Abstract
Traditional bridge inspection methods, such as visual inspections and basic non-destructive tests, remain foundational but face limitations due to labor intensity and dependency on inspector expertise, often resulting in inconsistencies. Machine learning (ML) tools offer transformative solutions by enabling the analysis of large datasets to enhance damage detection accuracy, optimize maintenance schedules, and improve resource allocation. The objective of this study was to explore the integration of advanced inspection tools, such as Unmanned Aerial Vehicles (UAVs), digital imaging, and fiber optic sensors, with ML models to enhance data accuracy, decision-making, and inspection rating efficiency. A state-of-the-art review is conducted in this study on the application of ML techniques in bridge inspection and maintenance, encompassing 60 articles and reports published in the last decade. The results of this study show that ML’s ability to integrate advanced inspection tools to improve data accuracy, decision-making, and inspection efficiency. However, the existing challenges persist in data quality, model generalization, and the need for standardization approaches across diverse bridge conditions. The review provides insights into current methodologies, benefits, limitations, and future directions, emphasizing ML’s pivotal role in modernizing bridge health monitoring for a safer and more sustainable infrastructure network. The findings of this study are expected to assist transportation agencies in planning and operating predictive maintenance strategies, focusing on damage detection, maintenance optimization, and resource allocation.
Keywords
Introduction
Bridges are an important component of the United States’ transportation system, connecting communities, facilitating the movement of goods, and enabling societal advancements by overcoming natural and manmade obstacles like mountains, railroads, and rivers (FHWA, 2023). Any damage to bridges can cause economic losses, service disruptions, and safety concerns for the traveling public (Pratama et al., 2024; Xiang et al., 2023). In the United States, the federal highway office of bridges and structures offers technical guidance and national policies regarding the design, construction, inspection, assessment, administration, and preservation of the country’s inventory of highway bridges, tunnels, culverts, walls, and other ancillary structures (FHWA, 2023). Approximately 42,000 bridges in the United States are structurally deficient, many of which are nearing the end of their designed lifespan (Meyersohn, 2024). The average age of bridges across the country is 44 years. The $1.2 trillion federal infrastructure bill passed in 2021, which allocates $110 billion for roads, bridges, and key infrastructure projects, could provide funding for essential bridge upgrades. The existing transportation statistics highlight the critical need for investment to enhance climate resilience and ensure the reliability and safety of the country’s bridge network (Argyroudis et al., 2022).
Bridges are complex structures, consisting of several elements, including decks, girders, piers, and foundations (Bian et al., 2013; Ellobody, 2023). Various construction materials (e.g., steel, concrete) and methods (e.g., reinforced, prestressed) contribute to different durability levels and maintenance needs (Bian et al., 2013; Cakir et al., 2021; Ellobody, 2023). Decks provide a travel surface for vehicular traffic and are typically made of reinforced concrete (RC), steel, open grating, or wood (Han, 2024). Concrete provides higher durability, whereas steel provides the capacity for a lighter design with a much greater load capacity than would otherwise be possible (Han, 2024). Girders, generally composed of steel, RC, or prestressed concrete (PC) materials, support the deck and act as the primary load-bearing element of the bridge superstructure (Ellobody, 2023). Abutments, piers, and foundations constitute the substructure part of the bridge and are primarily constructed from RC material (Han, 2024; Stefanidou and Kappos, 2023).
Bridge inspection is essential for ensuring public safety and infrastructure longevity, assisting in identifying and addressing structural deficiencies such as concrete cracks, steel corrosion, excessive deformations, and other forms of degradation (Shahin et al., 2024). Effective inspection allows for timely maintenance and repairs, preventing catastrophic failures and extending bridge life (Boone, 2023). The American Road & Transportation Builders Association indicated that approximately one in three U.S. bridges requires repair or replacement and that 42,400 bridges are rated in poor condition (ARTBA, 2024). The annual cost of additional maintenance and reduced production due to unplanned bridge failures in the United States is estimated at $750 billion, underscoring the critical need for robust and effective bridge management techniques (Wedel and Marx, 2022).
Background—Conventional inspection methods
Conventional bridge inspection methods include visual inspection, manual measurement, and basic non-destructive tests (NDT) like sounding and chain-dragging (FHWA, 2022). These methods are valued for their simplicity, cost-effectiveness, and ability to quickly identify visible defects like cracks, corrosion, and delamination (Lee and Kalos, 2015). However, the effectiveness of conventional bridge inspection is limited by its dependence on the inspector’s expertise, which can lead to variability in results. It is also labor-intensive, time-consuming, and can miss internal or subsurface defects that are not visible on the surface, potentially allowing hidden structural issues to go undetected (Shahin et al., 2024).
The inherent limitations of the conventional inspection methods highlighted the need for more advanced inspection technologies to complement and enhance traditional methods, to ensure a more thorough assessment of bridge safety and integrity (Alabama DOT, 2021; California DOT, 2017). Advancements in inspection technology, including advanced NDT, high-resolution imaging, unmanned aerial vehicles (UAVs), and robotic platforms, have enhanced the efficiency, expediency, and accuracy of inspection, allowing for detailed assessments and continuous monitoring (ARTBA, 2024).
Background—Advanced inspection methods
Advancements in the fields of data analytics, machine learning (ML), and artificial intelligence (AI) have greatly enhanced bridge inspection and maintenance (Stevens et al., 2020; Zhang et al., 2023b). These methods have several advantages, such as allowing the automated analysis and processing of massive amounts of inspection data, providing early warning of potential issues (Gagliardi et al., 2021; Neves et al., 2017), and enabling the creation of maintenance prediction models that result in optimal resource allocation and failure prevention (Ye et al., 2021). Additionally, digital twins and Building Information Modeling (BIM) provide real-time dynamic visualizations to develop bridge health (Yang and Xia, 2023).
