Railway Track Structural Health Monitoring: Identifying Emerging Trends and Research Agendas Using Bibliometric and Topic Modeling

Abstract

While railways are critical for transportation, their expansive networks spanning thousands of kilometers pose significant challenges for conventional structural health inspection and maintenance. Recent advancements in sensors and artificial intelligence technologies have led to a substantial growth in the body of research proposing innovative approaches for Railway Track Structural Health Monitoring (RTSHM) to enhance safety and operational efficiency. This work aims to synthesize the current RTSHM research landscape to explore mainstream and emerging directions and identify advancements, challenges, and opportunities in this field. Through the hybrid systematic review using bibliometrics analysis and topic modeling, core research themes emerged, from developing sensor and data acquisition techniques as the foundation, to be combined with AI-based methods for fault detection and prediction. These predictions are leveraged for predictive maintenance through degradation modeling, supplemented with information from dynamic response assessment and performance optimization, and the ultimate goal is integration of RTSHM for operational safety assessments and risk-based decision-making. While technologically advanced, current research predominantly focuses on detecting discrete defects, thereby neglecting the holistic management of the track system. This fragmentation contributes to a complex and often siloed landscape for infrastructure management, emphasizing that RTSHM remains in a critical developmental stage. Consequently, the development of smart railway, integrated with intelligent data collection devices, deep learning technologies, and innovative operational platforms, represents a challenging yet promising direction for future research. These advancements are anticipated to foster safer, more efficient, and sustainable railway systems worldwide.

1. Introduction

The railway industry plays a vital role in supporting economic growth, social development, and easing the burden on road infrastructure [1,2,3]. However, the increasing volume of rail traffic combined with the rising frequency of extreme weather events that can lead to rail buckling, component loss, or crossing failures, is hindering uninterrupted operations [3,4,5]. These risks underscore the need for frequent inspection and maintenance to prevent structural damage and accidents, ensuring safe and continuous service. However, the vast scale of railway networks, which span thousands of kilometers, makes such efforts difficult to sustain through conventional means. Inspection and maintenance in the railway sector still rely on traditional, periodic visual assessments, where personnel physically examine track conditions. These methods are inherently labor-intensive, time-consuming, costly, and prone to human error, particularly under harsh environmental conditions [3,6]. Moreover, the slow processing speeds of existing diagnostic equipment hinder the adoption of continuous, real-time data collection strategies [3,7,8]. As a result, conventional approaches often produce sparse data, fall short of modern automation demands, and introduce persistent safety risks [3,6,9].

In response to these challenges, adopting advanced Structural Health Monitoring (SHM) technologies has emerged as a crucial and forward-looking solution, marking a paradigm shift toward proactive infrastructure management [10]. Recent developments in predictive maintenance further address the limitations of conventional methods by leveraging sophisticated sensing technologies and advanced measurement devices [11]. These sophisticated systems incorporate Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) to automate data analysis, predict failures, and enhance system reliability [12,13]. Concurrently, wireless sensor networks have been deployed to enable automated monitoring and efficient data management [3]. Integrating these technologies with AI delivers superior performance and significant cost reductions compared to previous solutions. This integration facilitates real-time decision-making through defect classification, anomaly prediction, and continuous monitoring of train operations [14,15]. These advancements have solidified SHM’s potential, garnering substantial interest from the railway research community. This burgeoning interest is evident in the rapidly expanding body of literature and the increasing number of dedicated research initiatives focused on harnessing SHM for enhanced railway safety and operational efficiency.

Understanding the current research landscape in Railway Track Structural Health Monitoring (RTSHM) is paramount. While the rapidly expanding body of research on RTSHM necessitates systematic synthesis to identify key directions and challenges, existing review studies often provide a fragmented view, primarily focusing on specific issues rather than suggesting comprehensive approaches. For instance, research by Agustin, et al. [16], Wang, et al. [17], and Bosso, et al. [18] predominantly concentrates on surface defect detection, largely overlooking predictive methodologies and long-term risk mitigation. Similarly, there are studies exploring surface defect evaluation [3,19,20,21,22], deploying advanced monitoring tools such as sensors and uncrewed aerial vehicles [5,23,24,25,26,27,28], or applying broader assessment models [29,30], and these efforts frequently lack integration across the entire railway track.

