You are on page 1of 11

This article has been accepted for publication in IEEE Internet of Things Journal.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1

Wireless IoT Monitoring System in Hong Kong–


Zhuhai–Macao Bridge and Edge Computing
for Anomaly Detection
Xiaoyou Wang†, Wanglin Wu†, Yao Du, Jiannong Cao, Fellow, IEEE, Qianyi Chen, and Yong Xia

Abstract—The emergence of the Internet of Things (IoT) has I. INTRODUCTION


facilitated the development and usage of low-computational
microcontrollers at the edge of the network, which process data in
the proximity of data sources and thereby offload the pressure of
R ECENT years have witnessed an increase in the
construction of long-span bridges worldwide [1]–[3].
Monitoring the health condition of bridges during their service
data transmission. Recently, IoT is becoming a key technology for
structural health monitoring systems. This study designs a novel life is crucial to ensure infrastructure safety and economic
wireless IoT monitoring system for the Hong Kong–Zhuhai– development. The traditional structural health monitoring
Macao Bridge, the world longest sea-crossing bridge. The 5G (SHM) system typically consists of wired sensors and a
technology and edge computing are integrated to improve the centralized data acquisition system, with sensors strategically
system performance in sensor serviceability, data transmission, placed to continuously monitor structural vibrations, loads, and
time synchronization, and data quality control. Artificial environmental conditions [4][5]. However, challenges may
intelligent (AI) algorithm is embedded into the NVIDIA Xavier
arise during the installation of wired sensors in specific
NX edge computing boards to preliminarily detect data anomalies
caused by sensor faults, before uploading the massive data to the
positions of long-span sea-crossing bridges. The wire
cloud platform. As training AI models requires a large amount of connection and high maintenance costs of lines remain
labeled data and is always time consuming, a novel data anomaly complex issues in the wired SHM system [6]. Moreover, the
detection method is developed by transferring the model trained sensor layout in the SHM design stage may not be optimal.
from the other bridge to the target bridge. Given that pre-storing Moving or adding sensors in the existing SHM system is costly.
source data in edge devices consumes expensive storage resources, The next-generation Internet of Things (IoT) is becoming a
the source-free domain adaptation is developed by integrating key technology for SHM system with the advances of sensor
robust self-training mechanism and self-knowledge distillation technologies [7]. The wireless sensing network (WSN) has
strategy. Thus the model transfer is achieved cross bridges in the
revolutionized the way engineers monitor bridges. Sensors can
absence of source data. This study provides a valuable and
practical reference for developing a wireless IoT structural health
be flexibly added, installed or relocated at selected locations to
monitoring system for large-scale infrastructure and enabling collect data, which are transmitted wirelessly to the cloud
edge computing for data anomaly detection with high efficiency platform. Hence, the efficiency of bridge monitoring can be
and accuracy. improved, and the cost can be reduced. Many studies on
developing WSN for SHM have been conducted [8][9][10].
Index Terms—Internet of Things, edge computing, structural However, in the conventional cloud computing, data producers
health monitoring, anomaly detection, source-free domain generate raw data and send them to the cloud, whereas data
adaptation. consumers send data processing requests and receive processed
results from the cloud. This unidirectional data flow is
unsuitable with the increasing data quantity at the edge,
considering that tremendous data will lead to huge unnecessary
bandwidth and computing resource requests [11]. SHM
This work was supported by the Key-Area Research and Development
Program of Guangdong Province (Project No. 2019B111106001) and systems generate huge amounts of data every data. As the
RGC-GRF (Project No. 15217522). (†These two authors contribute equally to measurement data may inevitably suffer from noise and sensor
the work; Corresponding author: Yong Xia.) faults [12][13], data cleaning prior to data transmission can
Xiaoyou Wang and Yong Xia are with the Department of Civil and improve the data quality of SHM systems and offload the
Environmental Engineering and the Research Institute for Artificial
Intelligence of Things and, The Hong Kong Polytechnic University, Hong pressure of data transmission.
Kong (e-mail: xiaoyou.wang@connect.polyu.hk; ceyxia@polyu.edu.hk) The advances of IoT and edge devices, along with the
Wanglin Wu and Yao Du are with the Department of Civil and evolution of artificial intelligence (AI), has boosted the edge
Environmental Engineering, The Hong Kong Polytechnic University, Hong
Kong (ceewanglin.wu@polyu.edu.hk;duyao.du@polyu.edu.hk) computing [11][14]. In contrast to traditional cloud computing,
Jiannong Cao and Qianyi Chen are with the Department of Computing and edge computing allows data computation on the edge devices,
the Research Institute for Artificial Intelligence of Things, The Hong Kong which can be regarded as the downstream data on behalf of the
Polytechnic University, Hong Kong (e-mail: csjcao@comp.polyu.edu.hk;
19074424r@connect.polyu.hk). cloud and the upstream data on behalf of the deployed edge
algorithms [11][15]. Edge devices act as both data consumers