The synergy in improvements of hardware and software has resulted in exponential advancements in structural health monitoring (SHM) and their adoption in bridge inspection (Assaad and El-adaway, 2020; Gomez-Cabrera and Escamilla-Ambrosio, 2022). Predictive maintenance is a method in which the service life of a bridge is predicted based on inspection or diagnosis using an array of sensors to continuously collect real-time data. The sheer volume and complexity of the data collected can be difficult to interpret using traditional methods such as visual inspection (Zinno et al., 2023). ML models offer a promising solution to this problem (Gomez-Cabrera and Escamilla-Ambrosio, 2022).
Despite the progress made, the integration of ML techniques within bridge SHM remains limited, as many bridge authorities continue to rely on traditional methods (ARTBA, 2024). ML models such as ANNs, SVMs, and RF have been used in SHM, accounting for 27.84%, 17.72%, and 13.92% of existing applications respectively (Rashidi Nasab and Elzarka, 2023). Similarly, ML techniques: convolutional neural network (CNN) and long short-term memory (LSTM) networks are effective in processing complex datasets, detecting cracks, and predicting structural responses (Assaad and El-adaway, 2020; Catbas and Malekzadeh, 2016; Jeong et al., 2022; Meddaoui et al., 2024). The use of sensors, particularly temperature (60.71%) and vibration (46.42%) types, is integral to collecting real-time data for ML analysis (Catbas and Malekzadeh, 2016; Chalouhi et al., 2017; Darsono and Torbol, 2017).
Significance and objectives
List of common abbreviations and acronyms.
Conventional and advanced bridge inspection methods
Several methods are employed in bridge inspection. Conventional methods, as outlined in the FHWA document “Methods and Equipment for Bridge Inspection,” include using binoculars, flashlights, and digital cameras for capturing visual data, dimensional measurements, and photographic records of cracks, corrosion, deformations, and overall structural health Figure 1 (Dues et al., 2021; Rossow et al., 2012). Underwater inspections utilize diving gear, underwater cameras, and ROVs to examine submerged elements, marine growth, and structural damage. NDT tools like hammer sounding and ultrasonic testing are used to detect delamination and voids, and assess concrete quality through physical samples. State DOTs have used conventional methods because they are well-established and easy to implement with the right expertise. These methods have been the foundation of bridge inspection programs for decades (Alabama DOT, 2021; California DOT, 2017). Sample of traditional bridge inspection methods.
However, these traditional approaches have limitations. They are labor-intensive and time-consuming, often requiring huge manpower and resources to cover large or complex structures (Civera et al., 2022; Zhang et al., 2023a). Additionally, the accuracy and reliability of these inspections are highly dependent on the skill and judgment of the inspectors, which can lead to inconsistencies in data and potential oversights, especially for internal or subsurface issues that are not visible during routine inspections (Civera et al., 2022; Zhang et al., 2023a). Additionally, the increase in demand for bridge inspections and workforce disruptions, such as that caused by the COVID-19 pandemic led several states to face shortages in human inspectors (Assaad and El-adaway, 2020; Ye et al., 2021).
Advanced bridge inspection tools include digital image correlation, laser scanning, UAVs, fiber optic sensors, and acoustic emission testing, Figure 2, (Bassier et al., 2022; Catbas and Malekzadeh, 2016; Wedel and Marx, 2022). These techniques provide continuous, high-resolution data that enhance monitoring capabilities by detecting strain distribution, temperature effects, active cracks, and internal defects that might go unnoticed with traditional methods (Gagliardi et al., 2021). For example, UAVs allow for the inspection of hard-to-reach areas without the need for extensive scaffolding or traffic disruptions, while fiber optic sensors enable real-time monitoring of critical structural health indicators (Haupt et al., 2019). For example, UAVs equipped with visual sensors collect images, videos, and GPS data, Figure 3 (Perry et al., 2020). The data is processed using Structure-from-Motion (SfM) techniques to create a 3D as-built bridge information model (AB-BrIM). ML algorithms are then used for bridge element segmentation and defect detection, identifying damages such as cracks and spalling. Defect information, including location and severity, is mapped to specific bridge elements in the model. The system offers a 3D visualization of damage, suitable for bridge rating according to AASHTO and FHWA standards (Perry et al., 2020). Sample of advanced bridge inspection tools. Process of using UAVs in bridge inspection.
Conventional and advanced methods for bridge inspection and data inquiry (FHWA, 2022).
Condition assessment and load rating
FHWA bridge condition ratings based on 2022 NBIS guidelines.
Load rating analysis is another conventional evaluation method, assessing the bridge’s load-carrying capacity and structural integrity against permeant and transient traffic loads (Sun et al., 2022). The American Association of Transportation and Highway Officials (AASHTO) Manual for Bridge Evaluation (MBE) uses the following equation to determine the load rating factor (RF) for in-service bridges and culverts (AASHTO, 2017):
Machine learning techniques in bridge inspection
ML has emerged as a transformative tool in numerous engineering fields, including SHM and bridge maintenance (Civera et al., 2022; Favarelli and Giorgetti, 2021). Traditional methods of monitoring and maintenance such as visual inspections, NDT, and underwater inspections, involve periodic inspections and reactive measures, often resulting in high costs and potential safety risks due to unforeseen failures (Sun et al., 2020). By leveraging ML, engineers can predict and identify potential issues before they become critical, thereby optimizing maintenance schedules and improving overall bridge health management (Gomez-Cabrera and Escamilla-Ambrosio, 2022; Sun et al., 2020). ML algorithms can be categorized into supervised, unsupervised, and RL, each serving distinct purposes in the context of bridge monitoring and maintenance (Gomez-Cabrera and Escamilla-Ambrosio, 2022). In addition to the classic paradigms of supervised, unsupervised, and RL, current bridge inspection applications leverage semi-supervised and weakly supervised learning methods to mitigate the challenge of limited labeled data.