While considerable advancements have been made in the field of RTSHM, prior research has predominantly focused on the development of technologies for detecting individual defects [31,32]. Consequently, these efforts have resulted mainly in discrete solutions, overlooking the pressing need for a comprehensive solution. This deficiency curtails the formulation of holistic management strategies, impeding the practical application of advanced technologies for failure prediction. This is particularly critical as the body of knowledge concerning holistic management issues continues to expand, yet it has not been systematically synthesized. Therefore, this study’s primary motivation is to address this gap by systematically evaluating and consolidating the existing corpus of literature about RTSHM. A bibliometric review of current practices, capabilities, and limitations within the existing literature is essential for a deeper understanding of the research landscape. It is a foundational step in bridging the theory and practical application gap. This will support researchers and decision-makers in selecting the most effective intervention measures.

This study aims to consolidate the fragmented knowledge within the RTSHM field. By meticulously analyzing methodologies and dominant research topics, it maps the existing intellectual landscape to identify key challenges, underexplored areas, and strategic directions for future investigation, thereby equipping researchers and decision-makers with a comprehensive overview. To achieve these objectives, the paper is organized as follows. Section 2 outlines the comprehensive methodology for paper selection, subsequent bibliometric analysis, and topic modeling. Building on this foundation, Section 3 presents the bibliometric findings, encompassing publication trends, key contributions, and the intricate relationships within the existing body of literature. Section 4 then delves into a detailed content analysis of the topics derived from this study. Subsequently, Section 5 discusses the identified research gaps and proposes directions for future investigations. Finally, Section 6 provides concluding remarks.

2. Methodology

This study utilizes a hybrid review approach, including PRISMA protocol, bibliometric analysis, and topic modeling to provide a standardized structure for identifying, selecting, and evaluating relevant studies, enhancing the reproducibility of the outcomes. In addition, the PRODLDA algorithm for topic modeling processes text data, extracts latent topic clusters with high accuracy, and uncovers insights that are difficult to achieve with manual methods or basic analysis tools. Therefore, this integrated methodology streamlines data preprocessing, minimizes reviewer bias, and delivers quantifiable, reproducible results. The research process is illustrated in Figure 1. The workflow process involves defining search criteria and systematically filtering literature via the PRISMA protocol to finalize a corpus study. This corpus then undergoes a dual analysis: VOSviewer is utilized for overall research landscape observation and keyword co-occurrence clustering, while the PRODLDA algorithm performs in-depth topic modeling. The topic mapping, constructed from the comparative analysis and integration of findings from both methodologies, subsequently undergoes in-depth content analysis to identify research gaps and guide future research agendas.

2.1. Data Collection Following the PRISMA Protocol

A search query on the Web of Science, IEEE, and Scopus database used Boolean operators, combining (“Structural health monitoring” OR “SHM” OR “Maintenance” OR “Monitoring”) with (“Railway” OR “Rail” OR “Railway infrastructure” OR “Tracks” OR “Railway track”) via the AND operator to focus on railway-related studies. This search strategy guaranteed the inclusion of all research relevant to RTSHM, concurrently excluding unrelated studies. Eligible studies were English-language, peer-reviewed journal or conference publications from 2004 to 2025. Non-empirical works (e.g., book reviews and editorials) were excluded to prioritize new empirical data. The search, conducted in October 2025 via the database, initially retrieved 2503 records. After excluding 1185 non-journal/conference papers, 1318 remained. Further screening removed 96 non-English or out-of-range records, leaving 1222. Title, abstract, and keyword evaluation excluded 591 irrelevant records, resulting in 631. The full-text review excluded 26 inaccessible records and 131 that lacked direct RTSHM relevance. Ultimately, 474 studies were included for bibliometric and topic modeling analyses, forming an RTSHM dataset.

