Complex & Intelligent Systems (2025) 11:140

[Link]

ORIGINAL ARTICLE

A generalized diffusion model for remaining useful life prediction with

uncertainty
Bincheng Wen1 · Xin Zhao1 · Xilang Tang2 · Mingqing Xiao1 · Haizhen Zhu1 · Jianfeng Li3

Received: 25 June 2024 / Accepted: 29 December 2024 / Published online: 15 January 2025
© The Author(s) 2025

Abstract
Forecasting the remaining useful life (RUL) is a crucial aspect of prognostics and health management (PHM), which has
garnered significant attention in academic and industrial domains in recent decades. The accurate prediction of RUL relies on
the creation of an appropriate degradation model for the system. In this paper, a general representation of diffusion process
models with three sources of uncertainty for RUL estimation is constructed. According to time-space transformation, the
analytic equations that approximate the RUL probability distribution function (PDF) are inferred. The results demonstrate that
the proposed model is more general, covering several existing simplified cases. The parameters of the model are then calculated
utilizing an adaptive technique based on the Kalman filter and expectation maximization with Rauch-Tung-Striebel (KF-EM-
RTS). KF-EM-RTS can adaptively estimate and update unknown parameters, overcoming the limits of strong Markovian
nature of diffusion model. Linear and nonlinear degradation datasets from real working environments are used to validate the
proposed model. The experiments indicate that the proposed model can achieve accurate RUL estimation results.

Keywords Remaining useful life · Kalman filter · General diffusion model · Prognostic

Introduction Physics-based models necessitate knowledge of the equip-

ment’s degradation mechanism and the development of an
Research background and motivation appropriate mathematical model to depict the system’s degra-
dation process. Nevertheless, due to the fact that degradation
Estimating RUL is a highly important topic in the subject of mechanisms vary between devices, an increase in device
PHM, as it has the potential to enhance the reliability, avail- complexity also increases the difficulty of analyzing degra-
ability, and security of degrading systems. This has garnered dation mechanisms, which reduces the generalizability of the
significant interest from both the academic and industrial sec- constructed physical models and, in many instances, makes
tors [8]. Over the last twenty years, a multitude of estimation it challenging to construct a physical model of failure [10].
methods have been introduced and put into practice. They In contrast to physical-based models, data-driven models do
are typically classified as physics-based models [4, 19], data- not necessitate knowledge of system failure mechanisms.
driven models [20, 22], stochastic process models [5, 39], and By analyzing the degradation data, data-driven models are
hybrid models [2, 30] based on their underlying concepts. able to generate corresponding input–output models, which
significantly reduces the difficulty of developing equipment
B Bincheng Wen degradation models. However, due to its high requirements
1255292204@[Link] on data quality and data, it is often difficult to meet them
Xin Zhao in practical situations, which greatly limits its application in
atsl2023@[Link] specific projects [40]. Hybrid models are commonly used to
combine the physical model with the data-driven model in the
1 ATS Lab, Air Force Engineering University, Xi’an 710038, hopes of developing a new deterioration model that incorpo-
China
rates the benefits of both sides. However, as previously stated,
2 Equipment Management and UAV Engineering College, Air the failure physical model is challenging to implement, and
Force Engineering University, Xi’an 710038, China
the data-driven model necessitates a substantial volume of
3 Beijing Aeronautical Technology Research Center, Beijing data. Consequently, the primary obstacle encountered by the
100000, China

123
140 Page 2 of 21 Complex & Intelligent Systems (2025) 11:140

fusion model is the integration of these two models while moment may show two different degradation trends, increas-
mitigating their deficiencies [14]. ing or decreasing, and this degradation sequence uncertainty
The stochastic process model, in contrast to the afore- must be accounted for in the modeling process. Finally, for
mentioned approaches, on the one hand, does not necessitate the degradation data collection system, we have an example
comprehension of the degradation mechanism and, on the of solenoid valve degradation experiment, and the collec-
other, has less stringent data requirements than the data- tion device is as shown in Fig. 2. The data acquisition system
driven modeling approach. Simultaneously, due to its advan- employs a PXI6224 ADC acquisition board with a 16-bit
tages in physical interpretations and mathematical properties acquisition accuracy. The deterioration data acquired by the
and have been extensively used due to the efficacy of RUL acquisition system itself contains measurement errors due
estimation [48]. [26] have conducted a comprehensive study to the constraints of the acquisition precision. This uncer-
of typical stochastic process models, including inverse Gaus- tainty caused by the acquisition precision can be referred to
sian processes [3, 31], Gamma processes [15, 18], and as measurement uncertainty.
diffusion processes [1]. However, the inverse Gaussian and Inspired by real-world engineering examples and to
Gamma processes can only simulate monotonic degradation address the problems of the diffusion model itself, this paper
pathways, which severely restricts their application. Because plans to construct a generalized diffusion degradation model
it can handle monotonic and non-monotonic problems, the that incorporates three sources of uncertainty. The model
diffusion processes is commonly employed to describe indus- incorporates the three common uncertainties and extends
trial product degradation. The majority of previous research existing research models such as [11, 27, 45]. However, since
on diffusion degradation models falls into one of two cate- the generalized model is more complex than the existing
gories: linear degradation models or nonlinear degradation model [11, 27, 45], deriving the RUL probability density
models. Both instances are present in the actual process function under the condition of incorporating three sources
of apparatus degradation, as illustrated in Fig. 1. Existing of uncertainty is one of the challenges. Meanwhile, overcom-
research on diffusion degradation models, such as [11, 27, ing the strong Markovianity present in the diffusion model
45], is focused on analyzing a specific object or class of with model parameter updating is also a problem to be solved.
objects; that is, linear degradation or nonlinear degradation
cases are considered separately. Furthermore, no proposal Literature review
has been made for a generalized model that can incorporate
the existing research models simultaneously. On the other Stochastic process models can be classified into two distinct
hand, the diffusion degradation model exhibits strong Marko- categories based on the variation in the degradation path of
vianity, in which the degradation state of the system depends the research object: linear degradation models and nonlinear
only on the state value at the previous moment, which is degradation models. A linear diffusion model was proposed
obviously not consistent with the actual degradation process. by [13, 36, 41] as a means of estimating the time-to-failure
Long dependence characteristics are frequently observed in distribution. Nevertheless, nonlinearity is prevalent in practi-
actual degradation systems. This means that the degradation cal applications such as bearings, turbine engines, and fatigue
data is not only influenced by the previous moment state cracks. This issue is addressed in [28, 32, 47], which develop
value, but also has a connection to the historical degradation a nonlinear diffusion model from which a time-space trans-
process [29]. Enhancing the generalizability of the model formation is used to approximate the RUL PDF.
and overcoming the strong Markovianity problem constitute However, for both linear and nonlinear degenerate models
a critical area that demands additional research in the realm mentioned above, they are limited by the strong Markovian-
of diffusion models. ity of the diffusion model. Due to the Markovian nature of the
Meanwhile, taking the collected data on the degradation conventional diffusion model, the RUL PDF of the degrada-
of solenoid valves as an example, as shown in Fig. 1a, there tion process depends solely on the observed degradation data
are obvious differences in the degradation process between at the present time. The information provided in the histori-
different individuals, indicating that individual uncertainty cal data is ignored. In actuality, the degradation processes of
exists widely in the system degradation process. Assuming numerous industrial units exhibit lengthy dependence [29],
that the degradation process is the same for each individual and the historical data should be considered when investigat-
without accounting for uncertainty is clearly unreasonable ing the probability distribution of RUL. Several scholars have
and has a significant impact on RUL estimation findings. included historical data into RUL estimation by constructing
Furthermore, as seen in Fig. 1a, the degradation process of state space models. [25] proposed a linear diffusion model
the actual device is not absolutely monotonic, i.e., the degra- with KF algorithm to relax the Markovian assumption. [46]
dation route of the device exhibits apparent oscillations. The developed an adaptive nonlinear diffusion model, which was
fluctuation of the degradation path results in degradation data applied to the lithium-ion battery degradation data. [44] intro-
temporal uncertainty, i.e., the degradation process at the next duced a generalized diffusion model with recursive filtering

123
Complex & Intelligent Systems (2025) 11:140 Page 3 of 21 140

Fig. 1 Degradation path of 1

industrial equipment SV6
SV7
SV8
0.8

Degradation states
0.6

0.4

0.2

0
0 100 200 300 400 500 600 700 800 900 1000
Cycle
(a) Linear degradation paths of solenoid valve

