Damage Detection on Real Bridges Using Machine Learning Techniques: A Systematic Review

Abstract

Preventive maintenance efforts for bridge infrastructure have proven to mitigate early deterioration and reduce the probability of severe damage. Modern research has focused on the employment of online data directly collected within the structures, provided by several novel devices that feed machine learning approaches that continuously measure structural health. However, several issues remain within the related fields. The constant evolution of ML techniques, for example, provides new potential lines of research. Furthermore, widespread validation through real-world test cases and the use of diverse bridge typologies (which can be interesting considering their distinct behaviors) remain limited. This article seeks to examine the advancements in structural health monitoring (SHM) employing machine learning methods for identifying structural damage in bridges over a 7-year period, with a particular focus on studies employing real bridge data. Present challenges and future research directions are assessed. This study offers valuable insights to researchers and academics conducting research in this field, providing a thorough summary of current developments and combining ML methods with the four most-investigated bridge types in case studies.

1. Introduction

Monitoring the structural health of civil structures requires regular evaluations to ensure preventive and proactive maintenance [1]. Historically, this task relied on visual inspections carried out by experienced personnel, involving presence on site and visual inspection of possible signs of damage [2]. Vibration-based damage detection emerged in the 1970s–1980s, initially used in the oil industry for offshore rigs [3]. This approach has spread to aerospace and civil engineering, offering advantages such as vibration pattern extraction and timely anomaly detection, revealing structural damage.

Contemporary techniques have been developed, tailored for various types of structures and referred to as advanced structural health monitoring (SHM) in the literature. However, in recent years, the scientific and technological community has increasingly focused on structural health monitoring applied specifically to bridges. Advances in sensor design, machine learning techniques, and their implementation have resulted in a greater comprehension of bridge performance and safety, as well as elevated efficiency in the early identification and forecasting of anomalies and the long-term structural behavior of bridges. Taking this into account, this paper seeks to review advances in SHM using machine learning techniques to detect, locate, classify, and predict structural damage in bridges, especially considering work that uses real bridge data.

Currently, AI algorithms, particularly those based on ML, have established themselves as versatile tools that are applicable to various tasks. The structural health monitoring of bridges has seen increasing adoption of machine learning methods, including damage detection, localization, classification, and prediction, over the past few years. A key factor in the widespread use of these algorithms is their capability to adapt to different situations through the use of diverse data types, including vibration signals, images, and environmental factors. In this regard, analyzing the specific purpose for which these algorithms have been used in the SHM context is important. For example, early detection, prognosis, and reconstruction of missing data, as well as the type of information used for their training and validation, are crucial. This characterization enables the identification of trends, gaps, and opportunities in the design and application of AI-based bridge monitoring solutions.

On the other hand, we consider a fundamental aspect that has been explored in relatively little depth up to now: the differentiation of structural monitoring methodologies based on the type of bridge to which they are applied. Each type of bridge—suspension, arch, box girder, cable-stayed, or other structure—not only has unique geometric and structural characteristics, but is also subject to different load conditions, dynamic patterns, and deterioration mechanisms. For example, suspension bridges are often more exposed to wind-induced vibrations, whereas arch bridges can concentrate stresses at specific points that respond differently to cyclic or thermal loads. This structural, functional, and environmental variability directly impacts the effectiveness and suitability of damage detection techniques based on machine learning. Therefore, this systematic review proposes to classify and analyze recent literature according to bridge type, to identify patterns, strengths, and limitations of the methodologies applied in each context, thus offering a more accurate and useful vision for the development of intelligent and adaptive monitoring systems.

However, despite the recent development of highly promising algorithms with remarkable results in damage detection and diagnosis tasks, a significant proportion of these approaches have been validated using simulated data or highly idealized case studies. Although this type of validation is useful for demonstrating a method’s theoretical potential, it can generate uncertainty regarding its practical applicability in real contexts, where active civil structures are subject to much more complex and variable conditions, such as environmental noise, changing weather conditions, operational interference, and instrumentation limitations. Analyzing studies that have used data from real bridges allows us to approach a scenario closer to the effective deployment of these technologies, facilitating a more rigorous evaluation of the proposed techniques’ performance, robustness, and scalability. Therefore, this review focuses on identifying and examining studies that validated their methods using real field data, to highlight approaches that have already taken steps toward practical implementation in operational infrastructure.

Numerous comprehensive studies have delved into the realm of machine learning and deep learning techniques applied to civil structures, offering a broad overview of these methods [4,5,6,7]. However, while informative, these works may appear rather generalized in terms of a specific focus on bridge domain applications. In contrast, other studies have meticulously examined low-cost sensing techniques used to monitor civil structures [8,9,10,11,12]. However, these studies did not delve into the machine learning techniques underpinning these proposals and did not particularly emphasize their applicability to bridges.

Turning our attention to studies centered on vibration-based machine learning techniques, various research endeavors have explored damage detection methods for structures through these parameters [13,14,15,16]. However, these studies might be slightly outdated, given that most research in this field has been conducted recently. Similarly, several studies concentrated on bridge monitoring as a specific subject [17,18,19]. They distinguished between theoretical, numerical, and experimental studies, which is similar to the focus of our present work on experimental investigations. Although these studies encompassed reviews of structural health monitoring for bridges, the ML techniques employed in the systems were not explicitly emphasized. Finally, a handful of studies have focused specifically on the machine learning techniques employed by structural health monitoring systems applied to bridges [20,21]. However, these works lack an exhaustive coverage of the numerous bridge types used and do not particularly emphasize field studies.

After conducting a general review of various studies that analyzed the latest advances in this field and considering the growing number of publications in this area, it is necessary to update these reviews, considering a more specific study that comprehensively addresses a more specific area, such as monitoring the structural condition of bridges using ML algorithms. This article will examine the four types of bridge that have been most widely studied to date (box girder bridges, suspension bridges, arch bridges, and cable-stayed bridges), with a special emphasis on those studies that used real bridge data. This study also analyzes the applied ML algorithms, their use in the SHM context, and the types of data used as input for these techniques.

1.1. Objectives and Research Questions

Based on the motivations and gaps previously discussed, this study aims to address the following specific objectives:

• Classify the most commonly studied bridge types in SHM studies that involved ML techniques.
• Analyze machine learning algorithms applied in the context of bridge condition monitoring.
• Identify recent and emerging approaches in the field, highlighting their key contributions and characteristics.
• Characterize the types of data used to train and validate machine learning models in SHM applications.
• Propose future lines of research based on the reviewed literature and the limitations identified.

To guide the review process and meet the objectives mentioned above, the following research questions were formulated:

• RQ1: For what specific purposes have machine learning algorithms been applied to perform SHM tasks?
• RQ2: What types of data have been used to train and validate these algorithms?
• RQ3: How do machine learning applications differ between different types of bridges?
• RQ4: Are the technical aspects related to the application of machine learning algorithms in real-world case studies reported and discussed?
• RQ5: What are the main gaps in research and future directions suggested in the current literature?

1.2. Structure of the Article

This article is structured as follows: Section 2 reviews the methodologies used in detail. Section 3 categorizes and reviews the types of bridges, focusing on the applied ML techniques. Section 4 reviews current challenges and future research directions. Finally, Section 5 concludes the article.

2. Methodology

This section presents the methodology used to perform the systematic review, which provided a structured framework for this study. Drawing on the recommendations of the PRISMA guide [22] and adapting the approach used by [23] to the specific requirements, the procedure is shown in Figure 1.

2.1. Search Strategy

The first step was selecting academic databases to retrieve indexed articles. Given their extensive usage within the scientific community, the authors opted to use the SCOPUS and Web of Science databases. Subsequently, the time frame for the search was defined in relation to scientific advances and discoveries, restricting it to between 2017 and 2023. This time window encompasses a substantial body of research, which will be evident in the subsequent sections.

A set of key terms was first defined to initiate the search for academic articles, covering three main conceptual domains: structural health monitoring (e.g., “SHM”, “structural monitoring”, “damage detection”), algorithmic approaches (e.g., “artificial intelligence”, “machine learning”, “deep learning”), and the type of infrastructure under consideration (“bridge”). All possible combinations of these three concept groups were generated and connected using the AND operator (see

Table 1. Search queries used.

Query
“SHM” AND “artificial intelligence” AND “bridge”
“SHM” AND “machine learning” AND “bridge”
“SHM” AND “deep learning” AND “bridge”
“damage detection” AND “artificial intelligence” AND “bridge”
“damage detection” AND “machine learning” AND “bridge”
“damage detection” AND “deep learning” AND “bridge”
“structural monitoring” AND “artificial intelligence” AND “bridge”
“structural monitoring” AND “machine learning” AND “bridge”
“structural monitoring” AND “deep learning” AND “bridge”

for the resulting queries). Each query was entered into the database search engines, targeting occurrences in the title, abstract, and keywords fields. This search process resulted in an initial pool of 512 articles, which were subsequently subjected to inclusion and exclusion criteria to identify the final subset of studies for analysis.

2.2. Eligibility Criteria

Regarding the inclusion and exclusion criteria,

Table 2. Studio selection criteria used in the study.