Applications categorized based on ML types
Supervised learning
Supervised learning algorithms such as SVM, ANNs, and RF have been adopted in SHM systems for their ability to learn from labeled data and make accurate predictions (Sonbul and Rashid, 2023). SVMs, powerful classification tools that find the optimal hyperplane separating different classes in the feature space, are effective in high-dimensional spaces and are used to classify structural states, distinguishing between healthy and damaged conditions (Noori Hoshyar et al., 2023). Random forests, which are ensembles of DT, improve predictive performance by averaging multiple trees’ outputs, making them robust against overfitting (Samatas et al., 2021). In bridge maintenance, random forests help in predicting degradation patterns and identifying factors contributing to structural wear (Samatas et al., 2021). DT works by splitting data into subsets based on feature values, providing interpretable results useful for decision-making in maintenance planning and identifying critical risk factors in bridge structures (Almarahlleh et al., 2024). Lastly, K-NN, which classifies data points based on the majority class of their nearest neighbors, is a straightforward algorithm used in SHM to compare new sensor readings with historical data, facilitating the identification of abnormal conditions and damage (Assaad and El-adaway, 2020).
Unsupervised learning
Unsupervised learning algorithms such as c-means, K-means, and GMMs play a crucial role in SHM by analyzing unlabeled data to detect patterns and anomalies (Gagliardi et al., 2021; Nick et al., 2015). Fuzzy c-means clustering allows data points to belong to multiple clusters with varying degrees of membership, providing flexibility in handling uncertainty in data and clustering structural conditions that may not have clear boundaries (Silva et al., 2016). K-means clustering partitions data into k-clusters by minimizing variance within each cluster and is employed in SHM to group similar structural states and identify common patterns in sensor data (Nick et al., 2015). GMMs, which assume that data is generated from a mixture of several Gaussian distributions, are effective for modeling complex data distributions in SHM and detecting subtle changes in structural behavior (Silva et al., 2016). Association analysis discovers interesting relationships and correlations between variables in large datasets, different types of sensor readings, and structural conditions. Finally, techniques like ICA, used in blind source separation, separate multivariate signals into additive, independent components, which is useful in SHM for isolating different sources of structural vibrations or identifying specific damage signatures from mixed sensor data (Malekloo et al., 2022).
Reinforcement learning
RL is another promising area in the application of ML for bridge maintenance and SHM (Zhou et al., 2022). RL involves training an agent to make a sequence of decisions by rewarding desirable outcomes and penalizing undesirable ones. This approach is useful for developing adaptive maintenance strategies where the agent learns optimal actions through trial and error (Zhou et al., 2022). For example, an RL model can be used to determine the best times for maintenance activities by balancing the cost of maintenance with the risk of bridge failure. Over time, the RL agent can improve its strategy by learning from the outcomes of past actions, leading to more efficient and cost-effective maintenance plans. The ability of RL to adapt to changing conditions and learn from continuous feedback makes it a powerful tool for enhancing the sustainability and safety of bridge infrastructure (Fan et al., 2021).
Recent studies have demonstrated the increasing value of RL in optimizing bridge inspection and maintenance strategies. For instance (Lei et al., 2025), developed an RL-based decision-making framework that dynamically adjusts inspection frequencies based on structural condition transitions and resource availability. Similarly (Lai et al., 2024), applied deep reinforcement learning (DRL) to evaluate bridge maintenance actions under budget constraints, adding improvements in lifecycle cost savings and reliability. Moreover (Lei et al., 2023), proposed a multi-agent RL approach for real-time coordination of inspection resources across a network of aging bridges, demonstrating the ability of RL not only to optimize performance but also to reduce carbon footprint and economic costs in long-term infrastructure management. These studies confirm the growing potential of RL as a state-of-the-art technique for predictive and adaptive infrastructure management.
Deep learning
DL has emerged as a transformative approach in bridge inspection and SHM due to its ability to automatically extract hierarchical features from complex, high-dimensional data. Among the most widely used architectures, CNNs have shown excellent performance in detecting surface-level defects such as cracks, spalls, and corrosion in image-based datasets. For instance, CNNs have been successfully trained to identify concrete damage with high precision using UAV-captured imagery and thermal images (Babu et al., 2023). In applications involving time-series sensor data (e.g., acceleration, strain), RNNs and their variants such as LSTM networks have been adopted to capture temporal dependencies and degradation trends (Li et al., 2023). With innovations such as transfer learning from architectures like GoogLeNet, and attention modules like SE blocks and hourglass-shaped depthwise separable convolutions, deep learning models are now capable of detecting fine-scale defects effectively and efficiently (Hajializadeh, 2023; Song et al., 2024). These techniques demonstrate potential for automating bridge inspection processes when paired with drone-based data collection.
Semi-supervised and weakly supervised learning
Semi-supervised learning uses a small set of labeled data with a larger set of unlabeled data. For example (Zhang et al., 2023b), created a semi-supervised Mask R-CNN architecture that employed inspector input to increase segmentation accuracy on UAV-captured bridge data while eliminating the requirement for manual annotation. In addition, weakly supervised learning provides alternative, efficient paradigms by using unrefined or imprecise labels, such as image-level tags or region-level annotations, rather than entirely pixel-level masks (Zhu and Song, 2020) developed a weakly supervised network for crack detection on asphalt concrete bridge decks, employing an autoencoder-based feature extractor, k-means clustering, and CNNs. Their model achieved a segmentation accuracy of 98% across various crack types under challenging lighting and environmental conditions.
Recent developments and literature reviews
Recent literature on ML applications in bridge monitoring and maintenance reveals a growing focus on leveraging advanced algorithms for SHM. Jia and Li (2023), conducted a comprehensive review of deep learning techniques in SHM, covering the period from 2017 to 2023. This study highlights the latest data, algorithms, applications, challenges, and emerging trends in the field. Similarly, Niyirora et al. (2022), provided a systematic review of intelligent damage diagnosis in bridges using vibration-based monitoring and ML, spanning from 2011 to 2022. Fan et al. (2021), focused on the application of ML in the design and inspection of reinforced concrete bridges, identifying resilient methods and new applications from 2011 to 2022.