9250

9200
Degradation states

9150

9100

9050

9000
0 50 100 150 200 250 300 350 400
Cycle
(b) Nonlinear degradation paths of aeroengine

algorithm, which is an extension of [25, 46]. In addition, Therefore, when establishing a degradation model, temporal
the recursive filtering algorithm can be used to estimate and uncertainty, measurement uncertainty, and UTU uncertainty
update the model parameters online based on the state space are three crucial elements to be considered. [12, 39, 42]
model. This is also notably unique from other diffusion mod- investigated the UTU uncertainty in the diffusion degradation
els that require estimation of parameters based on historical model and obtained the analytical solution for the RUL PDF.
data [9, 28, 38]. However, the preceding work does not com- [6, 33] considered the measurement uncertainty in the diffu-
pletely account for the uncertainty of the degraded system, in sion model and proposed real-time estimation methods based
[25, 44, 46] only unit-to-unit(UTU) uncertainty is taken into on the state space. However, [12, 39, 42] considered the UTU
account. Other influencing factors, like measurement uncer- uncertainty and ignored the measurement uncertainty. [6,
tainty, are not considered. 33] consider measurement uncertainty while ignoring UTU
As analyzed previously with the example of a real uncertainty.
degraded device, there are several variables that typically Upon reviewing the extant research literature, it becomes
influence the uncertainty of RUL estimates. Include temporal evident that while some studies have attempted to resolve
uncertainty, measurement uncertainty and UTU uncertainty. the strong Markovianity with uncertainty problem, each of

123
140 Page 4 of 21 Complex & Intelligent Systems (2025) 11:140

Fig. 2 Solenoid valve degradation test platform

the aforementioned investigations is constrained to a spe- (4) A simulation and two actual datasets are used to validate
cific model of degradation. For example, [13, 25, 33, 36, 41] the generalizability of the proposed model.
only focuses on the case with linear degeneracy. [6, 12, 28,
32, 39, 42, 46, 47] only focuses with the nonlinear degra- Paper organization
dation case. A comprehensive stochastic degradation model
that addresses three-source uncertainty and partially miti- The subsequent sections of this work are organized in the
gates Markovianity has yet to be proposed by any researcher. following manner: “Model construction” section provides
A generalized stochastic degradation model that integrates an overview of the diffusion model’s broad statement. The
multi-source uncertainty is proposed in this paper in response state space model is used to develop the analytic version of
to the aforementioned literature. the RUL distribution with three-source uncertainty in “Esti-
mation of RUL with three sources of uncertainty” section.
Key contributions “Adaptive parameter estimation” section presents a method-
ology for estimating parameters using the KF-EM-RTS
Based on the above literature review, a generalized stochastic architecture. “Case study” section includes both nonlinear
process model is proposed in this research. The major objec- and linear scenarios to verify the effectiveness of the sug-
tives of this study are to construct a generalized diffusion gested approach. “Conclusion” section offers an in-depth
model that can estimate and update the model parameters summary of the entire text.
and RUL distribution online, partially overcome the strong
Markovian nature of diffusion model, and account for three-
source uncertainty. This work can be viewed as an extension Model construction
of the work of [27, 45]. The following is a summary of the
main contributions of this paper: Due to its well-defined mathematical features, straight-
forward statistical interpretation, and capacity to model
non-monotonic degradation, the diffusion model is exten-
(1) A generalized expression of the diffusion degradation sively used in degradation modeling. The universal diffusion
model is proposed, and its special cases include the degradation model is represented by the stochastic process
existing linear model[27], nonlinear model [45], and {X (t), t ≥ 0}, where X (t) represents the degraded state at
time-varying drift coefficient and diffusion coefficient time t. The degradation process can be mathematically rep-
model [11]. resented as the sum of a constant term, a drift term, and a
(2) Three sources of uncertainty are simultaneously consid- Brownian stochastic integral.
ered.
 t  t
(3) Based on the KF framework, its parameters and RUL
distribution can be estimated and updated online. X (t) = x0 + (τ ; θ )dτ + (τ ; θ )d B(τ ) (1)
0 0

123
Complex & Intelligent Systems (2025) 11:140 Page 5 of 21 140

Fig. 3 The flowchart of proposed model

In differential notation of model: satisfied the following function.

d X (t) = (t; θ )dt + (t; θ )d B(t) (2) x

(τ ; θ ) du
= c1 (τ ) + c2 (τ )
(τ ; θ ) (τ ; θ )
where B(t) denote a standard Brownian motion(BM). (t; θ ) z
is the drift coefficient function, and (t; θ ) is the diffusion x  
d 1
coefficient function. θ and θ are parameters of the coef- − du (5)
dt (τ ; θ )
ficient function. we assume that when t = 0, X (0) = 0. z
Based on the definition of first hitting time(FHT), lifetime is
characterized as: Then there exists a time-space transformation

T = inf {t : X (t) ≥ ω|X (0) < ω} (3)   t 

x̃ = (x, t) = (k1 ) 1/2
exp − c2 (τ )dτ
t0
The probability density function of the lifespan is denoted  x
1
as f T (t). The symbol ω represents the failure threshold of · dy
z (t; θ )
the system. An expert-defined constant or variable based  t   τ 
on industry norms and knowledge[7]. Nevertheless, RUL is −(k1 ) 1/2
c1 (τ ) exp − c2 (u)du dτ + k2
commonly employed in practical applications. The RUL of t1
 
t0

 t τ
system at time tk , given the degradation state xk , is defined
t˜ = ϕ(t) = k1 exp −2 c2 (u)du dτ + k3 (6)
as follows: t2 t0

lk = inf {lk : X (tk + lk ) ≥ ω|X (tk ) < ω} (4) The degradation process {X (t), t ≥ 0} is transformed into
standard BM by (6). In (6), z is an arbitrary value in the
where lk is RUL with the PDF represented by f L k (lk ). Note {X (t), t ≥ 0}. ti ∈ [0, ∞), i = 0, 1, 2, k1 , k2 , k3 are arbi-
that the drift coefficient given in (1) is not a constant. There- trary constants with k1 > 0.
fore, deriving the analytical formulation of its related RUL
is challenging. In order to determine the PDF of RUL, the Lemma 2 [11] For the degradation process X (t), if (τ ; θ )
following Lemma is first introduced. is continuous. When {X (t), t ≥ 0} crosses the threshold ω,
the PDF of time is defined as
Lemma 1 [23] For the degradation process in (1), the drift
coefficient function is (τ ; θ ), and the diffusion coefficient   dϕ(t)
function is (τ ; θ ). If there exist functions c1 (τ ) and c2 (τ ), p X (t) (ω, t) = p B(t˜) S B (t˜), t˜ , (7)
dt

123
140 Page 6 of 21 Complex & Intelligent Systems (2025) 11:140

  
where p B(t˜) S B (t˜), t˜ is PDF of FHT. It can be expressed as l
ω − xk − 0k (τ + tk ; θ )dτ (lk + tk ; θ )
× + 2
lk 2  (lk + tk ; θ )
  0  (τ + tk ; θ )dτ
  1 S B (t˜) d S B (t˜) ⎡ ⎤
p B(t˜) S B (t˜), t˜ = √ − l
(ω − xk − 0k (τ + tk ; θ )dτ )
2
2π t˜ t˜ d t˜ × exp ⎣− ⎦ (13)
l
2 0k 2 (τ + tk ; θ )dτ
S B2 (t˜)
exp − (8)
2t˜