ID	Category	Criteria
IN1	Inclusion	Article published between 2017 and 2023
EX1	Exclusion	Type of article (scientific journals and conference papers)
EX2	Exclusion	Abstract analysis (general analysis of the article)
EX3	Exclusion	No clearly identifiable type of bridge
EX4	Exclusion	No clearly identifiable application of ML algorithm
EX5	Exclusion	Detailed analysis

summarizes the conditions applied to filter the search results. First, duplicates were removed, and only articles published in scientific journals and conference papers were considered, excluding other types of document such as books, state-of-the-art reviews, and book chapters. Next, a general analysis of the abstract was conducted to rule out studies that had a different focus, either on SHM or ML, or that stated that they had used data extracted from structures other than bridges. Each remaining article was examined to determine the specific type of bridge to which the proposed method was applied. Consequently, we excluded articles lacking clarity in this aspect or those addressing bridge typologies other than the four initially considered (Box-Girder, Suspension, Arch, and Cable-Stayed). In addition, we excluded articles in which the application of an ML technique to support the SHM process could not be clearly detected. Note that this filtering process was solely applied to non-state-of-the-art review papers. The subset subjected to in-depth analysis comprised 76 articles.

2.3. Data Extraction and Analysis

Relevant information was extracted for each of the 76 selected studies using a structured data extraction template. The following attributes were recorded: publication year, bridge typology, case study used, data types used (e.g., vibration, strain, images), applied machine learning algorithm(s), SHM task addressed (e.g., damage detection, feature extraction), and whether real-bridge data were used for validation.

The extracted data were then coded into thematic categories, allowing classification by bridge type, algorithm family, SHM objective, and used data types. A descriptive analysis was conducted to identify patterns, frequency of use, and trends over time. All data were manually processed by a reviewer and verified by the team of authors to ensure consistency and reduce bias.

3. SHM Using ML from a Bridge-Type Perspective

Bridges are vital components of national and urban transportation and connectivity infrastructure throughout the world [24]. They come in various types, each possessing distinctive attributes and features tailored to specific scenarios and environments. For example, suspension bridges with lightweight structures excel at spanning wide rivers and valleys, effectively supporting substantial loads. In contrast, arch bridges utilize curved designs to distribute and bear bridge loads uniformly. Box-girder bridges, another type, are ideal for shorter to medium-span crossings, offering simplicity and cost-effectiveness in construction. In essence, each type of bridge has its own set of distinctive qualities and characteristics that make them suitable for different situations and environments.

Machine learning has gained substantial traction in bridge design, construction, and monitoring in recent years. Various ML techniques have been deployed across a diverse spectrum of bridge designs to enhance decision-making precision, optimize material utilization, facilitate early damage detection, and reduce expenditure. Figure 2 shows the increasing interest in this research domain and its publication trends over the past seven years.

The data-driven approach is one of the most prevalent strategies for applying ML techniques in the realm of structural health monitoring systems. This approach relies on the analysis of substantial datasets to reveal patterns and correlations within the behavior of bridges. Sensors strategically placed on real bridges and external sources, such as geographical maps, weather records, and traffic data, can acquire data. Taking advantage of this wealth of data, ML algorithms can discern intricate patterns and trends that are invaluable in optimizing designs and enhancing bridge performance. Figure 3 shows an illustration of a sensor deployment similar to that employed by [25], showcasing the use of uniaxial accelerometers to capture the bridge’s dynamic response.

The model-driven methodology is another prevalent approach in which simulation models are used to analyze bridge behavior across various scenarios and conditions. These models are constructed using design parameters, material specifications, and environmental variables, enabling them to forecast bridge performance under a range of loading and environmental conditions. The use of ML within this framework enhances bridge construction, and maintenance precision and efficiency, while facilitating problem identification and decision-making. Figure 4 illustrates this approach, as shown in [26], where an FEM was used to simulate the behavior of the China Tsing Ma suspension bridge.

Furthermore, the literature review indicated a higher concentration of studies on specific bridge types, notably box-girder bridges, cable-stayed bridges, arch bridges, and suspension bridges, as illustrated in Figure 5. This section examines these predominant bridge categories and highlights the diverse techniques employed, alongside their corresponding achievements and outcomes.

ANNs, support vector machines (SVMs), ensemble algorithms, and deep learning (DL) techniques are notable ML techniques employed in this domain. Figure 6 illustrates the distribution of these techniques, emphasizing the prevalent use of ANN and DL algorithms, which collectively constitute more than 50% of the total techniques reviewed in this study.

The applications of ML techniques within the realm of bridge structural health monitoring are discussed in the subsequent subsections. Particular attention is paid to projects that used data collected from actual bridges and field studies. Furthermore, these applications are categorized according to the four bridge typologies most frequently investigated in the literature. A concise table that summarizes the ML technique used, the input data sources, the bridge under examination, and the intended objectives of the implementation is provided for each typology.

3.1. Box-Girder Bridges

Beam bridges and box girder bridges are two types of bridges that mainly differ in their structure and design. Single-girder bridges with a horizontal main girder span between two piers or supports. The main girder can be made of reinforced concrete, steel, wood, or other resistant material. Beam bridges are suitable for short- to medium-distance crossings, and their construction is relatively simple and economical. However, they have limited load-carrying capacity and are not suitable for long distances or heavy loads.

Box-girder bridges have a box-shaped cross-section. The cross-section of the box is composed of two side plates and a top and bottom plate that form a hollow box. Due to their hollow-box design, which provides greater rigidity and strength, box-girder bridges are suitable for spanning longer distances and carrying heavier loads than beam bridges. In addition, box-girder bridges have better fatigue resistance and can better withstand dynamic loads, such as heavy traffic. In summary, beam bridges are more suitable for short-to-medium-distance crossings with light-to-moderate loads, whereas box girder bridges are more suitable for covering longer distances and supporting heavier loads. A summary of the latest applications of ML algorithms using box-girder bridges can be found in

Table 3. Summary of SHM applications using box-girder bridges.

References	Year	Structure	Algorithm	ML Purpose	ML Input
[27]	2017	Z24 Bridge	GA + EM + GMM	Hyperparameter tuning and damage detection	Vibration
[28]	2018	Sishui Bridge	RSNB	Damage detection	Vibration
[29]	2018	Hanxi bridge	PCA + EEMD + ICA	Signal analysis	Deflection
[30]	2019	Z24 Bridge	GMM	Damage detection	Vibration
[31]	2019	W180 Bridge	PSD	Damage detection	Vibration
[32]	2019	Z24 Bridge	Active Learning	Damage detection	Vibration
[33]	2019	n/a	RF	Seismic damage assessment	Bridge and earthquake parameters
[34]	2020	n/a	LSTM	Data reconstruction	Temperature, strain and displacement
[35]	2020	TCRF Bridge Dataset	CNN + GRU	Damage detection	Vibration
[36]	2020	n/a	Kringing model (GPR)	Performance prognosis	Natural frequencies (vibration)
[37]	2020	Z24 Bridge	ANN	Feature extraction and damage detection	Vibration
[38]	2020	Fuchang Bridge	DCGAN	Data reconstruction	Strain
[39]	2021	D Bridge	CNN	Image processing and crack detection	Images
[40]	2021	Z24 Bridge	AE	Feature extraction and damage detection	Vibration
[41]	2021	Z24 Bridge	K-Medoids	Damage detection	Vibration
[42]	2021	Z24 Bridge	Quefrency Techniques + PCA	Feature extraction and damage detection	Vibration
[43]	2021	TCRF Bridge Dataset	CNN + GRU	Damage detection	Vibration
[44]	2022	Itacaiúnas Bridges	Mahalanobis Algorithm	Damage detection	Vibration
[45]	2022	Z24 Bridge	TDNN	Damage detection	Vibration and audio
[46]	2022	TCRF Bridge Dataset	CNN + (CSTA)	Damage detection and feature extraction	Vibration
[47]	2022	Chongqi Bridge	RF	Performance prognosis	Temperature and wind parameters
[48]	2022	Z24 Bridge	CNN	Data compression, modal identification and damage detection	Vibration
[49]	2022	Z24 Bridge	RNN + CNN	Damage detection	Vibration
[50]	2022	Lieshihe Bridge	GAN + LSTM	Data reconstruction	Strain and displacement
[51]	2022	Z24 Bridge	Active Learning	Damage detection	Vibration
[52]	2022	Z24 Bridge	EVT	Damage detection	Vibration
[53]	2022	Z24 Bridge	SEV	Damage detection	Vibration
[54]	2022	TCRF Bridge Dataset	ESN + MSCNN	Damage detection and feature extraction	Vibration
[55]	2022	Bolte bridge and Sky rail of Victoria	CNN	Corrosion detection	Images
[56]	2022	n/a	XGBoost	Seismic damage assessment	Bridge and earthquake parameters
[57]	2022	n/a	ML Interpretability Techniques	Evaluation of the interpretability of ML models	Bridge and earthquake parameters
[58]	2023	Z24 Bridge	BiLSTM	Data reconstruction	Vibration
[59]	2023	Z24 Bridge	AE	Damage detection	Vibration

Note: n/a: Not applicable (data from different bridges was used or the name of the bridge was not reported).

.