Xu et al. (2023a), reviewed the SHM of concrete and steel bridges, offering insights into ML application choices over the extensive period from 2000 to 2023. Gomez-Cabrera and Escamilla-Ambrosio (2022), examined various ML techniques applied to SHM systems for both building and bridge structures, covering research from 2011 to 2022. Additionally, Sonbul and Rashid (2023), provided a systematic literature review on algorithms and techniques for SHM of bridges, analyzing studies from 2016 to 2023. These reviews underscore the advancements and ongoing research efforts in employing ML for enhancing the monitoring and maintenance of bridge infrastructures. They highlight the diverse algorithms and methodologies being explored, from traditional supervised learning techniques to more advanced deep learning approaches, all aimed at improving the accuracy, efficiency, and reliability of structural health assessments.
Previous reviews which focused on either data acquisition techniques or ML algorithmic development in isolation. However, this study provides an integrated overview that connects advanced data collection tools (e.g., UAVs, LiDAR, sensors) with ML models customized to different bridge components and data types (1D, 2D, 3D). It classifies ML applications based on bridge elements (e.g., decks, piers, cables, mechanical systems) and data techniques, highlighting practical use-cases for superstructure and substructure inspections. Furthermore, the study covers developing ML paradigms such as semi-supervised, weakly supervised, and deep learning, which have been underrepresented in previous reviews. By doing so, it provides a decision-focused framework for selecting appropriate ML tools based on dataset features, inspection targets, and component types, something that earlier evaluations lacked. These perspectives offer actionable guidance to both researchers and transportation agencies aiming to modernize their inspection and maintenance workflows using AI.
Summary of tools and techniques
Commonly used ML models for bridge assessment.
Previous literature reviews on ML applications in bridge inspection.
Systematic literature review methodology
Research questions
The review was guided by the following research questions: (1) How to use ML tools in bridge inspection/maintenance? (2) When to use the ML methods in bridge inspection/maintenance? (3) How to select an appropriate ML method (selection criteria)? (4) What recommendations and trends can be extracted for future research in this area?
To address the research questions, a systematic search of the literature was performed using several academic databases, including IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar. The search terms included combinations of keywords such as “ML”, “predictive maintenance”, “bridges”, “structural health monitoring”, “substructure”, “superstructure”, “scour risk”, and “deformation monitoring”. The distribution of published papers on the research topic across various journals is shown in Figure 4, with Sensors and Automation in Construction being the leading sources of publications, followed by SHM and Transportation Research Procedia, indicating the widespread interest in diverse fields such as sensor technology, automation, and infrastructure monitoring. List of journals housing research on ML applications for bridge inspection.
Literature search and inclusion criteria
The inclusion criteria for this review were peer-reviewed journal articles and conference papers published between 2014 and 2024, written in English, and focusing on the application of ML techniques to the maintenance and monitoring of bridge infrastructures. The exclusion criteria were studies that do not explicitly focus on bridges, articles not available in full text, and non-peer-reviewed sources such as editorials, opinion pieces, and dissertations. The initial search identified many studies. Screening was conducted based on the title, abstract, and keywords to filter out irrelevant studies. This step ensured that only studies relevant to the research questions were retained. The remaining studies were then assessed for document type (journal papers) and publication year (2014–2024). Figure 5 illustrates the increase in the number of papers published on the topic of research over the years. There was a rise in the number of publications in 2021, 2022, and 2023, indicating growing interest and advancements in the field. The trend shows that the topic has gained substantial traction in recent years, with 2022 experiencing the highest number of published papers, followed closely by 2023. This trend highlights the relevance and the increasing focus of researchers on your area of study. Temporal distribution of surveyed articles related to ML applications in bridge inspection (n = 60).
Data extraction and synthesis process
For each selected study, the following data were extracted: title, authors, and year of publication; objectives and research questions; ML algorithms used; bridge components targeted (e.g., substructure, superstructure); key findings and conclusions; and identified research gaps. The checklist included criteria such as the clarity of research objectives, appropriateness of the methodology, robustness of the data analysis, and relevance of the findings. The extracted data were synthesized to provide a comprehensive overview of the state of research on ML applications in bridge maintenance. The synthesis involved categorizing studies based on the bridge components they addressed (e.g., substructure, superstructure); summarizing the ML techniques and models used in each study; identifying common themes, patterns, and trends in the research; highlighting findings and contributions of each study; and discussing the identified research gaps and suggesting areas for future research.
The findings of the literature review were reported in a structured format, aligned with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. The report includes an introduction outlining the background and importance of the review; a detailed description of the methodology used for the review; a results section presenting the synthesized findings; a discussion section highlighting the implications of the findings and identifying research gaps; and a conclusion summarizing the key insights and recommendations for future research. The PRISMA chart in Figure 6 illustrates the flow of studies through the review process. Flowchart showing procedure followed in designing the literature review.
Applications of ML tools by bridge elements and data type
The categorization of the studies was based on two primary criteria: the bridge elements monitored, and the data types and collection methods. The first category, bridge elements, includes both superstructure and substructure components. As shown in Figure 7, Superstructure elements encompass the parts of the bridge above the abutments, such as girders, beams, cables, and the bridge deck itself, which are critical for supporting traffic loads and ensuring the overall stability of the bridge. Substructure elements include the foundation, piers, and abutments, which support the superstructure and transfer loads to the ground. The second category pertains to the data types and the methods used to collect this data. Studies utilized a variety of data types, such as structural vibrations, strain measurements, and displacement data, collected through various methods including sensor networks, visual inspections, and remote sensing technologies. ML enhances the process by automating data analysis, identifying patterns, and detecting anomalies that may not be immediately visible (Favarelli and Giorgetti, 2021). It enables predictive maintenance by analyzing data like structural vibrations and strain measurements to forecast potential failures, making it an essential tool for transforming raw data into actionable insights (Wedel and Marx, 2022). Schematics showing key bridge anatomy.
The classification in this section is not intended to imply that ML tools are tied to one structural domain or material type, but it is used to place the data inputs, computational approaches, and decision-making frameworks that have proven effective in practice. For example, vibration-based ML models used for scour detection in foundations differ substantially from vision-based models applied to deck crack segmentation, even if both use CNNs. Similarly, the type of bridge (e.g., cable-stayed, arch, beam) and accessibility of components (e.g., bearing systems vs pier walls) shape both the data collection strategy and the ML algorithm selected. The tables and subsections that follow reflect this integrated understanding, highlighting how bridge type, material, damage visibility, and inspection goal collectively influence the choice of ML tools and data sources.