If we want to obtain the PDF of FHT, we need to transform

X (t) into the standard BM. So the functions c1 (τ ) and c2 (τ ) Estimation of RUL with three sources of
need be determined firstly. Without loss of generality, we set uncertainty
k1 = 1, k2 = k3 = 0, t0 = t1 = t2 = 0, z = 0, and there
exist the function c1 (τ ) and c2 (τ ) at time t, satisfying In order to incorporate these three types of uncertainty into
the degradation modeling, the drift coefficient function is
x
(τ ; θ ) du denoted as (t; θ ) = a(t; b), where the parameter a is
c1 (τ ) = − c2 (τ ) used as the random effect to represent individual differences
(τ ; θ ) (τ ; θ )
0 and the parameter b is a fixed value [45]. The degradation
x   state X (t) and the proportional parameter a are jointly taken
d 1
+ du to be the hidden states. According to (1), degradation pro-
dt (τ ; θ )
0 cess of system containing three-sources uncertainty can be
 (τ ; θ ) ln((0; θ )) reconstructed as state space models.
c2 (τ ) = − − (9)
(τ ; θ ) t  
tk tk
xk = xk−1 + ak−1 (τ ; b)dτ + (τ ; θ )d B(τ )
which make (5) hold. Thus, the time-space transformation tk−1 tk−1
can be specified as ak = ak−1 +ζ
 t yk = xk + ε (14)
x̃ = (x, t) = x(t) − (τ ; θ )dτ
0
 t (10) where a0 ∼ N (μa , σa2 ). ε ∼ N (0, σε2 ), ζ ∼ N (0, σζ2 ), υk =
t˜ = ϕ(t) = 2 (τ ; θ )dτ tk
0 tk−1 (τ ; θ )d B(τ ) are i.i.d. nosie sequences. To apply KF,
(14) can be further rearranged into matrix form.
Correspondingly, the threshold ω is transformed into
z k = Fk z k−1 + ηk
 t (15)
S B (t˜) = (ω, ϕ −1 (t˜))=ω − (τ ; θ )dτ (11) yk = H z k + εk
0
    tk
According to Lemma 2, the PDF of lifetime is derived as xk υ 1 (τ ; b)dτ
where z k = , ηk = k , Fk = tk−1 ,
(12). Similarly, when t = tk and X (tk ) = xk , the PDF of ak ζ 0 1
RUL is derived as (13).  T
1
H = . Because the state noise is independent of each
0
f T |θ ,θ (t|θ , θ ) other, ηk obeys a two-dimensional normal distribution, i.e.
2 (t; θ ) ηk ∼ N (0,  Q k ). the 
covariance matrix of ηk can be expressed
∼
= χk 0
2π 0t 2 (τ ; θ )dτ as Q k = .
 0 σζ2
ω − 0t (τ ; θ )dτ (t; θ ) Given sequences y1:k = (y1 , y2 , . . . , yk ) of measure-
t 2 + 2
0  (τ ; θ )dτ
 (t; θ ) ments available up to the current time tk , the variance and
⎡ ⎤ expectation of the hidden states z k are defined as ẑ k|k and
2
(ω − 0t (τ ; θ )dτ )
× exp ⎣− ⎦ (12) Pk|k . Similarly, the one-step predicted expectation and vari-
2 0t 2 (τ ; θ )dτ ance of the hidden state z k can be defined as ẑ k|k−1 and Pk|k−1 .
f L k |xk ,θ ,θ (lk |xk , θ , θ ) Thus, based on the KF algorithm, the hidden state can be
updated. The initial values of  the hidden
 state are assumed
∼ 2 (lk + tk ; θ )  
= 0 0
l as ẑ 0|0 = μa and P0|0 = .
2π 0k 2 (τ + tk ; θ )dτ 0 σa2

123
Complex & Intelligent Systems (2025) 11:140 Page 7 of 21 140

Due to the linear Gaussian property of the KF algorithm, where,

the hidden states z k still obey two-dimensional Gaussian
distribution with z k ∼ N (ẑ k|k , Pk|k ). According to the ω1 = 2 (lk + tk ; θ )ω
characteristics of two-dimensional Gaussian distribution, the  lk
following expressions can be derived. − ak  (lk + tk ; θ )
2
(τ + tk ; b)dτ
0
 lk
ak |y1:k ∼ N (âk|k , ϑa,k
2
), xk |y1:k ∼ N (x̂k|k , + ak (lk + tk ; b) 2 (τ + tk ; θ )dτ
(16) 0
ϑx,k
2
), xk |ak , y1:k ∼ N (η̂xk |a,k , σx2k |a,k )  lk
(21)
ω2 = ω − ak (τ + tk ; b)dτ ,
where 0
A1 = 2 (lk + tk ; θ ), B1 = 1
ϑx,k  lk
η̂xk |a,k = x̂k|k + ξk (ak − âk|k ),
ϑa,k C1 = 2 (τ + tk ; θ )dτ
(17) 0
ϑxa,k
2
σx2k |a,k = ϑx,k
2
(1 − ξk2 ), ξk =
ϑx,k ϑa,k Due to ak |y1:k ∼ N (âk|k , ϑa,k 2 ), the expressions for

f L k |y1:k (lk |y1:k ) can be derived from (20) similarly. The

The hidden state xk and ak follow the Gaussian distribu- derivation is shown in (22).
tion. To calculate f L k |y1:k (lk |y1:k ), the Lemma 3 is introduced.
f L k |y1:k (lk |y1:k )
 
Lemma 3 If Z ∼ N (μ, σ 2 ), and ω1 , ω2 , A, B ∈ R, C ∈ = E ak |y1:k f L k |ak ,y1:k (lk |ak , y1:k )
 
R + . The following formula can be obtained. ω3 (B22 ϑa,k2 + C ) − A (B ω ϑ 2 + â C )
2 2 2 4 a,k k|k 2
=   3 (22)
(ω2 − B Z )2 C2 2π B22 ϑa,k 2 +C
E Z (ω1 − AZ ) exp(− ) 2
2C 
 (ω4 − B2 âk|k )2
C Bω2 σ 2 +μC exp −
= (ω1 − A ) (18) 2(B22 ϑa,k2 +C )
2
B2σ 2 +C B2σ 2 + C

(ω2 − Bμ)2 where,
× exp −
2(B 2 σ 2 + C)
ω3 = 2 (lk + tk ; θ )ω − x̂k|k 2 (lk + tk ; θ )
By the law of total probability, ϑx,k 2
+âk|k ξk  (lk + tk ; θ )
ϑa,k
f L k |y1:k (lk |y1:k ) ϑx,k
 +∞ ω4 = ω − x̂k|k + âk|k ξk
= f L k |z k, y1:k (lk |z k , y1:k ) pz k |y1:k (z k |y1:k )dz k ϑa,k
−∞ ϑx,k 2
   A 2 = ξk  (lk + tk ; θ )
= E ak |y1:k E xk |ak ,y1:k f L k |ak ,xk ,y1:k (lk |ak , xk , y1:k ) ϑλ,k
 lk
(19)
+2 (lk + tk ; θ ) (τ + tk ; b)dτ
0
According to the Lemma 3, xk |ak , y1:k ∼ N (ηxk |a,k , σx2k |a,k ), −σx2k |a,k (lk + tk ; b)
 
and the E xk |ak ,y1:k f L k |ak ,xk ,y1:k (lk |ak , xk , y1:k ) is derived  lk
from (13). The derivation is shown in (20). −(lk + tk ; b) 2 (τ + tk ; θ )dτ
0
  
E xk |ak ,y1:k f L k |ak ,xk ,y1:k (lk |ak , xk , y1:k ) lk ϑx,k
  B2 = (τ + tk ; b)dτ + ξk , C2
= E xk |ak ,y1:k f L k |ak ,xk (lk |ak , xk ) 0 ϑλ,k
 lk
=  
1
3 = σx2k |a,k + 2 (τ + tk ; θ )dτ (23)
lk
2π 0 2 (τ +tk ;θ )dτ 0
  
C1 (ω2 −B1 ηx |a,k )2 The x̂k|k and parameters in f L k |y1:k (lk |y1:k ) can be updated
B12 σx2 +C1
exp − 2(B 2 σ 2 k +C )
k |a,k 1 xk |a,k 1
by the KF algorithm when a new measurement is obtained.
 
B1 ω2 σx2 |a,k +ηxk |a,k C1 Specifically, if the drift coefficient is assumed as linear,
× ω1 − A1 B2σ 2
k
+C
(20) l
i.e., a 0k (τ + tk ; b)dτ = alk . The diffusion coefficient is
1 xk |a,k 1

123
140 Page 8 of 21 Complex & Intelligent Systems (2025) 11:140

lk
constant, i.e. 0 (τ + tk ; θ )d B(τ ) = σ B B(lk ). We then Table 1 Parameter setting of simulation
obtain
Parameter μa b σB σε σa σζ T
Value 0.50 0.03 0.80 0.002 0.002 0.001 200
f L k |y1:k (lk |y1:k )
 