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

and producers. Leveraging edge computing, some data pseudo-labels in source-free domain adaptation [30].
processing tasks are directly performed locally instead of In the context of edge computing, conducting source-free
sending requests and receiving contents from the cloud. As a domain adaptation for the anomaly detection of SHM data is of
result, the data transmission time and network complexity are great significance. The knowledge learned from other bridges
significantly reduced. AI algorithms can be embedded into the (i.e., source data) can be transferred to the target bridge for
edge devices to help pre-clean measurement data and detect the decision making. Only the model pre-trained source data need
anomalies. However, the supervised AI methods require to be stored in edge devices, without accessing the source data.
sufficient labeled data for model training, which is difficult for However, the research on this topic is currently lacking.
most engineering problems [16][17]. This study designs a novel wireless IoT SHM system for the
In application to civil engineering problems with limited and 55-km-long Hong Kong–Zhuhai–Macao Bridge (HZMB),
precious labeled data, performing cross-domain model transfer which is the longest sea-crossing bridge worldwide. Edge
in training AI models has been a trend [18][19]. The model computing and 5G technology are integrated to overcome the
learned from the labeled source domain can be intelligently data transmission, time synchronization, and data quality
transferred to the target domain through feature distribution control issues encountered in traditional WSNs. A novel data
alignment. The traditional domain adaptation methods require anomaly detection method is developed and embedded into the
the coexistence of source and target domain datasets, wherein edge devices for data pre-processing prior to transmission. The
the source datasets are generally in a large volume. Storing self-knowledge distillation is integrated with the
large amounts of source data is space-intensive and not self-supervised learning in the developed method to mitigate
cost-effective for the edge devices [20]. In addition, training AI overconfident pseudo-label predictions and improve model
models jointly with source and target datasets consumes large generalizability.
computational resources. Recently, source-free domain The remainder of this paper is organized as follows. Section
adaptation algorithms have been developed for computer vision II introduces the wireless IoT system designed for the HZMB.
tasks (e.g., image classification and segmentation) [21][22], Section III describes the developed source-free domain
where the source data are inaccessible during adaptation and adaptation algorithm for SHM data anomaly detection
only the model trained from source data is provided. embedded in the edge device. Section IV applies the proposed
Self-supervision is a main branch in source-free domain method to cross-bridge knowledge transfer and presents the
adaptation. Pseudo labeling [23], mutual information data anomaly detection results of the SHM data from the
maximization [24], entropy minimization [25], and contrastive HZMB. Section V draws the conclusion.
learning strategies [26] have been introduced to aid
self-training. For example, Liang et al. [24] proposed to II. WIRELESS SHM SYSTEM AND EDGE COMPUTING OF THE
maximize the mutual information across domains and applied QINGZHOU BRIDGE
the nearest prototype classifier to improve the accuracy of The HZMB consists of three cable-stayed bridges, a series of
pseudo-labels for self-supervision. Yang et al. [27] developed continuous viaducts, an undersea tunnel and four artificial
the neighborhood clustering regularization strategy to classify islands. The Qingzhou Bridge (Fig. 1), one of the three
target data by promoting label consistency among data with cable-stayed bridges, has a main span of 458 m and a total
high local affinity. Xia et al. [26] developed an adaptive length of 1,150 m. The steel box girder of the bridge is
adversarial network with a contrastive category-wise matching supported by two H-shaped towers, two auxiliary piers, and two
module. The main concern of these studies is to obtain robust transitional piers. The H-shaped tower consists of two concrete
representatives (e.g., pseudo-labels and prototypes) for domain legs, one concrete transom in the lower position, and one steel
adaptation. transom in the shape of a Chinese knot in the upper position
The limited computational capacity and memory space of [31][32].
low-computational microcontrollers (e.g., smartphones and
portable and edge devices) also pose challenges in the
deployment of deep complex AI models. Knowledge
distillation [28] is proposed to train a simplified small network
(i.e., student) to mimic the complex over-parameterized large
network (i.e., teacher) for knowledge transfer. Existing studies
have shown that the performance of student model may even
overpass the teacher model [28]. Recently, the knowledge Fig.1. Configuration of the Qingzhou Bridge
distillation technology has been improved to self-knowledge A. Wireless SHM system
distillation [29], where the model is taught to be its own teacher
A wired SHM system is designed to monitor the bridge’s
and the past predictions are used as soft labels for the next
structural responses at critical locations, environmental factors
prediction to mitigate the model collapse and improve the
(e.g., temperature, wind, humidity, and corrosion), and external
generalizability. Hence, in addition to compressing complex
loads (e.g., traffic) since the bridge opened to the public in 2018.
models, knowledge distillation can be utilized in the
The 3D layout of the wired SHM system is shown in Fig. 2,
self-distillation manner to enhance the accuracy of
which consists of 262 sensors, including hygro-thermometers,

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3

thermometers, anemometers, GPS rovers, accelerometers, One pyranometer facing up measures the direct solar radiation,
liquid leveling systems, displacement transducers, cable whereas the other one facing down measures the reflected solar
tensiometers, strain gauges, corrosion sensors, and reaction radiation from the sea. The thermal images measure the road
dynamometers. All sensors are connected to a central station to surface temperature directly. The system incorporates 5G, edge
store data in the centralized wired SHM system. computing, and efficient DL algorithms to solve data
Hong Kong
communication, time synchronization, decentralized
computing, and data quality control issues. The designed
wireless system can be regarded as a new paradigm for future
wireless SHM systems.
B. IoT paradigm demonstration
WSNs are prone to hardware or communication
Anemometer (4) malfunctions and failures, particularly in harsh environmental
Hygro-Thermometer (11)
Accelerometer (17)
conditions. Additionally, transmitting and synchronizing vast
GPS (3) amounts of data can be arduous in a centralized WSN. To
Liquid Leveling System (24)
Zhuhai Displacement Transducer (4) address these issues, the wireless SHM system for Qingzhou
Cable Tensiometer (24)
Strain Gauge (102)
Bridge incorporates emerging technologies such as 5G and
Thermometer (102) edge computing [33], which help alleviate data transmission
Corrosion Sensor (9)
Reaction Dynamometer (4) and network complexity challenges commonly faced by
Fig.2. Layout of wired sensors of the Qingzhou Bridge traditional wireless SHM systems.
The use of accelerometers serves as an example to illustrate
However, the wired SHM system has certain limitations. The
the full module composition within this advanced system and
primary concern is the high data processing load in the central
its workflow is illustrated in Fig. 4. To be specific, the
server. Besides, the installation and maintenance of wired
components include PCB393B31 accelerometer, HTeC-D3000
sensors are challenging in long-span sea-crossing bridges. For
data acquisition unit, Hongdian Z2 5G gateway, NVIDIA
example, the off-line data processing shows that a small
Xavier NX edge computing board, and Alibaba cloud platform.
number of sensors may be mis-connected to the data acquisition
Particularly, NVIDIA Xavier NX edge computing boards are
unit, causing difficulties in the later analysis results. In addition,
integrated to perform local data processing tasks, promising
the existing SHM system does have pyranometers to measure
data security, reducing data transmission load, and enabling
the solar radiation, which are necessary for obtaining the
timely responses in specific situations. The 5G timing module
temperature distribution of the bridge [31]. Installing and
is connected to each sensor locally to ensure time
connecting new sensors to the current SHM system are costly.
synchronization. Moreover, the adaptation of 5G
With the advances in wireless communication technology,
communication enhances data transmission speed and capacity,
WSNs have been used in some SHM systems with the
due to its low latency and broad bandwidth. All processed data
advantage of being easier in manipulation, more economic and
will be transmitted automatically to a central cloud platform for
efficient than the wired one. Wireless SHM systems can be
evaluation and long-term data storage, which is also more
designed in a decentralized manner, which reduces data
easily scalable and maintainable. Authorized users can
transmission and processing workloads. In this connection, this
configure the edge device remotely from the cloud.
study develops a novel wireless SHM system as a complement
to the above wired system, as displayed in Fig. 3.
Data transmission

Query

Cloud platform 5G gateway


...