2.2. Bibliometric Analysis Using the VOSviewer Software

Following the selection of 474 articles in Section 2.1, a bibliometric analysis was conducted to evaluate the dataset. The included records were fed into bibliometric analysis to identify four aspects: (1) publication trend, (2) leading institutions and key contributors, (3) keyword co-occurrence networks, to map interconnections among core concepts in the RTSHM field, and (4) topic mapping. These steps were performed using VOSviewer 1.6.20, with a minimum occurrence threshold of five for each keyword or entity, ensuring statistical significance and representativeness. In the topic mapping aspect, thematic clusters derived from the PRODLDA algorithm will integrate with the keyword co-occurrence network from VOSviewer to create a comprehensive overview.

2.3. Topic Modeling Using the PRODLDA Model

PRODLDA, introduced by Srivastava and Sutton [33], is a variant of Latent Dirichlet Allocation that performs word-level topic mixtures in the natural parameter space, avoiding the multinomial simplex constraint of traditional LDA. This approach enhances topic coherence, making PRODLDA a superior alternative for topic extraction. The implementation of PRODLDA in this study followed a structured four-step process, as depicted in Figure 1. The process began with text pre-processing the 474 selected studies to generate a clean corpus by removing stop words and standardizing data. Following this, hyperparameter tuning was conducted, using a Bayesian optimization approach implemented in the Optuna module. The key parameters are summarized as follows: hidden layer size = 128–512, dropout rate = 0.1–0.2, batch size = 32, number of training steps = 5000, and learning rate = 0.001–0.01 (log-uniform). The subsequent step involved training the PRODLDA model (Supplementary Materials) on the processed dataset to produce an optimized model capable of extracting latent topics.

In the final step, topic assignment was performed, using coherence and silhouette scores to evaluate topic consistency and separation. Based on Normalized Pointwise Mutual Information, the coherence score measures the semantic relatedness of words within a topic, as expressed in Formula (1) [34]. The Silhouette Coefficient for a sample is $(b-a)/max(a,b)$ [35]. To clarify, b is the distance between a sample and the nearest cluster the sample is not a part of. A Silhouette score close to +1 indicates that a data point is well clustered, near 0 suggests potential overlap with another cluster, and near −1 implies misclustering. In this model, the number of topics was determined empirically by evaluating coherence scores and topic interpretability across a range of k values from 2 to 20. The chosen configuration achieved the highest coherence while maintaining semantic distinctness between topics. The remaining hyperparameters were tuned automatically within the specified ranges to balance convergence speed and generalization.

These metrics ensure the selection of an optimal topic number that balances coherence and distinctiveness. The results of the topic assignment will be analyzed to extract the latent topics deeply. Topics from PRODLDA were mapped into meaningful clusters, each interpreted based on dominant keywords and research context, providing detailed insights into key research trends. These results were cross-referenced with the keyword co-occurrence network from the prior bibliometric analysis to identify relationships between PRODLDA topics and established keyword clusters. This process culminated in an integrated mapping diagram, reflecting the interplay and complementarity between topics and core RTSHM concepts.

$$v_{ij}=NPMI{\left({\omega }_i,{\omega }_j\right)}^{\gamma }={\left({\displaystyle \frac{\log \frac{P\left({\omega }_i,{\omega }_j\right)+\epsilon }{P\left({\omega }_i\right)·P\left({\omega }_j\right)}}{-\log \left(P\left({\omega }_i,{\omega }_j\right)+\epsilon \right)}}\right)}^{\gamma },$$

(1)

where $P\left({\omega }_i,{\omega }_j\right)$ is the joint probability of the co-occurrence of the two words ${\omega }_i$ and ${\omega }_j$; $P\left({\omega }_i\right)$ and $P\left({\omega }_j\right)$ are the individual probabilities of occurrence, and $\epsilon$is added to avoid the logarithm of zero.