Studies are also classified based on bridge type to better understand how structural material and system complexity influence ML tool selection. Figure 8 illustrates the distribution of reviewed studies by bridge material type. This distribution reflects a research focus toward conventional bridge types (e.g., concrete and steel), which often provide more accessible datasets and standardized inspection protocols. However, studies on more complex or less common bridge types—such as cable-supported and movable bridges—highlight emerging needs for tailored ML models that accommodate unique inspection challenges, such as joint movement, bearing wear, and unpredictable element behavior. Bridge types have been incorporated as part of our classification, aiming to provide more context-sensitive insights into how ML methods are matched to structural behavior, data availability, and inspection requirements. Distribution of reviewed studies by bridge material type.
Bridge elements used in ML studies
Superstructure elements in bridges
Articles on ML applications for maintenance of bridge superstructure elements.
Beams and girders are critical components of the bridge superstructure, providing the primary means of support and load distribution (Tran-Ngoc et al., 2019). Beams typically run horizontally and are designed to carry vertical loads, while girders are the main horizontal supports that often carry the loads transferred from the beams (Noori Hoshyar et al., 2023). Accurate prediction of the effects of temperature and other environmental factors on these elements is essential for their maintenance. Wang et al. (2022), developed advanced methods to predict the cross-sectional effective temperature and vertical temperature difference in flat steel box girders of long-span bridges using a combination of on-site environmental data and ML techniques. This study demonstrated that the multifactor RF model outperforms traditional single-factor models, improving predictive performance by 60%–90% (Wang et al., 2022). Additionally, studies have utilized ML for damage detection in beams, such as the work by Darsono and Torbol (2017), who applied deep learning techniques for damage detection in bridge structures, showcasing the potential for high accuracy in identifying structural weaknesses.
The bridge deck is the surface layer that carries traffic and distributes loads to the underlying structural elements (Kong et al., 2022). Deterioration of bridge decks can affect the overall health of the bridge (Almarahlleh et al., 2024). ML approaches have been utilized to enhance the prediction of bridge deck deterioration (Althaqafi and Chou, 2022). For instance Rashidi Nasab and Elzarka (2023), developed a framework for optimizing ML algorithms to predict bridge deck conditions in Ohio. Their study demonstrated that ensemble ML algorithms, such as random forest and XGBoost, predicted deck conditions more accurately than single models when using “optimal” features. Assaad and El-adaway 2020, introduced a computational data-driven asset management system that evaluates and predicts bridge deck conditions using ANNs and KNNs, achieving a prediction accuracy of 91.44%. Similarly, Almarahlleh et al. (2024), utilized hyperparameter optimization to develop accurate ML models for classifying bridge deck conditions and predicting deterioration trajectories.
In cable-stayed and suspension bridges, cables are key structural elements that bear loads and provide stability (Zhou et al., 2022). The fundamental frequency of these cables, influenced by environmental factors such as temperature and wind, is crucial for maintenance. Predicting temperature variations accurately helps in planning maintenance activities and ensuring the structural integrity of the bridge (Zhou et al., 202); Li et al., 2023; Li et al., 2023), developed a real-time intelligent prediction model using LSTM networks to model the non-linear relationships between temperature changes and the fundamental frequency of bridge cables. This model outperformed traditional linear regression and backpropagation neural network models, providing more precise predictions with reduced error rates (Li et al., 2023).
Movable bridges, such as drawbridges, have mechanical components like gearboxes, motors, and rack and pinion systems that require constant monitoring to ensure proper operation. Catbas and Malekzadeh (2016), introduced a ML-based algorithm for processing massive data collected from the mechanical components of movable bridges such as temperatures, vibration data from the accelerometers, and acoustic data from microphones. This approach uses robust regression analysis for long-term operational monitoring and has shown satisfactory performance in detecting common damage scenarios such as leakage and insufficient oil in the gearbox, as well as bolt removal from the rack and pinion system. This study highlights the potential of ML techniques in automating the monitoring and maintenance of mechanical components in movable bridges (Catbas and Malekzadeh, 2016).
Additionally, Darsono and Torbol (2017), found that key components of bridges, such as the deck, superstructure, substructure, and mechanical elements in movable bridges, are vulnerable to damage such as cracks, corrosion, deformation, and mechanical wear. Data for damage detection was collected using vibration sensors, strain gauges, and displacement sensors, with sensor networks transmitting real-time data. This data was processed using techniques like modal analysis, finite element modeling, and autoregressive models. ML, particularly deep neural networks trained with historical data, was used to predict and classify damage. The ML process involved data preprocessing, feature extraction, and model training through backpropagation, allowing for early damage identification and more effective maintenance strategies.
Similarly, Chalouhi et al. (2017), found that the primary type of damage evaluated was related to structural anomalies in the railway bridge, such as changes in stiffness and integrity of the bridge components. These anomalies could be caused by issues such as material degradation, fatigue cracks, or other structural failures that affect the overall performance of the bridge. The study focused on detecting deviations in the bridge’s response to train crossings, where abnormalities in acceleration measurements indicated potential damage or degradation in the bridge’s structure. This study confirmed the potential of ML techniques in real bridge health monitoring by effectively detecting structural anomalies. The data was collected using accelerometers and thermocouples installed on the bridge. Pre-processing of the acceleration data involves extracting features related to the crossing train, such as speed and number of axles. ANNs and Gaussian processes were then trained on data from the bridge’s reference condition to predict accelerations under various operational and environmental conditions. The model was validated using data from the San Michele Bridge in Italy, where the algorithm successfully detected structural anomalies by identifying discrepancies between predicted and measured responses (Chalouhi et al., 2017).
