Intelligent damage diagnosis in bridges using vibration-based monitoring approaches and machine learning: A systematic review

1. Introduction

Bridge structures play paramount contributions in fostering the economic growth of a nation and facilitating daily life for the world population. During its intended service life deteriorates progressively due to being susceptible to environmental and operational conditions [1,2]. Based on the American Association of State Highway and Transportation Officials Load Factor Resistance Design, the design life of a bridge structure is 75 years [3]. Nowadays, the world's focus shifted from building new bridges to sustaining old ones using emerging technologies, resources, and progressive construction techniques to meet this challenge. Damage diagnosis of bridge structures has attracted the interest of scientiﬁc researchers [4,5]. Achieving this aim requires the early diagnosis of the deteriorated bridge by integrating vibration-based damage detection (VBDD) with machine learning (ML) approaches to enhance its performance [6]. These approaches reduce the rehabilitation cost and increase the life cycle of the structures and structural integrity [7]. According to the 2021 America infrastructure report card, 7.5% of America's 617,000 bridges are structurally vulnerable. Bridge repairs cost $125 billion, whereas rehabilitation costs rose from $14.4 to $22.7 billion annually, causing economic loss [8].

Based on numerous literature, Vibration-based damage detection is a subset of vibration-based monitoring and machine learning is a subset of artificial intelligence which plays a significant role in providing fruitful information based on bridge dynamic response characteristics [9,10]. To assess change in the dynamics of structures stiffness reduction, damage degree identification, R-value, and computation efficiency are mostly used. These approaches have shown tremendous promise in structural health monitoring (SHM) of bridge structures by providing support for structural damage diagnosis and maintaining the structure's performance. Unfortunately, the forward and inverse problems are two significant challenges faced while applying the VBDD method. As illustrated by Zhan et al. [11] in modeling structures and establishing a robust inverse algorithm to diagnose the damage. Traditional methods in structural damage detection, such as non-destructive testing (NDT) and visual inspection, have proven limitations in their usage, such as the inability to identify damage within structures and non-structural component damage while modern approaches come up with several benefits such as Automation, Risk monitoring, high efficient, and computer vision in diagnosing damage [12,13].

To date, diverse VBDD approaches have been elaborated to be used in SHM and bridge health monitoring (BHM) to assess the structural physical change, which leads to the corresponding alteration in the vibration characteristics of the structure. Methods used in VBDD are natural frequencies, mode shapes, modal curvatures, and dynamically measured flexibilities. Natural frequency and mode shape are the most often utilized owing to can be quickly recorded from a few positioned sensors [14]. To conduct a rapid condition of large-scale cable-stayed bridges [15],Utilized a data-driven approach to examine the structural damage diagnosis using hybridization of support vector machine and enhanced feature extraction techniques, to analyze efficiency and robustness of approaches the following damage sensitive parameters were used damage level, location, Traffic loading, Sensor location, damping and stiffness [16]. The issues faced when using model-based damage-sensitive features as damage identification strategies in the real world include the incompetence of machine learning algorithms (MLA) for diagnosing failure, environmental and operational uncertainties [17]. As depicted in Fig. 1, VBDD methods are global methods that assess the alteration of vibration characteristics between the health and damaged states of the bridge structure to examine the vulnerability.

Although some of the research works have touched on digital revolution applications in bridge damage detection but a comprehensive review is lacking in this area [18]. Incorporating machine learning with Vibration-based damage detection has revealed considerable promise in real-world applications, particularly in bridge structures. Yan et al. [19] Researched to reveal the current trend of vibration-based structural damage diagnosis approaches appertaining to mathematical modeling software such as MATLAB, Python, etc. To overcome the existing drawbacks of previous studies, this research aims to provide a broader and comprehensive review to take into account the application of VBDD combined with ML approaches for anomaly detection in bridge structures. The author's attention to this paper is based on two primary factors. Firstly, to identify the suitability of these approaches in structural damage detection. Secondly, to present the extensive literature on VBDD methods that have been published and research gaps. This review paper can be the starting point for those curious about applying the VBDD approach in bridge damage detection.