Embed

Artificial
Sensor Data acquisition Edge device Intelligence
system

Fig. 4. Smart wireless accelerometer module

Edge computing serves as a prominent feature in the


Qingzhou Bridge's wireless SHM system, offering numerous
Fig.3. Layout of wireless sensors of the Qingzhou Bridge advantages by processing data at the network's edge prior to
transmission. This approach, in contrast to traditional cloud
The wireless system consists of 19 accelerometers, 2
computing, brings several key benefits that enhance overall
pyranometers, and 2 thermal imagers. The accelerometers are
system performance. First and foremost, edge computing
used to measure the vibration of the bridge deck and towers.
enables a faster response time, as data processing occurs closer

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4

to the source, reducing latency. This real-time processing using ANSYS, as shown in Fig. 7. The entire FE model consists
capability is crucial for critical applications in the wireless of 520,422 elements and 493,941 nodes. The solid elements are
SHM system, where timely decision-making and intervention used for the steel box girder, the asphalt concrete layer, tower
can prevent potential infrastructure damage or failure. Secondly, legs, lower tower transoms, and piers. The shell elements are
edge computing contributes to lower battery consumption and used for diaphragms, U-ribs, and upper tower transoms. In
bandwidth costs. By handling data processing at the network's particular, SOLID45, SOLID65, and SHELL181 are assigned
edge, the need for continuous data transmission to remote cloud to the solid and shell elements. Detailed information about the
servers is significantly reduced. This results in decreased power FE model is given in [31][32]. The first five frequencies and
usage and a more efficient utilization of available bandwidth, mode shapes of the FE model are calculated, and compared
ultimately leading to cost savings. Additionally, edge with the measured ones in Table I. The measured frequencies
computing enhances data safety and privacy. Localized data are very similar to the FE model analysis results, verifying the
processing minimizes the risk of data breaches or interception accuracy of the FE model and validating the serviceability and
during transmission, ensuring that sensitive information effectiveness of the designed wireless SHM system.
remains secure. The successful implementation of intelligent
edge computing in the Qingzhou Bridge's wireless SHM
system can be attributed to the NVIDIA Xavier NX, a widely
popular edge computing device for machine learning
applications. Its powerful processing capabilities, energy
efficiency, and compact design make it an ideal choice for
integrating advanced analytics and real-time monitoring into
the SHM system, ultimately contributing to the overall safety
and reliability of the bridge infrastructure.
C. Monitoring data and structural analysis
Structural vibration characteristics including frequencies and
mode shapes are widely used in SHM problems [34]. For
example, changes in the frequencies or mode shapes of a
structure before and after the damage occurs can be used to
Fig. 5. One-hour acceleration data from Sensors ACC1X and ACG1X
locate and quantity the damage [35]. They can also be used to
verify the accuracy of the numerical finite element model. If the
model is inaccurate, its frequencies and mode shapes may differ
from the measured counterparts. The model then needs to be
updated so that its frequencies and mode shapes match the
measures ones in an optimal manner. This technique is referred
to as model updating [36].
The modal analysis is conducted using the ARTeMIS Modal
Pro software based on the measurement data. Fig. 5 shows
one-hour time-domain acceleration data from the wireless
monitoring system. The frequency domain decomposition
method is used to calculate the spectral density matrix of the
measured acceleration data. Afterward, the singular value Fig. 6. Spectrum analysis based on acceleration data
decomposition is performed on the spectral density matrix to
approximately decompose the acceleration response into a set
of independent single-degree-of-freedom systems. The
self-spectral density of these single-degree-of-freedom systems
are calculated in the modal coordinates. The peaks in the
spectral density correspond to the natural vibration frequencies
of the structure, and the corresponding eigenvectors represent
structural vibration mode shapes. The spectral analysis result is
shown in Fig. 6. The peaks are distinct and significant, which
indicates the high quality of measurement data. The first five
frequencies and mode shapes are then estimated and shown in
Table I.
The modal analysis results are further compared with the
counterparts calculated from the finite element (FE) model. The Fig. 7. Refined FE model of the Qingzhou Bridge
3D refined FE model of the Qingzhou Bridge is established

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5

TABLE I
MODAL ANALYSIS AND COMPARISON
Resources First five modes

Analysis
based on
monitoring
data

(1) 0.269 Hz (2) 0.372Hz (3) 0.497 Hz (4) 0.577 Hz (5) 0.644 Hz

Analysis
based on
the FE
model

(1) 0.266 Hz (2) 0.379Hz (3) 0.446 Hz (4) 0.580 Hz (5) 0.648 Hz

as
III. EDGE COMPUTING FOR DATA ANOMALY DETECTION 𝐾

The measurement data may inevitably suffer from noise and ℒ𝑠𝑟𝑐 = −𝔼𝑥𝑠 𝜖𝑋𝑠 ∑ 𝑦𝑘 log 𝜎𝑘 (𝜙𝑠 (𝑥𝑠 )) (1)
anomalies. Based on the designed new wireless system, AI 𝑘=1

algorithms are embedded into the NVIDIA edge computing where 𝜙𝑠 (𝑥𝑠 ) = ℎ𝑠 (𝑓𝑠 (𝑥𝑠 )) is the output of network, 𝜎(∙)
board to detect data anomalies caused by sensor faults. denotes the softmax processing, 𝜎𝑘 (𝑧) = 𝑒 𝑧𝑘 ⁄∑𝐾 𝑧𝑖
𝑖=1 𝑒 denotes
Afterward, only data of good quality will be transmitted to the the kth element in the softmax output of a K-dimensional vector
central server. z, and 𝑦𝑘 is the one-of-K encoding vector where the kth element
In the context of edge computing, this study develops a equals 1 and the rest are all 0. However, such a training manner
source-free domain adaptation method for data anomaly makes the network prone to overfitting and shows weak
detection, which transfers the AI model trained from the other adaptation performance in cross-domain tasks. To address this
bridge with sufficient labeled data to the target Qingzhou issue, the label-smoothing technique [24] is introduced to
Bridge without any labeled data. The method mitigates the improve the model generalizability in down-scaling tasks and
problem of lacking labeled data in the traditional supervised further facilitate the source-free domain adaptation as specified
training and avoids the requirement of pre-storing source data later. Specifically, the original cross-entropy function is
in the edge computing in the traditional domain adaptation. The modified to add a uniform distribution 1/K over the
framework of the developed method is illustrated in Fig. 8. ground-truth one-hot label 𝑦𝑘 . The cross-entropy loss is then
Specific techniques are outlined in the following sections to aid calculated between the modified target label 𝑦̂𝑘 and network’s
the domain adaptation in the situation that source domain data softmax output as
𝐾
are unavailable.
𝑙𝑠
Source Bridge (Labeled) Feature extractor Classifier
ℒ𝑠𝑟𝑐 = −𝔼𝑥𝑠 𝜖𝑋𝑠 ∑ 𝑦̂𝑘 log 𝜎𝑘 (𝜙𝑠 (𝑥𝑠 )) (2)
Resnet-18 𝑘=1
F B C Lce
where 𝑦̂𝑘 = (1 − 𝛼)𝑦𝑘 + 𝛼/𝐾, and 𝛼 is the smoothing factor
C N
... with the default value of 0.1 [23]. The label-smoothing strategy
Initialization encourages the extracted features to be equidistantly and tightly
Lim
Student clustered according to categories [24].
F B
Target Bridge (Unlabeled) C N
C Lkd B. Information maximization for source-free domain
Lsl adaptation
EMA