In the context of structural health monitoring (SHM) systems, ref. [60] distinguished between modal parameter and direct signal-based approaches, with the latter categorized into time, frequency, and time–frequency domains [61]. An innovative algorithm for assessing damage on box-girder bridge columns was introduced in the frequency domain [31]. It employs power-spectral density to process structural responses and introduces a novel damage index using the least square distance metric. The approach was validated using a numerical model of the W180 box-girder bridge, a structure previously studied by [62], and other similar bridges. Unlike the algorithm evaluation, a numerical model was not provided for the application. Alternatively, a method for detecting, localizing, and quantifying damage was proposed in [28]. It combines rough set theory and naive Bayes for classification using dynamic signatures linked to modal parameters as sensitive damage features. The four-step process involves extracting modal parameters, computing dynamic signatures, localizing damage, and quantifying damage severity using the RSNB classification algorithm. Validation employed a FEM model of a 50-meter-long box-girder bridge located in Shenyang, China, comparing results with various ML techniques.

A hybrid option that blends the strengths of both methods can be considered in addition to the two main approaches to structural health monitoring (data-driven and model-driven). This approach can be addressed using model updating techniques, where data from a real structure are compared with FEM-predicted responses. Significant discrepancies between simulated and measured responses indicate structural-damage-associated changes in the bridge dynamics. The FEM must first undergo a calibration step (commonly referred to as model updating) to ensure effectiveness, in which parameters related to material properties, geometry, and loading conditions are adjusted to achieve the closest possible match with the real structure. To optimize this calibration process, computational methods, such as convolutional neural networks (CNN) [63,64], Bayesian models [65], and deep reinforcement learning (DRL) [66], have been successfully employed.

Another way to implement a hybrid approach is through data augmentation techniques (increasing the number of test cases analyzed by the algorithm by creating new cases artificially). This integration is exemplified by [30], where data from the normal conditions of the Z-24 bridge in Sweden and simulations of a finite element model were combined. This integration aimed to detect structural damage under variable conditions, using the Gaussian Mixture Modeling (GMM) algorithm for classification. The performance of the proposed algorithm was compared using SHM system data, FEM data, and combined data sources. Additionally, the natural frequencies of the bridge served as sensitive damage features for input into the classification algorithm. It is important to note the extensive use of this dataset within this field. More details on data collection can be found in [67], which describes the SIMCES project. Refer to [68,69,70] for additional information on the bridge.

The shortage of labeled data for training supervised algorithms is a challenge in SHM [71]. Approaches such as active learning, fuzzy labels, and semi-supervised learning have been used to overcome these limitations [72]. A probabilistic method based on active learning was introduced in [32] for the classification of structural damage. It adapts as new data with varying distributions arrive, actively discovering new classes. This was validated using acoustic emission [73], Gnat aircraft datasets [74], and the Z-24 bridge. A variation was proposed in [51], altering the data querying based on an information criterion. An automated anomaly detection model for structural damage using modal parameters and the one-class convolutional neural network (OCCNN) algorithm on the Z-24 bridge was presented in [37]. In [59], an autoencoder was used to reconstruct undamaged acceleration data sequences, with macrosequences helping to establish damage thresholds.

Extreme value theory (EVT), a branch of statistics that deals with stochastic deviations, was used to statistically address data scarcity in [75]. A previous study [52] presented an anomaly detection method using probabilistic ML with EVT as a core component, validated on the Z-24 and Tianjin-Yonghe bridges. Comparisons were made with classical techniques, such as GMM, singular value decomposition (SVD), and principal component analysis (PCA). An unsupervised algorithm based on semiparametric extreme value (SEV) theory was proposed to tackle unknown parameter estimation challenges in [53]. Furthermore, a method for estimating thresholds using the generalized extreme value (GEV) theory [76], incorporating K-medoids classification and a new damage indicator to improve detection accuracy, was introduced in [41].

Hybrid algorithms have emerged to address structural health monitoring challenges. The combination of CNNs and RNNs has been explored [35]. Convolutional neural networks (CNNs) extract spatial relations, whereas RNNs, such as GRU, capture temporal relations. The proposed model, HCG (Hierarchical CNN and GRU framework), merges CNN and GRU for joint spatial-temporal modeling of damage detection. Validation used the TCRF Bridge dataset. Similarly, ref. [49] sequentially combined CNN and GRU to extract the characteristics of vibration data and classify damage. Meanwhile, [43] introduced the PCBG model, a parallel architecture of CNN and GRU (Figure 7), which achieved robust results compared with other approaches.

Effective feature extraction improves damage detection algorithms [77]. A two-level strategy for vibration-based damage detection was presented in [40]. An initial identification of the modal parameters is used to estimate the natural frequencies using stochastic subspaces [78]. Then, stacked AEs are applied to improve damage detection classifiers, particularly using a Gaussian mixture model (GMM) with expectation-maximization (EM-GMM) [79]. This approach was compared with PCA, AANN, and KPCA. A genetic algorithm (GA) was employed for parameter optimization to address the sensitivity of the EM algorithm to parameter choice [27], enhancing the performance of EM-GMM in distinguishing between structural states.

Furthermore, a novel architecture convolutional neural network, named the Channel-Spatial-Temporal Attention (CSTA)-based network, was proposed in [46]. This architecture employs three sequential attention blocks to enhance the importance of selected features, while disregarding those that are deemed uninformative. Each attention mechanism weighs certain features in each of the three domains (channel, spatial, and temporal). The process begins with a squeeze-and-excitation (SE) attention mechanism, as described in [80]. Subsequently, two new attention mechanisms are introduced for computing spatial attention (local and global attention) and temporal attention (grouped self-attention). Experiments were conducted using the TCRF dataset mentioned above (Figure 8), and the network was compared with other architectures such as MLSTM_FCN [81], HGC [35], PCBG [43], and ResNet [82].

Common methods for extracting damage-sensitive features often operate in the frequency or time domains, with a subset focusing on the time–frequency domain [83]. A novel model in [54] combined Echo State Networks (ESN) for time-domain features and a Multiscale Convolutional Neural Network (MSCNN) for frequency-domain features. This model simultaneously extracts damage-sensitive features from the time–frequency domain. The model employs data augmentation to divide the sensor data into equal-length batches. A Fourier transform is applied to each batch copy to represent the data in the time and frequency domains. The proposed model demonstrated improved performance when compared with existing neural network architectures.

In contrast, ref. [42] introduced a quefrency-domain-based feature extraction method for damage sensitivity, a well-explored approach in acoustics [84]. The advantage of the quefrency domain is its low-dimensional representation, which enhances the speed of the damage detection algorithms. They employed PCA for cepstral coefficients to extract damage-sensitive features, to minimize the impact of external variables. Similarly, ref. [45] combined acceleration and acoustic data in the quefrency domain for a novel approach to monitoring structural health. Their four-phase damage detection framework included data preparation, feature extraction (using cepstral coefficients), model training, testing, and validation. A transfer learning strategy incorporating vibration and acoustic data was adopted, as described in [85]. Data normalization was highlighted due to significant frequency differences. A time delay neural network (TDNN) was recognized for shift invariance, and context modeling for damage detection [86].

Given the challenging conditions sensors endure in bridge monitoring systems, sensor failures leading to data loss are common [87]. Such failures can disrupt the algorithms processing the data. Techniques for reconstructing lost data using information from functional sensors are essential. Ref. [34] presented an ML-based model that establishes relationships among environmental variables (temperature, deformation, dynamic displacement) to bridge global and local responses. This knowledge helps in data recovery. A memory-enhanced short-term LSTM network was employed to map three relationships: temperature to deformation, vehicle-induced deformation to dynamic displacement, and dynamic displacement to vehicle-induced deformation.

Additionally, this issue can be split into two scenarios: recreating data from a faulty sensor using similar sensors (STS) and attempting to recreate data using different sensors (DTS), requiring inter-domain relationship extraction. Various studies have tackled this challenge using different approaches. These include combining CNN with GAN [38], employing GAN and LSTM separately for STS and DTS cases [50], and using a BiLSTM [58]. The proposals were validated using different datasets, such as the Fuchang and Lieshihe bridges in China and the Z-24 box-girder bridge in Switzerland.

However, data loss can also result from factors such as network congestion due to the high dimensionality of sensor data. This aspect remains relatively unexplored in the SHM context. To address this, ref. [48] introduced a three-step damage detection approach involving data compression and recovery, modal parameter identification, and final classification. A model-assisted rakes-based compressed sensing technique (MRAK-CS) was employed in the first phase to alleviate network congestion [88]. Two neural network architectures were compared: a one-class classifier neural network (OCCNN) and an auto-associative neural network (AANN).

With the advancement of computer vision techniques, several crack and corrosion detection approaches based on images captured directly from the structure have emerged [89]. For example, ref. [90] proposed an efficient lightweight encoder–decoder network that integrates attention mechanisms and residual blocks to detect pavement cracks. Similarly, ref. [91] introduced a framework for designing optimized neural network architectures for real-time crack detection. Moreover, ref. [92] presented a novel model that enhances the well-known YOLO computer vision algorithm by incorporating transformer-based components, achieving effective crack detection in the concrete of various bridges.

In addition, two studies focusing on using images as input for the ML algorithm, validated with images of the beam of box-girder bridges, are worth noting [39,55]. The first study tackled the challenge of working with images captured by unmanned aerial vehicles (UAVs). External factors, such as vibration, wind, and UAV movement, often affect these images. Techniques are employed to evaluate and enhance image quality, followed by inputting the processed images into a CNN model for damage classification. Specifically, the CNN model of [93] was used for damage detection using preprocessed images obtained through the mentioned algorithms, leading to better damage detection outcomes.

Conversely, corrosion can cause performance degradation, as it responds to operational and environmental loads, and over time, it can cause more severe failures [94]. Thus, in [55], a hierarchical CNN-based classifier was developed for detecting corrosion in images. A public dataset of UAV images captured in Australia was used to validate the proposed method. Moreover, the Bolte Bridge and the Victoria Sky Rail were selected for evaluation.