3. Bibliometric Analysis

3.1. Publication Trend

3.1.1. Annual Publications, Citations, and Leading Journals

Figure 2 reveals a remarkable evolution in RTSHM research activity from 2004 to October 2025, with an annual publication trend of R² = 0.7718, and correlation for citation accumulation of R² = 0.9821. In its early phase (2004–2012), the field was relatively dormant (Figure 2a), producing only 1–4 papers per year with fewer than 100 citations annually—an outcome of technological constraints and limited scholarly engagement. A turning point came in 2013, after which publications surged steadily, climbing from 11 papers in 2013 to 90 in 2024, signaling the rapid establishment of RTSHM as a vital research domain. Although citation counts peaked in 2017 and have since tapered, this trend is mainly attributable to the natural lag in recognizing newer work. Early 2025 data already record 29 publications and 11 citations within ten months, suggesting that growth will continue. Journal contributions underscore this momentum (Figure 2b): the top ten outlets collectively account for 167 papers and 3374 citations. Among them, Applied Sciences-Basel (30 papers, 296 citations) and Sensors (28 papers, 336 citations) emerge as the leading publication venues, offering accessible platforms for applied studies, particularly in sensing and monitoring. By contrast, Measurement (14 papers, 581 citations) and Mechanical Systems and Signal Processing (14 papers, 531 citations) dominate in citation impact, reflecting their central role in advancing theoretical frameworks and methodological innovation within the field.

3.1.2. Countries’ Collaboration Networks

Figure 3 presents the number of contributions by country and collaborative network among the top 10 contributing countries. Figure 3a clearly shows that China is the leading country in terms of the number of research contributions with 172, far ahead of the UK (54 contributions) and the US (35 contributions). To analyze the collaborative ties more closely, Figure 3b illustrates the international collaboration among top 10 contributors. Interestingly, China and the USA emerge as hubs within this focused network. China’s prominence is highlighted by its extensive collaborations with the USA and the UK. This finding may be explained by its massive high-speed rail network investments. In contrast, despite a lower publication count, the USA maintains an influential position through its diverse international partnerships. The figure also reveals that European regional countries, such as the UK, the Netherlands, Sweden, Portugal, Spain, and Italy, form a tightly knit regional collaboration network, likely facilitated by shared technical standards and joint railway projects within the European Union. Furthermore, Australia, characterized by long railway lines and challenging geographical conditions, frequently collaborates with China to leverage its expertise in monitoring complex railway systems. These intricate collaborative networks underscore the importance of knowledge and technology exchange in addressing the growing global demand for safe and efficient railway infrastructure.

3.2. Leading Institutions and Key Contributors

Figure 4 and Figure 5 reveal distinct patterns in institutional and individual contributions to RTSHM research. Among the top ten institutions, all based in Asia and Europe, Asian universities dominate in publication volume, led by Beijing Jiaotong University (38 articles) and Southwest Jiaotong University (36 articles), reflecting the scale of regional investment in railway research. In contrast, European institutions show greater citation impact. Delft University of Technology (20 articles, 773 citations) and the University of Birmingham (21 articles, 523 citations) stand out, highlighting a clear divide between Asia’s strength in research output and Europe’s influence in scholarly impact. At the researcher level, three trajectories emerge: early pioneers such as Dollevoet R and Núñez A, with steady contributions from 2010 to 2018; emerging scholars like Ni YQ and Aela P, who have been active since 2018; and long-term leaders exemplified by Wang P, whose productive output peaked in 2021–2022. Collaboration networks further emphasize this structure, with Wang P at the center of a central cluster linked to Núñez A and Li ZL, while the sustained partnership of Aela P and Jing GQ anchors a second cluster. These institutional and individual patterns underscore how RTSHM research is shaped by regional strengths, scholarly influence, and collaborative hubs that continue to drive the field forward.

3.3. Keyword Co-Occurrence Networks

Figure 6 presents the RTSHM keyword co-occurrence network, comprising 112 keywords, with the following parameters: 112 keywords were selected (minimum occurrence = 5), grouped into five clusters with a total link strength of 2253. The first cluster is characterized by keywords such as “sensor,” “damage,” “diagnosis,” and “monitoring.” This revealed a foundational research direction aimed at leveraging modern data acquisition devices to detect and diagnose structural damage, enabling continuous and effective monitoring. The second cluster highlights the rapid emergence of ML and computer vision. Key terms like “computer vision,” “deep learning,” “digital image correlation,” “classification,” and “defect detection” are central to this theme. This demonstrates a significant trend towards automating image data analysis for high-precision defect identification. The third cluster addresses the practical aspects of railway management and upkeep. Its focus on “track maintenance,” “fault diagnosis,” and “rail inspection” underscores the research community’s efforts to enhance maintenance strategies through continuous condition assessment and fault progression forecasting, thereby improving system safety. The fourth cluster is prominent, with keywords such as “behavior,” “simulation,” “track stiffness,” and “vehicle-track interaction,” revealing a strong focus on using numerical simulations to assess the interplay between train speed, track properties, and vehicle–track interaction. Finally, the fifth cluster is represented by terms like “deterioration,” “degradation,” and “stochastic model.” It emphasizes the application of probabilistic models to evaluate structural decay, thereby supporting the development of adaptive strategies to maintain track quality and reliability.