... Teacher The traditional adversarial domain adaptation aims to match


F B C
the feature distributions of source and target domains (i.e.,
C N Without gradient
𝑓𝑠 (𝑥𝑠 ) and 𝑓𝑡 (𝑥𝑡 ) ), so the same classifier trained from the
Fig. 8. Framework of the developed method
labeled source domain can be applied to the unlabeled target
domain. However, this study supposes that the source data are
A. Source domain training not accessible in the domain adaptation and that only the model
The network, composed of a feature extractor 𝑓𝑠 and a learned from the source domain is available. Hence, the
classifier ℎ𝑠 , is first trained by the labeled source domain data. essential challenge is how to adapt the feature distributions of
The commonly used cross-entropy loss function is formulated source and target domains in the absence of source data.

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

Following the idea of adversarial domain adaptation, the where 𝛿(∙,∙) measures the cosine distance between two items,
source and target domain are designed to share the same ‖∙‖2 denotes the L2-norm. In the ith iteration, the class
classifier to help match the feature distributions. That is, the centroids and pseudo-labels are then updated as follows:
classifier for the target domain is directly cloned from the
(𝑖) ∑𝑥 𝜖𝑋 𝜉(𝑦̃𝑡(𝑖−1) = 𝑘) 𝑓̃𝑡 (𝑥𝑡 )
source domain and kept fixed ( ℎ𝑡 = ℎ𝑠 ). Only the feature 𝜇𝑘 = 𝑡 𝑡 (8)
extractor is updated (𝑓𝑡 ≠ 𝑓𝑠 ) during the adaptation. As the ∑𝑥𝑡𝜖𝑋𝑡 𝜉(𝑦̃𝑡(𝑖−1) = 𝑘)
(𝑖)
classifier has been trained to classify the source domain whose (𝑖) 𝑓̃𝑡 (𝑥𝑡 ) ∙ 𝜇𝑘
labels are represented by the one-hot encoding, if the feature 𝑦̃𝑡 = arg min (1 − (𝑖)
) (9)
𝑘 ‖𝑓̃𝑡 (𝑥𝑡 )‖2 ‖𝜇𝑘 ‖2
distributions of the source and target domains have been
matched, the predicted labels of the target domain should be where 𝜉(∙) is an indicator that equals 1 when the argument is
close to one-hot encoding and discriminative across different true. Eqs. (3) and (4) are iterated for multiple rounds. With the
classes. In corresponding, the loss function ℒ𝑖𝑚 consists of two self-estimated pseudo-labels, the loss function is given by:
𝐾
items, ℒ𝑒𝑛𝑡 and ℒ𝑑𝑖𝑣 . Minimizing ℒ𝑒𝑛𝑡 impels the softmax
output to approach the one-hot encoding label, and minimizing ℒ𝑠𝑙 = −𝔼𝑥𝑡 𝜖𝑋𝑡 ∑ 𝑦̃𝑡 log 𝜎𝑘 (𝜙𝑡 (𝑥𝑡 )) (10)
ℒ𝑑𝑖𝑣 promotes the diversity of the predicted labels. The loss 𝑘=1