In the aftermath of seismic events, local authorities commonly conduct inspections to determine the safety of civil structures such as bridges for public use. Rapid assessment of structural integrity at a regional scale poses a significant challenge [95]. A rapid methodology for evaluating post-seismic structural damage was presented in [33]. This approach employs a set of 21 parameters that incorporate geometric, material-related features [96], and the characteristics of seismic motions [97]. Various ML algorithms were evaluated, including discriminant analysis, k-nearest neighbors (KNN), Naive Bayes, Decision Tree, and Random Forest. In particular, Random Forest achieved the highest accuracy. In a separate work, ref. [56] introduced a novel data-driven strategy to regionally assess the condition of reinforced concrete bridges using machine learning. The inputs encompassed the physical, geometric, and material parameters related to the bridges, combined with seismic motion data from [97]. Six ML techniques were evaluated, with XGBoost emerging as the optimal choice after hyperparameter optimization.

Typically, traditional bridge inspection methods involve visual evaluations and dynamic and static load tests [98]. Although the first two methods are relatively affordable, they may not capture all of the relevant structural health aspects. In contrast, static load tests can address these limitations, but their expense poses a challenge. To address this issue, ref. [36] employed the Kriging spatial inference technique to predict the results of costly static load tests. The process involves developing an initial finite element model of the structure, which is refined using dynamic load test results to better match the actual structure. Then, this enhanced model is used to predict the results of the static load test.

Vibration control systems (VCSs) play a vital role in bridges, with four main types: passive, active, semi-active, and hybrid. Tuned mass dampers (TMDs) are a common passive solution for managing wind-induced vibrations in bridges [99]. Although the optimal TMD design has been well studied, few studies have assessed its effectiveness. To address this, ref. [47] introduced a machine-learning-based method to evaluate the efficacy of TMD. Seven algorithms were compared: Random Forest, ANN, DT, KNN, GBRT, AdaBoost, and XGB. Random Forest excelled in predicting TMD vibrations using wind and temperature-related input features. These studies stand out for their machine learning approaches for predicting bridge or component performance, contrasting with previous applications focused on damage detection or monitoring.

Deflection, a critical metric in bridge assessment [100], can be negatively influenced by the interplay of operational and environmental factors, reducing the precision of the safety assessment methods relying on it. In response, ref. [29] devised an integrated machine learning algorithm based on ensemble empirical mode decomposition (EEMD), Principal Component Analysis (PCA), and Independent Component Analysis (ICA). The algorithm separates the observed signals from the individual deflection components. The approach was theoretically validated using a Hanxi bridge finite element model and a liquid-level sensing system installed on the same bridge in 2015 (Figure 9).

However, a notable challenge with numerous ML models lies in their opaqueness, often operating as “black-box” systems that yield outcomes based on inputs, without unveiling the inherent data relationships learned. Addressing this concern, ref. [57] underscored the potential for better decision-making in preventive maintenance strategies for structures if the predictions from these models could be scrutinized and understood. Their study investigated three model interpretability techniques: Partial Dependence Plot (PDP), Accumulated Local Effects (ALE), and Shapley Additive Explanations (SHAP). These methods were rigorously explored within a seismic engineering context, and their applicability was scrutinized in two case studies: one focused on component-level damage estimation (reinforced concrete walls) and the other on regional damage assessment for inclined bridges in California.

Furthermore, an innovative mobile application for structural health monitoring and damage detection was introduced that embraced advances in smart sensing technology [44]. The app exploits the accelerometers inherent within mobile devices to capture structural accelerations, discern modal parameters, and use the Mahalanobis distance to perform unsupervised structural damage detection. The experimentation began by assessing the app’s efficacy in estimating natural frequencies. Performance comparisons were conducted between the app and other sensor devices, including piezoelectric accelerometers, in controlled laboratory settings and real bridge environments. Remarkably, the study’s findings underscore the application’s remarkable accuracy and substantial cost advantage compared with a comprehensive instrumentation system.

3.2. Suspension Bridges

Suspension bridges are engineering structures consisting of two support towers and a main cable that stretches between them. Secondary cables, also known as stay cables, extend from the main cable to the bridge deck, where the load is supported. Suspension bridges are well suited for crossing large bodies of water or deep valleys because they can span much longer distances than other types of bridges. They can also support heavy loads such as trucks and buses. Compared with other types of bridges, suspension bridges have several distinct characteristics. For example, they are considerably thinner and lighter than beam or arch bridges, making them more elegant and aesthetically appealing. They can also be more resistant to certain types of forces, such as high winds, due to their aerodynamic design. However, suspension bridges are more expensive to construct than other types of bridges, due to their design and construction complexity. Suspension bridges currently account for approximately 10% of the world’s longest and most iconic bridges. Some notable examples include the Golden Gate Bridge in San Francisco, the Akashi Kaikyo Bridge in Japan, and the Hangzhou Bay Bridge in China. Although suspension bridges represent a small portion of the total number of bridges in the world, they have been important to civil engineering and remain a popular choice for bridge construction projects over long distances.

Table 4. Summary of SHM applications using suspension bridges.

References	Year	Structure	Algorithm	ML Purpose	ML Input
Lu et al. [101]	2017	n/a	SVR	Fatigue assessment	Axle wights
Goulet and Koo [102]	2018	Tamar Bridge	BDLM	Signal analysis	Temperature, vehicle load and natural frequencies
Lo et al. [103]	2019	Nova Scotia Bridge	ANN	Prediction residual life	Strain
Silva et al. [104]	2019	Tamar Bridge	DPCA	Data processing	Natural frequencies
Yang et al. [105]	2019	n/a	PCA-CP + Video Process	Fatigue assessment	Images
Deng et al. [106]	2020	n/a	CNN	Fatigue assessment	Images
Jang and Smyth. [107]	2020	n/a	ANN	Signal analysis	Temperature
Ni et al. [108]	2020	n/a	CNN + AE	Feature extraction and data compression	Vibration
Bao et al. [109]	2020	Xiamen Haicang Bridge	ANN	Data compression and reconstruction	Vibration
Wang et al. [110]	2020	n/a	CNN	Cracks detection	Images
Deng et al. [111]	2021	Nanxi Bridge	SVM	Fatigue assessment	Vehicle load
Sun et al. [112]	2022	25 de Abril Bridge	GBRT	Fatigue assessment	Temperature, vehicle load and wind parameters
Sun et al. [113]	2022	n/a	SVM	Fatigue assessment	Strain and temperature
Chen et al. [114]	2023	Huangpu Bridge	DL-AR-ATT	Data reconstruction	Strain and temperature
Sun et al. [115]	2023	25 de Abril Bridge	RF	Interpreting cumulative displacement	Temperature, vehicle load and wind parameters

Note: n/a: Not applicable (data from different bridges was used or the name of the bridge was not reported).

provides a summary of the latest applications of ML algorithms using suspension bridges.

Diverse ML applications focusing on fatigue assessment for suspension bridge analysis have been proposed. To estimate cable tension, ref. [105] employed video processing and principal component analysis (PCA), integrating cable vibration concepts for stress determination. An innovative fatigue estimation method for stays or hangers and using support vector machines (SVM) was introduced to model daily fatigue damage from traffic loading parameters [111]. In [113], a three-step process was presented to assess fatigue by classifying the types of trains that cross a bridge using an SVM for high prediction accuracy. For hot spot analysis in steel beams, ref. [101] employed SVR to replace FEM updates, whereas [112] employed GBRT with monitored parameters for fatigue prediction. The influence of train load on suspension bridge fatigue was emphasized and validated using data from the SHM system of the 25 de Abril Bridge (see [116] for more details).

Similarly, ref. [115] presented an approach to quantitatively interpret the effects of the load on the cumulative displacement, which is vital for assessing the service life of a bridge. This method analyzed correlations between environmental loads, traffic, and bridge response using Shapley additive explanations (SHAP) and random forest (RF) data from the April 25th Bridge. Train loading was the most influential factor in the cumulative displacement. Wind and temperature had a minor impact on a structure’s failure. A comparative study of logic data analysis (LAD), artificial neural networks (ANN), and proportional hazard models (PHM) was conducted [103]. NASA Turbofan data and Nova Scotia bridge SHM data were employed. Although the LAD and ANN models excelled in accuracy, PHM was preferred for scarce or non-representative data. To identify fatigue-induced cracks in welded steel bridge joints, a computer vision-based strategy was introduced [110]. This method detects and segments cracks using IoT image capture, panoramic generation, and convolutional networks, and its efficacy was validated on a real 20-year-old steel suspension bridge.

Numerous studies have focused on data analysis, processing, and compression, as exemplified by [102]. Their study empirically validated the Bayesian dynamic models for SHM to separate the measured structural responses into traffic, temperature, and other components. This decomposition allowed each component to be specifically analyzed, thereby improving the results. The work verified the Bayesian model by considering Tamar Bridge data (the system monitoring and data processing details can be found in [117,118,119]). Modal properties are vital as damage indicators reflecting material changes. In [107], Jang and Smyth explored the impact of temperature on modal parameters, to distinguish damage-induced changes from environmental effects, enhancing the efficacy of the SHM system. Their study evaluated three techniques, including an artificial neural network, to replicate the effects of temperature on natural frequencies.