3.4. Topic Mapping

Figure 7 presents the Coherence and Silhouette score calculations for the number of topics from the result of the PRODLDA model. The model achieves a peak Coherence score of 0.493 with six topics (Figure 7a), complemented by an average Silhouette score of 0.610 at this level (Figure 7b). The analysis of keywords derived from the PRODLDA topic models six themes shaping the research landscape of RTSHM (see Appendix A,

Table A1. Keywords for each topic from the PRODLDA topic modeling.

Theme	Topic Modeling
Sensor technology in defect detection	(0.045*“sensor” + 0.045*“detection” + 0.035*“fiber” + 0.025*“measurement” + 0.046*“defect” + 0.029*“surface” + 0.027*“signal” + 0.022*“condition “ + 0.031*“damage” + 0.017*“bragg” + 0.024*“magnetic” + 0.022*“test” + 0.026*“fastener” + 0.017*“grate” + 0.016*“acceleration “ + 0.015*“measure” + 0.025*“crack” + 0.017*“vibration” + 0.026*“image”)
Data processing techniques in condition detection	(0.022*“measurement” + 0.032*“geometry” + 0.037*“condition” + 0.028*“detection” + 0.016*“measure” + 0.022*“inspection” + 0.033*“vehicle” + 0.029*“defect” + 0.029*“ learn “ + 0.015*“sensor” + 0.024*“corrugation” + 0.029*“fault” + 0.016*“algorithm” + 0.016*“irregularity” + 0.019*“wear” + 0.014*“quality” + 0.025*“machine” + 0.017*“degradation” + 0.031*“network”)
Application of an AI model in RTSHM	(0.013*“inspection” + 0.019*“detection” + 0.014*“defect” + 0.018*“condition” + 0.044*“ wheel” + 0.025*“vehicle” + 0.029*“fault” + 0.027*“image” + 0.036*“measurement” + 0.015*“geometry” + 0.028*“sensor” + 0.015*“surface” + 0.022*“learn” + 0.021*“process” + 0.017*“time” + 0.014*“acceleration” + 0.020*“approach” + 0.020*“deep” + 0.019*“artificial”)
Predictive maintenance strategies	(0.018*“wheel” + 0.024*“condition” + 0.017*“vehicle” + 0.028*“damage” + 0.033*“geometry” + 0.018*“ballast” + 0.027*“wear” + 0.025*“contact” + 0.024*“force” + 0.024*“sensor” + 0.023*“prediction” + 0.028*“defect” + 0.014*“network” + 0.031*“measurement” + 0.013*“analysis” + 0.018*“study” + 0.021*“impact” + 0.017*“approach” + 0.020*“load”)
Track performance and operational efficiency	(0.051*“ballast” + 0.040*“sleeper” + 0.030*“slab” + 0.020*“sensor” + 0.021*“speed” + 0.017*“high” + 0.022*“load” + 0.025*“condition” + 0.018*“measurement” + 0.023*“fiber” + 0.021*“study” + 0.014*“inspection” + 0.019*“defect” + 0.016*“stiffness” + 0.021*“test” + 0.016*“structural” + 0.020*“increase” + 0.020*“temperature” + 0.020*“stress”)
Safety management and risk mitigation	(0.058*“detection” + 0.034*“defect” + 0.037*“network” + 0.030*“learn” + 0.037*“image” + 0.035*“obstacle” + 0.026*“feature” + 0.017*“fastener” + 0.029*“inspection” + 0.026*“surface” + 0.025*“algorithm” + 0.020*“performance” + 0.021*“improve” + 0.024*“accuracy” + 0.019*“sensor” + 0.014*“information” + 0.020*“deep” + 0.017*“high” + 0.019*“time”)

Note: The asterisk (*) denotes the weight (or probability) of the specific keyword within the topic.