2. Artificial intelligence in bridge monitoring

2.1. Application of artificial intelligence

In recent decades, Research on the identification of structural damages have been of great concern worldwide due to the deterioration and degradation of the structure having numerous undesirable effects on users and society [20]. Integrating the digital revolution empowered by improvements in artificial intelligence (AI) and big data (BD) in damage detection brings a potential transformation in damage diagnosis [21]. As depicted in Fig. 2, Various components of AI have been elaborated. Huang et al. [22]illustrated the various applications of artificial intelligence in civil engineering structures and this field focuses on design optimization, maintenance, and management of the structure. This provides the stakeholder with information that facilitates decision-making during the design, performance, and management of bridges within their lifespans [23]. The trend of AI in the bridge engineering field brings significant benefits like automation, risk mitigation, high prediction accuracy, digitalization, knowledge representation, information fusion, intelligence optimization, and data mining [7]. Owing to the rising sophisticated in the design of up-to-date structures and digitization in the construction industry, AI techniques are more practicable and accessible [24].

2.2. Structural health monitoring

SHM has emerged as the most efficient tool for diagnosing damage in structures [25,26]. Early damage diagnosis increases the structures' lifecycle and decreases rehabilitation costs [27]. There are two main types of SHM such as sensor-based approaches, to obtain raw data sensors placed on the monitored structures to record data, and vision-based methods that use advanced optical instruments [10], R-value, Root mean square, and computation time parameter has been used. Nowadays, more effort has been placed into developing damage detection techniques based on SHM [28] as depicted in Fig. 3. Four main structural damage identification levels have been described. Therefore, it is necessary to look for accurate, efficient, reliable, and sophisticated methods to enhance structural performance by ensuring public safety [20]. Extensive studies used the SHM approach to detect damages in diverse structures such as bridges [29], tunnels [30], buildings [31], Offshore platforms [32], etc. Optimal monitoring periods and the extension of the lifecycle of the welds on a steel bridge deck has determined, and the findings revealed that the decision on short-term monitoring is the most worthwhile SHM methods [33]. To maintain the structural integrity of bridges, it is necessary to make optimum decisions about bridge operation and maintenance.

2.3. Bridge health monitoring

A nation's infrastructure relies heavily on bridges. Monitoring the health of a bridges structure allows deterioration to be detected in real time [34]. BHM typically comprises BD, which needs the applicability of ML to be processed [35]. Novel bridge health monitoring that utilized a multimodal hybrid bridge energy harvester integrated with piezoelectric and electromagnetic conversion has been proposed by Iqbal and Khan. [36], to test its efficiency in the health monitoring of bridge structures. The bandwidth, number of resonant frequencies, voltage generation capacity, and acceleration level were used. The low-frequency bandwidth of 11–43 Hz can gather both bridge vibrations and ambient wind concurrently. Wireless bridge health monitoring using the internet of things and optical sensors have been utilized to detect damage in the deteriorated bridge [37]. A combination of the Kohonen neural network and long short-term memory (LSTM) neural network has been used in the comprehensive analysis of the state of a bridge and reveal the capacity of early warning to detect damage [38]. Here in, Khuc and Catbas [39]. Presented a vehicle modeling framework utilizing computer vision in structural identification-based bridge health monitoring. While Kwon et al. [40] Detected anomalies in bridge structures under complex loads using generated industry foundation classes to enhance safety and reduce maintenance costs. Developing a long-term BHM strategy to handle environmental and operational factors improves monitoring efficiency.