A novel approach, introduced in [104], adapted deep principal component analysis (DPCA) to capture nonlinear relationships within structural responses, improving the performance of normalization and damage detection. In [108], a one-dimensional CNN was combined with an autoencoder to compress and reconstruct SHM data (Figure 10). The method achieved accurate anomaly identification and minimal reconstruction errors at a compression rate of 10%. A deep-learning-based framework, DL-AR-ATT, integrating an autoregressive model and an attention mechanism for the precise reconstruction of correlated structural responses was proposed in [114]. The approach outperformed hybrid models based on one-dimensional CNNs. Furthermore, the reconstruction of compressive sensing data using neural networks was addressed in [109]. The proposed approach was validated through experiments involving sine waves and wireless data from the Xiamen Haicang Bridge, demonstrating a highly accurate reconstruction. In particular, network parameters offered transparent mapping, ensuring robust interpretability for the neural network used in the compressive-sensing data reconstruction.

3.3. Cable-Stayed Bridges

Cable-stayed bridges are engineering structures consisting of one or more support masts and suspension cables that extend from the mast to the bridge deck. Unlike suspension bridges, cable-stayed bridges do not have a main cable extending between the support towers. Instead, suspension cables support the bridge deck directly from the mast. Cable-stayed bridges can span long distances and are suitable for crossing large bodies of water or deep valleys. In addition, they can support heavy loads and are a popular choice for vehicular traffic bridge construction.

Cable-stayed bridges have several distinct characteristics compared with other types of bridges. For example, their design is simpler and requires fewer construction materials than suspension or arch bridges. Furthermore, they are easier to maintain than the main cables on suspension bridges because they are easier to inspect and repair. However, cable-stayed bridges may be less resistant to certain types of forces, such as high winds, because they are not as aerodynamic as suspension bridges. Cable-stayed bridges currently represent about 20% of the world’s longest and most iconic bridges. Some notable examples include the Biscayne Bay Bridge in Spain, the Peace Bridge in Thailand, and the Millennium Bridge in the United Kingdom. A detailed summary of the latest advances in the application of different ML algorithms, using cable-stayed bridges as a case study, can be found in

Table 5. Summary of SHM applications using cable-stayed bridges.

References	Year	Structure	Algorithm	ML Purpose	ML Input
Pan et al. [120]	2018	Manavgat Bridge	SVM	Damage detection and feature extraction	Vibration
Zhu et al. [121]	2018	n/a	UDS + SVR	Fatigue assessment and computational simulation support	Vehicle load, wind and wave parameters
Lu et al. [122]	2021	Yangtze River Bridge	ASVR	Damage detection in cables and fatigue assessment	Vehicle load, cable and bridge parameters
Entezami et al. [123]	2020	Tianjin-Yonghe Bridge	AE	Damage detection, feature extraction and dimensionality reduction	Vibration
Fang et al. [124]	2020	n/a	SVR	Computational simulation support	Wind and waves parameters
Li et al. [125]	2020	n/a	CNN	Damage detection	Deflection
Li et al. [126]	2020	n/a	CNN	Damage detection	Deflection
Arul et al. [127]	2020	(IPC-SHM)	RF	Anomaly detection	Vibration
Zhang et al. [128]	2021	(IPC-SHM)	LSTM	Damage detection in cables	Cable forces and cable forces ratio
Pamuncak et al. [129]	2021	Suramadu Bridge	CNN	Fatigue assessment	Temperature
Dan et al. [130]	2021	Tianjin-Yonghe Bridge	FESN (RNN)	Damage detection	Vibration
Rastin et al. [131]	2021	Tianjin-Yonghe Bridge	GAN	Damage detection	Vibration
Rastin et al. [132]	2021	Tianjin-Yonghe Bridge	CAE	Damage detection	Vibration
Liu et al. [133]	2021	n/a	DNN	Modal identification	Vibration
Dang et al. [134]	2021	Vasco da Gama Bridge	CNN	Damage detection	Vibration
Mao et al. [135]	2021	n/a	GAN + AE	Anomaly detection	Vibration
Chamangard et al. [85]	2022	Tianjin-Yonghe Bridge	CNN	Damage detection	Vibration
Chou et al. [136]	2022	(IPC-SHM)	CNN	Anomaly detection	Vibration
Liu et al. [137]	2022	(IPC-SHM)	GAN + CNN	Anomaly detection	Vibration
Ye et al. [138]	2023	Sutong Bridge	RF	Performance prognosis	Wind parameters
Shu et al. [139]	2023	Tianjin-Yonghe Bridge	MTDNN	Modal identification	Vibration

Note: n/a: Not applicable (data from different bridges was used or the name of the bridge was not reported).

.

A CNN-based method was employed in [129] to estimate cable tension from temperature fluctuations using data from the Suramadu Bridge in Indonesia. Comparison with other techniques (ANN and linear regression) confirmed the superior cable force estimation capability of the proposed CNN. Furthermore, a probabilistic approach was presented to assess the accumulation of fatigue damage in coastal bridge welds [121]. Stochastic load modeling, followed by finite element analysis aided by uniform design sampling (UDS) and Support Vector Regression (SVR), was employed to manage the computational complexity. The approach is consistent with [124], who used SVR and Latin Hypercubes Sampling (LHS) to replace resource-intensive FEM updates. Experiments using a cable-stayed bridge model revealed the impact of wind on beam displacements and the influence of waves on the shear response of the foundation base.

Regarding cable-stayed bridge cable damage detection, ref. [122] employed a finite element model (FEM) of a real Yangtze-River-spanning bridge to evaluate the reliability of the cable system against fatigue and atmospheric corrosion. Using adaptive support vector regression (ASVR), nonlinear response functions of key components were approximated for cable rupture prediction. The reliability assessment employed the $\beta$-unzipping method, which incorporates an event tree for estimations of varied structural reliability. A hybrid LSTM-CNN architecture for cable damage detection was proposed using a different approach in [128]. The input data included individually measured cable forces and force ratios between cable pairs. Combining these scenarios improved the identification of probable damaged cables, contributing to the achievement of first-place in the IPC-SHM 2020 competition.

Various machine learning techniques are used in SHM for cable-stayed bridges, often using ANNs to detect structural damage via vibration data or other sources. Four deep learning algorithms (Multilayer Perceptron, LSTM, 1D, and 2D CNN) were evaluated for structural damage detection from raw time series data [134]. The distinction between 1D and 2D CNNs lies in the input format: raw signals vs. time–frequency data derived from short-time Fourier transform (STFT). The validation used four case studies, including data from the Vasco da Gama Bridge instrumentation system in Lisbon, Portugal.

In [131,132], the problem of the limited availability of structural data was addressed using two GAN variants for unsupervised identification, quantification, and localization of damage. A deep convolutional GAN (DCGAN) detected and quantifies damage, whereas a conditional GAN localized it. Similarly, in [132], Convolutional Autoencoders (CAE) identified and quantified damage using vibration signals, without damage data. To combat the scarcity of training data, transfer learning was used in [85]. Two CNN components, “source domain” and “target domain”, leveraged numerical model and bridge data, respectively, yielding strong results. In contrast, a previous study [125] used deflection data from a cable-stayed bridge scaled for structural damage detection using a deep CNN. This study notably outperformed classical ML algorithms and stands as one of the few that used deflection data for analysis.

Ref. [120] introduced an innovative SHM approach using SVM with radial kernels in cable-stayed bridges. Advanced feature extraction methods, including wavelet transform, Hilbert–Huang Transform, and Teager–Huang Transform. The effectiveness of the method was evaluated through a numerical simulation using the Manavgat cable-stayed bridge, with wavelet transform showing robustness to noise, while THH and THT demonstrated sensitivity to damage. In another study, ref. [138] explored the behavior of a cable-stayed bridge in high-speed winds, devising an early warning model based on wind variables. The dynamic response of the bridge was predicted using wind turbulence data and random forest vibration amplitude forecasting. This approach was validated on the Sutong cable-stayed bridge, generating early warning alerts using a vibration indicator and threshold.

In a different approach to modal identification from data collected on cable-stayed bridges, a method that employs a customized deep neural network (DNN) for modal shape identification was presented [133]. Using raw vibration data from a real cable-stayed bridge, their approach demonstrated a performance comparable to that of various advanced techniques for identifying modal parameters, underscoring its effectiveness. A time-efficient procedure was proposed for identifying modal parameters using ML [139]. Their multitask deep neural network efficiently extracts independent modes from structural dynamic responses, yielding modal frequencies and damping coefficients. The proposed method provides accurate modal parameter identification, while significantly reducing processing times compared with existing techniques.

A novel model for damage detection using unlabeled data and DL techniques was introduced [123]. Their approach integrates feature extraction through an autoregressive (AR) model, dimensionality reduction using a deep autoencoder, and anomaly detection via the Mahalanobis distance. The data were obtained from the structural health monitoring (SHM) system of the Tianjin-Yonghe cable-stayed bridge (Figure 11) in China (see [140] for more information). The study addressed two key SHM challenges: (1) the high dimensionality of input data, which can impede damage detection algorithms, and (2) the scarcity of labeled structural damage data, which is often required for supervised methods. Despite these challenges, the results highlighted the efficacy of the proposed approach in detecting early damage.

