Mechanical Systems and Signal Processing 160 (2021) 107879

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

Switching Bayesian dynamic linear model for condition

assessment of bridge expansion joints using structural health
monitoring data
Yi-Ming Zhang a,b, Hao Wang a,⇑, Yu Bai b, Jian-Xiao Mao a, Xiang-Yu Chang a, Li-Bin Wang c
a
Key Laboratory of C&PC Structures of Ministry of Education, Southeast University, Nanjing 211189, China
b
Department of Civil Engineering, Monash University, Clayton, VIC 3800, Australia
c
School of Civil Engineering, Nanjing Forestry University, Nanjing 210037, China

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

Article history: Age-related deterioration and premature failure have been primary concerns for bridge
Received 22 November 2020 expansion joints. It is essential to improve the understanding of their operational perfor-
Received in revised form 13 February 2021 mance. The existing approaches mainly formulate the deterministic/probabilistic
Accepted 16 March 2021
temperature-displacement relationship (TDR) model to assess the structural condition of
Communicated by Spilios Fassois
bridge expansion joints. Nevertheless, it is not easy to guarantee a strong correlation
between representative temperature and displacement. Rather than establishing the TDR
model, this work uses the displacement response to evaluate the expansion joint condition
Keywords:
Expansion joints
by combining the Bayesian dynamic linear model (BDLM) with Markov-switching theory.
Condition assessment The external temperature effect on the displacement is modeled by the superposition of
Bayesian dynamic linear model harmonic components in BDLM. The expectation–maximization (EM) algorithm initialized
Markov-switching theory by the subspace method is employed to optimize the initial parameters. The presented
Structural health monitoring approach is validated through the simulated data, and then it is applied to the expansion
Long-span bridges joint of a long-span bridge. Results show that EM with the subspace method involves high
computational accuracy and efficiency in estimating unknown parameters compared to the
Newton-Raphson approach. The switching BDLM successfully identifies the degradation
process of expansion joints and offers the transition probability from the normal to other
states.
Ó 2021 Elsevier Ltd. All rights reserved.

1. Introduction

Bridge expansion joints are designed to accommodate the movements under various excitations, including but not
limited to the thermal effect, wind action, traffic vibration, and extreme events (such as earthquakes and hurricanes)
[1,2]. It is vital that each joint works efficiently to ensure bridge safety and functionality. However, the long-term perfor-
mance of bridge expansion joints has been a concern because of the repetitive traffic loads and constant environmental
exposure conditions, especially for long-span bridges characterized as flexible and vibration-sensitive [3,4]. It is reported
that the average service life of bridge expansion joints is much lower than expected, and some failures occurred after only

⇑ Corresponding author.
E-mail addresses: [email protected], [email protected] (H. Wang).

https://doi.org/10.1016/j.ymssp.2021.107879
0888-3270/Ó 2021 Elsevier Ltd. All rights reserved.
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

a three-year service for long-span bridges [5–7]. Since failures of expansion joints could result in serviceability interruption
and even structural failure, it is important to improve the understanding of joint system performance [3].
Structural health monitoring (SHM) allows for real-time tracing bridge performance by various sensors [8,9]. It serves as a
powerful tool to gain insight into the behavior of expansion joints. Guo et al. [6] analyzed the expansion joint displacement
of a long-span suspension bridge with viscous dampers based on long-term monitoring data. Hu et al. [10] presented the
longitudinal displacement data of a bridge expansion joint and discussed its relationship with temperature and traffic den-
sity. Also, the monitoring data offer a way to prognosis the possible failure of bridge expansion joints at an early stage [4,11].
The displacement of bridge expansion joints is highly responsive to temperature, and this correlation could be used to eval-
uate structural conditions [12]. Three types of representative temperatures, including the effective temperature, mean tem-
perature, and principal component of the temperatures, are commonly used to characterize the relationship between
temperature and displacement [13,14]. Ni et al. [11] exploited the dependence between the effective temperature and
expansion joint displacement through Bayesian linear regression. An anomaly index for damage alarm of bridge expansion
joints was then derived from the regression model and reliability theory. Huang et al. [13] used the proposed representative
temperature to establish the temperature-displacement relationship (TDR) model for the expansion joint. Subsequently,
they developed a performance-alarming method by formulating a mean value control chart for the TDR model error. The
accuracy and reliability of TDR model-based approaches directly depend on the selected temperature. Inappropriate temper-
ature input could undermine the modeling accuracy [13]. Besides, the literature that focuses on the performance alarm of
bridge expansion joints with long-term monitoring data is still limited.
Data-driven methods are intensively attractive in condition assessment or anomaly detection due to their ability to track
patterns in measurements [15–17]. A variety of regression methods are usually implemented for this purpose. Extensively
used approaches include autoregressive (AR) model [18], neural networks [19,20], support vector regression (SVR) [21],
Gaussian process regression (GPR) [22], and regression trees [23]. These approaches generally consist of two steps [24].
The first step is to define patterns of normal structural behavior derived from the measurements. Then, newly available
observations are evaluated to identify the deviations from the computed patterns. Even though the AR model involves
the advantages of computational simplicity and efficiency, its prediction capacity is relatively limited [25]. The main draw-
back of neural networks is their lack of interpretation ability, which is criticized as a ‘‘black box” [26]. In terms of kernel-
based methods (such as the SVR and GPR), it is nontrivial to select an appropriate kernel function that directly affects the
modeling flexibility and efficiency [22]. The Kalman filter (KF) is usually performed to diagnose structural health state
because of its two merits [27,28]. Firstly, the KF model takes into account both observation and latent state noise. Secondly,
the model inference involves the dynamic state update by considering the prediction error. However, the KF model only
characterizes a single regime, which could result in unreliable state estimates. Switching state-space model, also known
as the switching Kalman filter (SKF), overcomes the limitation of the KF [29–31]. It assumes that KF models operate on a
mixture of regimes and enable mixing proportions to be determined [32]. Recently, the SKF has been applied to model
the machine degradation and detect abnormal behavior of dams [33,34].
Bayesian dynamic linear model (BDLM), which is usually referred to as the state-space model, shows a promising appli-
cation in the field of SHM [35–37]. The increasing applications of BDLM can be attributed to its two admirable merits. First,
the BDLM combines various components (e.g., trend, season, and regression components) to capture time-series features and
ensure high modeling flexibility. Second, the BDLM is a probabilistic model with time-varying coefficients. It allows for
dynamic uncertainty quantification in the modeling process. In previous studies, the BDLM is extensively employed to fore-
cast various structural responses based on the monitoring data. Wang and Ni [38] adopted the BDLM involving model class
selection to forecast structural strain responses in a real-time manner. Ma et al. [39] used an improved BDLM to predict the
static strains and then separate the load effects (dead load, temperature load, and audience load) of a revolving auditorium.
Fan and Liu [40,41] applied the BDLM with first/second-order trend and season components to predict the bridge extreme
stresses. To improve the computational accuracy, Wang et al. [37] introduced the AR component to the ordinary BDLM and
forecasted the temperature-induced strains of a long-span bridge. Comparatively, little work has been conducted on anom-
aly detection of civil structures based on the BDLM. Zhang et al. [42] presented an online method to detect anomalies in the
SHM data utilizing the maximum likelihood estimation (MLE)-based BDLM. Goulet and his colleagues [35,43] proposed a
BDLM framework to estimate the hidden states of structural response and the external effects. Later, they [34] combined this
approach with the SKF to detect the abnormal dam behavior using long-term monitoring data. Newton-Raphson (NR) algo-
rithm initialized by the empirical values was employed to estimate model parameters. In practice, the NR algorithm is par-
ticularly sensitive to its initial starting points that directly affect modeling accuracy. It is nontrivial to choose appropriate
initial parameters when lacking the engineering heuristics and a large volume of data.
This study employs BDLM with Markov-switching to evaluate the status of bridge expansion joints. Specifically, a
weighted combination of BDLMs is implemented to estimate the expansion joint condition at each time step [44,45]. The
switching between models indicates a change in the latent state, meaning that abnormal behavior may occur in the system.
Different from the existing methods, the switching BDLM considers only the displacement data to assess the condition of
bridge expansion joints without formulating the TDR model. It adopts the harmonic components in the BDLM to model
the temperature effect on displacement. In particular, the switching BDLM can consider the transition probability between
different states and separate the structural responses into hidden state variables. These features render this approach suit-
able for reliable evaluation of the bridge expansion joint condition with monitoring data. Instead of using the NR algorithm,
the expectation–maximization (EM) algorithm initialized by the subspace method is adopted to estimate the unknown
2
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

