threshold value, the
mechanism structure will reach a critical state, which can be considered as the failure. The fatigue-crack
growth data in four test specimens of 2017-T4 aluminum alloy provided in [32] are obtained under a
stress level of 200 MPa. For each sample, ten crack levels are recorded, and the monitoring interval is
100000 cycles until the end of the experiment. Figure 6 depicts the evolving process of fatigue-crack
growth data of 2017-T4. From Figure 6, it can be found that the degradation trajectories for these four
test specimens of 2017-T4 aluminum alloy exhibit the nonlinearity to some extent.
Similarly, the offline parameter estimators for the above three methods can be acquired by MLE
utilizing all the fatigue-crack growth data of 2017-T4, which are summarized in Table 3. After that,
2017-T4 aluminum alloy #3 is regarded as the research object for RUL prediction. Accordingly, the
failure threshold is set to be the last fatigue crack data (w = 5.6). We utilize the proposed sequential
Bayesian method in Section 3 to update the parameters of the stochastic vector λ. Furthermore, the
PDFs and the corresponding predicted values of the RUL from 1.5×105 cycles to 2.3×105 cycles for these
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:14

7
#1
#2
6
#3

Fatigue-crack length of 2017-T4 (mm)

#4
5

0
1.4 1.6 1.8 2.0 2.2 2.4
Condition monitoring time (105 cycles)

Figure 6 (Color online) Fatigue-crack growth data of 2017-T4.

Table 3 The estimated values of all the parameters for the three methods
2 2
µλ1 µλ2 σλ 1
σλ 2
ς θ σB
Our model 1.15E−9 1.49E−4 7.13-11 3.43E−5 7.64E−11 4.4107 0.7535
Method 1 1.50E−4 – 3.531E−5 – – 4.4402 0.7788
Method 2 1.42E−9 1.63E−4 7.83E−11 3.47E−5 – 4.3998 0.7623

Predicted RUL 60 Predicted RUL

60
True RUL True RUL

40 40
PDF

PDF

20 20

0 0
2.3 2.3
Mo 2.2 2.1 2 Mo
2.2
2.1 2
nito 2 1.5 nito 2
ring 1.9 ring 1.9 1.5
tim 1.8 1.7 1 s) tim 1.81.7 1
e (1
0 5 c 1.6 1.5 0.5 5 cycle e (1 0.5 5 ycl
es)
ycl 0 (10 0 5 c 1.61.5
0 (10 c
es) RUL ycle
s) RU L

Figure 7 (Color online) The PDFs of the RULs under the Figure 8 (Color online) The PDFs of the RULs under the
proposed method with the fatigue crack data of 2017-T4 alu- Method 1 with the fatigue crack data of 2017-T4 aluminum
minum alloy. alloy.

three methods are provided in Figures 7–9. It can be seen from these figures that the proposed method
and Method 2 can predict RUL for 2017-T4 aluminum alloy more accurately compared with Method 1.
Because the updating mechanism in the proposed method and Method 2 can guarantee the accuracy of
the stochastic parameters utilizing the online CM data. Moreover, the prediction effect of the proposed
method is slightly better than that of Method 2.
For sake of further investigating the prediction performance quantitatively, we provide the REs and
MSEs of the RULs for these three methods in Figures 10 and 11. It can be observed that all the REs and
MSEs of the RULs at each monitoring point by our proposed method are less than those by Methods
1 and 2. Especially at the 2.3 × 105 cycles, the RE and MSE of the predicted RUL for the proposed
method are 3.84% and 8.41 × 10−4 respectively, which indicates the effectiveness of the proposed method
in RUL prediction. What is more, the box plots of the predicted RUL at the 6th and 8th monitoring
points for the three methods are depicted in Figure 12. From the figure, it can be found that the median
of the RUL for the proposed method is closer to the true RUL compared with other methods. And it can
be concluded that the dispersion degree of the proposed method is slightly less than that of Method 2,
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:15

60

Proposed method
Predicted RUL 50 Method 1
40
True RUL Method 2
30 40
PDF

RE (%)
20
30
10
20
0
2.3
Mo 2.22.1 2 10
nito 2 1.5
rin
g ti 1.91.8 1 s)
me ycle
(10 5 1.71.6 0.5 05 c
cyc 1.5 0 L (1 0
les
) RU 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3
Monitoring time (105 cycles)

Figure 9 (Color online) The PDFs of the RULs under the Figure 10 (Color online) The comparison of RE for these
Method 2 with the fatigue crack data of 2017-T4 aluminum three methods.
alloy.

0.08
0.7
RUL (105 cycles)

Proposed method True RUL

0.07 0.6
Method 1
0.06 Method 2
0.4
0.05
0.2
Proposed method Method 1 Method 2
MSE

0.04
6th monitoring point
RUL (105 cycles)

0.03 0.3 True RUL

0.02
0.2
0.01
0.1
0 Proposed method Method 1 Method 2
1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 8th monitoring point
Monitoring time (105 cycles)

Figure 11 (Color online) The MSE comparison of the pre- Figure 12 (Color online) Box plots of the predicted RUL at
dicted RULs for three methods. the 6th and 8th monitoring points.

which is consistent with the PDF curves in Figures 7 and 9.