Finally, contemporary DL techniques were applied in [130]. This study addressed the gap that many classification algorithms focus on single-type time series data rather than leveraging combined multivariate sensor data. The Functional Echo State Network (FESN), a variant of Recurrent Neural Networks (RNNs), was introduced to bridge this gap. FESNs, part of the Echo State Network (ESN) family, have broad applications in engineering, telecommunications, and various fields involving signal processing. In a pioneering approach, this neural network architecture was augmented with the Particle Swarm Optimization (PSO) algorithm to optimize hyperparameters and enhance accuracy, yielding rates exceeding 97%. The study is a rare example of the integration of metaheuristics into a framework.

Several studies have introduced techniques using the IPC-SHM dataset in the context of anomaly detection due to sensor failures in SHM systems [127,135,136,137]. A method that combines “Shapelet Transform” with Random Forest for anomaly classification was proposed in [127]. Shapelet Transform uses time series data shape for representation, finding Shapelets that best match the data to create a new feature space. Despite automatic anomaly detection, challenges related to dataset specifics were noted, suggesting potential solutions through preprocessing.

In [135], GANs and AEs were used to transform sensor data into Gramian Angular Field images using advanced image processing. With an average accuracy of 99% for anomaly differentiation, the proposed approach demonstrated high accuracy and robustness. Moreover, various NN models, such as ANN, CNN-based spectrogram, and temporal history models, and an image-based time–frequency hybrid CNN (GoogLeNet), have been explored [136]. GoogLeNet performed exceptionally well; however, confusion due to class similarity was acknowledged. To address this, an ensemble model that combines the four methods was introduced for improved results. Finally, a model comprising three modules was proposed in [137]. The first transforms acceleration signals into a 3-channel space using time, frequency (via Fourier transform), and Gramian Angular Field representations. GANomaly is employed for feature extraction in imbalanced SHM data. Finally, ResNet-18 is used for anomaly classification.

3.4. Arch Bridges

Arch bridges are characterized by a curved arch structure that spans from one end of the bridge to the other. Arch bridges are capable of supporting heavy loads due to their arch design, which distributes the load of the bridge evenly across its structure. Arch bridges can be constructed from stone, brick, reinforced concrete, or steel. Unlike suspension and suspension bridges, arch bridges do not require suspension cables or support towers to hold the bridge deck in place, allowing for a simpler and more economical construction. Another distinguishing feature of arch bridges is their elegant and often ornate appearance. Arches can be single or multiple curves and decorated with sculptures, mosaics, reliefs, and other decorative elements. Arch bridges can be built at various heights and lengths, making them suitable for various settings and uses.

As shown in

Table 6. Summary of SHM applications using arch bridges.

References	Year	Structure	Algorithm	ML Purpose	ML Input
Anaissi et al. [141]	2017	Sydney Harbour Bridge	SVM	Damage detection	Vibration
Chalohi et al. [142]	2017	San Michele Bridge	ANN + GP	Damage detection	Vibration and temperature
Khoa et al. [143]	2017	Sydney Harbour Bridge	TD + KMeans + SVM	Damage and anomaly detection	Vibration
Alamdari et al. [144]	2017	Sydney Harbour Bridge	K-Means	Damage detection	Vibration
Anaissi et al. [145]	2018	Sydney Harbour Bridge	SVM	Hyperparameter tuning and damage detection	Vibration
Ruffels et al. [146]	2020	n/a	ANN	Damage detection	Vibration
Ye et al. [147]	2021	Jiubao Bridge	GP	Forecast of wind speed	Wind parameters
Gardner et al. [148]	2022	KW51 Bridge	DA-GMM	Damage detection	Vibration
Feng et a. [149]	2022	Jiangnan Arch Bridge	SVM	Damage detection in cables	Cable parameters
Civera et al. [150]	2022	n/a	K-Means	Damage detection and modal identification	Vibration

Note: n/a: Not applicable (data from different bridges was used or the name of the bridge was not reported).

, the most commonly used bridge for validating works that used data from an arch bridge was the Sydney Harbour Bridge, described above, although these studies were condensed between 2017 and 2018. In turn, a recurrent use of the KMeans clustering technique can also be seen; although, as will be seen below, they were used for different reasons.

Although SVMs and the KMeans clustering algorithm have been extensively used in an arch bridge SHM context, they have also been applied in different approaches. An SVM with a Gaussian kernel was employed in [141], where a performance problem due to the sensitivity of the $\sigma$ parameter of the Gaussian kernel was highlighted. To address this issue, a new algorithm called Appropriate Distance to Enclosing Surface (ADES) was proposed, which allows adaptive tuning of the Gaussian model parameter. Data obtained from the Sydney Harbour Bridge were used to validate the damage detection capabilities and parameter tuning of this method, in addition to various other datasets. The proposed algorithm can identify suitable sigma parameter values when data are high-dimensional.

Simultaneously, ref. [145] extended this research by including additional experiments. In [143], a data tensor decomposition (TD) algorithm was proposed that allows a machine learning algorithm to process the temporal, spatial, and feature attributes of data simultaneously. In addition, SVM is employed to detect and evaluate the severity of the damage progress, and KMeans is employed to locate the structural damage by exploiting the performed tensor decomposition. The proposed method was validated using structural data collected from the Sydney Harbour Bridge, and the results included the correct detection of damage and the estimation of different severity levels, as well as the detection of spatial anomalies associated with sensor and instrumentation problems.

In [144], the authors proposed a novel approach based on spectral moments (SM). These moments contain information on various frequencies and are useful for structural damage detection. In addition, a variant of the KMeans algorithm was used to group the results and determine which bridge arches showed anomalous responses. The method was validated in two stages. In the first stage, data from the dynamic responses of bridge arches with known problems were input, and the method was able to correctly identify them. In the second stage, the amount of data from different arches was expanded, and the arches identified as damaged by the model were found to have some type of unknown structural damage during testing through extensive investigation and evaluation.

Moreover, a new multistage clustering algorithm for automatic operational modal analysis (AOMA) was presented in [150], which aims to detect and evaluate structural damage. This method was applied to a scale model of a historic masonry-built arch bridge in a laboratory environment. Operational modal analysis (OMA) is an output-only technique that uses modal parameters as features and damage indicators. The proposed algorithm includes a modal identification component based on clustering algorithms, particularly the KMeans algorithm. The feasibility of applying the proposed algorithm in civil structures was demonstrated through experimental validation, although the focus of the work was on masonry-built arch bridges.

Finally, a model designed for cable damage detection using the Jiangnan Arch Bridge as a case study was presented in [149]. However, a finite element model of the structure was employed to validate the model. The model aims to investigate the limits of cable damage levels due to fire and corrosion effects. Several machine learning techniques, including a backpropagation neural network (BPNN), a radial basis function neural network (RBFNN), and an adaptation of the support vector machine using least squares (LS-SVM), were tested. Among the obtained results, LS-SVM achieved a higher accuracy in predicting cable damage under the combined effects of fire and corrosion.

Techniques such as artificial neural networks (ANN), and others based on the Gaussian Process (GP) and GMM, have been reported. An ANN was used in [146] to predict the acceleration of an undamaged structure (a scale model of an arch bridge in a laboratory). After establishing a healthy reference state, damage-sensitive features were identified by calculating the root mean square error between the actual measured data and those predicted by the ANN. Outliers were used as possible indicators of damage. In addition, these outliers were detected using two methods: the Kolmogorov–Smirnov test and the Mahalanobis distance. The results showed that the Mahalanobis distance was more effective in detecting gradual damage than the Kolmogorov–Smirnov test was for identifying small isolated damages.

Following an approach similar to [142], the authors detected and located damage to the historic San Michele railway bridge in northern Italy (Figure 12) using a four-step method that used data such as ambient temperature and dynamic responses of the bridge deck measured during traffic. The first step of the method involved obtaining reference data corresponding to the optimal state of the bridge for use in training and data on the current condition. Then, preprocessing was performed to extract the characteristics of the trains that cross the bridge. In the third stage, a neural network and a GP model were trained for each installed accelerometer. Finally, in the last stage, the acceleration of each sensor was predicted, and the bridge was damaged if the difference from the actual responses exceeded a predefined threshold.

Furthermore, data obtained from the structural health monitoring system (specifically, wind records) installed on the Jiubao Bridge in China were used by [147] to develop an emulator model based on a Bayesian approach and a GP, to perform a probabilistic prediction of wind speed. This variable is of great importance, because it can influence a structure’s dynamic response and act as an external stress factor, independently of operational variables. The Bayesian emulator used various covariance functions, such as the squared-exponential (SE), Matern (MA), periodic (PE), and composite functions. These functions were selected because they were sensitive parameters of the process and could provide accurate results. Moreover, a novel approach to SHM, known as population-based structural health monitoring (PBSHM), arose from an issue existing in most data-driven SHM systems: these systems are typically constrained by data from a limited number of structural health states, restricting their ability to detect new types of damage.

Recognizing this limitation, PBSHM expands a labeled health dataset by considering the population structure. This approach enhances the likelihood of having more labeled information available, thereby increasing the diagnostic potential of new structural health states [151]. In light of this, ref. [148] proposed a new method based on the domain-adapted Gaussian mixture model (DA-GMM), which performs inference through a linear mapping that transforms the target structure data into a GMM inferred from data from another structure. Three case studies were examined to validate the model, one of which corresponded to a heterogeneous population of bridges. In the case, data from the Z-24 (box girder bridge) and the KW51 (arch bridge) were used as part of the population.