parameters. The performance of NR and EM algorithms is then thoroughly compared by simulated data concerning accuracy
and efficiency. Subsequently, the switching BDLM optimized by the EM algorithm is applied to expansion joints of a long-
span cable-stayed bridge.
To summarize, the main contributions of this work are three-fold: (1) unlike the existing methods, the present switching
BDLM employs only the displacement data to assess the bridge expansion joint condition without dependence on the TDR
model. The temperature effect on displacement is modeled by harmonic components in the BDLM; (2) the switching BDLM
considers the transition probability from the normal to an abnormal state and interprets structural responses through dis-
composing them into latent state variables; and (3) the subspace method is implemented to initialize the EM algorithm to
mitigate the numerical instability and reduce the computational expense. The rest of this paper is organized as follows. In
Section 2, formulations of switching BDLM and the parameter estimation are briefly introduced. Section 3 verifies the fea-
sibility of switching BDLM using data from a simulated example. Data collected from the expansion joint of a long-span
bridge is adopted to examine its condition by utilizing the switching BDLM in Section 4. Finally, Section 5 ends with the con-
clusions drawn from this study.

2. Bayesian dynamic linear model with switching theory

The formulations of switching BDLM and estimation of unknown parameters are presented in this section. BDLM is firstly
described for introducing its combination with switching theory.

2.1. Bayesian dynamic linear model

BDLM uses the linear state-space model to estimate the hidden states [46]. It is comprised of the observation equation
and system equation written as
Observation equation : Yt ¼ Ft ht þ v t ; vt  Nð0; Rt Þ; ht  Nðmt ; Vt Þ ð1Þ

System equation : ht ¼ Gt ht1 þ wt ; wt  Nð0; Q t Þ ð2Þ

where t denotes time; Yt ht is the observation vector; ht is the hidden state vector with the expected matrix mt and covari-
ance matrix Vt; Ft and Gt are the observation and evolution (transfer or system) matrices; and vt and wt are the observation
and evolution errors, respectively, which are internally and mutually independent.
BDLM combines different components to characterize time-series features [37].
The components, including the trend component, periodic (harmonic) component, and AR component, are briefly pre-
sented as follows.

(1) Trend component

The zeroth-order trend describes the baseline of the time series, which is defined as

2 
Y 0TR
t ¼ h0TR
t þ v 0TR
t ; v 0TR
t N 0; rR;0TR
t

2  ð3Þ
h0TR
t ¼ h0TR
t - 1 þ wt ; wt
0TR 0TR
 N 0; rQ;TR
t

By comparing Eq. (3) with the observation equation [Eq. (1)] and system equation [Eq. (2)], the canonical form of Eq. (3)
can be expressed as
 2  2
F 0TR
t ¼ 1; G0TR
t ¼ 1; R0TR
t ¼ rR;0TR ; Q 0TR
t ¼ rQ ;TR ð4Þ
where the superscripts of the vectors in Eqs. (3) and (4) denote the zeroth-order trend.
The first-order trend component in the state-space model is known as the constant speed model. It can be written as

2 
Y 1TR
t ¼ F 1TR
t ht
1TR
þ v 1TR
t ; v 1TR
t N 0; rR;1TR
t

2  ð5Þ
h1TR
t ¼ G1TR
t h1TR
t - 1 þ wt
1TR
; w1TR
t  N 0; rQ;1TR
t

h iT
h_ t
1TR
The hidden states can be denoted as h1TR
t ¼ h1TR
t
, that is, the response and its change rate. Then, the system equa-
tion of the first-order trend becomes [47]

_ 1TR
h1TR
t t - 1 þ Dt ht
¼ h1TR - 1 þ w1TR
t
ð6Þ
h_ t ¼ h_ t - 1 þ w
1TR 1TR
_ 1TR
t

3
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

where Dt is the time interval.

Based on Eqs. (5) and (6), the observation matrix and evolution matrix are expressed as

1 Dt
F 1TR
t ¼ ½ 1 0 ; G1TR
t ¼ ; ð7Þ
0 1
The evolution error can be written by the following Eq. (8) [31,48]
Z   " #
h  1TR T i Dt
Dt  s Dt  s T
2   Dt3 =3 Dt2 =2
Q 1TR
t ¼ E w1TR
t wt ¼ r Q ;TR
t ds ¼ r Q;TR 2
ð8Þ
0 1 1 Dt2 =2 Dt

where the superscripts of variables in Eqs. (5)–(8) denote the first-order trend.
The second-order trend component in the state-space model is called the constant acceleration model, which is expressed
as

2 
Y 2TR
t ¼ F 2TR
t h2TR
t þ v 2TR
t ; v 2TR
t N 0; rR;2TR
t

2  ð9Þ
h2TR
t ¼ G2TR
t h2TR
t - 1 þ wt
2TR
; w2TR
t  N 0; rQ;2TR
t

h iT
h_ t €
2TR 2TR
The state vector is h2TR
t ¼ h2TR
t ht , and the system equation is given as below [48]

_ 2TR
þ 12 Dt 2 €ht
2TR
h2TR
t t - 1 þ Dt ht
¼ h2TR - 1 - 1 þ w2TR
t

h_ t ¼ h_ t - 1 þ Dt €ht - 1
2TR 2TR 2TR
_t
þw
2TR ð10Þ
€h2TR ¼ €h2TR þ € 2TR
w
t t - 1 t

Then, the observation and evolution matrices of the second-order trend are written as
2 3
1 Dt Dt 2 =2
6 7
F 2TR
t ¼ ½ 1 0 0 ; G2TR
t ¼ 40 1 Dt 5 ; ð11Þ
0 0 1
Like the Eq. (8), the evolution error of the second-order trend can be expressed as [31,48]
2 32 3T 2 5 3
h Z ðDt  sÞ2 =2 ðDt  sÞ2 =2 Dt =20 Dt4 =8 Dt 3 =6
 2TR T i Dt
6 76 7

2   26 7
Q 2TR
t ¼ E w2TR
t wt ¼ 4 Dt  s 54 Dt  s 5 rQ;TR
t ds ¼ rQ ;TR 4 Dt 4 =8 Dt3 =3 Dt 2 =2 5 ð12Þ
0
1 1 Dt 3 =6 Dt2 =2 Dt
where the superscripts of variables in Eqs. (9)–(12) denote the first-order trend.