Substructure elements in bridges
Recent studies have leveraged ML techniques to improve the monitoring and maintenance of bridge substructure elements, addressing the limitations of traditional methods in terms of efficiency, accuracy, and cost. Traditional methods for predicting scour depth at bridge piers primarily relied on empirical or semi-empirical formulas, such as the HEC-18 formula commonly used in the United States (Chou and Nguyen, 2022; Dong et al., 2020). These methods were based on physical models that incorporate hydraulic parameters like flow depth, velocity, and sediment characteristics (Dong et al., 2020). However, these approaches often struggle to account for the complex interaction between various environmental factors, leading to less accurate predictions of scour depth in real-world scenarios (Dong et al., 2020). To enhance prediction accuracy, modern studies have integrated ML techniques with advanced sensors for data acquisition (Dong et al., 2020). Common sensors include GPR, accelerometers, and GPS, which are used to monitor structural deformation and displacements (Chou and Nguyen, 2022). Additionally, remote sensing technologies, such as SAR, have been employed to track surface deformations over time (Gagliardi et al., 2021). These sensors, combined with ML models like CNNs and DBNs, allow for more accurate and dynamic predictions of scour depth (Dong et al., 2020). In particular, studies utilizing PSI for satellite-based assessments have demonstrated the potential for large-scale, efficient monitoring of infrastructure (Gagliardi et al., 2021).
Data collection for scour prediction and bridge monitoring primarily involved both laboratory-controlled experiments and field data. For instance, Chou and Nguyen (2022), utilized a laboratory dataset containing 151 instances of scour depth measurements from various experimental setups, focusing on different pier geometries and hydraulic conditions. This data was instrumental in developing and validating ML models like the LSSVR, RBFNN, and FBI algorithms. Additionally, field data was collected from 79 bridge sites across 17 U.S. states, providing real-world environmental conditions, such as varying water velocities, sediment types, and pier structures (Chou and Nguyen, 2022). The combination of field and laboratory data allows for comprehensive analysis and prediction. In another study, PSI was used to collect satellite-based data over a 3-year observation period, which helped monitor deformation patterns on bridge structures (Gagliardi et al., 2021). Advanced sensors such as GPR, accelerometers, and GPS were employed to gather real-time data on pier displacement and scour depth (Gagliardi et al., 2021).
Articles on ML applications for maintenance of bridge piers.
Data types used in ML studies
Summary of Tools and Data Types Extracted from Previous Studies on bridge health monitoring.
Databases such as NBI and the FHWA compile extensive bridge attributes, including age, structure length, roadway width, deck width, design load, and ADT (Kong et al., 2022; Zhang et al., 2023b). The AssetWise Ohio Department of Transportation database integrates typical bridge attributes with information on deck width, deck geometry, and core element conditions of the superstructure and substructure (Ohio DOT, 2021; Rashidi Nasab and Elzarka, 2023). Additionally, inspection reports from the French National Railway Company (SNCF) offer detailed structural condition insights, including flow type, riverbed slope, flood flow, pier shape, foundation type, and scour history (Wang et al., 2023). In addition to that, many U.S. states have developed their own systems to manage and assess the structural integrity of bridges. For example, Alabama DOT uses advanced techniques such as drones, ultrasonic testing, and ground-penetrating radar to inspect its bridges and has allocated federal funds to repair structurally deficient structures (Alabama DOT, 2021). Alaska DOT, facing unique geographic challenges, relies on the AASHTOWare Bridge Management system, which compiles extensive data from bridge inspections and helps prioritize rehabilitation projects across its vast, remote areas (Alaska DOT, 2023). California’s bridge inspections, overseen by Caltrans, incorporate both visual assessments and advanced technologies such as drones and AI-based imagery analysis to enhance the speed and accuracy of inspections (California DOT, 2017).
While many ML applications in bridge inspection focus on 1D (sensor signals) and 2D (image-based) data, recent advancements in 3D data processing and ML have expanded bridge inspection. For example, (Xia et al., 2022), demonstrated effective semantic segmentation of LiDAR-derived bridge point clouds, enabling automated identification of structural components, and applied a semantic segmentation algorithm that extracts local geometric features (e.g., curvature, density) to differentiate between elements such as decks, girders, piers, and railings. Similarly, (Yang et al., 2022), combined PointNet with GNN to classify bridge elements (such as deck, pier, abutment, bearing, etc.) from a dataset consisting of full-scale 3D scans of concrete bridges captured with terrestrial laser scanning with over 99% accuracy, showcasing scalable digital twin capabilities. Moreover (Bolourian et al., 2023), developed a surface-normal-enhanced PointNet++ model to detect and segment 3D defects—cracks and spalls—with high reliability (≥92% recall). These studies collectively illustrate the promise of 3D deep learning architectures (e.g., PointNet, 3D-CNN, GNN) for comprehensive bridge modeling, structural segmentation, and defect detection, thereby enabling more holistic and geometry-aware inspection strategies.
Critical insights from literature review
Data collection
The findings from the review of various tools and data types used in bridge monitoring and inspection (shown in Table 8) highlight a comprehensive approach to SHM. Different sensors, such as ammeters, thermometers, and strain gauges, are employed to gather electrical data, temperature, humidity, displacement, and vibration data, which are crucial for identifying improper performance, temperature effects, and structural strain. Solar radiation sensors and ultrasonic anemometers further contribute by measuring environmental factors like solar radiation, wind speed, and direction. Accelerometers and microphones capture essential vibration and acoustic data to assess the integrity of bridge structures. In addition to these sensors, advanced methods such as UAVs and LiDAR are used to collect high-resolution images and 3D point-cloud data, facilitating detailed visual inspections and damage mapping. Radar systems, specifically synthetic aperture radar, measure structural displacements and subsidence, providing insights into large-scale changes. Moreover, field surveys, annual inspections, and sophisticated systems like AssetWise and the Bridge Inspection Application System (BIAS) are integral for gathering and managing critical data on bridge age, design load, traffic, and structural condition. Finite Element Analysis is also widely utilized to simulate structural responses, offering detailed insights into potential displacement and deformation under various conditions. These tools and data types form a robust framework for detecting damage, monitoring environmental influences, and ensuring the longevity of bridges.