4. Discussion

The following section discusses the most relevant findings identified in this systematic review. First, a general comparison is made between the most commonly used algorithms, discussing their main advantages and limitations, which can serve to guide new researchers in choosing the type of algorithm to use depending on the problem to be addressed. In addition, some challenges related to the implementation of an automated structural condition monitoring system are analyzed. Finally, each of the previously developed research questions is addressed.

4.1. Brief Description of Some Commonly Used ML Algorithms

Table 7. Overview of some of the most commonly used ML algorithms.

Algorithm	Frequency of Use	Main Data Types	Main SHM Tasks	Advantages	Limitations
CNN	17	images, vibration, deflection	image analysis, data compression, feature extraction, damage detection	Automatic feature extraction	Lower effectiveness with sequential data
SVM	11	Tabular data (cable, wind, wave parameters, etc)	fatigue assessment, computational simulation support, hyperparameter tuning	Good performance with few and high dimensionality data	Sensitivity to hyperparameter and kernel selection
AE	6	vibration	anomaly detection, damage detection, feature extraction	Unsupervised anomaly detection	Sensitivity to noise and unrepresentative data
ANN	6	strain, temperature, vibration	damage detection, feature extraction	Flexibility for various types of data and problems	Risk of overfitting with limited data
GAN	5	strain, displacement, vibration	data reconstruction, data augmentation	Ability to learn complex distributions	High computational demand
Random forest	5	Tabular data (bridge, earthquake, wind parameters, etc)	prognosis (damage, performance, etc)	Robustness against noise and overfitting	Heavier and slower models for predictions
LSTM	4	strain, displacement, vibration	data reconstruction	Suitable for sequential data	High computational cost

provides a comparative overview of the most frequently used ML algorithms in bridge SHM. This section summarizes their frequency of use, main data types, SHM tasks, and key advantages and limitations. CNN-based approaches stand out due to their strong performance in image-based tasks and automatic feature extraction, while SVM and random forest remain popular for tabular datasets and scenarios with limited data. Emerging methods, such as GANs and LSTMs, are gaining traction, particularly for data reconstruction and sequential signal analysis, although they often require higher computational resources. This comparison highlights the versatility of DL methods and their trade-offs in terms of computational cost and data requirements.

4.2. Challenges of Implementing an Automated SHM System

Automated monitoring systems are generally preferred over traditional visual inspection techniques. Among their various advantages, such systems enable continuous real-time monitoring of a bridge, whereas the periodic visiting of experts to assess structural conditions limits visual inspections. Moreover, the evaluation process becomes more objective by relying on computational techniques, minimizing the dependence on the experience and subjective perception of inspection personnel. Another significant advantage is the potential reduction in long-term costs, primarily by avoiding the operational interruptions required for manual inspections and by enabling early detection, thus reducing the likelihood of costly repairs after major damage has occurred.

However, the implementation of an automated SHM system involves several challenges that must be carefully addressed to ensure its proper and reliable operation. Some of these challenges are as follows:

• Device selection: A crucial aspect of implementing an SHM system is deciding which variables should be monitored to ensure that the algorithms analyzing the data have access to the most comprehensive information possible. For example, if the dynamic response of the bridge (e.g., vibration) is of interest, accelerometers would need to be installed, and a decision must be made regarding the appropriate type (uniaxial, biaxial, or triaxial) to provide sufficient information for the vibration analysis algorithms to detect structural changes caused by damage or deterioration. Monitoring certain environmental variables (e.g., temperature, wind speed, or humidity) may also be beneficial, as these parameters can enhance the performance of ML models by providing them with more context for making informed inferences.
• Sensor costs: In practice, the acquisition and installation of sensors in bridge-type structures requires a considerable initial investment. Therefore, the goal is to obtain the largest possible amount of information using the smallest number of sensors. The sensor locations must be carefully selected to maximize the collected information. Therefore, an entire area of research is dedicated to optimizing sensor placement to minimize implementation costs and maximize the quantity and quality of the information captured by the monitoring system. More information on this topic can be found in [152,153].
• Data transmission: Once the sensors have been installed in the structure, they generate an immense volume of data, highlighting the need for big data techniques. These data must be transmitted from the structure to a dedicated server for storage and processing, allowing structural health assessment tasks to be performed. However, the transmission process is challenging due to the sheer volume of information. Therefore, data compression methods [154,155] and missing data reconstruction [156,157,158] are essential to ensure efficient and reliable data handling.
• Data quality: The performance of ML algorithms is highly dependent on the quality of the available data. Without reliable data, achieving acceptable results is generally not feasible. Data collected from real operational structures (as opposed to controlled laboratory settings) often exhibit various sources of noise, which may arise from sensor sensitivity, data transmission processes, or environmental variables. Consequently, applying preprocessing techniques aimed at improving data quality before they are fed into ML algorithms is essential. This step can significantly improve accuracy and enable the monitoring system to effectively outperform traditional visual inspections.
• Hardware requirements: As briefly mentioned in Section 4.1, some algorithms are computationally intensive, requiring a sufficiently powerful infrastructure to execute the demanding processes involved in automated monitoring. This often entails additional costs associated with specialized computing services for running ML techniques. In some cases, these methods may be deployed directly on the devices installed on the structure, which further limits the available computational capacity. Therefore, when proposing or implementing new techniques, hardware constraints must be carefully considered.

Key challenges, such as device selection, sensor costs, data transmission, data quality, and hardware requirements, strongly influence the successful implementation of automated SHM systems. These factors affect the reliability and scalability of monitoring solutions and determine their feasibility for large-scale deployment. Future research should focus on developing cost-effective sensor placement strategies, robust data preprocessing pipelines, and lightweight yet accurate ML models that can operate under limited computational resources.

4.3. Discussion of Research Questions

The following subsection addresses each of the research questions outlined in the introduction, providing a synthesis of the findings and a critical discussion of how the reviewed literature contributes to answering them.

4.3.1. RQ1: For What Specific Purposes Have Machine Learning Algorithms Been Applied to Perform SHM Tasks?

As discussed previously, each ML technique has distinct advantages and limitations, which makes careful selection of the most suitable approach essential, depending on the specific problem to be addressed. As discussed in Section 3, various ML algorithms have been applied to several SHM tasks in bridges. These tasks include some of the core functions of SHM systems, such as fatigue assessment, structural damage detection, crack identification, and detection of anomalous structural response behaviors or abnormal patterns.

Moreover, the integration of ML techniques and Big Data tools introduces additional challenges that must be addressed to ensure the effectiveness of automated SHM systems. For instance, each ML algorithm involves a set of hyperparameters that can critically influence its performance, leading to the application of optimization techniques, including other ML-based methods, to fine-tune these parameters. Additionally, because ML models typically require large amounts of data to achieve reliable performance, new data handling requirements emerge, including data transmission (e.g., compression of sensor data and reconstruction of missing data) and preprocessing tasks (e.g., feature extraction, data augmentation, and signal processing). Specialized ML algorithms are often employed for these complementary tasks, enhancing the overall efficiency and accuracy of SHM pipelines.

Although significant progress has been made in applying ML to a wide range of bridge SHM tasks, there is substantial room for improvement, particularly in increasing the robustness of these techniques under real-world operational and environmental conditions and in bridging the gap between research prototypes and field-deployed systems.

4.3.2. RQ2: What Types of Data Have Been Used to Train and Validate These Algorithms?

Moreover, the input variables used in the ML models were declared in the reviewed literature, thereby gaining valuable insights. Damage identification based on vibration data is one of the most widely studied research avenues. This is mainly due to the prevalence of vibration as the variable chosen for three of the four bridge typologies examined (Figure 13). This damage identification approach detects structural-damage-induced alterations in certain modal properties [13]. This research strand has maintained substantial interest within the scientific community over the years [159,160] and appears poised to continue its prominence.

However, in the suspension bridge typology, the emphasis on vibration data was less pronounced (Figure 13). This could be attributed to the fact that the application of ML techniques in this type of bridge does not seem to have been primarily focused on damage detection. As illustrated in Figure 14, studies on this typology have mainly aimed to investigate the variables associated with the bridge’s degradation throughout its operational life. Therefore, objectives such as interpreting cumulative displacement, predicting residual life, and assessing fatigue collectively represent a larger proportion of the goals of employing ML models for suspension bridges.

However, it should be noted that simply measuring structural vibrations is insufficient, because these can be influenced by environmental or operational variables [161]. These environmental variations are often attributed to factors such as humidity, wind, and temperature [68]. As shown in Figure 13, a concerted effort is currently underway to incorporate these variables into the studies being conducted. This is done to improve the results obtained, by distinguishing changes caused by external variables from those resulting from structural damage, ultimately achieving improved outcomes.

4.3.3. RQ3: How Do Machine Learning Applications Differ Between Different Types of Bridges?

Not all studies apply ML techniques for the same purpose, as shown in Figure 14. Given the ability of ML techniques to analyze large volumes of data and extract hidden patterns [6], diverse objectives should be considered when applying them. Here, structural damage detection is currently the primary reason for employing ML techniques in bridge monitoring, except for suspension bridges. However, common objectives for the four bridge typologies studied generally include feature extraction, fatigue assessment, anomaly detection, data reconstruction, and modal identification.