(2) Harmonic component

Periodic behavior is evident for the displacement of expansion joints due to the temperature effect. The harmonic com-
ponent in BDLM could well describe this pattern, which is expressed as below [46]
(    2 !)
n o cosð2p=cÞ sinð2p=cÞ  2 rQ;SE 0
F SE SE SE SE
t ; G t ; Rt ; Q t ¼ ð1; 0Þ; ; rR;SE ;  2 ð13Þ
sinð2p=cÞ cosð2p=cÞ 0 rQ ;SE
where c is the period of displacement of the bridge expansion joint; rR;SE denotes the standard deviation of observation error;
and rQ ;SE is the standard deviation of system error in the harmonic component.

(3) AR component

The AR model reflects the relationship of observations between previous and current time steps. It models the residuals
that could not be described by other components of BDLM. The AR (p) model, where p is the order of the AR model, is
described as follows:
X
p
Y AR
t ¼ /j Y tj þ v AR
t ð14Þ
j¼1

t is the measurement at time t; /j is the autoregressive coefficient; and v t

where Y AR AR
is the observational error.
The state-space formulation of the AR component can be written by the following Eq. (15) [46]

4
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

8 0 1 9
> /1 /2    /p1 /p >
>
> >
>
>
> B 1 0  0 0 C >
>
n o >< B
B
C
C  R;AR 2
 Q ;AR 2 > >
=
F AR ; GAR
; R AR
; Q AR
¼ ð1; 0;   Þ; B 0 1  0 0 C; r ; r ; 0; 0;    ; 0 ð15Þ
t t t t
> B C >
>
> B .. .. .. .. .. C >
>
>
> @ . . . . . A >
>
>
: >
;
0 0 0 1 0
where r R;AR
represents the standard deviation of observation error; and rQ ;AR is the standard deviation of system error.

2.2. Switching Bayesian dynamic linear model

BDLM with switching can be considered as a mixture of BDLMs. It enables the transition of the dynamical process from
one operating regime to another. The Markov-switching variable (St) with the transition matrix (Z) is formulated to deter-
mine the probability of BDLM to be applied at time t. In the switching BDLM, its parameters rely on the current state, that is,
Ft = F(St), Gt = G(St), Rt = R(St), and Qt = Q(St).
Define the posterior probability (the weight of each model) as PrðSt jY 1:T Þ that can be obtained in forward and backward
steps. This paper selects the KF and Kalman smoother (KS) to perform the forward and backward passes [29]. The details of
KF and KS are given in the Appendix. A. Specifically, KF is performed in the forward step as follows:
ðiÞj ðiÞj ðiÞj j j
ðmtjt ; V tjt ; V t;t1jt ; LiðjÞ i i j j
t Þ ¼ Filterðmt1jt1 ; V t1jt1 ; Y t ; F t ; G t ; Rt ; Q t Þ ð16Þ

where the superscript (i)j denotes that one state (St-1 = i, i = 1, 2,. . ., N) switches to another state (St = j, j = 1, 2,. . ., N); N is the
number of BDLMs; V t;t1jt is the cross variance; and Lt is the likelihood of BDLM. The first- or second- order polynomial com-
ponent is usually adequate for most practical situations [31,33,46]. Higher-order trend components with rare applications
are not discussed here. Thus, the number of BDLMs in this paper is less than 3, i.e., N 6 3.
The mixture of Gaussians is collapsed into a single Gaussian density by the moment matching. The collapse step can be
written as
iðjÞ iðjÞ
ðmjtjt ; V jtjt ; M jtjt Þ ¼ Collapseðmtjt ; V tjt ; W iðjÞ
t Þ ð17Þ

Eq. (17) is calculated from

Z iðjÞ ¼ PrðSt¼j jSt1 ¼ iÞ ð18Þ

iðjÞ LiðjÞ iðjÞ i

t Z M t1jt1
M t1;tjt ¼ PrðSt1 ¼ i; St ¼ jjY 1:t Þ ¼ P P iðjÞ iðjÞ i ð19Þ
i j Lt Z M t1jt1

X iðjÞ
M jtjt ¼ PrðSt ¼ jjY 1:t Þ ¼ Mt1;tjt ð20Þ
i

iðjÞ
W iðjÞ j
t ¼ PrðSt1 ¼ ijSt ¼ j; Y 1:t Þ ¼ M t1;tjt =M t jt ð21Þ
X iðjÞ
mjtjt ¼ mtjt W iðjÞ
t ð22Þ
j

X X

>
iðjÞ iðjÞ iðjÞ
V jtjt ¼ W tiðjÞ V tjt þ W iðjÞ
t mtjt  mjtjt mtjt  mjtjt ð23Þ
j j

iðjÞ
where W iðjÞ
t is the state switching probability; M jtjt is the state probability at time t; M t1;tjt is joint probability; and Z iðjÞ is the
transition probability.
The backward pass is performed as
ðjÞk ðjÞk ðjÞk jðkÞ
ðmtjT ; V tjT ; V tþ1;tjT Þ ¼ Smoothðmktþ1jT ; V ktþ1jT ; mjtjt ; V jtjt ; V ktþ1jtþ1 ; V tþ1;tjtþ1 ; Gkt ; Q kt Þ ð24Þ

Then, the collapse step is expressed as

jðkÞ jðkÞ
ðmjtjT ; V jtjT ; M jtjT Þ ¼ CollapseðmtjT ; V tjT ; W tkðjÞ Þ ð25Þ

Eq. (25) is obtained by

Mjtjt Z jjk
U jðkÞ
t ¼ PrðSt ¼ jjStþ1 ¼ k; Y 1:T Þ ¼ P j jjk ð26Þ
j M t jt Z

5
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

jðkÞ
M t;t þ 1jT ¼ U jðkÞ k
t M tþ1jT ð27Þ
X jðkÞ
M jtjT ¼ PrðSt ¼ jjY 1:T Þ ¼ M t;tþ1jT ð28Þ
k

jðkÞ
W kðjÞ
t ¼ PrðStþ1 ¼ kjSt ¼ j; Y 1:T Þ ¼ Mt;tþ1jT =M jtjT ð29Þ
X iðjÞ
mjtjT ¼ mtjT W kðjÞ
t ð30Þ
j

X X

>
iðjÞ iðjÞ iðjÞ
V jtjT ¼ W kðjÞ
t V t jT þ W tkðjÞ mtjT  mjtjT mtjT  mjtjT ð31Þ
j j

where k denotes the state (St = k, k = 1, 2, . . ., N) and T represents the sample size.
Three BDLMs including trends with different orders, daily cycle, seasonal cycle, and AR components are expressed as fol-
lows. Note that the time interval (Dt) of expansion joint displacements is set as one hour. As a result, the periods for daily and
seasonal components are 1 day (24 h) and 365 days (365  24 h), respectively.