). Theme 1 focuses on advanced sensor technologies, which effectively support rail defect detection. Theme 2 highlights data processing techniques that facilitate precise fault prediction. Theme 3 delves into AI applications in RTSHM, prominently featuring learn, deep, neural, convolutional, and algorithmic models. Theme 4 examines maintenance strategies and rail damage forecasting, reinforced by prediction, degradation, fault, inspection, and condition, which align with its predictive and upkeep-oriented scope. Theme 5 targets enhancing rail performance and operational efficiency, emphasizing its commitment to optimizing track dynamics and operational resilience. Lastly, Theme 6 prioritizes comprehensive safety management and risk-mitigated decision-making. These keyword distributions, extracted from PRODLDA, provide a compelling foundation for the thematic nomenclature, ensuring each theme accurately mirrors the evolving priorities within RTSHM research.

A comparative analysis was performed against the keyword co-occurrence network generated by VOSviewer (as detailed in Section 3.3) to ensure the validity of these algorithmically derived themes. This validation process confirmed an alignment between the PRODLDA themes and the VOSviewer keyword clusters, substantiating their thematic consistency. However, a critical insight from this analysis was the profound interrelationship between Theme 1 and Theme 2 of PRODLDA topic modeling. These two themes represent sequential and mutually dependent phases of the research process, a synergy that VOSviewer captured by consolidating these concepts into a single, unified cluster. Therefore, to enhance conceptual clarity and align with the empirically observed structure, Themes 1 and 2 were consolidated into a single, cohesive theme designated “Sensing and data acquisition techniques for RTSHM”.

The refined topic mapping is presented in Figure 8, which maps the consolidated PRODLDA results onto the VOSviewer co-occurrence network. This visualization illustrates the multidimensional and intricate structure of the RTSHM field, where each distinct circle defines a core research theme. At the same time, the overlapping regions highlight the interdisciplinary nature and complex interconnections among sub-domains. The congruence between the two analytical approaches underscores the comprehensive management of RTSHM, from initial data collection and analysis, through proactive degradation modeling and performance optimization, to operational safety assessment. Based on this validated framework, five core themes are identified for subsequent in-depth analysis: (1) sensing and data acquisition techniques for RTSHM, (2) AI approaches for fault detection and prognosis, (3) predictive maintenance and degradation modeling for railway track infrastructure, (4) dynamic response analysis and performance optimization of railway tracks, and (5) integrating RTSHM for operational safety assessment and risk-based decision-making.

4. Content Analysis

4.1. Sensing and Data Acquisition Techniques for RTSHM

Detecting various types of rail damage, such as cracks, geometric irregularities, material degradation, and subsurface defects, helps reflect vital information on track stability, alignment, and load-bearing capacity. Of interest here is that capturing such information can be conducted through various sensing techniques such as Inertial Sensors (IS), Optical Sensors (OS), Non-Destructive Testing Sensors (NDTS), and Fusion Sensors (FS). Firstly, IS, such as prominent accelerometers and track geometry cars, are used to measure vibrations, acceleration, and geometric parameters to monitor real-time track conditions and detect component defects. Secondly, OS such as cameras, lasers, and LiDAR, collect 3D point clouds, images, and geometric profiles for high-precision measurement of track geometry and component integrity. Third, NDTS, like eddy current and Hall effect sensors, analyzes magnetic fields and strain to detect surface cracks and rail stress. Finally, FS integrates multiple technologies, such as lasers, IMUs, and Fiber Bragg Grating (FBG) sensors, to comprehensively assess track geometry, deformation, and specific defects through advanced data processing. These sensor groups collectively enable robust data acquisition for effective RTSHM, addressing diverse structural health challenges.