functions are formulated as follows: D. Self-knowledge distillation


𝐾
Self-knowledge distillation is further utilized to mitigate
ℒ𝑒𝑛𝑡 = −𝔼𝑥𝑡𝜖𝑋𝑡 ∑ 𝜎𝑘 (𝜙𝑡 (𝑥𝑡 )) log 𝜎𝑘 (𝜙𝑡 (𝑥𝑡 )) (3) overfitting or overconfident predictions and improve the model
𝑘=1 generalizability. The teacher and student models are designed
𝐾
with the same architecture, which are initialized based on the
ℒ𝑑𝑖𝑣 = ∑ 𝑝̅𝑘 log 𝑝̅𝑘 (4)
parameters learned from the source domain. In the training
𝑘=1
stage, the teacher model is used to estimate the pseudo-labels of
ℒ𝑖𝑚 = ℒ𝑒𝑛𝑡 + ℒ𝑑𝑖𝑣 the unlabeled target domain. The pseudo-labels are then used to
(5)
calculate the training loss, based on which the student model is
optimized. Afterward, the parameters of the teacher model are
where 𝑝̅𝑘 = 𝔼𝑥𝑡 𝜖𝑋𝑡 [𝜎𝑘 (𝜙𝑡 (𝑥𝑡 ))] is the mean output of the
updated by referring to the student model using an exponential
entire target domain.
moving average (EMA) strategy [37]. Hence, the pseudo-labels
C. Self-supervised prototype learning are estimated as
(𝑖)
The information maximization technique helps classify the exp[𝛿(𝑓𝑡𝑇𝑒𝑎 (𝑥𝑡 ), 𝜇𝑘 )⁄𝑇]
target domain through feature distribution alignment. However, 𝑦̅𝑡 = (11)
𝑇𝑒𝑎 (𝑥 ), (𝑖) ⁄
∑𝐾𝑘=1 exp[𝛿(𝑓𝑡 𝑡 𝜇𝑘 ) 𝑇 ]
a few target data may be forcibly driven to the false category
and then significantly deviate from the true label [23]. For where T denotes the temperature scaling parameter, and the
example, the target data from the first category may have the overconfident estimations are mitigated when 𝑇 > 1. The loss
network output [0.3, 0.4, 0.2, 0.1] during training, but will be function is then estimated as
𝐾
forced to approach [0, 1, 0, 0] with the information
maximization strategy. To alleviate this problem, ℒ𝑘𝑑 = −𝔼𝑥𝑡 𝜖𝑋𝑡 ∑ 𝑦̅𝑡 log 𝜎𝑘 (𝜙𝑡 (𝑥𝑡 )) (12)
𝑘=1
self-supervised prototype learning is additionally introduced to
further improve the accuracy of pseudo-labels and supervise E. Summary
the target data encoding training. The prototype (i.e., centroid) The overall model, consisting of a feature extractor 𝑓𝑠 and a
is less sensitive to outliers and can better characterize the classifier ℎ𝑠 , is first trained using the source domain data to
distribution of different categories, thus improving the minimize Eq. (2). Following, the classifier is fixed (ℎ𝑡 = ℎ𝑠 ),
robustness of the classifier. The class-wise prototype in the and the feature extractor 𝑓𝑠 is used to initialize the feature
target domain is firstly calculated following the idea of extractors 𝑓𝑡𝑆𝑡𝑢 and 𝑓𝑡𝑇𝑒𝑎 in the student and teacher models for
weighted k-means clustering: the target domain. The teacher model is used to estimate the
(0)
∑𝑥𝑡𝜖𝑋𝑡 𝜎𝑘 (𝜙̃𝑡 (𝑥𝑡 )) 𝑓̃𝑡 (𝑥𝑡 ) pseudo-labels. Afterward, 𝑓𝑡𝑆𝑡𝑢 is updated to minimize the loss
𝜇𝑘 = (6)
∑𝑥𝑡𝜖𝑋𝑡 𝜎𝑘 (𝜙𝑡 (𝑥𝑡 )) function ℒ, which is formulated as
where 𝜙̃𝑡 = 𝑓̃𝑡 ∘ ℎ𝑡 denotes the model learned in the previous ℒ = ℒ𝑖𝑚 + 𝜆ℒ𝑠𝑙 + 𝛾ℒ𝑘𝑑 (13)
iteration with ℎ𝑡 unchanged. In the beginning, 𝜙̃𝑡 is the model
trained from the source data. The pseudo-label is then estimated where 𝜆 and 𝛾 are the trade-off parameters of self-supervised
via the classifier with the nearest distance: pseudo-labels and self-knowledge distillation pseudo-labels,
(0) (0)
𝑦̃𝑡 = arg min (𝛿(𝑓̃𝑡 (𝑥𝑡 ), 𝜇 )) 𝑘
respectively. 𝑓𝑡𝑇𝑒𝑎 is then updated with an EMA of parameters
𝑘
(0) in 𝑓𝑡𝑆𝑡𝑢 . The teacher model has no gradient backpropagation, as
𝑓̃𝑡 (𝑥𝑡 ) ∙ 𝜇𝑘 (7)
= arg min (1 − ) indicated in Fig. 8. The classifier remains fixed during the
(0)
𝑘 ‖𝑓̃𝑡 (𝑥𝑡 )‖2 ‖𝜇𝑘 ‖2 training stage.

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7

IV. CROSS-BRIDGE MODEL TRANSFER caused by sensor fault, as shown in Fig. 10. The descriptions of
all patterns are specified in Table II. The quantities of each
A. Source bridge
pattern in the source training and test data are given in Table III.
A long-span cable-stayed bridge in China provided by The The training dataset has a total of 28,272 samples and the test
First International Project Competition is used as the source dataset has 26,448 samples in total.
bridge [38], as shown in Fig. 9. The SHM system of the bridge
consists of multiple sensors. Only acceleration data are used in
this study for anomaly detection. The data from a total of 38
channels are used with a sampling frequency of 20 Hz. The
measurement data are divided into hourly segments without
overlapping. One-month data (January 1–31, 2021) are used as
the source training dataset. The data in February 2012
(February 1–29, 2021) are used as a blind dataset to test the
model performance. Fig. 10. Data patterns in the source bridge

TABLE II
DESCRIPTION OF THE ANOMALY PATTERNS
Patterns Descriptions
Normal Data are collected in the normal state of sensors
Missing Data are absent with the magnitude of zero or a fixed value
Minor Data have a quite tiny amplitude relative to the normal data
Outlier One or more outliers appears in the dataset
Fig. 9. Sensor layout of the source bridge Square Data oscillates violently to a fixed magnitude
The entire source training and test dataset contains seven Trend Data are non-stationary with monotonous trend
data patterns, including the normal pattern and six abnormal Drift Data are non-stationary with random drift
patterns (i.e., missing, minor, outlier, square, trend, drift)

TABLE III
SOURCE TRAINING AND TEST DATASET
Datasets Patterns Normal Missing Minor Outlier Square Trend Drift Total
Quantity 13,575 2,942 1,775 527 2,996 5,778 679 28,272
Training
(%) 48.02% 10.41% 6.28% 1.86% 10.60% 20.44% 2.40% 100%
Quantity 12,897 2,967 1,650 331 3,214 4,558 831 26,448
Test
(%) 48.76% 11.22% 6.24% 1.25% 12.15% 17.23% 3.15% 100%