(1) Regime with the zeroth-order trend

The evolution matrix is given by

G0t ¼ blkdiagðG0TR Day

t ; Gt ; GSeason
t ; GAR
t Þ ð32Þ
with
2 3
1 0 0 
6 7 cosðp=12Þ sinðp=12Þ
G0TR ¼ 40 0 0 5; GDay ¼
t t
sinðp=12Þ cosðp=12Þ
0 0 0

cos½p=ð365  12Þ sin½p=ð365  12Þ
GSeason ¼ ; GAR
t ¼ /AR ð33Þ
t
sin½ðp=ð365  12Þ cos½p=ð365  12Þ
The observation matrix is written as

F 0t ¼ ½ 1 0 0 1 0 1 0 1  ð34Þ

The evolution error matrix is


Q 0t ¼ blkdiag Q 0TR Day
t ; Qt ; Q Season
t ; Q AR
t ð35Þ

with
2 3
1 0 0 "  #
  6 7 rDay 2 0
¼ r
2
Q 0TR
t
0TR
4 0 0 0 5; Q t ¼
Day
 Day 2
0 0 0 0 r
ð36Þ
"  #
rSeason 2
0  AR 2
Q Season  Season 2 ; Q t ¼ r
AR
t ¼
0 r
The observation error matrix is
 2
R0t ¼ rR ð37Þ

(2) Regime with the first-order trend

The evolution matrix is written as

02 3 1
1 Dt 0
1 B6 7 C
Gt ¼ blkdiag@4 0 1 0 5; GDay
t ; GSeason
t ; GAR
t A ð38Þ
0 0 0

6
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

The observation matrix is given by

F 1t ¼ ½ 1 0 0 1 0 1 0 1  ð39Þ

The evolution error matrix is

2 3
0 3 1
Dt =3 Dt 2 =2 0
B   6 7 C
Q 1t ¼ blkdiag@ r1TR 4 Dt 2 =2
2 Day Season
Dt 0 5; Q t ; Q t ; Q AR
t A ð40Þ
0 0 0
The observation error matrix is expressed as
 2
R1t ¼ rR ð41Þ

(3) Regime with the second-order trend

The evolution matrix is given by

02 3 1
1 Dt Dt 2
B6 7 Day Season AR C
G2t ¼ blkdiag@4 0 1 Dt 5 ; G t ; G t ; Gt A ð42Þ
0 0 1
The observation matrix is written by

F 2t ¼ ½ 1 0 0 1 0 1 0 1  ð43Þ

The evolution error matrix is

2 50 3 1
Dt =20 Dt 4 =8 Dt3 =6
B   26 7 C
Q 2t ¼ blkdiag@ r2TR 4 Dt4 =8 Dt 3 =3 Dt2 =2 5; Q Day
t ; Q Season
t ; Q AR
t A ð44Þ
Dt3 =6 Dt 2 =2 Dt
The observation error is expressed as
 2
R2t ¼ rR ð45Þ

where the abbreviation blkdiag ðÞ represents the block diagonal matrix.

2.3. Parameter estimation

The parameters (H) to be estimated include F jt ; Gjt ; Rjt ; Q jt . The maximum likelihood estimation is adopted to determine the
parameters from the observations. The log-likelihood function of switching BDLM is obtained as

P
T
LðHÞ ¼ lnpðY t jY 1:t1 Þ
t¼1
" #
P
T P
N P N
¼ ln pðY t ; St ¼ j; St1 ¼ ijY 1:t1 Þ
t¼1 j¼1 i¼1
" #
P
T P
N P
N
¼ ln LiðjÞ
t PrðSt ¼ j; St1 ¼ ijY 1:t1 Þ ð46Þ
t¼1 j¼1 i¼1
" #
P
T N P
P N
¼ ln LiðjÞ iðjÞ
t Z PrðSt ¼ jjY 1:t1 Þ
t¼1 j¼1 i¼1
" #
P
T N P
P N
¼ ln LiðjÞ iðjÞ
t Z M t jt
i

t¼1 j¼1 i¼1

The optimal parameters can be determined by maximizing the log-likelihood function as

H ¼ argmaxLðHÞ ð47Þ
The Newton-Raphson (NR) algorithm can be employed to solve Eq. (47) [49]. Nevertheless, this algorithm particularly
depends on the initial values of model parameters and easily results in a local maximum. Considering the EM algorithm also
could get stuck in local maxima, the starting points are generated by the subspace method to mitigate such risk [42,49,50].
7
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

The subspace method is an alternative to the maximum likelihood training. The primary benefits of the subspace method are
that it avoids the convergence difficulties of the EM and does not involve the initialization process. Therefore, this method
can be adopted to initialize the EM algorithm to accurately estimate the BDLM parameters [50].
Given the observations Yt, suppose D is a d-observation Hankel matrix of size s as follows [51]
2 3
Y1 Y2 Y3  Ys
6 . .. .. .. .. 7
D¼6
4 .. . . . .
7
5 ð48Þ
Yd Y dþ1 Y dþ2    Y dþs1
The estimation of the hidden state is generated through singular value decomposition (SVD). The input D is obtained by

D  URH S T ð49Þ
with the estimation given by
F¼U ð50Þ

^hS ¼ RH S T ð51Þ

where U and S have the orthonormal columns; ^hS is denoted as the extended observability matrix and represents the state
estimation of BDLM; and RH contains the singular values. The least-squares of G can be expressed as
^ ¼ argmink Gh0:s1  h1:s k2 ¼ h1:s hy
G ð52Þ
F 0:s1

where k  kF and y represent the Frobenius norm and the pseudo inverse, respectively. Given the estimation of F and G, the
covariances V and W can be obtained directly from residuals.
The E-step in EM algorithm is to compute the expectation of the log-likelihood function using the KF algorithm. Param-
eters are then identified by maximizing the expected log-likelihood function in the M-step. Murphy [29] has derived the ana-
lytical formulation of maximization as follows.
! !1
X
T X
T
Git ¼ W it P t;t1 W it P t1 ð53Þ
t¼2 t¼2

! !
1 X
T X
T
Q it ¼ PT i
W it P t  Gi W it P Tt;t1 ð54Þ
t¼2 W t t¼2 t¼2

! !1
X
T X
T
F it ¼ W it Y t hTt W it P t ð55Þ
t¼1 t¼1

! !
1 X
T

Rit ¼ PT i
W it Y t Y >t  Fi ht YTt ð56Þ
t¼1 W t t¼1

with

P t ¼ V tjT þ htjT hTtjT ð57Þ

P t;t1 ¼ V t;t1jT þ htjT hTt1jT ð58Þ

Note that in Eqs. (57) and (58), htjT ; V tjT , and V t;t1jT are obtained from the KS algorithm. The E and M steps are iteratively
repeated until the change of likelihood between consecutive time steps is below a certain threshold.

3. Illustrative example

3.1. Simulated data

The simulated data is generated to illustrate the effectiveness of switching BDLM. Because of the temperature effect, the
displacement of expansion joints exhibits daily and seasonal cycles, which can be modeled by the harmonic components in
BDLM. The time interval is set as one-hour for one year (a total of 8760 observations). The simulated displacement data have
three regimes with different trend orders, including zeroth-order, first-order (176.5th day to 182nd day, and a total of 132
data points), and second-order (263rd to 264.5th day, and a total of 36 data points). As mentioned in describing the BDLM,
the zeroth-, first-, and second-order can be considered as the baseline, speed, and acceleration components, respectively.
8
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Three BDLMs are utilized to model the simulated displacement. To make the system remain in its own state (terms on the
main diagonal tends to be 1), the transition matrix is assumed as
2 3 2 3
Z 11 Z 12 Z 13 0:98 0:01 0:01
6 7 6 7
Z ¼ 4 Z 21 Z 22 Z 23 5 ¼ 4 0:01 0:98 0:01 5
Z 31 Z 32 Z 33 0:01 0:01 0:98
The initial state for each BDLM is
T
m0 ¼ ½ 40 0 0 8 8 30 30 0:01  ; V 0 ¼ I
where I denotes the identity matrix with size 8  8.
The formulations of BDLMs are presented in Eqs. (32)–(45). The unknown parameters are assumed to be

H ¼ /AR ; r0TR ; r1TR ; r2TR ; rDay ; rSeason ; rAR ; rR ¼ ½0:85; 0:1; 0:02; 0:02; 0:03; 0:01; 0:1; 0:1
As a result, the simulated data for this example is shown in Fig. 1.