Overall, the results in these two experimental studies imply that our proposed sequential Bayesian
method is superior to Methods 1 and 2. Therefore, the proposed method has potential application
prospects for the RUL estimation issues of the engineering equipment.

6 Conclusion
A sequential Bayesian updated diffusion process model is developed for adaptive RUL prediction. First, a
generalized diffusion process-based degradation modeling framework is constructed to describe the health
performance of stochastic degraded equipment under the complex conditions. Then, we utilize the MLE
method to estimate the initial model parameters by analyzing the historical degradation information.
Furthermore, the stochastic parameters in generalized diffusion process can be updated recursively based
on the sequential Bayesian method for particular equipment in service. Hence, the RUL distribution is
updated adaptively by incorporating the obtained posterior estimates. Finally, we provide two practical
case studies associated with the gyroscope and 2017-T4 aluminum alloy to demonstrate the effectiveness
and superiority of the proposed sequential Bayesian method. The improvements for degradation modeling
and parameters updating are especially significant for decision making related activities including spare
parts ordering and maintenance scheduling.
Despite the fact the generalized diffusion process-based degradation model for adaptive RUL prediction
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:16

via the sequential Bayesian method can achieve satisfactory results in the experimental studies and exhibit
the superiority in RUL prediction performance, some issues are still available in this paper and require
the additional research. First, the hidden or partially observed degradation cases may be encountered in
the practical case, which deserves the further exploration. Second, when the stochastic vector does not
follow multidimensional normal distribution, this will cause challenge for derivation. Thus, it is necessary
to research the non-Gaussian case of the stochastic vector.
Acknowledgements The work was supported by National Natural Science Foundation of China (Grant Nos. 61833016, 61922089,
61773386, 61573365, 61573366, 61903376, 61773389, 61673311), Shaanxi National Science Foundation (Grant Nos. 2020JQ-489,
2020JM-360, 2020JQ-298), and National Key R&D Program of China (Grant No. 2018YFB1306100). The authors would like to
thank the anonymous reviewers for their valuable suggestions.