The data samples are converted into grayscale figures and where TP, FP, FN, and TN denote the true positive, false
used as the input to the feature extractor 𝑓𝑠 , which adopts the positive, false negative, and true negative, respectively.
pre-trained ResNet18 model. As the gray image has only one Accuracy represents the correct prediction of the entire datasets.
channel, rather than three channels of RGB images, an For each category, precision and recall are defined, where the
additional convolutional layer with one input channel is added former equals the ratio of correctly predicted positives over all
before the ResNet18 module. Besides, an additional batch positives, and the latter equals the ratio of correctly predicted
normalization layer and a fully connected layer are placed after positives over all samples of this category.
the ResNet18 module, as illustrated in Fig. 8. The classifier ℎ𝑠 The training accuracy over epochs is plotted in Fig. 11(a).
is composed of a fully connected layer and a weight The well-trained model is applied to the blind test dataset for
normalization layer, with the final output size of seven. As performance validation. The confusion matrix of the test
introduced previously, the label-smoothing technique is dataset is plotted in Fig. 11(b). The precision and recall of each
employed to train the model through back-propagation. The category are calculated in Table IV. The precision and recall of
network is updated by using the stochastic gradient descent the outlier and drift patterns are relatively low compared with
with a momentum of 0.9 and weight decay of 10 –3. The learning the rest patterns, which is probably due to the insufficient
rate of the feature extractor and the classifier is set to 10–3 and training samples and muti-pattern anomalies [39]. Nevertheless,
10–2, respectively. The batch size is 76 in the source dataset and the model achieves an acceptable overall accuracy in
the epoch is set to 50. Several indices are used to evaluate the classifying data anomalies. The well-trained model is used in
model performance, which are defined as the following section for domain adaptation.
𝑇𝑃 + 𝑇𝑁 TABLE IV
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (14) CLASSIFICATION RESULTS
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 Patterns Normal Missing Minor Outlier Square Trend Drift
𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (15) Precision 99.0% 99.8% 94.6% 80.4% 99.7% 97.6% 83.2%
𝑇𝑃 + 𝐹𝑃 Recall 99.3% 99.9% 95.6% 79.5% 98.8% 96.3% 87.1%
𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = (16) Accuracy 98%
𝑇𝑃 + 𝐹𝑁

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 8

The missing pattern has both high precision and recall, and the
remaining patterns either have low precision or recall. The
performance of the trained anomaly detection model degrades
when applied to the new bridge.
The proposed source-free domain adaptation method is then
applied. The hyper-parameters in Eq. (13) are set to 𝜆 = 0.3
and 𝛾 = 0.1. The temperature scaling parameter in Eq. (11) is
set to 𝑇 = 2. The epoch is set to 30. The accuracy over epochs
is plotted in Fig. 14. The blue marker denotes the model
(a) accuracy in each epoch. The red marker denotes the accuracy
by further updating the predicted labels using the
self-supervised prototype learning and self-knowledge
distillation. In each epoch, the model parameters remain
unchanged in the step of estimating red marker, but the
obtained pseudo-labels will participate in the model training in
the next epoch, as defined in Eqs. (10) and (12). In general,
self-supervised learning and self-knowledge distillation
improve the accuracy of the obtained labels. After training, the
obtained model is applied to the target datasets, and the results
are shown in Fig. 13(b) and Table VI. Compared with using the
(b)
Fig. 11. Source model training. (a) Accuracy over epochs; (b) Confusion matrix
model trained from source bridge directly, the cross-domain
of source test dataset. adapted model achieves more accurate classification results,
validating the effectiveness of the developed method.
B. Target bridge
The target bridge is the Qingzhou Bridge of the HZMB.
Similar to the source bridge, only the acceleration data are
studied. There are five uniaxial accelerometers and seven
biaxial accelerometers in total, as shown in Fig. 3, resulting in a
total of (5+7×2) = 19 channels. The sampling frequency is 50
Hz. One-month data collected from February 1–28, 2023 are
studied. Similarly, data samples are generated based on the
non-overlapped hourly responses. Data anomalies occur in four
channels only in February, that is, two biaxial accelerometers
installed on the tower. Hence, only acceleration data from these (a)
four channels form the target dataset for data anomaly detection.
Five patterns are plotted in Fig. 12. The specific dataset
information is given in Table V.

normal missing outlier

(b)
Fig. 13. Confusion matrix of the target dataset. (a) Without adaptation; (b) After
trend drift adaptation.
Fig. 12. Data patterns in the HZMB
TABLE V
TARGET DATASET
Dataset Patterns Normal Missing Outlier Trend Drift Total
Quantity 2,450 57 24 118 39 2,688
Target
Percentage 91.15% 2.12% 0.89% 4.39% 1.45% 100%

The model trained from the source dataset is first directly


applied to the target dataset. The confusion matrix is shown in
Fig. 13(a). The accuracy of the entire target dataset is 58%, with
the precision and recall of each category specified in Table VI. Fig. 14. Variation of accuracies over epochs.

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9

TABLE VI D. Further analysis


CLASSIFICATION RESULTS BEFORE AND AFTER DOMAIN ADAPTATION
Methods Patterns Normal Missing Outlier Trend Drift
Despite the improvement in anomaly detection accuracy,
Precision 96.3% 95% 3.6% 30.7% 2.9% some samples are misclassified. For example, 165 samples with
Without
Recall 57% 100% 95.8% 40% 23.1% the normal pattern are misclassified as outlier and 85 samples
adaptation
Accuracy 57% are misclassified as trend. The misclassified samples are
Precision 98.7% 100% 9.5% 49.2% 65% compared with the ground-truth samples. Fig. 16 gives the
After
Recall 89.2% 100% 75% 76.3% 38.5%
adaptation misclassified examples that are supposed to be normal data.
Accuracy 88%
The peak in the measurement data caused by passing vehicles is
C. Comparative study misclassified as outlier. The reason may be that the traffic
Several state-of-the-art methods are selected for the volume on the source bridge within an hour was high, but the
comparative study. The algorithms are described as follows: traffic volume on the HZMB was relatively low in February
a) Source-only. The ResNet18 model [40] without any 2023. Besides, when manually labeling the target datasets, only
knowledge transfer is selected as the baseline model, that is, samples with a severe trend are labeled as trend, and the rest
the model trained from the source bridge is directly applied with a slight trend are labeled as normal. As a result, such target
to the target bridge for anomaly detection. The method is patterns are not seen (in other words, such knowledge is not
referred to as source-only. learned) in the source bridge and are misclassified as abnormal
b) 3C-GAN. Li et al. [41] developed a Collaborative Class patterns that are more similar in appearance.
Conditional Generative Adversarial Network (3C-GAN) to
match the target distribution with the unseen source data
through adversarial learning. The objective function
incorporates the clustering-based regularization item to
generate the decision boundary in the low-density region.
c) SHOT. Source HypOthesis Transfer (SHOT) is proposed by
(a) (b)
Liang et al. [23] for unsupervised domain adaptation in the
absence of source data and only a trained source model is
available. SHOT has been widely used in many cases
including closed-set, partial-set and open-set adaptation.
d)w/o IM. The ablation study is conducted to investigate the
contribution of information maximization mechanism by
removing ℒ𝑖𝑚 from the right-hand-side of Eq. (13). The (c) (d)
method is referred to as without Information Maximization Fig. 16. Misclassified patterns and the ground-truth samples. (a) Outlier; (b)
(w/o IM). Normal data misclassified as outlier; (c)Trend; (d)Normal data misclassified as
Trend.
e) w/o PL. The model performance without self-supervised
Prototype Learning (w/o PL) is also investigated by setting Considering that a limited number of anomalies may be
collected and labeled in the bridge’s initial operation stage, the
𝜆 = 0 in Eq. (13).
source dataset is updated by adding a small number of manually
Fig. 15 compares the accuracies of six different methods.
labeled target data samples (i.e., few-shot samples).
The accuracy is the averaged results of five implementations.
Specifically, 50 normal samples misclassified as outlier, 50
For the easy comparison, the ResNet18 pre-trained on
normal samples misclassified as trend, 20 trend samples
ImageNet [42] is employed as the backbone module of the
misclassified as normal, and 10 drift samples misclassified as
feature extractor in all algorithms. Fig.15 demonstrates that the
minor are labeled with their ground-truth labels and then added
proposed method has the highest accuracy. Besides, the
into the source dataset. The developed source-free domain
accuracy decreases significantly after removing information
adaptation method is similarly applied. The classification result
maximization (from 88% to 64%) or prototype learning (from
and the t-distributed stochastic neighbor embedding (t-SNE)
88% to 79%), validating the effectiveness of these two
[43] feature visualization result are shown in Figs. 17(a) and (b).
mechanisms. In summary, our method is superior to the other
The precision, recall and accuracy indices are listed in Table
five methods in the data anomaly detection task.
VII. As the model is pre-taught to learn the related knowledge
from the few-shot samples, the performance is further
improved compared with the results in Table VI. Hence, with
more labeled data samples collected in the future, the model
will be continuously improved.
TABLE VII
CLASSIFICATION RESULTS AFTER IMPROVEMENT
Method Patterns Normal Missing Outlier Trend Drift
Precision 99.2% 100% 26.7% 78.9% 88.5%
After
Recall 96.2% 100% 83.3% 89% 59%
Fig. 15. Accuracy comparison of different methods in anomaly detection task. improvement
Accuracy 95%