2.4. Big data

Big data (BD) approaches refer to software or algorithms for data acquisition, storage, computing, and analysis. Which is firstly proposed by lecture John Mashey in 1998 [41] and are used chiefly to clarify enormous datasets. The rapid development of BD techniques gains high potential application by distinguished governmental institutions, scientific researchers, and other fields [42]. Commonly used data analysis methods are knowledge discovery in the database, data mining, machine learning, pattern recognition, and statistics [43]. The report released by the international data corporation revealed that global data increased significantly over the past 20 years [44]. Teradata became the first economically feasible parallel database system. BD acquisition has four steps such as data production, data gathering, data storage, and data processing [45]. The capability of extracting useful information from large datasets is called big data mining. New mining techniques are required due to the large size and complexity of datasets' volume, variability, and velocity [46]. To ensure BD accessibility and reliability, measure should be taken on its storage and management. The process of BD analysis is depicted in Fig. 4.

3. Vibration-based damage detection by machine learning

Vibration-based monitoring approaches focused on investigating the bridge's condition and providing helpful knowledge for future bridge designs. The utilization of ML in damage detection of different structures is becoming of great interest to researchers owing to its ability to detect damage and plays a significant part in SHM [47]. ML approaches are split into four broad classifications: supervised, unsupervised, semi-supervised, and reinforcement learning [48]. For decades, adapted computer vision approaches have been expeditiously raised to conquer the shortcomings of conventional visual inspection and contact sensor-based methods like temperature effects, wind, and other uncertainties [49]. As illustrated in Fig. 5, various approaches utilized in vibration-based damage identification have been presented. Flah et al. [10] reviewed the implementation of ML in structural health monitoring. Structural damage identification approaches accessible for SHM applications are based on data processing, feature extraction, and classification. A novel wireless sensor network(WSN) in reliance on one-dimensional convolutional neural networks(CNN) has been established [50]. Hybrid Cuckoo search (CS) with ANNenhanced the training parameter of ANN by reducing the dissimilarity between real and desired outputs. The findings prove that ANN incorporated with CS is accurate and efficient and needs a lower computational time than ANN [51].

Recently, bridge operation and maintenance decision-making has been changing towards a data-driven manner where sensing techniques were adopted [52]. Those involved in bridge operation, repair, rehabilitation, and maintenance revealed that the dataset is the key to success. ML has been widely used in bridge damage detection. Tran-Ngoc et al. [53] demonstrated an approach that reduces the drawback caused by the utilization of backpropagation algorithms appertaining to gradient descent methods, which are faced with local minimum and minimize the accuracy and effectiveness of ML. A novel MLA has been proposed to diagnose damage using data on load's position, magnitude, and speed acquired from a bridge weigh-in-motion system to enhance the accuracy of the approach [54]. A novel two-stage ML method for bridge anomaly detection utilizing the responses measured on a passing vehicle has been suggested by Malekjafarian et al. [49]. Applications Machine Learning (ML) tools are more feasible, efficient, accurate, and faster in damage detection [15]. Numerous studies have been conducted, as depicted in Table 1.

Table 1. Some studies for Vibration based damage detection by machine learning.

Reference	Bridge structure	Input	Algorithm	Output
(Figueiredo et al. 2019)	Post-tensioned concrete box girder bridge(Z-24 Bridge)	natural frequencies	Gaussian mixture model	stiffness Reduction
(Yazdani-chamzini, Razani, and Haji 2013; Malekjafarian et al. 2019)	Beam bridge	Acceleration, speed and position of vehicle	Incorporating an Artificial Neural Network model and a Gaussian process	Root mean square, Damage index
(Tran-Ngoc et al. 2019)	Truss bridge	Natural frequency and mode shape.	Combination based on an artificial neural network and cuckoo search algorithm	Root mean square, Damage degree identification
(Tran-Ngoc et al. 2020)	Truss bridge	Natural frequency	Cuckoo search algorithm	R-Values, Root mean square, computation time
(Gonzalez and Karoumi 2015)	Railway bridges	Strain and accelerations	Artificial neural network	Damage index
(Pan et al. 2018)	Cable stayed bridge	Stiffness degradation	Support vector machine	Damage index