This study’s classification based on the different bridge typologies to which ML techniques have been applied also provided intriguing insights. First, some scenarios may have techniques tailored to a specific type of bridge. For example, significant attention has been paid to the evaluation of seismic damage applied to box-girder bridges, driven by the prevalence of this type of bridge compared to others [162]. This prevalence enables a more extensive assessment of the seismic response in numerous structures of a single type. In addition, there are situations in which techniques should function independently of the bridge typology. Consequently, greater emphasis should be placed on the simultaneous validation of future proposals across multiple bridge types. Ultimately, the accessibility of data from diverse bridge types does not seem to impede the exploration of technique robustness across various typologies.

Finally, considering the findings presented in Section 3.1, highlighting the multitude of studies that have employed the Z-24 bridge dataset for damage detection, and the observations made in Section 3.3 regarding the recent surge in research focused on anomaly detection, the increased availability of sufficiently large high-quality datasets is conducive to stimulating research interest in a particular field. This underscores the imperative to provide researchers and academics with new datasets derived from diverse complex structures, targeting various objectives (damage detection, anomaly detection, modal identification, damage localization, damage classification, etc.). The provision of such datasets aims to foster a more comprehensive and well-rounded development of this research domain.

4.3.4. RQ4: Are the Technical Aspects Related to the Application of Machine Learning Algorithms in Real-World Case Studies Reported and Discussed?

After a thorough review of all the compiled work, several notable findings that merit attention emerged from the analysis. Specific details on the implementation of these methods from a more technical perspective were often missing. These details can be of significant relevance, because they not only offer insights into potential areas of improvement, but also serve as valuable resources for newcomers to the field by providing specific implementation information.

An interesting aspect to analyze, though often absent in research work, pertains to the execution times of the training stage, data processing, and operational phases (e.g., damage detection, localization, and classification). The latter two phases are particularly significant because SHM systems are expected to interact with real-time data [163]. Tasks such as feature extraction, modal identification, and damage detection must be performed within a short timeframe [164]. This aspect must be considered because it may render some proposed techniques impractical for real-world applications, due to their processing time. However, this also presents an opportunity for future research focusing on optimizing these algorithms.

Another aspect worth considering is the hardware requirements, such as the processing units (CPU, GPU, RAM, etc.), for data analysis and ML algorithm execution. This holds significance because computational resources greatly vary; for instance, a portable computing unit such as a Raspberry Pi with a 1.2 GHz processor and 1 GB of RAM differs substantially from a dedicated server housing last-generation processors and more than 128 GB of RAM. Consequently, this disparity can influence processing times and the volume of data that can be handled, potentially limiting the applicability of proposed techniques and impeding their implementation for comparative analysis with other algorithms. Unfortunately, most of the reviewed works tended to overlook this aspect, underscoring the need for future research efforts to provide detailed specifications in this regard.

4.3.5. RQ5: What Are the Main Gaps in Research and Future Directions Suggested in the Current Literature?

Following this brief analysis, we discuss some lines of research that were identified during the examination of the studies. These directions may encompass new avenues for exploration and extensions of existing lines that have not yet been thoroughly investigated.

• Data Fusion: As illustrated, most studies typically relied on a single type of data to feed the employed ML models (e.g., acceleration). Various environmental variables (collected by instrumentation systems) can influence structural responses in certain cases [29,102,107], potentially leading to low-quality outcomes. Consequently, real-time analysis of these alterations could differentiate between structural damage and environmental variations. In this regard, data fusion techniques can be employed to improve an algorithm’s insights, by integrating as much information about the structure as possible. For instance, ref. [165] proposed an unsupervised approach that fuses heterogeneous sources—dynamic response, strain measurements, and environmental variables (temperature, wind speed, and wind direction). Leveraging multimodal data substantially increased the robustness of the model and reduced false positives in damage detection.
• Novel SHM Approaches: An intriguing trend that recently emerged is the exploration of novel approaches to address SHM challenges, particularly regarding the scarcity of available data, both in healthy (typically abundant) and damaged states. These approaches encompass population-based SHM [166], transfer learning-based SHM [167], drive-by bridge monitoring [19], and prognosis-based damage detection, a less-explored concept. Each of these approaches offers distinct avenues for further research.
• Prognosis-Based Damage Detection: As mentioned earlier, this emerging approach received relatively little attention in the literature reviewed and warrants a separate discussion. Training ML models with data collected from sensors installed on healthy structures is the core of prognosis-based damage detection. The goal is for these algorithms to learn to predict a healthy structure’s dynamic responses. After the training phase, these models are used to continuously predict and compare these responses with actual sensor-acquired responses. The differences are then computed using predefined metrics. If the calculated metric value exceeds a predefined threshold, the structure is likely to have undergone some degree of structural damage, because the predicted response statistically deviates from the actual response. One advantage of this approach is that it does not require labeled data, which can be scarce. Furthermore, model training is often conducted at the sensor level (with one model per sensor) or for substructures, aiding in the approximate localization of the damaged section.
• Hybrid Models: Given the persistent challenge of limited availability, whether this pertains to labeled data or data from diverse structural states, it is imperative to explore alternatives that can yield a greater volume of data for training ML models. In this regard, a contemporary approach involves hybrid models that combine the strengths of data- and model-driven methodologies. These hybrid models enable the complementary integration of these two perspectives. This field is still in its nascent stage, providing ample room for innovation and novel contributions, as evident from the discussions. A comprehensive review of this topic can be found in [168].
• Applicability to Different Bridge Typologies: Following an analysis of the ML techniques applied to various bridge typologies, relatively few studies encompassing multiple bridge types in their validation were obvious. This observation suggests a potential avenue for investigation, specifically regarding the adaptability of the proposed algorithms across diverse bridge typologies. Exploring modifications that could enhance the results when considering a heterogeneous population of bridges is important. This could lead to a more comprehensive understanding of these techniques’ generalizability across different structural forms.

One major limitation of this review is the concentration of studies on well-known bridges, particularly the repeated use of the Z24 bridge as a case study. This overrepresentation may introduce selection bias and limit the generalizability of the findings to other types of bridges and real-world conditions. Furthermore, future research could consider analyzing bridge materiality alongside structural typology, which may allow for a more refined disaggregation and categorization of each study.

Finally, although this review followed a systematic and transparent methodology, it was neither preregistered nor accompanied by a formal protocol. Potential biases arising from post hoc decisions during the study selection and synthesis phases may have been introduced. Furthermore, heterogeneity in the manner in which the reviewed studies reported their methods and results, combined with the wide range of SHM-related tasks addressed, hindered the implementation of a standardized quality assessment or a meaningful quantitative synthesis, which may have impacted the comparative analysis’s comprehensiveness. Finally, future research in this area could greatly benefit from reporting the performance of the proposed techniques after a trial phase in real and operational environments (e.g., an active bridge). Such real-world validation would provide the most reliable evidence of the practical capability and robustness of methods presented under realistic operational and environmental conditions.

5. Conclusions

This article reviewed recent advances in SHM that applied machine learning ML algorithms using real bridge data, with a focus on studies that considered any of the four most studied types of bridges: box-girder bridges, suspension bridges, arch bridges, and cable-stayed bridges. This emphasis underscores the importance of developing research directions geared toward industry and real-world problem solving.

A systematic review process was conducted, and 79 studies, covering almost 30 different bridges around the world, met the inclusion criteria. Box-girder bridges made up 41% of the corpus, cable-stayed 29%, suspension 18%, and arch 12%. Vibration was the most commonly used input data for box girder, cable-stayed, and arch bridges (54.8%, 48.1%, and 72.7%, respectively), whereas vehicle load and temperature were the most commonly used input data (45.4% in total) for suspension bridges. Similarly, damage detection was the main objective of using ML algorithms for box girder, cable-stayed, and arch bridges (with 50%, 34.6%, and 61.5%, respectively), whereas fatigue assessment was more studied (35.3%) for suspension bridges.

This review highlights the ML algorithms, their applications, and input data used in SHM studies on real bridges, emphasizing the novelty of the latter in-depth analysis. Research on ML-based SHM has expanded rapidly; however, there has been a particular focus on box girder structures, which may be due to a limited set of reference data for other types of bridge. The evidence base is skewed toward a single bridge that has been repeatedly analyzed (Z-24), introducing a potential selection bias. Future studies should broaden bridge types, improve validation, and standardize reporting.

Additionally, several lines of research that emerged from the analysis of the reviewed works have been presented. These lines can serve as starting points for researchers entering this field or as a foundational basis for ongoing studies. Furthermore, in this data-driven era, technological advancements will always go hand in hand with the availability of data to test and validate the various proposals in real case studies. This was evidenced by the clear surge in research dedicated to specific areas, such as anomaly detection, a task increasingly sought-after by the industry. This surge was fueled by the 1st International Project Competition for SHM in 2020, which aimed to encourage the scientific community to develop methods and techniques in the context of SHM.

Real case studies were conducted to validate the robustness, fault tolerance, and sensitivity of the proposed algorithms to variations arising from operational and environmental variables. This validation is essential to ensure the applicability of these techniques to real-world industrial problems, thereby preventing them from remaining solely as state-of-the-art methods in academia.

A comprehensive overview of the most significant advancements in academia was provided to summarize the contributions of this work. Furthermore, machine learning techniques are extending their reach across various domains within the field of structural health monitoring. Hence, it is anticipated that, in the future, not only will these domains be further explored in depth but also new lines of research will emerge, as presented in this study. Finally, the review conducted in this work will be of great benefit to researchers and academics who wish to dive into this area of engineering, offering a perspective rooted in applied sciences.