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 10

[5] Y. Ni, Y. Xia, W. Liao, and J. Ko, "Technology innovation in developing


the structural health monitoring system for Guangzhou New TV Tower,"
Struct. Health Monit., vol. 16, no. 1, pp. 73-98, 2009.
[6] J. P. Lynch, and K. J. Loh, “A summary review of wireless sensors and
sensor networks for structural health monitoring,” Shock vibr. dig., vol.
38, no. 2, pp. 91-130, 2006.
[7] M. Mishra, P. B. Lourenço, and G. V. Ramana, “Structural health
monitoring of civil engineering structures by using the internet of things:
A review,” J. Build. Eng., vol. 48, pp. 103954, 2022.
[8] M. Abdulkarem, K. Samsudin, F. Z. Rokhani, and M. F. A Rasid,
"Wireless sensor network for structural health monitoring: A
contemporary review of technologies, challenges, and future direction,"
Struct. Health Monit., vol. 19, no. 3, pp. 693-735, 2020.
(a) [9] A. Sofi, J. J. Regita, B. Rane, and H. H. Lau, "Structural health
monitoring using wireless smart sensor network–An overview," Mech
Syst Signal Process, vol. 163, p. 108113, 2022.
[10] Y. Wang, J. P. Lynch, and K. H. Law, “A wireless structural health
monitoring system with multithreaded sensing devices: design and
validation,” Struct. Infrastruct. Eng., vol. 3, no. 2, pp. 103-120, 2007.
[11] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, "Edge computing: Vision and
challenges," IEEE Internet Things J., vol. 3, no. 5, pp. 637-646, 2016.
[12] Y. Bao, Z. Tang, H. Li, and Y. Zhang, "Computer vision and deep
learning–based data anomaly detection method for structural health
monitoring," Struct. Health Monit., vol. 18, no. 2, pp. 401-421, 2019.
[13] Y. Zhang, X. Wang, Z. Ding, Y. Du, and Y. Xia, "Anomaly detection of
sensor faults and extreme events based on support vector data
(b) description," Struct. Control Health Monit., vol. 29, no. 10, p. e3047,
Fig. 17. Model performance improved by adding labeled target samples. (a) 2022.
Updated results; (b) t-SNE visualization. [14] G. Premsankar, M. Di Francesco, and T. Taleb, "Edge computing for the
Internet of Things: A case study," IEEE Internet Things J., vol. 5, no. 2,
pp. 1275-1284, 2018.
V. CONCLUSIONS [15] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, "Edge
This study develops a novel wireless IoT SHM system for intelligence: Paving the last mile of artificial intelligence with edge
computing," Proc. IEEE, vol. 107, no. 8, pp. 1738-1762, 2019.
the HZMB. A 5G timing module is connected to each sensor [16] J. Jiao, M. Zhao, J. Lin, and K. Liang, "A comprehensive review on
locally to ensure time synchronization, and the NVIDIA Xavier convolutional neural network in machine fault diagnosis,"
NX edge computing board is integrated to detect abnormal data Neurocomputing, vol. 417, pp. 36-63, 2020.
[17] L. Sun, Z. Shang, Y. Xia, S. Bhowmick, and S. Nagarajaiah, "Review of
at the edge of network to reduce the data transmission pressure. bridge structural health monitoring aided by big data and artificial
A novel data anomaly detection method is developed through intelligence: From condition assessment to damage detection," J. Struct.
cross-bridge knowledge transfer. Different from the traditional Eng., no. 5, p. 04020073, 2020.
methods that require the coexistence of source and target [18] P. Gardner, X. Liu, and K. Worden, "On the application of domain
adaptation in structural health monitoring," Mech Syst Signal Process, vol.
datasets for model adaptation, the source-free domain 138, p. 106550, 2020.
adaptation is deployed to reduce data storage consumption of [19] X. Wang and Y. Xia, "Knowledge transfer for structural damage
edge devices. The developed method is tested on the one-month detection through re-weighted adversarial domain adaptation," Mech Syst
Signal Process, vol. 172, p. 108991, 2022.
monitoring data from the HZMB. The results show that the [20] Y. Zhang, Z. Ren, K. Feng, K. Yu, M. Beer, and Z. Liu, "Universal
deployed self-learning and self-knowledge distillation source-free domain adaptation method for cross-domain fault diagnosis
strategies improve the accuracy of estimated pseudo-labels. of machines," Mech Syst Signal Process, vol. 191, p. 110159, 2023.
[21] Y. Liu, W. Zhang, and J. Wang, "Source-free domain adaptation for
Compared with directly using the model trained from other semantic segmentation," in Proc IEEE Comput Soc Conf Comput Vis
bridges, the cross-bridge adaptation obviously improves the Pattern Recognit, 2021, pp. 1215-1224.
model performance. Overall, data anomaly detection integrated [22] J. N. Kundu, N. Venkat, and R. V. Babu, "Universal source-free domain
adaptation," in Proc IEEE Comput Soc Conf Comput Vis Pattern
with edge computing offers an excellent wireless IoT system Recognit, 2020, pp. 4544-4553.
for health monitoring of long-span bridges. [23] J. Liang, D. Hu, and J. Feng, "Do we really need to access the source data?
source hypothesis transfer for unsupervised domain adaptation," in Conf.
Mach. Learn., 2020, pp. 6028-6039.
REFERENCES [24] J. Liang, D. Hu, Y. Wang, R. He, and J. Feng, "Source data-absent
[1] W. Huang, M. Pei, X. Liu, and Y. Wei, "Design and construction of unsupervised domain adaptation through hypothesis transfer and labeling
super-long span bridges in China: Review and future perspectives," Front. transfer," IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 11, pp.
Struct. Civ. Eng., vol. 14, pp. 803-838, 2020. 8602-8617, 2021.
[2] N. Memisoglu Apaydin, A. C. Zulfikar, and O. Cetindemir, "Structural [25] J. Jiao, H. Li, T. Zhang, and J. Lin, "Source-Free Adaptation Diagnosis
health monitoring systems of long-span bridges in Turkey and lessons for Rotating Machinery," IEEE Trans. Industr. Inform., 2022.
learned from experienced extreme events," J. Civ. Struct. Health Monit., [26] H. Xia, H. Zhao, and Z. Ding, "Adaptive adversarial network for
vol. 12, no. 6, pp. 1375-1412, 2022. source-free domain adaptation," in Proc. IEEE Int. Conf. Comput. Vis.,
[3] Y. Fujino, D. M. Siringoringo, Y. Ikeda, T. Nagayama, and T. Mizutani, 2021, pp. 9010-9019.
"Research and implementations of structural monitoring for bridges and [27] S. Yang, Y. Wang, J. Van De Weijer, L. Herranz, and S. Jui, "Generalized
buildings in Japan," Engineering, vol. 5, no. 6, pp. 1093-1119, 2019. source-free domain adaptation," in Proc. IEEE Int. Conf. Comput. Vis.,
[4] Y. L. Xu, and Y. Xia, Structural health monitoring of long-span 2021, pp. 8978-8987.
suspension bridges, CRC Press, 2011. [28] J. Gou, B. Yu, S. J. Maybank, and D. Tao, "Knowledge distillation: A
survey," Int. J. Comput. Vis., vol. 129, pp. 1789-1819, 2021.