3.1. Machine learning algorithm

Machine learning algorithms (MLA) have attracted the interest of various researchers and have been applied in design optimization, performance assessment, maintenance, SHM, damage detection, and construction, as depicted in Fig. 6 [10]. MLA is generally categorized into supervised and unsupervised learning algorithms. Supervised algorithms demand a dataset comprising human-labeled data and unsupervised learning algorithms utilize unlabeled data for training [55]. The task performed by the ML system is classification, regression, prediction, and clustering. Azimi et al. [25] conducted an in-depth review of the latest literature in SHM utilizing a deep learning(DL) based approach and furnished researchers with a comprehensive understanding of diverse SHM applications, challenges, and future trends. The most popular ML algorithms are decision trees, neural networks, ensemble methods, support vector machines, Gaussian mixture, Hidden Markov model, and a combination of two or more algorithms that have been implemented in bridge engineering to accelerate the learning process to solve a particular problem [56]. In this regard, Kohonen neural network algorithms and long short-term memory (LSTM) neural network algorithms have been used to handle BD [38]. The results revealed that Kohonen clustering can detect outliers and classify patterns. LSTM prediction algorithms can predict future deflection values with good accuracy and low mean square error.

With the outgrowth of BD technology, applying MLA in diverse sub-fields of bridge engineering has been hastened in recent periods. Vibration-based anomaly detection process together with four MLA has been used to diagnose structural vulnerability under operational and environmental changeability [57]. Based on Fig. 7 ML algorithm has been adopted in damage detection; four kernel-based machine-learning algorithms have been proposed for damage identification subject to fluctuating operational and environmental factors [35]. The suggested algorithm is applicable for damage detection, comparing the classification performance between these algorithms. For instance, Zhan et al. [11] numerically investigated the attainability of adopting a probabilistic neural network and FEM updating methods to explore the state of connection joints in steel truss bridges. The results revealed that the accuracy of preliminary damage localization could outstrip 90%. In concurrence with BD, ML has driven current breakthroughs concerning load-capacity assessment and failure mode prediction, among others. These up-to-date advancements in implementing ML signals have great potential for growth within the bridge engineering sector.

3.2. Neural network-based approach

The neural network is a model that originated from the connections and interactions between the human brain's nervous system by mimicking the person's ability to learn and decision-making [16,23]. Neural networks (NN), also called artificial neural network (ANN), is made of an input layer, hidden layer, and output layer, as depicted in Fig. 8. The CNN and radial basis neural networks are the most often used neural networks. Sometimes hybrid it with other algorithms to enhance learning accuracy and computation time [55,58,59]. The application of ANN is usually divided into four categories: prediction, classification, data association, and data filtering [31]. A neural network consists of three significant actions its pattern of connections among neurons, its approach to deciding the weights of connections, and its activation function [60]. In comparison, Artificial intelligence algorithms such as genetic algorithms are probabilistic optimization algorithms based on natural evolution. Appear as an advantageous concept from the domain of artificial intelligence and has been fruitfully used in modeling theoretical and practical challenges over the past decade, particularly those related to the mechanical behavior of composite materials and neural networks that can be utilized as a powerful regression tool [57].

3.3. Deep learning method in BDD

Deep learning (DL) is a part of the ML method appertaining to neural networks, which use multiple layers in the network. Nowadays, DL turned out to be the principal method for much progressing work in ML [25]. There are different types of deep learning, such as CNN, recurrent neural networks, deep autoencoder, deep restricted Boltzmann machines, deep belief networks, and generative adversarial networks [61]. DL is carried out through a deep learning application programming interface like TensorFlow, PyTorch, or Caffe, allowing high-dimensional datasets to be handled [62]. A novel method combining CNN with the gapped smoothing method to detect damage in the Bo Nghi bridge located in Vietnam. A Hybrid of gapped smoothing method and CNN provided effective results for being used for diagnosis and localization damage [63]. The findings were confirmed using ultrasonic pulse velocity testing, which yielded 96% accuracy and 97.79% specificity. Based on Bayesian optimization, a novel method of using three-level images has been proposed for inspecting the post-disaster reinforced concrete bridges [64]. A novel DL approach called DT-YOLOv3 has been proposed by Yu et al. [65] to diagnose bridge deterioration.