IS encompasses a broad range of devices designed to measure motion and forces. Foundational to this group are accelerometers and gyroscopes, which are instrumental in quantifying critical parameters like acceleration, vibration, and velocity for real-time track condition monitoring [36,37,38,39,40,41,42,43]. Exemplifying this, studies conducted by Obrien, et al. [36] and Bocciolone, et al. [37] effectively utilized these sensors on in-service trains to provide frequent updates on track conditions. To further enhance this efficacy, research by Avsievich, et al. [8] incorporated high-precision digital accelerometers, significantly improving railway safety. Another approach, in the context of increasingly extended track network lengths, is to use track geometry cars that incorporate multiple inertial measurements to capture precise geometrical parameters [28,44,45,46,47,48,49,50,51,52]. Notably, studies by Sresakoolchai and Kaewunruen [44] and Sánchez, et al. [46] applied this method and showed faster and more cost-effective results than traditional inspection techniques. Beyond geometry, IS plays a pivotal role in material integrity assessment. For instance, magnetic sensors leverage flux leakage to detect surface defects [53,54,55,56,57], while Pires, et al. [58] employed accelerometers to identify surface imperfections. Concurrently, more specialized IS technologies have emerged; piezoelectric sensors are crucial for crack detection using guided ultrasound, as demonstrated by Yang, et al. [59], while differential triple coil sensors, designed by Li, et al. [60], target rail bottom defects. However, despite this versatility, IS faces significant limitations that undermine its reliability. The main concerns are its vulnerability to measurement inaccuracies and environmental noise, as mentioned by Obrien, et al. [36], and Sresakoolchai and Kaewunruen [44], along with performance degradation due to temperature fluctuations and significant implementation costs, as noted by Li, et al. [60]. Furthermore, according to Yang, et al. [59] and Diaz, et al. [61], emerging IS technologies still lack extensive validation in diverse real-world railway conditions.

In contrast to motion-based sensing, OS operate by capturing and responding to electromagnetic radiation. Within RTSHM, techniques like structured light [62,63,64,65,66], and FBG sensors [67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88] are extensively utilized due to their capability to monitor defects and enable real-time, multi-parameter assessment by converting light wavelength shifts into data on strain, temperature, deformation, and vibration, thus providing a comprehensive insight into dynamic track behavior. FBG sensors, for example, convert light wavelength shifts into strain, temperature, and deformation data, providing comprehensive insight into dynamic track behavior. Their enhanced accuracy over conventional methods was demonstrated by Yüksel, et al. [77], while their capacity for non-disruptive, in situ monitoring was showcased by Laudat, et al. [87]. Extending the application of light-based sensing, a structured light sensor has been employed to monitor fastener defects [63], and to assess rail wear [66]. Furthermore, OS excels in high-precision geometric measurement. Camera systems provide images for 3D geometry and gauge assessment [1,89,90,91,92,93,94,95], while 3D point cloud data from a laser is critical for ensuring geometric accuracy [96,97,98,99,100,101,102]. Meanwhile, the application of Lidar to extract and reconstruct a 3D model of the railway track has also been carried out by Fedorko, et al. [103]. Another key method involves ultrasonic sensors for detecting internal rail defects, as explored by Bahati, et al. [104]. Nevertheless, the efficacy of OS is not without its caveats. Same as IS, OS is also significantly affected by environmental conditions like ambient light, dust, and moisture, which can cause signal loss and data errors [62,63,67,68,69,89,90]. Moreover, Cui, et al. [62] and Han, et al. [63] further emphasized operational complexity and low sampling rates in some systems, which may make them unsuitable for dynamic testing, as further reinforced by Skrickij, et al. [89] and Mohammadzadeh, et al. [90].

Shifting the focus from optical methods to subsurface and material evaluation, NDTS plays a role in identifying hidden flaws without inducing damage. Methods such as using alternating current field meters of Chacón Muñoz, et al. [105], differential eddy current sensors of Chandran, et al. [106], Shao, et al. [107], and Hall effect sensors of Ji, et al. [108], Heindler, et al. [109], and Gong, et al. [110] have proven effective in detecting surface and near-surface cracks. Beyond surface integrity, NDTS contributes significantly to assessing acoustic performance by measuring rail surface roughness to predict rolling noise, as studied by Jeong and Jeong [111] using displacement sensors. Furthermore, this group of sensors also demonstrated instrumental worth in monitoring internal stress, using resistance strain gauges, deployed by Liu, et al. [112], Zhu and Lanza di Scalea [113], and Lee, et al. [114], which provide direct insights into stress distribution. Extending its diagnostic reach, another type of sensor, Ground Penetrating Radar (GPR), which assesses the track substructure to pinpoint subsurface anomalies, was also analyzed by Goodarzi, et al. [115]. In addition, several studies have shown that NDTS accuracy is highly sensitive to deployment conditions, including sensor lift-off effects and platform instability pointed out by Chacón Muñoz, et al. [105], Chandran, et al. [106], Shao, et al. [107], and Jeong and Jeong [111]. Environmental factors also affect performance, and Liu, et al. [112] pointed out this limitation when the stress gauge is affected by uneven temperature distribution. In addition, operational limitations such as infrequent GPR data collection can affect the reliability of the prediction, which was also mentioned by Goodarzi, et al. [115] in their study.