Conventional methods are more widely used, papers focusing on traditional approaches such as strain gauges, accelerometers, annual inspections, and digital cameras for 2D images. These techniques have been established for decades and continue to play a critical role in bridge inspection due to their reliability and widespread adoption. In contrast, advanced methods, including the use of UAVs, LiDAR, Radar, and FEA, are represented in only seven papers. This indicates a gradual shift towards integrating modern technologies, such as 3D modeling and UAVs, into bridge inspections. While conventional methods remain dominant, the growing application of advanced techniques reflects the industry’s evolving approach to more efficient and accurate damage detection and structural monitoring.
Bridge components
The analysis of component types in the context of ML for predictive maintenance of bridges reveals a pronounced focus on superstructure components (Shown in Figure 9). As illustrated in the pie chart, most of the studies emphasize the superstructure elements, indicating their critical importance in bridge maintenance research. In contrast, the substructure components receive less attention, with an even smaller fraction of studies addressing both superstructure and substructure elements. This highlights the research gap which indicates the relative underrepresentation of substructure components in predictive maintenance studies. There is a pressing need to expand the scope of ML applications to include substructure maintenance. This gap identifies a critical area for future research, a balanced focus between superstructure and substructure elements for comprehensive bridge maintenance. Distribution of focus on bridge components, for research related to ML applications in bridge inspection.
Furthermore, a portion of the studies conducted on superstructure components are dedicated to bridge decks, beams & girders. However, mechanical components, such as gearboxes, motors, and rack and pinion systems, along with bridge bearings and bridge cables, are less frequently addressed. This disparity indicates several research gaps that need to be addressed to develop a comprehensive predictive maintenance strategy. Furthermore, bridge bearings, which play a crucial role in accommodating movements and forces within the bridge structure, are also underrepresented in the research. The lack of focus on these elements may lead to incomplete maintenance strategies, potentially overlooking critical stress points that can affect the bridge’s overall performance and longevity.
ML tools
Tables 6 and 7 show that there is an increasing use and success of specific ML algorithms in bridge maintenance, particularly XGBoost and ANNs. XGBoost has consistently demonstrated high accuracy in both substructure and superstructure applications, making it a prominent choice for predictive maintenance tasks. Its proven performance across various components, such as bridge piers and decks, highlights its effectiveness in handling large datasets while delivering accurate predictions (Kong et al., 2022; Li et al., 2024; Wang et al., 2023). Additionally, a key trend is the rise of hybrid models, which combine multiple ML techniques to boost performance. For example, ANNs integrated with evolutionary algorithms like PSO, and CNNs combined with GANs, have shown remarkable improvements in prediction accuracy (Munawar et al., 2022). These hybrid approaches capitalize on the strengths of different algorithms, resulting in more robust models capable of handling complex data (Munawar et al., 2022; Tran-Ngoc et al., 2019).
DT is also prominent, with six occurrences in the superstructure and one in the substructure, followed by Gradient Boosting (7 times), SVMs (8 times), and Random Forest (7 times). These algorithms effectively perform predictive maintenance, damage detection, and structural assessment tasks for components such as decks, beams, girders, and piers. Ensemble methods like Gradient Boosting and Random Forest provide high accuracy and robustness, making them valuable tools for bridge health monitoring.
Another notable observation is the variety of data types used in these models, including large-scale numerical datasets, dynamic signals, and image-based data. This shift toward integrating diverse data sources could enhance the comprehensiveness and accuracy of bridge health assessments. By incorporating various data inputs into advanced ML models, future research could create more precise and effective predictive maintenance systems, leading to better-informed decision-making in bridge health monitoring. The current studies in ML for bridge maintenance have several limitations that need to be addressed to enhance the effectiveness and applicability of these technologies. For instance, many studies primarily rely on specific data types or limited data sources. The reliance on a narrow range of sensor data or environmental factors can lead to incomplete models that do not fully capture the complexities of real-world bridge environments, restricting the generalizability of the findings across different bridge types and conditions.
As shown in Figure 10, to effectively select the appropriate ML tool for inspecting specific bridge components or conditions, several factors should be considered. These include the type of data, the desired performance level, the size of the dataset, and the computational efficiency of the algorithm. Recent studies have demonstrated the utility of advanced deep learning architectures in bridge inspection tasks. For example, ResNet50 has been widely used for image classification tasks such as crack detection, achieving high accuracy up to 97.5% (Santaniello and Russo, 2023). On the other hand, Cycle-GAN is typically applied in image-to-image translation tasks, such as enhancing low-quality inspection images or generating synthetic training data under varying lighting conditions (Munawar et al., 2022). These models serve different functions and are selected based on the specific needs of the inspection pipeline. In addition, sensor-based vibration monitoring, often used for cables and girders, benefits from temporal sequence models like LSTM networks, which effectively capture time-dependent patterns and achieve high accuracy, such as 93.96% in specific cases (Li et al., 2023). The most appropriate algorithms used with different bridge elements.
Selecting an ML tool for bridge inspections depends on many factors, such as the data types and the dataset size. Detailed performance metrics of ML algorithms and selection criteria are provided in the Supplemental Material section for readers who are interested in learning more about this topic.
Current studies emphasize short-term predictions and immediate damage detection, with less focus on long-term maintenance planning and lifecycle cost analysis. This short-term focus can lead to maintenance strategies that are reactive rather than proactive, missing opportunities for more efficient and cost-effective long-term solutions. There is a need for more robust validation and testing of ML models in diverse and challenging environments to ensure their reliability and effectiveness in varied real-world applications. There is a lack of standardized methodologies for integrating diverse data sources and ML techniques, resulting in varied performance and reliability across different studies. The application of ML models often faces challenges related to data quality, sensor accuracy, and environmental noise, which can affect the predictive accuracy of these models. Additionally, while some studies have demonstrated the potential of advanced algorithms like neural networks and RL, their practical implementation in real-world scenarios is often limited by computational complexity and the need for extensive training data.