4. Modal parameter-based method

To compare the dynamic properties of an intact and damaged structures, parametric techniques monitor structural deterioration. These approaches use structural damage identification sensing interfaces to provide vibration signals for determining the structural system's uncertain dynamic behavior [14]. Modal parameters include mass, damping, stiffness, mode shapes, and frequency. These parameters are damage-sensitive features under environmental and operational conditions. Global and local methods are two major approaches to structural damage detection techniques. Experimentally bridge deterioration was evaluated globally and locally using multi-sensor data and optimized functional echo state networks to identify multivariate time series data without requiring manual feature extraction [66]. For global damage detection, damage conditions are evaluated for the whole bridge while for local damage detection, structural damage is quantified on a small scale.

4.1. Modal strain energy-based damage index

Modal strain energy (MSE) has extensively gained adaptability in structural damage identification and is the most reliable method among vibration-based approaches [67]. Stubbs and Kim introduced the MSE-based damage index approach in 1996. This technique identifies the anomaly as a reduction in modal strain energy, which implies that the bridge structure's stiffness reductionprevents energy restoration according to Bernoulli-Euler beam theory [29]. The Damage index, modal strain energy change, cross-modal strain energy, and other iterative modal strain energy methods are the four categories of modal strain energy approaches [68]. A hybrid of the MSE-based damage index approach and an ANN have been utilized to diagnose damage in steel girder bridges [69]. Grouping modal strain energy was used as a new approach to detecting structural deterioration of marine structures [70]. In reliance on the research conducted by Tan et al. [71], modal strain energy can be symbolized by ‘U’ as given below in Equation (1):(1) $U = \int \frac{E I}{2} {(\frac{d^{2} y}{d x^{2}})}^{2} d x$ Where EI stands for flexural rigidity of the section and $\frac{d^{2} y}{d x^{2}}$ represents curvature, y stands for the vertical deflection, and $x$ stands for the distance measured along the length. In Equation (2), $β_{j i}$ express MSE between damage and intact state of the bridge structure and $β_{j i}$ is damaged indicator for $i^{t h}$ mode at location $j^{t h}$ . $ɸ_{j i}$ Represents the mode shape for the undamaged state, while $ɸ_{j i}^{*}$ represents the mode shape for the undamaged state of the bridge [72].(2) $β_{j i} = \frac{[{(ɸ_{j i}^{*})}^{2} + \sum {(ɸ_{j i}^{*})}^{2}] + [{\sum (ɸ_{j i})}^{2}]}{[{(ɸ_{j i})}^{2} + \sum {(ɸ_{j i})}^{2}] + ({[\sum ɸ_{j i}^{*}]}^{2})}$

4.2. The modal flexibility-based damage index

Modal flexibility is another method for structural damage identification according to natural frequency and mass normalized mode shapes. Firstly was proposed by Pandey and Biswas [73]. This method has been adopted by Wickramasinghe et al. [74] because it can detect deterioration in the main cables and hangers of a suspension bridge. Hybrid of modal curvature and modal flexibility approach to identify multiple cracks through-beam structure [75]. It is most often used for its efficiency and accuracy in structural health monitoring. Modal flexibility is symbolized by ‘F' in Equation (3).(3) $F = \sum_{i = 1}^{n} \frac{1}{{w_{i}}^{2}} ɸ_{i} {ɸ_{i}}^{T}$ where ɸi is the ‘ith’ mode shape; wi is ‘ith'modal frequency, and n is the number of degrees of freedom. The vulnerability of the bridge structure diminishes the structural stiffness, leading to an increase in the structure's flexibility. The variability in modal flexibility of structural failure (Δ[F]) can be written as shown in Equation (4):(4) $Δ [F] = [F ˆ (*)] - [F]$