3.2. Comparison of parameter estimation methods

The data sets of six months are employed to learn the unknown parameters since this period presents the full amplitude
of the seasonal cycles. In this example, the NR and EM algorithms are adopted to estimate the parameters independently, and
their performance is compared regarding both accuracy and efficiency. The parameter estimation is implemented on a com-
puter with an Intel i7-9700 processor and 32 GB of memory. The initial values of parameters for the NR approach are set as

H0NR ¼ /AR ; r0TR ; r1TR ; r2TR ; rDay ; rSeason ; rAR ; rR ¼ ½0:3; 0:1; 0:1; 0:1; 0:1; 0:1; 0:1; 0:1
Consequently, it consumes 844.05 s to learn the parameters from the training data set. The estimated parameters are
h i
HNR ¼ 0:75; 0:07; 0:01; 1:06  107 ; 0:03; 0:07; 0:1; 0:1

As an alternative approach, the EM algorithm is firstly implemented to estimate parameters of BDLM without switching.
The status of displacement is assumed to be normal most time, and it can be firstly modeled by BDLM with the zeroth-order
trend. As mentioned previously, the EM algorithm easily traps into the local maximum and involves numerical instability.
Therefore, the subspace method is performed to initialize the EM algorithm to address these issues [49]. The estimated
parameters of BDLM with the zeroth-order trend are then assigned to corresponding positions in the switching BDLM.
Different thresholds of the EM algorithm are discussed to establish the efficient switching BDLM. Euclidean distance is
adopted to quantitatively assess the accuracy of model parameters, which is defined as

^ Hk
d¼kH ð59Þ

where H^ and H are the estimated and real model parameters, respectively.
The computational time and Euclidean distance with different thresholds are displayed in Fig. 2. It is found that the com-
putational time dramatically increases when the threshold value is less than 103. Besides, the corresponding threshold of

Fig. 1. Simulated displacement of the expansion joint.

9
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 2. Comparison of the computational accuracy and efficiency with different thresholds.

the minimum distance for 102. Considering the accuracy and time consumption, the optimal threshold is set as 102 in the
simulated case.
As a result, it takes 15.47 s to estimate all parameters of BDLM with switching, and the optimal parameters are

HEM ¼ ½0:65; 0:06; 0:01; 0:02; 0:03; 0:01; 0:1; 0:1

The consumed time of NR (844.05 s) and EM (15.47 s) algorithms indicates that the EM algorithm initialized by the sub-
space method exhibits overwhelming superiority in computational efficiency. It can also be observed that the estimated
parameters of the EM algorithm are closer to the true parameter values than those of the NR approach. More specifically,
the estimated variances of the second-order trend and seasonal cycle components via the NR approach deviate significantly
from the true value.
Figs. 3 and 4 display the identification results of displacement through switching BDLM that optimized by NR and EM
algorithms, respectively. Note that Models 1, 2, and 3 represent the BDLM with zeroth-, first-, and second-order trends,
respectively. As shown in Fig. 3, for the switching BDLM optimized by the NR method, the probabilities of model classes dra-

Fig. 3. Diagnosis results of switching BDLM optimized by NR algorithm: (a) probability of model class; (b) model class.

10
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 4. Identification results of switching BDLM optimized by EM algorithm: (a) probability of model class; (b) model class.

matically fluctuate when the regime-switching occurs. Nevertheless, the model classes are incorrectly identified as Model 3
(between 160th day and 175th day) and Model 2 (between 263rd day and 264th day). By contrast, it is observed from Fig. 4
that the model classes are successfully assigned to their true state without triggering false alarms. Thus, it can be concluded
that the switching BDLM with the EM algorithm initialized by the subspace method owns better computational accuracy and
efficiency.
The confusion matrices are presented in Fig. 5 to clearly present the identification results. From Fig. 5(a), the following
observations are obtained: (1) one data point originally belonging to Model 2 is incorrectly identified as Model 1; (2) one
data point that belongs to Model 1 is arranged as Model 3; and (3) the overall identification accuracy is more than 97%.
For the switching BDLM optimized by the NR approach, as shown in Fig. 5(b), it can be found that all data points in Model
3 are mistakenly identified as Model 2. Besides, 182 and 357 data points originally belonging to Model 1 are incorrectly
arranged for Model 2 and Model 3, respectively. In summary, the switching BDLM that is optimized by EM with the subspace
method exhibits higher accuracy than the NR approach.
The hidden states of simulated data are discomposed using switching BDLM optimized by the EM algorithm. The sepa-
rated results are presented in Fig. 6. It can be found from Fig. 6(a) that the initial value of the separated baseline is

Fig. 5. Identification results by switching BDLM: (a) EM algorithm; (b) NR approach.

11
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 6. Hidden components of switching BDLM: (a) baseline component; (b) speed component; (c) acceleration component; (d) daily cycle component; (e)
seasonal cycle component; (f) AR component.

40 mm. This value is consistent with the simulated initial state of the zeroth trend order. Sudden variation can be observed
when the expansion joint condition changes and the baseline eventually decreases to nearly 26 mm on the 365th day. The
sudden change can also be found in zeroth-order (baseline), first-order (speed), second-order (acceleration), and AR compo-
nents [Fig. 6(b), (c), and (f)]. Note that the temperature effect modeled by the harmonic components is not affected by the
state switching of displacements [Fig. 6(d) and (e)]. In general, the initial values of all hidden components are close to the
simulated ones. This result verifies that the switching BDLM accurately decomposes the latent states of expansion joint
displacements.

4. Application

In this section, the monitoring displacements of the expansion joint collected from the Sutong Cable-Stayed Bridge (SCB)
are employed to examine the performance of the switching BDLM.

4.1. Sutong Cable-Stayed bridge and structural health monitoring system

The SCB with a main span of 1088 m is the longest cable-stayed bridge in China, as shown in Fig. 7. Two pylons of the SCB
with an inverted-Y shape have a height of 300 m. The pylons and bridge deck are connected by 272 symmetrically dis-
tributed cables. Two modular expansion joints are installed at the ends of the SCB, and their configuration is displayed in
Fig. 8. The center beam can slide on support bars to accommodate the longitudinal movements with an extension limitation
of 2.6 m [6].
A sophisticated long-term SHM system has been implemented to monitor the performance of the SCB after its completion
(June 2008). The SHM system comprises more than 1440 sensors of various types, including anemometers, temperature
meters, humidity sensors, displacement sensors, corrosion gauges, global positioning systems, and strain gauges. Two
draw-wire displacement gauges at each end of the bridge girder are utilized to monitor longitudinal movements of expan-
sion joints. Without loss of generality, two displacement gauges (denoted as DIS 01 and DIS 02) are selected to demonstrate
the switching BDLM. The hourly-averaged displacements of both gauges for six months (from January 1 to June 30, 2009) are
of concern. Thus, a total of 4320 h of data are utilized for modeling the switching BDLM, as shown in Fig. 9. It is found that the
displacement exhibits periodic behavior because of the temperature effect.
12
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 7. SHM system of the SCB.

Fig. 8. Configuration of the expansion joint of the SCB: (a) schematic diagram; (b) A-A section.

Fig. 9. Expansion joint displacement of the SCB.