ω5 (B32 ϑa,k2 + C ) − A (B ω ϑ 2 + â C )
3 3 3 6 a,k k|k 3
=    data. However, the KF-based adaptive state model includes
3
C3 2π B32 ϑa,k 2 +C
3 hidden variables. Therefore, the MLE is inadequate for
 resolving this issue. This work utilizes the EM algorithm
(ω6 − B3 âk|k )2 to perform adaptive parameter estimation. To begin with, the
× exp − (24)
2(B32 ϑa,k
2 +C )
3 joint likelihood function L k (ψ) may be generated using the
  measurement dataset y1:k and state dataset z 0:k .
ϑx,k
where ω5 = ω − x̂k|k + âk|k ξk ϑλ,k σ B2 , ω6 = ω − x̂k|k + ⎛ ⎞
ϑx,k ϑx,k 2 ϑx,k 
k
âk|k ξk ϑλ,k , A3 = ξk ϑλ,k σ B − σx2k |a,k , B3 = lk + ξk ϑλ,k , and L k (ψ) = ln ⎝ p(y j |z j , ψ)⎠ + ln p(z 0 |ψ)
C3 = σxk |a,k + σ B lk , which is the Eq. (52) in [27].
2 2 j=1
⎛ ⎞ (26)
If the drift coefficient is a nonlinear. The diffusin coef-
l 
k
ficient is constant, i.e. 0k (τ + tk ; θ )d B(τ ) = σ B B(lk ). + ln ⎝ p(z j |z j−1 , ψ)⎠
We then obtain j=1

f L k |y1:k (lk |y1:k ) The conditional probability distribution in Eq. (27) can be
 
expressed using the state function and measurement function
ω5 (B42 ϑa,k2 + C ) − A (B ω ϑ 2 + â C )
3 4 4 6 a,k k|k 3
=   from Eq. (15).
 3
C3 2π B42 ϑa,k 2 +C
3
   
0 0 0
 p(z 0 |ψ) ∼ N (ẑ 0|0 , P0|0 ) = N ,
μa 0 σa2
(ω6 − B4 âk|k )2
× exp − (25) p(z j |z j−1 , ψ) ∼ N (F j z j−1 , Q j )
2(B42 ϑa,k
2 +C )
3
p(= y j |z j , ψ) ∼ N (H j z j , σε2 ) (27)
ϑx,k 2 l
where A4 = ξk ϑλ,k σB + σ B2 0k
Using the idea of EM, our first step is to compute the con-
(τ + tk ; b)dτ − σx2k |a,k (lk + tk ; b) − (lk + tk ; b)σ B2 lk , ditional expectation of the joint likelihood function during
l ϑ
and B4 = 0k (τ + tk ; b)dτ + ξk ϑλ,k x,k
, which is the Eq. (11) the E-step:
derived in [17].
In summary, (22) is a more general formula that includes L k (ψ|ψ̂k(l) ) = E z (l) [L k (ψ)] (28)
0:k |y1:k ,ψ̂k
linear degradation processes with constant diffusion coeffi-
cients, age-dependent nonlinear degradation processes with The expansion is as (29):
constant diffusion coefficients, and age-dependent nonlinear
degradation processes with time-varying diffusion coeffi- (l)
L k (ψ|ψ̂k ) = E z |y ,ψ̂ (l) [L k (ψ)]
0:k 1:k k

k  2,(l)  1  (l)  k  (l) 

cients discussed in [11, 17, 27]. This is important since
exceptional cases are discussed in [11, 17, 27], and our = − ln σε,k  − ln P0|0,k  − ln Q k 
2 2 2
model would have to return to those cases in order to obtain ⎡ 2
⎤
1 k

accurate results. The derivations of the aforementioned equa- − E (l) ⎣ yj − Hj z j ⎦
2,(l) z 0:k |y1:k ,ψ̂k
tions involve several unknown parameters, represented as 2σε,k j=1
 !"
ψ=(z 0|0 , P0|0 , σε2 , σζ2 , b, θ ). These parameters need to be 1 (l) −1 (l) T (l)
− T r (P0|0,k ) E z |y ,ψ̂ (l) (z 0 − ẑ 0|0,k ) (z 0 − ẑ 0|0,k )
obtained before starting the KF algorithms for state pre- 2 0:k 1:k k

diction. The following section utilizes the EM algorithm to 1  k

(l) −1
− Tr (Q k ) E z |y ,ψ̂ (l)
estimate the value of ψ. 2
j=1
0:k 1:k k

!
(z j − F j z j−1 )T (z j − F j z j−1 ) (29)
Adaptive parameter estimation
In (29), Ez (l) (z j ), E z (l) (z j z j−1 ),
Typically, model parameters are determined using maximum 0:k |y1:k ,ψ̂k 0:k |y1:k ,ψ̂k

likelihood estimation (MLE) and historical deterioration Ez (l) (z 2j ) are the conditional expectation of state vari-
0:k |y1:k ,ψ̂k

123
Complex & Intelligent Systems (2025) 11:140 Page 9 of 21 140

ables. The Rauch-Tung-Striebel(RTS) smoothing is used for k  (l) 1  (l) 

∝ − ln σε2  − ln P0|0,k 
calculation[25]. Based on the RTS algorithm, the conditional 2 k 2
expectation of state variables can be calculated. k  (l)  1
− ln Q k  − (l)
2 2(σε2 )k
Ez (l) (z j ) = z̃ j|k ⎡ ⎤
0:k |y1:k ,ψ̂k k  
Ez (z j z Tj ) = z̃ j|k z̃ Tj|k + P̃ j|k ⎣ y 2j − 2y j H j z̃ j|k + H j (z̃ j|k z̃ Tj|k + P̃ j|k )H jT ⎦
(l)
0:k |y1:k ,ψ̂k
(30) j=1
Ez (l) (z j z Tj−1 ) = z̃ j|k z̃ Tj−1|k + M j|k #   T
0:k |y1:k ,ψ̂k 1 (l) −1 (l)
− T r (P0|0,k ) (z̃ 0|k z̃ 0|k
T
+ P̃0|k − ẑ 0|0,k z̃ 0|k
Ez (l) (z j−1 z Tj ) = z̃ j−1|k z̃ Tj|k + M j|k 2
$
0:k |y1:k ,ψ̂k     T
(l) (l) (l)
−z̃ 0|k
T
ẑ 0|0,k + ẑ 0|0,k ẑ 0|0,k )
By substituting the conditional expectation from (30) into
(29), we may derive (31).
1  (l) −1
k
− Tr (Q k ) z̃ j|k z̃ Tj|k + P̃ j|k
(l) 2
L k (ψ|ψ̂k ) = E z (l) [L k (ψ)] j=1
0:k |y1:k ,ψ̂k

Fig. 4 The degradation path of 250

simulation

200
Degradation state

150

100

50

0
0 20 40 60 80 100 120 140 160 180 200
Time

Fig. 5 Adaptive parameter 0.6 0.1 estimated

estimation process for Ture parameter estimated 1 B
a Ture parameter
simulation data
Value

0.4
Value

Value

0.05 0.5
0.2
estimated a Ture parameter
0 0 0
0 100 200 0 100 200 0 100 200
Cycle Cycle Cycle
0.3 0.03 0.4
estimated estimated b
Value

0.2
Value

Value

0.02
estimated 0.2
0.1 Ture parameter
0.01
Ture parameter
Ture parameter
0 0 0
0 100 200 0 100 200 0 100 200
Cycle Cycle Cycle

123
140 Page 10 of 21 Complex & Intelligent Systems (2025) 11:140

0.05 0.05 0.05 −F j (z̃ j−1|k z̃ Tj|k + M j|k ) − F jT (z̃ j|k z̃ Tj−1|k
0 0 0 !
0 100 200
0.1
0 100 200 0 100 200
+M j|k ) + F j (z̃ j−1|k z̃ Tj−1|k + P̃ j−1|k )F jT (31)
0.05 0.1
PDF of RUL

0.05 0.05
0 0 0
0 100 200 0 100 200 0 100 200 floatbelowskip20pt
0.2 0.2
1

The RUL PDF of NL1

0.8

0.6
The M-step can be
 estimated "based on the E-step, i.e.,
0.1 The RUL PDF of NL3
ψ̂k(l+1) = arg max L k (ψ|ψ̂k(l) ) . Optimize compute the
0.1 0.4

0.2

0 0 The ground-truth of RUL

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0 100 200 0 100 200 The RUL of NL1

ψ
0.4 The RUL of NL3
0.2
0.5 ±20% confidence interval likelihood function L k (ψ|ψ̂k(l) ) and obtain the estimate of
0
0 50 100 150
0
0 50 100 150 the unknown parameter vector ψ̂k(l+1) . The results can be
Cycle obtained.
(a) PDF of RUL estimation
 (l+1)
(l+1) (l+1)
100
z 0|0,k = z̃ 0|k , P0|0,k = P̃0|k σε2
±20% confidence interval k

1  2 
NL1 RUL predicted value k
80
NL3 RUL predicted value
= y j − 2y j H j z̃ j|k + H j (z̃ j|k z̃ Tj|k + P̃ j|k )H jT
RUL ground truth k
j=1