In the above equation, the superscript'*' stands for the damaged state. Therefore the absolute maximum value in every column of Δ[F] is taken as the modal flexibility change value Δ F for the relevant node. Based on the study conducted by Tan et al. [71], the % change of modal flexibility (ΔF%) was given in Equation (5):(5) $Δ F % = \frac{Δ [F]}{[F]} x 100 %$

4.3. Modal assurance criterion

The modal assurance criterion (MAC) is one of the most crucial modal indicators and is mostly used in bridge health monitoring to assess the change in mode shape [76]. MAC have been utilized to demonstrate the supporting status of measured mode shape for the long-span cable-stayed bridges in health monitoring [77]. The MAC has been adopted in health monitoring to evaluate uncertainty in the operational modal analysis [78]. As was illustrated above, a change in dynamic characteristics implies a structural change from a healthy state to a damaged state. MAC is a dimensionless indicator that demonstrates the correlation between two models. The ith-order formula is displayed in the equation below:(6) ${M A C}_{i} = \frac{{({ɸ_{i}}^{T} {ɸ_{i}^{*}})}^{2}}{\ \ \ ({ɸ_{i}}^{T} {{ɸ}_{i}}) {({ɸ_{i}^{*}}}^{T} {ɸ_{i}^{*}})}$ Where in ${{ɸ}_{i}}$ is the ith-order modal vibration mode, while ${ɸ_{i}^{*}}$ is the ith-order modal vibration mode of the damaged structure.

4.4. The modal curvature-based damage index

Modal curvature revealed several merits in damage detection, such as identifying damage in the damaged neighborhood and being more insensitive to environmental variability than natural frequency [79]. The central difference approximation approach may estimate the modal curvature at the measurement point j based on the mode shape displacement. The function for modal displacement is in Equation (7):(7) $ɸ_{i j}^{″} = \frac{ɸ_{i j + 1} + {2 ɸ}_{i j} + ɸ_{i j - 1}}{L^{2}}$

Here, ‘ $ɸ$ ’ is the modal displacement for ${t h e j}^{t h}$ mode shape at the measurement displacement ith, and ‘L’ is the distance between two successive measurements. As a result, the damaged index may be derived from the model form without using undamaged structure baseline data. The damage index for each mode is adjusted using the standard deviation to avoid spurious peaks due to measurement noise. The extent of the damage may therefore be determined by calculating the change in the mode shape curvature based on the study conducted by Altunışık et al. [80]:(8) $ɸ_{i j}^{″} = \sum_{j = 1}^{n} | {(ɸ_{i j}^{*})}^{″} + {(ɸ_{i j})}^{″} |$ Where the superscript ‘ $*$ ’ stands for damaged state and ‘n’ is the number of modes considered. A modal curvature and modal flexibility hybrid has been utilized to identify multiple cracks within the beam structure [80].

5. Vibration-based damage detection in various types of bridge

5.1. Damage detection in reinforced concrete bridge

Researchers have conceived various techniques to facilitate the repair and rehabilitation of reinforced concrete (RC) bridges. In this section, we tried to summarize the application of vibration-based monitoring approaches as illustrated in Table 2. Vibration characteristics and artificial neural networks are combined to develop a strategy to detect deterioration in composite bridgesmade of the reinforced concrete deck and I steel girder [71]. Numerical and experimental work has been conducted to evaluate the approach by reducing stiffness of members. The findings revealed that the approach is competent in detecting structural damage. They suggest further research to implement this approach on real bridges [29,81]. Damage assessment on the prestressed concrete bridge has been conducted using fiber optic sensors and artificial neural networks [82]. Aasim et al. [83]utilized vibration-based structural damage identification of reinforced concrete bridge structures in real life and were monitored using a potable vibration computation system with a load drop impact excitation. Cancelli et al. [84] proposed an approach for localizing and quantifying damage in a pretension concrete girder through stochastic subspace detection and particle swarm model updating. Root mean square and R-value parameters were assessed to evaluate robustness of the approach.

Table 2. Combination of VBDD approach and ML techniques in various type of bridge.