13
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

4.2. Construction of the switching BDLM

The switching BDLM for condition assessment of expansion joints consists of two model classes, including the first- and
second-order trends. Specifically, BDLM with the first- and second-order trend represents the normal and abnormal states of
the expansion joint, respectively [34]. Because six months’ data is selected as target quantities (including training and testing
data sets), the seasonal cycle is not considered in building the BDLM. The global matrices defining switching BDLM are intro-
duced as follows:

(1) Normal model class

The evolution matrix is defined as

02 3 1
1 Dt 0
B6 7 C
GNormal
t ¼ blkdiag@4 0 1 0 5; GDay
t ; GAR
t A ð60Þ
0 0 0
with

cosðp=12Þ sinðp=12Þ
GDay ¼ ; GAR
t ¼ /AR ð61Þ
t
sinðp=12Þ cosðp=12Þ
The observation matrix is expressed as

F Normal
t ¼ ½1 0 0 1 0 1 ð62Þ

The evolution error matrix is written as

2 30 3 1
Dt =3 Dt2 =2 0
B 1TR 2 6 2 7 Day AR C
Q Normal
t ¼ blkdiag@ r 4 Dt =2 Dt 0 5; Q t ; Q t A ð63Þ
0 0 0
with
" 2 #
rDay 0  2
Q Day
t ¼   ; Q AR
t ¼ rAR ð64Þ
0 r Day 2

The observation error is given by

 2
RNormal
t ¼ rR ð65Þ

(2) Abnormal model class

The evolution matrix is provided as

02 3 1
1 Dt Dt 2
B6 7 Day AR C
GAbnormal
t ¼ blkdiag@4 0 1 Dt 5 ; G t ; G t A ð66Þ
0 0 1
The observation matrix is expressed as

F Abnormal
t ¼ ½1 0 0 1 0 1 ð67Þ

The evolution error matrix is written as

2 5
0 3 1
Dt =20 Dt 4 =8 Dt3 =6
B   26 7 C
Q Abnormal
t ¼ blkdiag@ r2TR 4 Dt4 =8 Dt 3 =3 Dt2 =2 5; Q Day
t ; Q AR
t A ð68Þ
Dt3 =6 Dt 2 =2 Dt
The observation error is given by
 2
RAbnormal
t ¼ rR ð69Þ
The transition matrix is assumed as

14
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

 
Z 11 Z 12 0:98 0:01
Z¼ ¼
Z 21 Z 22 0:01 0:98
The unknown parameters to be estimated are grouped as follows:

H ¼ /AR ; r1TR ; r2TR ; rDay ; rAR ; rR ð70Þ

The first five months’ measurements are used as the training data to learn BDLM parameters, and the remaining one-
month data is chosen to validate the model performance. After its effectiveness has been proved in the previous section,
the EM algorithm initialized by the subspace method is chosen to optimize the unknown parameters. The convergence cri-
terion of the EM algorithm is that the log-likelihood variation between two iterations is below 103, or the number of iter-
ations reaches 2000. The estimated parameters for DIS 01 and DIS 02 are
h i
HDIS01 ¼ 0:856; 0:0846; 2:38  104 ; 1:45  104 ; 8:38  105 ; 2:81  105
h i
HDIS02 ¼ 0:859; 0:0697; 2:73  104 ; 1:53  104 ; 8:47  105 ; 2:84  105

where the ordering of parameters is identical with Eq. (70).

4.3. Condition evaluation of expansion joints

The switching BDLM with optimal parameters is subsequently implemented to evaluate the condition of bridge expan-
sion joints. One-month data is used to test the performance of switching BDLM. Fig. 10 presents the evaluation results of
DIS 01. Specifically, the ‘‘Normal model” (model 1) represents the normal condition of the expansion joint, and the ‘‘Abnor-
mal model” (model 2) denotes the abnormal condition. It is observed that the probabilities of the normal model are nearly 1
and the probabilities of the abnormal model are approximately 0 at each time step. Similar results can be found for DIS 02
(Fig. 11). Thus, it indicates that the expansion joints of the SCB are healthy consistent with the real state.
The degradation process of the expansion joint is considered to further verify the effectiveness of the switching BDLM.
The shift is introduced into the original data set to achieve this purpose as follows [11,13]:
Dde ¼ D  Dshift ð71Þ

Fig. 10. Identification results of original measurements of DIS 01 by switching BDLM: (a) model probability; (b) model classes.

15
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 11. Identification results of original measurements of DIS 02 by switching BDLM: (a) model probability; (b) model classes.

where Dde is the simulated displacement of the expansion joint suffered degradation; D is the original measurement; and
Dshift is the shift value that controls degradation severity. The shift values of 10 mm, 20 mm, and 30 mm are selected to sim-
ulate different degradation processes. We assume this process begins from the 12th day for DIS 01 and the 18th day for DIS
02 in the testing phase. Table 1 lists the diagnosis results of four deterioration cases. It can be concluded that the switching
BDLM can identify the real condition of the expansion joint even though the deterioration severity is small, e.g., 10 mm for
Case 2.
The detailed evaluation results for Case 2 are discussed below. It is obvious from Fig. 12 that the probability of the normal
model degrades to approximately 0 and the probability of the abnormal model reaches nearly 1 on the 12th day. Similarly,
this observation can be seen on the 18th day in Fig. 13. The results illustrate that the switching BDLM correctly identifies the
state change when the shift value is 10 mm. Then, the hidden states of expansion joints displacement are estimated and dis-
played in Figs. 14 and 15. As shown in Fig. 14(a), the baseline displacement dramatically decreases by 10 mm on the 12th day
for the validating data, which is identical to the simulated deterioration. The sudden change can also be found from the speed
component [Fig. 14(b)], acceleration component [Fig. 14(c)], and the AR component [Fig. 14(e)], except the daily cycle com-
ponent [Fig. 14(d)]. Note that the AR component can model time-independent errors that cannot be captured by other com-
ponents. It is observed from Fig. 14(e) that the AR component mainly varies between 0.1 mm and 0.1 mm. This small
amplitude indicates that the estimated parameters of switching BDLM are adequate. Similar observations can be found in
Fig. 15. In summary, the presented approach exhibits excellent performance in identification of the bridge expansion joint
condition.

Table 1
Diagnosis results of the expansion joint with four deterioration cases.

Case Deterioration severity (mm) DIS 01 condition DIS 02 condition

1 0 Healthy Healthy
2 10 Degradation Degradation
3 20 Degradation Degradation
4 30 Degradation Degradation

16
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 12. Diagnose results of original measurement of DIS 01 by switching BDLM: (a) model probability; (b) model classes.

Fig. 13. Diagnose results of original measurement of DIS 02 by switching BDLM: (a) model probability; (b) model classes.

17
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

Fig. 14. Hidden states of displacement for DIS 01: (a) baseline component; (b) speed component; (c) acceleration component; (d) daily cycle component;
(e) AR component.

Fig. 15. Hidden states of displacement for DIS 02: (a) baseline component; (b) speed component; (c) acceleration component; (d) daily cycle component;
(e) AR component.