References
1 Pecht M. Prognostics and Health Management of Electronics. Hoboken: Wiley, 2008
2 Jardine A K S, Lin D, Banjevic D. A review on machinery diagnostics and prognostics implementing condition-based main-
tenance. Mech Syst Signal Process, 2006, 20: 1483–1510
3 Lee J, Wu F, Zhao W, et al. Prognostics and health management design for rotary machinery systems-Reviews, methodology
and applications. Mech Syst Signal Process, 2014, 42: 314–334
4 Tsui K L, Chen N, Zhou Q, et al. Prognostics and health management: a review on data driven approaches. Math Problems
Eng, 2015, 2015: 1–17
5 Lei Y, Li N, Guo L, et al. Machinery health prognostics: a systematic review from data acquisition to RUL prediction. Mech
Syst Signal Process, 2018, 104: 799–834
6 Si X S, Li T M, Zhang Q. A general stochastic degradation modeling approach for prognostics of degrading systems with
surviving and uncertain measurements. IEEE Trans Rel, 2019, 68: 1080–1100
7 Zhai Q, Ye Z S. RUL prediction of deteriorating products using an adaptive Wiener process model. IEEE Trans Ind Inf, 2017,
13: 2911–2921
8 Lall P, Lowe R, Goebel K. Prognostics health management of electronic systems under mechanical shock and vibration using
kalman filter models and metrics. IEEE Trans Ind Electron, 2012, 59: 4301–4314
9 Si X S, Li T M, Zhang Q, et al. Prognostics for linear stochastic degrading systems with survival measurements. IEEE Trans
Ind Electron, 2020, 67: 3202–3215
10 Xi X P, Chen M Y, Zhou D H. Remaining useful life prediction for multi-component systems with hidden dependencies. Sci
China Inf Sci, 2019, 62: 022202
11 Yu Y, Si X S, Hu C H, et al. Online remaining-useful-life estimation with a Bayesian-updated expectation-conditional-
maximization algorithm and a modified Bayesian-model-averaging method. Sci China Inf Sci, 2021, 64: 112205
12 Jin X, Sun Y, Que Z, et al. Anomaly detection and fault prognosis for bearings. IEEE Trans Instrum Meas, 2016, 65:
2046–2054
13 Si X S, Li T M, Zhang Q. Optimal replacement of degrading components: a control-limit policy. Sci China Inf Sci, 2021, 64:
209205
14 Si X S, Hu C H, Li T M, et al. A joint order-replacement policy for deteriorating components with reliability constraint. Sci
China Inf Sci, 2021, 64: 189203
15 Verbert K, de Schutter B, Babuska R. A multiple-model reliability prediction approach for condition-based maintenance.
IEEE Trans Rel, 2018, 67: 1364–1376
16 Hu C H, Pei H, Si X S, et al. A prognostic model based on DBN and diffusion process for degrading bearing. IEEE Trans
Ind Electron, 2020, 67: 8767–8777
17 Gebraeel N, Lawley M, Liu R, et al. Residual life predictions from vibration-based degradation signals: a neural network
approach. IEEE Trans Ind Electron, 2004, 51: 694–700
18 Garcı́a-Nieto P J, Garcı́a-Gonzalo E, Sánchez-Lasheras F, et al. Hybrid PSO-SVM-based method for forecasting of the
remaining useful life for aircraft engines and evaluation of its reliability. Reliability Eng Syst Saf, 2015, 138: 219–231
19 Guo L, Lei Y, Li N, et al. Machinery health indicator construction based on convolutional neural networks considering trend
burr. Neurocomputing, 2018, 292: 142–150
20 Deutsch J, He D. Using deep learning-based approach to predict remaining useful life of rotating components. IEEE Trans
Syst Man Cybern Syst, 2018, 48: 11–20
21 Zhang J X, Hu C H, He X, et al. A novel lifetime estimation method for two-phase degrading systems. IEEE Trans Rel, 2019,
68: 689–709
22 Wang D, Zhao Y, Yang F, et al. Nonlinear-drifted Brownian motion with multiple hidden states for remaining useful life
prediction of rechargeable batteries. Mech Syst Signal Process, 2017, 93: 531–544
23 Ling M H, Ng H K T, Tsui K L. Bayesian and likelihood inferences on remaining useful life in two-phase degradation models
under gamma process. Reliability Eng Syst Saf, 2019, 184: 77–85
24 Ye Z S, Chen N. The inverse Gaussian process as a degradation model. Technometrics, 2014, 56: 302–311
25 Doksum K A, Hoyland A. Models for variable-stress accelerated life testing experiments based on Wiener processes and the
inverse gaussian distribution. Theor Probab Appl, 1993, 37: 137–139
26 Tseng S T, Tang J, Ku I H. Determination of burn-in parameters and residual life for highly reliable products. Naval Res
Logistics, 2003, 50: 1–14
27 Tseng S T, Peng C Y. Optimal burn-in policy by using an integrated Wiener process. IIE Trans, 2004, 36: 1161–1170
28 Elwany A, Gebraeel N. Real-time estimation of mean remaining life using sensor-based degradation models. J Manufacturing
Sci Eng, 2009, 131: 051005
29 Si X S, Wang W, Chen M Y, et al. A degradation path-dependent approach for remaining useful life estimation with an exact
and closed-form solution. Eur J Operational Res, 2013, 226: 53–66
30 Gebraeel N Z, Lawley M A, Li R, et al. Residual-life distributions from component degradation signals: a Bayesian approach.
IIE Trans, 2005, 37: 543–557
31 Whitmore G A, Schenkelberg F. Modeling accelerated degradation data using Wiener diffusion with a time scale transforma-
tion. Lifetime Data Anal, 1997, 3: 27–45
 Pei H, et al. Sci China Inf Sci June 2022 Vol. 65 162206:17

32 Si X S, Wang W, Hu C H, et al. Remaining useful life estimation based on a nonlinear diffusion degradation process. IEEE
Trans Rel, 2012, 61: 50–67
33 Si X S, Ren Z Q, Hu X X, et al. A novel degradation modeling and prognostic framework for closed-loop systems with
degrading actuator. IEEE Trans Ind Electron, 2020, 67: 9635–9647
34 Tang S J, Guo X S, Yu C Q, et al. Real time remaining useful life prediction based on nonlinear Wiener based degradation
processes with measurement errors. J Cent South Univ, 2014, 21: 4509–4517
35 Wang Z Q, Hu C H, Wang W, et al. An additive Wiener process-based prognostic model for hybrid deteriorating systems.
IEEE Trans Rel, 2014, 63: 208–222
36 Zhang Z X, Hu C H, Si X S, et al. Stochastic degradation process modeling and remaining useful life estimation with flexible
random-effects. J Franklin Institute, 2017, 354: 2477–2499
37 Wang Y E, Ma W M, Chow T W S, et al. A two-step parametric method for failure prediction in hard disk drives. IEEE
Trans Indust Inform, 2014, 10: 419–430
38 Ebenezer R H P, Susan E, Srinivasan R, et al. Template-based gait authentication through Bayesian thresholding. IEEE/CAA
J Autom Sin, 2019, 6: 209–219
39 Fang H Z, Tian N, Wang Y B, et al. Nonlinear Bayesian estimation: from Kalman filtering to a broader horizon. IEEE/CAA
J Autom Sin, 2018, 5: 401–417