Application	Approach	Results	Reference
Reinforced concrete bridge	Modal curvatures gapped smoothing method combined with Convolutional Neural Network	Combination of these approaches are more accurate in damage detection for concrete girder bridge.	(Nguyen et al., 2020)
	Deep convolutional denoising autoencoder utilized to extract damage feature.	This technique is adequate for damage detection under environmental variability.	(Shang et al., 2021)
	Hybrid of stochastic subspace identification approach with particle swarm optimization algorithm.	These approaches are efficiency and use less computation time to assess damage in pretension concrete girder bride.	(Cancelli et al., 2020)
Steel bridge	Artificial Neural Network technology incorporated with modal flexibility and modal strain energy based damage indices	A safe and effective functioning of arch-type bridges may be ensured by this method.	(Jayasundara et al., 2020)
	Combine deep auto-encoder with support vector machine	The approach are more stable, reliable and robust for structural damage detection in steel truss bridge.	(Wang and Cha 2021)
	Modal flexibility based damage index	This approach is appropriate to monitor the condition of suspension bridge and counteract their premature collapse.	(Wickramasinghe, Thambiratnam, and Chan 2020)

5.2. Damage detection in steel bridge

To sustain and preserve the service life of steel bridge structures, it is essential to conduct early monitoring of structural damage [26,85]. Corrosion and fatigue are the prominent time-dependent degradation and aging effects in steel, arising from a complicated process influenced by material properties and environmental conditions. Numerous researchers have utilized the vibration-based approach and machine learning tools to diagnose anomalies in steel bridges. Jayasundara et al. [86] Developed a dual-criteria approach relying on vibration features to detect and identify anomalies in steel arch bridges. VBDD approach comes up with the solution of dealing with the limitation faced with non-destructive techniques for local damage diagnosis [87,88]. The economic impact of these benefits is remarkably pertinent for steel bridges because of their prevalent utilization in various countries globally [11]. As depicted in Fig. 9, the combination of the Vibration-based approach and the artificial neural network has been utilized to detect some micro-cracks that appear around the connection joints, which is fatigue failure due to operational loading [89]. Ma et al. [90] Assessed the fatigue performance of innovative shallow buried modular bridge expansion joint. The results revealed that this joint exhibits a good fatigue performance under the standard vehicle load conditions [91]. Monitoring data can detect structural changes and deterioration in cable-stayed bridges, providing a key basis for bridge structural assessments [92]. This study used spatiotemporal graph convolutional networks to detect novelty in cable-stayed bridges by analyzing spatiotemporal correlations between cable forces from different cable dynamometers, this enables the data-based method suitable for long-term continuous monitoring [93].

5.3. Damage detection in timber bridge

Timber is the oldest construction material that garners greater attention in green infrastructure development. Around 200 timber bridges have been built in Norway in the last 30 years, comprising the world's strongest Kjøllsaeter timber bridge and the world's longest timber highway bridge, the Tynset Bridge, whose span is 125 m [94]. Proposed methods for using timber substitute for other, more emission construction materials. The outcome confirmed that the timber bridge emits much less pollution than the concrete bridge in all impact categories, with substantial environmental benefits from material end-of-life rehabilitation. Large-scale timber bridge designs should be explored for future road developments. According to this study, Rodrigues et al. [95]demonstrated that the sustainability of timber-concrete composite Bridge decks is examined from environmental, economic, and social aspects. Timber-concrete composite decks have a lower environmental impact and are more cost-effective than concrete decks. Summing up existing SHM and NDT methods in timber bridge structures, Palma and Steiger [12] discussed the appropriate monitoring techniques while moisture content, decaying, insect attack, delamination and cracks, deformations, and stresses are all critical considerations to monitor in wood structures, as depicted in Fig. 10. Rashidi et al. [96]provided a summary of different vulnerability mechanisms of timber bridges, their preventive actions together with feasible remedial preferences for management and maintenance of timber bridges which is of uppermost essential to maintain and prolong the lifecycle of timber bridges.