Conclusions and recommendations
This study conducts a systematic review of the application of ML techniques in the inspection and maintenance of bridges. The analysis encompassed 60 key studies published between 2015 and 2024. The studies were categorized based on bridge elements monitored, data types and collection methods employed, and ML algorithms used for data analyses and prediction.
The study finds that NNs, SVMs, and RF are the most widely applied ML algorithms in bridge SHM, accounting for 27.84%, 17.72%, and 13.92% of applications, respectively. These algorithms have shown high effectiveness in detecting structural damage, predicting maintenance needs, and improving overall bridge safety. The research emphasizes that ML models have enabled a shift from reactive maintenance practices to proactive approaches, allowing for more efficient maintenance planning and reduced costs due to fewer unexpected failures. However, despite the progress, the study concludes that more work is needed to address challenges such as enhancing data quality, improving model generalization, and ensuring that diverse bridge conditions are represented in the models.
Practical implications
The use of sensors and ML algorithms enables continuous monitoring, allowing for timely interventions that can prevent major structural issues, thus ensuring the longevity and reliability of bridge systems. The study offers several recommendations for action items in industry and transportation agencies. First, there should be a concerted effort to standardize data collection and integration methodologies to ensure the reliability of ML models across different bridge systems. This includes developing industry-wide protocols for sensor installation, data collection, and data sharing. Second, the industry should invest in advanced sensor technologies and data analytics tools to facilitate continuous monitoring and real-time data processing. Implementing state-of-the-art sensors and robust data infrastructure will enhance the precision and utility of the collected data.
Third, training and capacity-building programs should be established to equip engineers and maintenance personnel with the necessary skills to implement and manage ML-based SHM systems effectively. These programs should cover the fundamentals of ML, data analysis, and the practical aspects of integrating these technologies into existing maintenance workflows. Fourth, ML results should be connected with decision-making systems, such as Bridge Management Systems (BMS), to enable automated inspection and repair predictions. Finally, model performance must be continuously monitored and evaluated, with expert comments used to adjust the models over time. For example, ML models trained in historical inspection, inventory, and environmental data exceed traditional methods in predicting future bridge conditions or deterioration rates. Predicting time needed to failure, condition rating, or risk scores can be integrated into BMS optimization tools (e.g., PONTIS, AASHTOWare BrM) to assist with life-cycle cost analysis and intervention timing. Agencies can start by exporting the existing BMS data, training ML models externally, and then bringing the results back into the system to help with scheduling and managing the expenses. Another integration can be performed over time by using software to connect live BMS data to real-time ML services, allowing for automated updates and dynamic decision-making. This solution enables agencies to move from reactive maintenance to proactive, data-driven asset management, while ensuring ML insights are accessible, interpretable, and embedded within existing workflows.
An example of a project done by (Kulhandjian, 2023), it integrates drones equipped with both infrared and high-resolution optical cameras, along with an onboard minicomputer running ML algorithms to perform fully autonomous inspections of roads and potentially bridges. The system demonstrated an accuracy of 84.6% using optical images and 95.1% using thermal images in detecting roadway defects and incorporates real-time navigation and defect localization without human intervention. Although the prototype was primarily tested on roads due to airspace restrictions, the modular framework has clear potential for bridge inspection automation. This approach represents a major step toward scalable, cost-effective infrastructure monitoring using AI-powered aerial systems (Turkan and Xu, 2019) developed a comprehensive inspection framework combining UAVs, computer vision, and Bridge Information Modeling (BrIM). In their case study on an Oregon bridge, UAV-captured imagery was processed to detect surface-level defects like cracks, which were then automatically mapped to corresponding 3D BrIM elements, enabling centralized, cloud-based access to up-to-date inspection data for all stakeholders. This integration not only improved inspection accuracy and efficiency but also allowed inspection outputs to be directly linked with lifecycle decision modules in the BMS. Together, these examples illustrate how transportation agencies can harness ML-driven technologies to automate data collection, improve condition assessments, and enhance data-informed maintenance planning within modern BMS frameworks.
Recommendations and future research
For future research, the study recommends expanding the focus to include both superstructure and substructure elements to develop a more comprehensive maintenance strategy. Current research has focused on superstructure components, leaving a critical gap in the understanding and maintenance of substructure elements like foundations, piers, and scour protection measures. Additionally, exploring innovative data fusion techniques that combine different types of sensor data, such as vibrations, strain, temperature, and environmental factors, will provide a more holistic understanding of bridge health. Further research should also address the challenges related to data quality, sensor accuracy, and environmental noise to enhance the generalizability and reliability of ML models.
Also, hybrid modeling approaches, which integrate ML algorithms with physics-based simulations, can improve forecast interpretability and reliability, particularly when the data is limited or noisy. In addition, multi-sensor data fusion methods, such as combining visual imagery, LiDAR, acoustic emission, vibration, and thermal monitoring, can provide a more complete picture of structural behavior. This multi-source integration has the potential to overcome the limitations of single-source datasets, allowing for more accurate and robust prediction models.
Finally, emerging sensor technologies such as fiber optic sensors, UAV-based LiDAR, multispectral imaging, and digital twins show a promise for improving data collection during bridge inspections. These techniques can increase the quality, resolution, and frequency of data input into ML models, resulting in more accurate and timely assessments. However, real-world implementation presents obstacles such as data diversity, environmental instability, scalability, and interoperability among sensor systems. It is necessary to explore the generalization capability of ML models across various bridge types, geographic locations, and environmental conditions. Most current models are created and validated on isolated datasets, limiting their adaptability. Future research should emphasize the generation of standard datasets that cover a variety of structural types and deterioration changes. Advanced methods, including transfer learning, domain adaptation, and meta-learning, can also be used to increase the scalability and resilience of ML applications in real-world bridge management systems.
Supplemental Material
Supplemental Material - State-of-the-art review of machine learning applications for bridge inspections
Supplemental Material for State-of-the-art review of machine learning applications for bridge inspections by Amr Ashmawi, Phuong Nguyen, and Akram Jawdhari in Advances in Structural Engineering
Footnotes
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Supplemental Material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