18
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

5. Concluding marks

An approach based on the Bayesian dynamic linear model (BDLM) and Markov-switching theory is presented for
condition assessment of bridge expansion joints. To be specific, a weighted combination of BDLMs is performed to
estimate the expansion joint status at each time step. The states (normal and abnormal) are identified by detecting
the switching between models. Different from the previous literature that models the temperature-displacement rela-
tionship (TDR) to perform the condition assessment, the switching BDLM employs the displacement data itself without
depending on the quality of TDR. The switching BDLM also considers the prior probability of normal states and tran-
sition probability between different states. The temperature effect on the displacement is described by the superposi-
tion of harmonic components of BDLM. Although the Newton-Raphson (NR) approach can estimate parameters in the
switching BDLM, it is not easy to provide suitable initial parameter values that directly affect the modeling perfor-
mance. To mitigate such an issue, the EM algorithm initialized by the subspace method is utilized to optimize the
switching BDLM parameters.
The effectiveness of the switching BDLM for condition assessment of expansion joints is demonstrated by an illustra-
tive example and a real-world long-span bridge. The following observations are obtained: (1) the EM algorithm initial-
ized by the subspace method owns superiority over the NR method in terms of computational accuracy and efficiency;
(2) the switching BDLM offers the transition probability from the normal state to the deterioration state. The present
approach can identify a small shift of displacement (e.g., 10 mm in this study) and is robust towards false alarms in
a real application; and (3) the switching BDLM can decompose the displacement into different subcomponents. The state
change can be easily observed in the subcomponents, including zeroth-order (baseline), first-order trend (speed), second-
order (acceleration), and AR components. The daily and seasonal cycle subcomponents are less affected by the state
change of displacements whose periodic behavior is mainly induced by the external temperature. In summary, the
switching BDLM is applicable for the condition assessment of bridge expansion joints and offers a way to model its real
status over time. Future work regarding developing this switching BDLM approach for quantitative condition assessment
of bridge expansion joints is expected.

CRediT authorship contribution statement

Yi-Ming Zhang: Conceptualization, Investigation, Software, Methodology, Writing - original draft. Hao Wang: Supervi-
sion, Funding acquisition, Writing - review & editing. Yu Bai: Supervision, Writing - review & editing. Jian-Xiao Mao: Val-
idation, Writing - review & editing. Xiang-Yu Chang: Writing - review & editing. Li-Bin Wang: Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgments

The authors gratefully acknowledge the support from the National Natural Science Foundation of China (Grant
Nos. 51978155 and 51722804), the National Ten Thousand Talent Program for Young Top-notch Talents (Grant
No. W03070080), the Jiangsu Provincial Key Research and Development Program (Grant No. BE2018120), the Funda-
mental Research Funds for the Central Universities (2242020k1G013), and the Jiangsu Health Monitoring Data Center
for Long-Span Bridges. The first author would also like to acknowledge the support from the China Scholarship
Council (201906090073) and the Scientific Research Foundation of Graduate School of Southeast University (YBPY
2017).

Appendix A

A.1 Details of Kalman filter

The Kalman filter (KF) can be expressed as

ðmtjt ; V tjt ; V t;t1jT ; Lt Þ ¼ Filterðmt1jt1 ; V t1jt1 ; Y t ; F t ; Gt ; Rt ; Q t Þ ðA:1Þ
The KF can be divided into two steps: prediction and measurement steps. The prediction step is defined as follows.

mtjt1 ¼ Gt mt1 ; V tjt1 ¼ Gt V t1 GTt þ Q t ; ðA:2Þ

In the measurement step, the prediction error (e), innovation covariance matrix (H), the Kalman gain matrix (K), and the
likelihood (L) are firstly computed as

19
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

et ¼ Y t  F t mtjt1 ;
H t ¼ F t V tjt1 F >t þ Rt ;
ðA:3Þ
K t ¼ V tjt1 F >t H1
t ;
Lt ¼ Nðet ; 0; H t Þ:
Then, the mean, variance, and cross variance can be updated as
mtjt ¼ mtjt1 þ K t et ;
V tjt ¼ V tjt1  K t F t V tjt1 ; ðA:4Þ
V t;t1jt ¼ Gt V t1jt1  K t F t Gt V t - 1jt1

A.2 Details of Kalman smoother

The Kalman smoother (KS) can be written as

ðmtjT ; V tjT ; V tþ1;tjT Þ ¼ Smoothðmtþ1jT ; V tþ1jT ; mtjt ; V tjt ; V tþ1jtþ1 ; V tþ1;tjtþ1 ; Gtþ1 ; Q tþ1 Þ ðA:5Þ
The KS is initialized with the last step of the KF. From the KF, it can be obtained the following quantities

mtþ1jt ¼ Gtþ1 mtjt ; V tþ1jt ¼ Gtþ1 V tjt GTtþ1 þ Q tþ1 ðA:6Þ

The smoother gain matrix is calculated as

J t ¼ V tjt GTtþ1 V 1
tþ1jt ðA:7Þ

Finally, the mean, variance, and cross variance can be updated as

 
mtjT ¼ mtjt þ J t mtþ1jT  mtþ1jt ;
 
V tjT ¼ V tjt þ J t V tþ1jT  V tþ1jt J Tt ; ðA:8Þ
 
V tþ1;tjT ¼ V tþ1;tjtþ1 þ V tþ1jT  V tþ1jtþ1 V 1
tþ1jtþ1 V tþ1;t jtþ1

References

[1] E.J. Ancich, G.J. Chirgwin, S.C. Brown, Dynamic anomalies in a modular bridge expansion joint, J. Bridge Eng. 11 (5) (2006) 541–554.
[2] E. McCarthy, T. Wright, J.E. Padgett, R. DesRoches, P. Bradford, Development of an experimentally validated analytical model for modular bridge
expansion joint behavior, J. Bridge Eng. 19 (2) (2014) 235–244.
[3] R.J. Dexter, C.B. Osberg, M.J. Mutziger, Design, specification, installation, and maintenance of modular bridge expansion joint systems, J. Bridge Eng. 6
(6) (2001) 529–538.
[4] Y.Q. Ni, X.G. Hua, K.Y. Wong, J.M. Ko, Assessment of bridge expansion joints using long-term displacement and temperature measurement, J. Perform.
Constr. Facil 21 (2) (2007) 143–151.
[5] J. Marques Lima, J. de Brito, Inspection survey of 150 expansion joints in road bridges, Eng. Struct. 31 (5) (2009) 1077–1084.
[6] T. Guo, J. Liu, Y. Zhang, S. Pan, Displacement monitoring and analysis of expansion joints of long-span steel bridges with viscous dampers, J. Bridge Eng.
20 (9) (2015), 04014099.
[7] Z. Sun, Y. Zhang, Failure mechanism of expansion joints in a suspension bridge, J. Bridge Eng. 21 (10) (2016), 05016005.
[8] C.R. Farrar, K. Worden, Structural health monitoring: a machine learning perspective, John Wiley & Sons, 2012.
[9] J.L. Beck, K.-V. Yuen, Model selection using response measurements: Bayesian probabilistic approach, J. Eng. Mech. 130 (2) (2004) 192–203.
[10] J. Hu, L. Wang, X. Song, Z. Sun, J. Cui, G. Huang, Field monitoring and response characteristics of longitudinal movements of expansion joints in long-
span suspension bridges, Measurement 162 (2020), 107933.
[11] Y.Q. Ni, Y.W. Wang, C. Zhang, A Bayesian approach for condition assessment and damage alarm of bridge expansion joints using long-term structural
health monitoring data, Eng. Struct. 212 (2020), 110520.
[12] Q.i. Xia, Y. Xia, H.-P. Wan, J. Zhang, W.-X. Ren, Condition analysis of expansion joints of a long-span suspension bridge through metamodel-based
model updating considering thermal effect, Struct. Control Health Monit. 27 (5) (2020), e2521.
[13] H.-B. Huang, T.-H. Yi, H.-N. Li, H. Liu, New representative temperature for performance alarming of bridge expansion joints through temperature-
displacement relationship, J. Bridge Eng. 23 (7) (2018), 04018043.
[14] H.F. Zhou, Y.Q. Ni, J.M. Ko, Constructing input to neural networks for modeling temperature-caused modal variability: mean temperatures, effective
temperatures, and principal components of temperatures, Eng. Struct. 32 (6) (2010) 1747–1759.
[15] J. Mao, H. Wang, B.F. Spencer Jr., Toward data anomaly detection for automated structural health monitoring: Exploiting generative adversarial nets
and autoencoders, Struct. Health Monit. (2020), 1475921720924601.
[16] M. Trapp, C. Bogoclu, T. Nestorović, D. Roos, Intelligent optimization and machine learning algorithms for structural anomaly detection using seismic
signals, Mech. Syst. Sig. Process. 133 (2019), 106250.
[17] H. Sarmadi, A. Karamodin, A novel anomaly detection method based on adaptive Mahalanobis-squared distance and one-class kNN rule for structural
health monitoring under environmental effects, Mech. Syst. Sig. Process. 140 (2020), 106495.
[18] R. Yao, S.N. Pakzad, Autoregressive statistical pattern recognition algorithms for damage detection in civil structures, Mech. Syst. Sig. Process. 31
(2012) 355–368.
[19] M. Schlechtingen, I. Ferreira Santos, Comparative analysis of neural network and regression based condition monitoring approaches for wind turbine
fault detection, Mech. Syst. Sig. Process. 25 (5) (2011) 1849–1875.
[20] S.-C. Kuok, K.-V. Yuen, Model-free data reconstruction of structural response and excitation via sequential broad learning, Mech. Syst. Sig. Process. 141
(2020), 106738.
[21] R. Kromanis, P. Kripakaran, Performance of signal processing techniques for anomaly detection using a temperature-based measurement
interpretation approach, J. Civil Struct. Health Monit. 11 (1) (2021) 15–34.