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3300073

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 11

[29] S. Yun, J. Park, K. Lee, and J. Shin, "Regularizing class-wise predictions


via self-knowledge distillation," in Proc IEEE Comput Soc Conf Comput
Vis Pattern Recognit, 2020, pp. 13876-13885.
[30] G. Yang et al., "Transformer-based source-free domain adaptation," arXiv
preprint arXiv, 14138, 2021.
[31] Y.S. Shan, L.F. Li, Q. Xia, W.B. Gao, Q. Jing, Y. Xia, “Temperature
behavior of cable-stayed bridges. Part I — global 3D temperature
distribution by integrating heat-transfer analysis and field monitoring
data,” Adv. Struct. Eng., vol. 26, no. 9, pp. 1579-1599, 2023.
[32] Y.S. Shan, Q. Jing, L.F. Li, W.B. Gao, Z.L. Xia, Y. Xia, “Temperature
behavior of cable-stayed bridges. Part II — temperature actions by using
unified analysis”, Adv. Struct. Eng., vol. 26, no. 9, pp. 1600-1620, 2023.
[33] Q. Chen, J. Cao, and Y. Xia, "Physics-enhanced pca for data compression
in edge devices," IEEE Trans. Green Commun. Netw., vol. 6, no. 3, pp.
1624-1634, 2022.
[34] X. Wang, L. Li, J. L. Beck, and Y. Xia, “Sparse Bayesian factor analysis
for structural damage detection under unknown environmental
conditions,” Mech Syst Signal Process, vol. 154, pp. 107563, 2021
[35] X. Wang, R. Hou, Y. Xia, and X. Zhou, “Structural damage detection
based on variational Bayesian inference and delayed rejection adaptive
Metropolis algorithm,” Struct. Health Monit., vol. 20, no. 4, pp.
1518-1535, 2021.
[36] M. Friswell, J.E. Mottershead, “Finite element model updating in
structural dynamics,” Springer Science & Business Media, vol.38, 1995.
[37] Z. Cai, A. Ravichandran, S. Maji, C. Fowlkes, Z. Tu, and S. Soatto,
"Exponential moving average normalization for self-supervised and
semi-supervised learning," in Proc IEEE Comput Soc Conf Comput Vis
Pattern Recognit, 2021, pp. 194-203.
[38] Y. Bao, J. Li, T. Nagayama, Y. Xu, B. F. Spencer Jr, and H. Li, "The 1st
international project competition for structural health monitoring
(IPC-SHM, 2020): a summary and benchmark problem," Struct. Health
Monit., vol. 20, no. 4, pp. 2229-2239, 2021.
[39] Y. Du, L.F. Li, R.R. Hou, X.Y. Wang, W. Tian, and Y. Xia,
"Convolutional neural network-based data anomaly detection considering
class imbalance with limited data," Smart Struct Syst., vol. 29, no. 1, pp.
63-75, 2022.
[40] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition.,” in CVPR, pp. 770–778, 2016.
[41] R. Li, Q. Jiao, W. Cao, H.-S. Wong, and S. Wu, “Model adaptation:
Unsupervised domain adaptation without source data,” in Proc. CVPR,
Jun. 2020, pp. 9641–9650.
[42] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A
large-scale hierarchical image database,” in CVPR, 2009, pp. 248–255.
[43] L. Van der Maaten, G. Hinton, “Visualizing data using t-SNE,” J Mach
Learn Res., vol. 29, no. 1, 2008.

Authorized licensed use limited to: UNIV OF ALABAMA-TUSCALOOSA. Downloaded on September 15,2023 at 12:03:04 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.

You might also like