60 11 (b(l+1) ) r 12 (b(l+1) )
r(l+1) k (l+1) k
RUL

21 (b(l+1) ) r 22 (b(l+1) )
r(l+1) k (l+1) k
40
z̃ j|k z̃ Tj|k + P̃ j|k − F j (z̃ j−1|k z̃ Tj|k + M j|k )
=
20 −F jT (z̃ j|k z̃ Tj−1|k + M j|k ) + F j (z̃ j−1|k z̃ Tj−1|k + P̃ j−1|k )F jT

1  11
k
(l+1) (l+1)
0 χk (θ ) = r(l+1) (bk )
120 130 140 150 160 170 180 190 200 k
j=1
Cycle
(b) RUL of two models  (l+1) 1  22
k
(l+1)
σζ2 = r(l+1) (bk ) (32)
k k
j=1
20
AE of NL1
AE of NL3

15
It is worth noting that bk(l+1) cannot be obtained directly.
The estimation of bk(l+1) can be achieved via maximizing
L k (ψ|ψ̂k(l) ) through "fminsearch" function in MATLAB.
AE

10  (l+1)
(l+1) (l+1)
However, if bk = 1, χk (θ ) = σ B2 , then σζ2
k
5 can be obtained directly. The PDF of RUL can be updated
based on the estimated parameters and KF framework.
0
The relatively simple nature of the KF and EM algorithms
120 130 140 150 160 170 180 190 200 results in a low computational complexity and high timeli-
Cycle
(c) AE of two models ness for the KF-EM-RTS. However, the corresponding KF &
EM algorithms are more demanding in terms of initial value
1
NL1
settings. If the initial parameters are not set appropriately, the
NL3 algorithms may not converge. Nevertheless, the existence of
0.8 historical datasets in the process of device degradation allows
for the iterative learning of model parameters based on these
0.6
data, which can then be set as the initial value of the algo-
CRA

rithm. Therefore, the impact of parameter settings on the

0.4
actual process is limited.
0.2

0
Case study
Simulation data

(d) CRA of two models In this section, linear and nonlinear degradation models, both
of which are specific examples of the proposed model, are
Fig. 6 Comparison of different evaluation metrics for RUL estimation. examined independently to validate the effectiveness of pro-
NL1 denotes the model proposed in this paper and NL3 denotes the
model of [24] posed model. (1) The simulation and C-MAPSS dataset are

123
Complex & Intelligent Systems (2025) 11:140 Page 11 of 21 140

Fig. 7 The degradation paths of 9250

aircraft engine

9200

Degradation states
9150

9100

9050

9000
0 50 100 150 200 250 300 350 400
Cycle

Fig. 8 Estimation and actual 9250

degradation path of FD01 The actual degradation path
The eatimated degradation path
Degradation states

9200

9150

9100

9050
0 20 40 60 80 100 120 140 160 180 200
Cycle

analyzed for the nonlinear model. (2) Two cases are inves- where RU L r eal is the RUL ground truth, and RU L pr edicted
tigated for the linear degradation model: a NASA research is the predicted RUL.
center battery dataset and an ATS Lab solenoid valve dataset.
Absolute error (AE) and cumulative relative accuracy(CRA)
are utilized to evaluate the prediction performance of RUL. Nonlinear degradation model
The specific implementation process is shown in Fig. 3.
Simulation case study

In this subsection, the parameter estimation and the accu-

AE = |RU L r eal − RU L pr edicted | racy of the RUL estimation are verified. The nonlinear drift
T   coefficient function (t; b) = bebt . First of all, we use Euler
1  AE i (33) approximation to generate D degenerate paths. Table 1 shows
CRA = 1−
T RU L r eal,i the simulation model’s true parameters:
i=1

123
140 Page 12 of 21 Complex & Intelligent Systems (2025) 11:140

Fig. 9 Adaptive parameter 0.1 20

estimation process for FD01 0.6 estimated a estimated a
estimated B

Value

Value

Value
0.4
0.05 10
0.2

0 0 0
0 100 200 0 100 200 0 100 200
Cycle Cycle Cycle
0.1 2 1
estimated 2 estimated b
estimated
Value

Value

Value
0.05 1 0.5

0 0 0
0 100 200 0 100 200 0 100 200
Cycle Cycle Cycle

xk = xk−1 + ak−1 (tk−1 ; b)(tk majority of the estimation findings for NL1 and NL3 are
− tk−1 ) + σ B (B(tk ) − B(tk−1 )) inside the confidence interval, as depicted in Fig. 6b. Nev-
(34) ertheless, the precision of NL1’s assessment surpasses that
ak = ak−1 +ζ
of NL3. In addition, the prediction results of both models
yk = xk + ε converge with the input of new data, although the pro-
posed model converges faster than NL3. Considering the
Based on the set parameters, we obtained degradation
uncertainty from three different sources might enhance the
path. It is shown in Fig. 4. The nonlinear degradation model
accuracy of estimating the RUL. To provide a more detailed
proposed by the [24] is used as a comparison to verify the effi-
comparison between NL1 and NL3, the computational results
cacy of proposed model, which only takes into account UTU
of AE and CRA can be shown in Fig. 6, specifically in panels
uncertainty. The model presented in this paper is denoted as
(c) and (d).
NL1. [24]’s model is denoted as NL3. By utilizing the KF-
Figure 6c illustrates that the proposed model has fewer
EM-RTS method stated in this research work, we are able to
AE than NL3. Moreover, the CRA of the proposed model is
acquire the parameters depicted in Fig. 5. Based on the infor-
greater than that of NL3, showing that the proposed model
mation provided in Fig. 5, it is evident that the degradation
can estimate the RUL more precisely.
parameters may be accurately learned using the KF-EM-
RTS method. As more data is collected, the estimates of the
parameters eventually converge, and the parameters can be Aircraft engine data
continuously updated as new data is received. For the given
error parameters σa , σε , σζ there are errors in the learning. To examine the impact of the degradation model on RUL
This is mainly due to their small values, which make it dif- estimation for nonlinearly degraded devices, experiments are
ficult to learn them accurately, but they do not have a large conducted using the C-MAPSS dataset in this study. Four
effect on the RUL estimation. Based on the proposed param- distinct degradation datasets are included in the C-MAPSS
eter estimation method, historical data are not required to dataset: FD001, FD002, FD003, and FD004. The FD001
acquire prior parameters in this research. dataset, comprising 100 degradation units, is exclusively
Figure 6a displays the outcomes of RUL PDF. By examin- utilized for analysis in the experiments detailed in this sub-
ing Fig. 6a, it is evident that the PDF of suggested model has section. The degradation trajectory of each unit is illustrated
a narrower distribution, indicating a smaller standard devia- in Fig. 7. Some degradation trajectories of FD001 exhibit
tion. This suggests that the suggested approach has the ability pronounced nonlinear characteristics, as illustrated by the red
to decrease the level of uncertainty in the estimation of RUL. degradation trajectories in Fig. 7. To facilitate the subsequent
Figure 6b displays the estimated results of RUL beside phase of the experiment, individuals exhibiting more pro-
the actual RUL. The blue region corresponds to a confidence nounced nonlinear degradation characteristics are chosen to
interval bounded by upper and lower limits of 20%. The conduct the subsequent experimental validation. Simultane-

123
Complex & Intelligent Systems (2025) 11:140 Page 13 of 21 140

Fig. 10 Comparison of different 0.02 0.02

0.02
evaluation metrics for B07 RUL 0.01 0.01 0.01
estimation. NL1 denotes the 0 0 0
0 100 200 0 100 200 0 100 200
model proposed in this paper, 0.02 0.02
0.04

PDF of RUL
NL2 denotes (t; b) = bt (b−1) , 0.01 0.01 0.02
NL3 denotes the model in [24], 0 0 0
0 100 200 0 100 200 0 100 200
NL4 denotes the model in [16] 0.04 The RUL PDF of NL1
0.05
0.02 The RUL PDF of NL2
0.02 0.01 The RUL PDF of NL3
0 0 0 RUL PDF of NL4
The
0 100 200 0 100 200 0 100 200
The ground-truth of RUL
0.05 0.05 The RUL of NL1
The RUL of NL2
0 0 The RUL of NL3
0 50 100 150 0 50 100 150 The RUL of NL4
Cycle ±20% confidence interval

(a) PDF of RUL estimation

200
±20% confidence interval
NL1 RUL predicted value
NL2 RUL predicted value
150 NL3 RUL predicted value
NL4 RUL predicted value
RUL ground truth