20
Yi-Ming Zhang, H. Wang, Y. Bai et al. Mechanical Systems and Signal Processing 160 (2021) 107879

[22] H.-P. Wan, Y.-Q. Ni, Bayesian multi-task learning methodology for reconstruction of structural health monitoring data, Struct. Health Monit. 18 (4)
(2019) 1282–1309.
[23] Y.-M. Zhang, H. Wang, J.-X. Mao, Z.-D. Xu, Y.-F. Zhang, Probabilistic framework with Bayesian optimization for predicting typhoon-induced dynamic
responses of a long-span bridge, J. Struct. Eng. 147 (1) (2021), 04020297.
[24] R. Kromanis, P. Kripakaran, SHM of bridges: characterising thermal response and detecting anomaly events using a temperature-based measurement
interpretation approach, J. Civil Struct. Health Monit. 6 (2) (2016) 237–254.
[25] H.-P. Wan, Y.-Q. Ni, Bayesian modeling approach for forecast of structural stress response using structural health monitoring data, J. Struct. Eng. 144 (9)
(2018), 04018130.
[26] R. Moghaddass, S. Sheng, An anomaly detection framework for dynamic systems using a Bayesian hierarchical framework, Appl. Energy 240 (2019)
561–582.
[27] B. Pourbabaee, N. Meskin, K. Khorasani, Sensor fault detection, isolation, and identification using multiple-model-based hybrid Kalman filter for gas
turbine engines, IEEE Trans. Control Syst. Technol. 24 (4) (2016) 1184–1200.
[28] K.-V. Yuen, P.-F. Liang, S.-C. Kuok, Online estimation of noise parameters for Kalman filter, Struct. Eng. Mech. 47 (3) (2013) 361–381.
[29] Murphy KP. Switching kalman filters. 1998.
[30] Z. Ghahramani, G.E. Hinton, Switching state-space models, Citeseer (1996).
[31] X. Zeng, R. Ghanem, Dynamics identification and forecasting of COVID-19 by switching Kalman filters, Comput. Mech. 66 (5) (2020) 1179–1193.
[32] J.F. Smith, A. Pillai, K. Chen, B. Horwitz, Identification and validation of effective connectivity networks in functional magnetic resonance imaging using
switching linear dynamic systems, Neuroimage. 52 (3) (2010) 1027–1040.
[33] C.K.R. Lim, D. Mba, Switching Kalman filter for failure prognostic, Mech. Syst. Sig. Process. 52 (2015) 426–435.
[34] L.H. Nguyen, J.-A. Goulet, Anomaly detection with the switching kalman filter for structural health monitoring, Struct. Control Health Monit. 25 (4)
(2018), e2136.
[35] J.-A. Goulet, Bayesian dynamic linear models for structural health monitoring, Struct. Control Health Monit. 24 (12) (2017), e2035.
[36] Y.W. Wang, Y.Q. Ni, X. Wang, Real-time defect detection of high-speed train wheels by using Bayesian forecasting and dynamic model, Mech. Syst. Sig.
Process. 139 (2020), 106654.
[37] H. Wang, Y.-M. Zhang, J.-X. Mao, H.-P. Wan, T.-Y. Tao, Q.-X. Zhu, Modeling and forecasting of temperature-induced strain of a long-span bridge using an
improved Bayesian dynamic linear model, Eng. Struct. 192 (2019) 220–232.
[38] Y.W. Wang, Y.Q. Ni, Bayesian dynamic forecasting of structural strain response using structural health monitoring data, Struct. Control Health Monit.
27 (8) (2020), e2575.
[39] Z. Ma, C.B. Yun, Y.B. Shen, F. Yu, H.P. Wan, Y.Z. Luo, Bayesian forecasting approach for structure response prediction and load effect separation of a
revolving auditorium, Smart Struct. Syst. 24 (2019) 507–524.
[40] X. Fan, Y. Liu, Use of monitored daily extreme stress data for performance prediction of steel bridges: dynamic linear models and Gaussian mixed
particle filter, Mech. Syst. Sig. Process. 121 (2019) 841–855.
[41] Y. Liu, X. Fan, A data assimilation method about Bayesian Fourier dynamic linear prediction of periodic extreme stresses for steel bridges, Mech. Syst.
Sig. Process. 128 (2019) 82–92.
[42] Y.M. Zhang, H. Wang, H.P. Wan, J.X. Mao, Y.C. Xu, Anomaly detection of structural health monitoring data using the maximum likelihood estimation-
based Bayesian dynamic linear model 1475921720977020 Struct. Health Monit. (2020).
[43] J.-A. Goulet, K.i. Koo, Empirical validation of bayesian dynamic linear models in the context of structural health monitoring, J. Bridge Eng. 23 (2) (2018),
05017017.
[44] C.-J. Kim, Dynamic linear models with Markov-switching, J. Econom. 60 (1-2) (1994) 1–22.
[45] R.H. Shumway, D.S. Stoffer, Dynamic linear models with switching, J. Am. Stat. Assoc. 86 (415) (1991) 763–769.
[46] M. West, J. Harrison, Bayesian Forecasting and Dynamic Models, Springer Science & Business Media, 2006.
[47] J.A. Goulet, Probabilistic Machine Learning for Civil Engineers, MIT Press, 2020.
[48] Y. Bar-Shalom, X.R. Li, T. Kirubarajan, Estimation with Applications to Tracking and Navigation: Theory Algorithms and Software, John Wiley & Sons,
2004.
[49] K.P. Murphy, Machine Learning: a Probabilistic Perspective, MIT press, 2012.
[50] D. Barber, Bayesian Reasoning and Machine Learning, Cambridge University Press, 2012.
[51] B. Boots, G.J. Gordon, S.M.A. Siddiqi, Constraint generation approach to learning stable linear dynamical systems, in: Advances in Neural Information
Processing Systems, 2008, pp. 1329–1336.

21