RUL
100

50

0
140 150 160 170 180 190 200 210 220
Cycle
(b) RUL of four models

120
AE of NL1
AE of NL2
100 AE of NL3
AE of NL4

80
AE

60

40

20

0
140 150 160 170 180 190 200 210 220
Cycle
(c) AE of four models
1
NL1
NL2
NL3
0.8 NL4

0.6
CRA

0.4

0.2

0
FD01 data
(d) CRA of four models

123
140 Page 14 of 21 Complex & Intelligent Systems (2025) 11:140

Fig. 11 The degradation paths 2.2

of batteries [Battery05]
[Battery06]
2 [Battery07]
[Battery18]

1.8

Capacity/A.h
1.6

1.4

1.2

1
0 20 40 60 80 100 120 140 160 180
Charge and discharge cycle

Fig. 12 Estimation and actual 0.5

degradation path of B07 The actual degradation path
The eatimated degradation path
0.4
Degradation states

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160
Cycle

ously, let (t; b) = bt (b−1) be used to represent the nonlinear parameters accurately reflect the actual degradation process.
degradation model in power exponent form, denoted NL2. Keeping the precision of parameter estimation and updating
NL4 represents the parameter estimation model that Liu [16] is a given.
has proposed as being based on the MLE. The model pro- The results of parameter estimation are depicted in Fig. 9.
posed in this paper is denoted as NL1. NL2, NL3, and NL4 It is evident that as more data is accumulated, the estimated
represent models of comparison. parameters tend to become stable. This suggests that as degra-
By applying the KF tracking algorithm described in dation progresses to a later stage, the estimation error of
“Adaptive parameter estimation” section, the estimated parameters decreases and the estimation accuracy of RUL
degradation trajectory can be obtained, as illustrated in Fig. 8. increases.
The result in Fig. 8 indicates that the proposed model can The RUL estimation results of different models are shown
more accurately monitor the actual degradation trajectory, in Fig. 10, from which it is obvious that the proposed model
ensuring that the estimation and update results for model NL1 achieves the best results. By comparing with NL2, it

123
Complex & Intelligent Systems (2025) 11:140 Page 15 of 21 140

Fig. 13 Adaptive parameter

estimation process for B07 10-3 10-6 10-4
5 2 1
estimated a 2 2
estimated a
estimated B

Value

Value

Value
1

0 0 0
0 100 0 100 0 100
Cycle Cycle Cycle
10-7 10-5
5
2 2
Value estimated estimated

Value
5

0 0
0 100 0 100
Cycle Cycle

can be found that different nonlinear degradation models Linear degradation model
have a great influence on the accuracy of RUL estima-
tion. In practical engineering, it is necessary to choose the The battery dataset
corresponding exponential or power exponential nonlinear
degradation model for the nonlinear characteristics of the The battery datasets used in this paper are from the NASA-
specific degradation data. The actual RUL probability distri- PCoE Centre. Multiple charges and discharge tests are used
bution is shown in Fig. 10a, from which it can be seen that the to determine the life cycle. and the battery is defined as failing
probability density distribution of the proposed NL1 model when its discharge capacity drops to 80% of the rated capac-
is the most concentrated, and the peak-to-peak values are sig- ity. The experiment yields four clusters of experimental data,
nificantly larger than those of the NL2, NL3, and NL4. The namely B05, B06, B07, and B18. Figure 11 is the degrada-
above results show that by incorporating the three sources tion path of the battery. The degradation process for all four
of uncertainty, the uncertainty of the obtained RUL estima- batteries are similar, so we chose B07 for the experiment.
tion results is greatly reduced. NL4 is a degradation model Similarly to [21], the battery data are transformed into
without online updating of parameters, and the comparison incremental form. The subsequent validation results are
with the online updating models NL1, NL3 also shows that based on the degraded state after the conversion. Accord-
the parameters estimated based on the historical data can not ing to the data on battery degradation, the rate of degradation
accurately describe the degradation trend of the new indi- is approximately constant. Consequently, the linear model
viduals, and it is necessary to update the parameters of the with a constant diffusion coefficient is applicable to these
model in the actual RUL estimation process. experimental results, which is a specific example of the pro-
posed model. The model presented in this paper is denoted
as L1. Correspondingly, Wang’s model [37] is used for com-
Discussions parison, which simplifies the consideration and only analyzes
the effect of UTU uncertainty on the model. Wang’s model is
The experimental analysis of the simulation dataset and denoted as NL2. Dai’s model only uses MLE for parameter
aero-engine degradation data indicates that the generalized estimation and does not consider uncertainty with realizing
stochastic degradation model proposed in this paper, which parameter updates, denoted as L3 [43].
incorporates three sources of uncertainty, can more accu- Figure 12 shows the real degradation trajectory and the
rately characterize the nonlinear degradation equipment in estimated degradation trajectory. From Fig. 12, it can be
operation. In the meantime, improved prediction outcomes found that the error between the estimated result and the
are achieved through the integration of uncertainty into the true value is very small. This demonstrates that the proposed
proposed model, thereby significantly mitigating the uncer- model is capable of precisely tracking the actual trajectory of
tainty associated with RUL estimation.

123
140 Page 16 of 21 Complex & Intelligent Systems (2025) 11:140

0.02 0.02
0.01
0.02 degradation, which aids in minimizing the error associated
0.01 0.01
0 0 0 with online parameter updates.
0 50 100 0 50 100 0 50 100
0.04 0.04 0.05 From Fig. 13, it can be found that the estimated parameters
PDF of RUL

0.02 0.02 of the model converge rapidly with the input of the battery
0 0 0
0 50 100 0 50 100 0 50 100 data. In addition, it is not difficult to discover that when the
0.02 0.05 The RUL PDF of L1
0.01 The RUL PDF of L2
B07 degradation state gives rise to unavoidable fluctuations,
0 0
The RUL PDF of L3
The ground-truth of RUL
this can be well captured by parameters such as μa , σ B2 , and
0 50 100 0 50 100
0.05
The RUL of L1 σε2 . This is also a sufficient indication that the proposed degra-
0.05 The RUL of L2
The RUL of L3 dation model accurately tracks the degradation trajectory and
0 0 ±20% confidence interval
0 50 100 150 0 50 100 150 is reflected in the parameter estimation process. Furthermore,
Cycle as observational data inputed, the
(a) PDF of RUL estimation
  unknown parameter vec-
tors ψ= z 0|0 , P0|0 , σε , σζ , σ B can be updated.
2 2 2

80 The PDFs of RUL are shown in Fig. 14a. The figure illus-
±20% confidence interval
70 L1 RUL predicted value trates that the estimation results obtained by the proposed
L2 RUL predicted value
60 L3 RUL predicted value model outperform those of L2 and L3. On the other hand,
RUL ground truth
the peak value of the L1 model is considerably greater than
50
those of the L2 and L3 models as a result of the compre-
RUL

40
hensive integration of the three sources of uncertainty. This
30
indicates that the uncertainty of the RUL estimation results is
20 diminished and the distribution becomes more concentrated
10 when the effect of uncertainty is accounted for.
0 The RUL estimation results are compared in Fig. 14b. The
60 70 80 90 100 110 120
Cycle RUL estimated by both models follows the same trend, how-
(b) RUL of three models ever the estimation precision of the proposed model is better
than that of L2 and L3. The RUL estimation fluctuates due
30 to fluctuating degraded data. However, when degraded data
AE of L1
AE of L2 accumulates, parameter estimations converge to ensure RUL
25 AE of L3
estimation precision. An alternate perspective is to evaluate
20 the RUL estimation precision of the two models using per-
formance indicators like AE and CRA. The AE is shown in
AE

15
Fig. 14c, from which it can be seen that the proposed model
10 outperforms L2 and L3, particularly in the case of long-term
prediction. The CRA quantifies the accuracy of the estima-
5
tion model. As the value of CRA increases, the precision of
0 the estimated results also increases. The proposed model can
60 70 80 90 100 110 120
Cycle
be found to obtain higher CRA values on B07 than L2 and
(c) AE of three models L3.

1
L1
L2 Solenoid valve dataset
L3
0.8

The second dataset used in this article is solenoid valve(SV)

0.6 dataset from ATS Lab, which is tested at room temperature
CRA

[35]. The Euclidean distance between the dynamic current

0.4
waveform and the template one is choosed as the health indi-
cator. Figure 15 shows the SV degradation process. At the
0.2
end of Fig. 15 there is a sudden severe rise, indicating that
0
a failure of the SV has occurred and the current is changing
Battery 07
severely. 0.7–0.9 is the distance value around the sharp rise.
(d) CRA of three models As a result, the failure threshold is set at 0.8 [34]. The data
volume of SV6 is 3.6207 ∗ 106 , and the The data volume of
Fig. 14 Comparison of different evaluation metrics for B07 RUL esti- SV8 is 2.3795 ∗ 106 . To reduce the calculation burden, we
mation. L1 denotes the model proposed in this paper, L2 denotes the
model in [37], and L3 denotes the model in [43] resampled the data. The resampling data is shown in Fig. 15.

123
Complex & Intelligent Systems (2025) 11:140 Page 17 of 21 140

Fig. 15 Degradation path of 1

solenoid valve [SV6]

Sudden Rise
0.8

d(Ik,Itemp)
0.6

0.4

0.2

0
0 500 1000 1500
Cycle
(a) SV6 degradation paths

0.9 [SV8]

0.8 Sudden Rise

0.7
d(Ik,Itemp)

0.6

0.5

0.4

0 100 200 300 400 500 600 700 800 900

Cycle
(b) SV8 degradation paths

The degradation trends of solenoid valves are similar. So the converges to the actual RUL over time, whereas proposed
SV8 is chosen for the subsequent verification. model has a faster convergence rate and greater precision.
Figure 16 shows that the proposed method can success- The results in Fig. 18c, d show AE and CRA, from which
fully estimate the degradation trajectory. The estimated we can see that L1 has smaller AE and higher CRA, which
parameters for proposed model are shown in Fig. 17. All the indicates that the degenerate model that considers the three-
model parameters are able to converge as the data accumu- source uncertainty helps to improve the RUL estimation
late. accuracy.
The probability density distribution of the RUL estimate
is shown in Fig. 18a, from which it can be simultaneously Discussions
found that the distribution of the proposed model L1 is more
centralized and has the highest peak-to-peak value. Also the As demonstrated by the comparison above, the method sug-
estimated value of RUL for L1 is closer to the true RUL. gested in this article has the potential to improve the precision
The RUL estimation results of SV8 are shown in Fig. 18b, of RUL estimation. By integrating the three sources of uncer-
and the RUL estimation results have some deviation from the tainty, the uncertainty associated with the RUL estimation
actual RUL in the initial stage. The predicted RUL gradually outcomes decreases, leading to a more centralized probability

123
140 Page 18 of 21 Complex & Intelligent Systems (2025) 11:140

Fig. 16 Estimation and actual 0.9

degradation path of SV8 The actual degradation path
The eatimated degradation path
0.8

Degradation states
0.7

0.6

0.5

0.4

0.3
0 100 200 300 400 500 600 700 800 900
Cycle

Fig. 17 The process of

parameter estimation for SV8 10-3 10-7 10-6
1 5 estimated 2 5 estimated 2
a B
Value

Value

Value
0
-1
-2 estimated a
0 0
0 500 0 500 0 500
Cycle Cycle Cycle
-7 -6
10 10
2 2
estimated 1 estimated
1
Value

Value

0.5 0.5

0 0
0 500 0 500
Cycle Cycle

distribution of the obtained RUL with greater peak-to-peak Conclusion

values. Moreover, this assists the PHM in making ulti-
mate maintenance decisions and decreases the maintenance This study constructed a general stochastic model with three
expenses. In the meanwhile, a comparative analysis with sources of uncertainty to implement RUL prediction for
other degenerate models demonstrates that the three-source degraded systems. The existing model [11, 17, 27] can be
uncertainty should be accounted for when constructing the considered as its particular case. Based on Lemma 1 and
degenerate model. Lemma 2, the degenerate model is transformed into a stan-
dard BM, and the approximate analytic form of the RUL
probability density function is derived based on the concept
of FHT. The unknown parameters are estimated and updated
by KF-EM-RTS. To validate the proposed model, nonlinear
simulation data and two actual data sets are utilized. The

123
Complex & Intelligent Systems (2025) 11:140 Page 19 of 21 140

Fig. 18 Comparison of different 10-3 10-3

5 5 0.01
evaluation metrics for SV8 RUL 0.005
estimation 0 0 0
0 500 1000 0 500 1000 0 500 1000
-3 -3
10 10
0.01

PDF of RUL
0.005 5 5
0 0 0
0 500 1000 0 500 1000 0 500 1000

0.01 0.01 The RUL PDF of L1

0.005 0.005 The RUL PDF of L2
0 0 The RUL PDF of L3
0 500 1000 0 500 1000 The ground-truth of RUL
The RUL of L1
0.02 The RUL of L2
0.01 0.02
The RUL of L3
0 0
0 50 100 150 0 50 100 150 ±20% confidence interval
Cycle
(a) PDF of RUL estimation

800
±20% confidence interval
700 L1 RUL predicted value
L2 RUL predicted value
600 L3 RUL predicted value
RUL ground truth
500

RUL
400

300

200

100

0
400 500 600 700 800 900 1000
Cycle
(b) RUL of three models

400
AE of L1
350 AE of L2
AE of L3
300

250
AE

200

150

100

50

0
400 500 600 700 800 900 1000
Cycle
(c) AE of three models

1
L1
L2
L3
0.8

0.6
CRA

0.4

0.2

0
SV8

(d) CRA of three models

123
140 Page 20 of 21 Complex & Intelligent Systems (2025) 11:140

experimental results demonstrate that the proposed model 3. Chen X, Sun X, Si X, Li G (2019) Remaining useful life prediction
provides more precise RUL estimation. based on an adaptive inverse gaussian degradation process with
measurement errors. IEEE Access 8:3498–3510
The existing work still has some limitations that need fur- 4. Daigle MJ, Goebel K (2012) Model-based prognostics with concur-
ther study: rent damage progression processes. IEEE Trans Syst Man Cybern
Syst 43(3):535–546
5. Duan S, Song W, Zio E, Cattani C, Li M (2021) Product technical
(1) As in previous studies, this paper assumes that the mea- life prediction based on multi-modes and fractional lévy stable
surement noise obeys a normal distribution and the motion. Mech Syst Signal Process 161:107974
measurement function is linear. This assumption may be 6. Feng L, Wang H, Si X, Zou H (2013) A state-space-based prog-
improper for some real-world applications. nostic model for hidden and age-dependent nonlinear degradation
process. IEEE Trans Autom Sci Eng 10(4):1072–1086
(2) A general stochastic model is proposed in this paper. 7. Gao H, Cui L, Kong D (2018) Reliability analysis for a wiener
However, only the case of linear drift coefficients with degradation process model under changing failure thresholds.
constant diffusion coefficients is considered in practical Reliab Eng Syst Saf 171:1–8
applications according to the characteristics of the actual 8. Heng A, Zhang S, Tan AC, Mathew J (2009) Rotating machinery
prognostics: state of the art, challenges and opportunities. Mech
data. The case of nonlinear and time-varying diffusion Syst Signal Process 23(3):724–739
coefficients needs to be verified in more detail. These 9. Huang T, Zhao Y, Coit DW, Tang LC (2021) Reliability assess-
issues will be explored in our following study. ment and lifetime prediction of degradation processes considering
recoverable shock damages. IISE Trans 53(5):614–628
10. Lei Y, Li N, Guo L, Li N, Yan T, Lin J (2018) Machinery health
prognostics: a systematic review from data acquisition to rul pre-
Funding Haizhen Zhu: National Natural Science Foundation of China, diction. Mech Syst Signal Process 104:799–834
Grant no. 72301293. Xilang Tang: National Natural Science Foundation 11. Li N, Lei Y, Guo L, Yan T, Lin J (2017) Remaining useful life pre-
of China, Grant no. 72201276. diction based on a general expression of stochastic process models.
IEEE Trans Ind Electron 64(7):5709–5718
Data availability Both the aero engine data and the battery data are pub- 12. Li N, Lei Y, Yan T, Li N, Han T (2019) A wiener-process-model-
lic data sets and are freely available for download online. The solenoid based method for remaining useful life prediction considering unit-
valve data is the experimental data set of this research group, and the to-unit variability. IEEE Trans Ind Electron 66(3):2092–2101
corresponding author can be contacted if necessary. 13. Li T, Pei H, Pang Z, Si X, Zheng J (2020) A sequential
Bayesian updated wiener process model for remaining useful life
Declarations prediction. IEEE Access 8:5471–5480. [Link]
ACCESS.2019.2962502
14. Liao L, Koettig F (2014) Review of hybrid prognostics approaches
Conflict of interest The authors declare that they have no known com-
for remaining useful life prediction of engineered systems, and an
peting financial interests or personal relationships that could have
application to battery life prediction. IEEE Trans Reliab 63(1):191–
appeared to influence the work reported in this paper.
207
15. Lin CP, Ling MH, Cabrera J, Yang F, Yu DYW, Tsui KL (2021)
Open Access This article is licensed under a Creative Commons Prognostics for lithium-ion batteries using a two-phase gamma
Attribution-NonCommercial-NoDerivatives 4.0 International License, degradation process model. Reliab Eng Syst Saf 214:107797
which permits any non-commercial use, sharing, distribution and repro- 16. Liu M, Dong Z, Shi H (2023) Multi-sensor information fusion and
duction in any medium or format, as long as you give appropriate credit multi-model fusion-based remaining useful life prediction of fan
to the original author(s) and the source, provide a link to the Creative slewing bearings with the nonlinear wiener process. Sustainability
Commons licence, and indicate if you modified the licensed mate- 15(15):12010. [Link]
rial. You do not have permission under this licence to share adapted 17. Liu S, Fan L (2022) An adaptive prediction approach for rolling
material derived from this article or parts of it. The images or other bearing remaining useful life based on multistage model with three-
third party material in this article are included in the article’s Creative source variability. Reliab Eng Syst Saf 218:108182
Commons licence, unless indicated otherwise in a credit line to the 18. Liu X, Matias J, Jäschke J, Vatn J (2022) Gibbs sampler for noisy
material. If material is not included in the article’s Creative Commons transformed gamma process: inference and remaining useful life
licence and your intended use is not permitted by statutory regula- estimation. Reliab Eng Syst Saf 217:108084
tion or exceeds the permitted use, you will need to obtain permission 19. Lui YH, Li M, Downey A, Shen S, Nemani VP, Ye H, VanElzen C,
directly from the copyright holder. To view a copy of this licence, visit Jain G, Hu S, Laflamme S, Hu C (2021) Physics-based prognostics
[Link] of implantable-grade lithium-ion battery for remaining useful life
prediction. J Power Sources 485:229327
20. de Oliveira da Costa PR, Akçay A, Zhang Y, Kaymak U (2020)
Remaining useful lifetime prediction via deep domain adaptation.
References Reliab Eng Syst Saf 195:106682
21. Pang Z, Si X, Hu C, Zhang Z (2022) An age-dependent and
1. Cai B, Fan H, Shao X, Liu Y, Liu G, Liu Z, Ji R (2021) Remaining state-dependent adaptive prognostic approach for hidden nonlin-
useful life re-prediction methodology based on wiener process: ear degrading system. IEEE/CAA J Automat Sin 9(5):907–921
Subsea Christmas tree system as a case study. Comput Ind Eng 22. Ragab M, Chen Z, Wu M, Foo CS, Kwoh CK, Yan R, Li X (2020)
151:106983 Contrastive adversarial domain adaptation for machine remaining
2. Chen L, Ding Y, Liu B, Wu S, Wang Y, Pan H (2022) Remaining useful life prediction. IEEE Trans Ind Inform 17(8):5239–5249
useful life prediction of lithium-ion battery using a novel particle 23. Ricciardi LM (1976) On the transformation of diffusion processes
filter framework with grey neural network. Energy 244:122581 into the wiener process. J Math Anal Appl 54(1):185–199

123
Complex & Intelligent Systems (2025) 11:140 Page 21 of 21 140

24. Si XS (2015) An adaptive prognostic approach via nonlinear 39. Wang P, Gao RX, Woyczynski WA (2020) Levy process-based
degradation modeling: application to battery data. IEEE Trans Ind stochastic modeling for machine performance degradation prog-
Electron 62(8):5082–5096 nosis. IEEE Trans Ind Electron 68(12):12760–12770
25. Si XS, Wang W, Hu CH, Chen MY, Zhou DH (2013) A wiener- 40. Wang S, Jin S, Bai D, Fan Y, Shi H, Fernandez C (2021) A critical
process-based degradation model with a recursive filter algorithm review of improved deep learning methods for the remaining useful
for remaining useful life estimation. Mech Syst Signal Process life prediction of lithium-ion batteries. Energy Reports 7:5562–
35(1):219–237 5574
26. Si XS, Wang W, Hu CH, Zhou DH (2011) Remaining useful life 41. Xie G, Li X, Peng X, Qian F, Hei X (2019) Estimating the probabil-
estimation - a review on the statistical data driven approaches. Eur ity density function of remaining useful life for wiener degradation
J Oper Res 213(1):1–14 process with uncertain parameters. Int J Control Autom Syst
27. Si XS, Wang W, Hu CH, Zhou DH (2014) Estimating remaining 17(11):2734–2745
useful life with three-source variability in degradation modeling. 42. Ye ZS, Chen N, Shen Y (2015) A new class of wiener process
IEEE Trans Reliab 63(1):167–190 models for degradation analysis. Reliab Eng Syst Saf 139:58–67
28. Si XS, Wang W, Hu CH, Zhou DH, Pecht MG (2012) Remaining 43. Yi D, Shu C, Qin-jie G, Tian-jian Y, Xun W, Fu-liang B (2021)
useful life estimation based on a nonlinear diffusion degradation Life prediction of Ni-Cd battery based on linear wiener process.
process. IEEE Trans Reliab 61(1):50–67 J Central South Univ 28(9):2919–2930. [Link]
29. Song W, Liu H, Zio E (2022) Long-range dependence and heavy s11771-021-4816-5
tail characteristics for remaining useful life prediction in rolling 44. Yu W, Shao Y, Xu J, Mechefske C (2022) An adaptive and general-
bearing degradation. Appl Math Model 102:268–284 ized wiener process model with a recursive filtering algorithm for
30. Song Y, Liu D, Yang C, Peng Y (2017) Data-driven hybrid remain- remaining useful life estimation. Reliab Eng Syst Saf 217:108099
ing useful life estimation approach for spacecraft lithium-ion 45. Yu W, Tu W, Kim IY, Mechefske C (2021) A nonlinear-drift-driven
battery. Microelectron Reliab 75:142–153 wiener process model for remaining useful life estimation consid-
31. Sun B, Li Y, Wang Z, Ren Y, Feng Q, Yang D (2021) An improved ering three sources of variability. Reliab Eng Syst Saf 212:107631
inverse gaussian process with random effects and measurement 46. Zhai Q, Ye ZS (2017) Rul prediction of deteriorating products
errors for rul prediction of hydraulic piston pump. Measurement using an adaptive wiener process model. IEEE Trans Ind Inform
173:108604 13(6):2911–2921
32. Sun H, Pan J, Zhang J, Cao D (2020) Non-linear wiener process- 47. Zhang Y, Yang Y, Li H, Xiu X, Liu W (2022) A data-driven mod-
based cutting tool remaining useful life prediction considering eling method for stochastic nonlinear degradation process with
measurement variability. Int J Adv Manuf Technol 107(11):4493– application to rul estimation. IEEE Trans Syst Man Cybern Syst
4502 52(6):3847–3858
33. Tang S, Yu C, Wang X, Guo X, Si X (2014) Remaining useful life 48. Zhang Z, Si X, Hu C, Lei Y (2018) Degradation data analysis and
prediction of lithium-ion batteries based on the wiener process with remaining useful life estimation: a review on wiener-process-based
measurement error. Energies 7(2):1–28 methods. Eur J Oper Res 271(3):775–796
34. Tang X, Xiao M, Hu B (2020) Application of Kalman fil-
ter to model-based prognostics for solenoid valve. Soft Comput
24(8):5741–5753
Publisher’s Note Springer Nature remains neutral with regard to juris-
35. Tang X, Xiao M, Liang Y, Hu B, Zhang L (2018) Application of
dictional claims in published maps and institutional affiliations.
particle filter technique to online prognostics for solenoid valve. J
Intell Fuzzy Syst 35(1):523–532
36. Usynin A, Hines JW, Urmanov A (2008) Uncertain failure thresh-
olds in cumulative damage models. In: 2008 Annual Reliability
and Maintainability Symposium, pp. 334–340
37. Wang D, Tsui KL (2018) Brownian motion with adaptive drift
for remaining useful life prediction: revisited. Mech Syst Signal
Process 99:691–701
38. Wang H, Ma X, Zhao Y (2020) A mixed-effects model of two-phase
degradation process for reliability assessment and rul prediction.
Microelectron Reliab 107:113622

123