SF represents a paradigm shift, integrating data from diverse sensors to overcome individual limitations and provide comprehensive insights into infrastructure health. A principal application is the holistic monitoring of track geometry. For this, systems combine laser scanners, cameras, IMUs, and GNSS to collect precise 3D data on parameters like gauge, superelevation, and alignment, enabling real-time monitoring of irregularities [116,117,118,119,120,121,122,123]. Secondly, SF is used for targeted defect detection by combining different physical principles. For instance, Pathak, et al. [124] integrating laser-activated ultrasonics for rail foot flaws, Velha, et al. [125] combining FBG and Raman sensing for deformation, Kudo, et al. [126] and Higgins and Liu [30] studying LiDAR-camera fusion to locate damage, and Sun, et al. [127] with PZT actuators with FBG sensors to detect enhanced damage. More than that, FS enables holistic operational monitoring and tracks condition assessment. Track recording vehicles, for instance, are deployed for both general geometric evaluation (line, level, gauge, and twist) in the studies of Counter, et al. [128] and Popov, et al. [129], and specific integrity checks on structures like slab tracks have also been considered [130,131,132,133]. Concurrently, Mahtani, et al. [134] combined LiDAR, IMU, and GNSS systems to help automate the identification of tracks, catenaries, and obstacles. At a more granular level, specialized fused sensors are used to analyze environmental impacts on displacement and to continuously monitor the condition of critical components like railway switches, as found in the results of Lu and He [135] and Del Álamo, et al. [136]. However, despite these advancements, FS faces inherent challenges that hinder optimal performance. As noted by Akpinar and Gulal [122] and Khosravi, et al. [123], complex calibration and integration demand extensive signal processing and error modeling, often resulting in unstable trajectories. Environmental sensitivities, including temperature and weather variations, as reported by Velha, et al. [125] and Lu and He [135], alongside GNSS signal loss in tunnels, restricts operational efficacy, particularly at higher speeds, as highlighted by Akpinar and Gulal [122]. Practically, impractical installations, as pointed out by Velha, et al. [125] and component deterioration as identified by Sadeghi and Askarinejad [130], poses logistical hurdles, while performance variability due to temperature shifts and limited training data, as observed by Del Álamo, et al. [136], compromises data reliability, necessitating reliance on visual inspections as suggested by Akpinar and Gulal [122].

4.2. AI Approaches for Fault Detection and Prognosis

AI has emerged as a transformative force in academia and industry, yet its application within the railway sector remains nascent compared to other domains. As Tang, et al. [137] summarized, AI can be pivotal in optimizing complex systems, detecting infrastructure defects, enhancing safety, and improving service quality. To address the multifaceted challenges of RTSHM, researchers have drawn from a diverse spectrum of AI paradigms. Prominent among these are DL models, particularly Convolutional Neural Networks (CNN) with architectures such as Fully Convolutional Networks (FCN), Deep Convolutional Networks (DCN), and You Only Look Once (YOLO), and Artificial Neural Networks (ANN). These have been instrumental in tasks requiring feature extraction and classification. Alongside these, unsupervised Clustering models, including algorithms like K-means, Fuzzy, and the Gaussian Mixture Model, have demonstrated significant efficacy in pattern recognition, often replacing traditional manual inspection methods. Furthermore, other models such as Margin-Based, Tree-Based, Statistical, Recurrent Neural Networks (RNN), and Optimization-Based approaches are carving out their niches. While their application might be less frequent than that of CNN and ANN, their unique algorithms offer novel pathways, either as standalone solutions or in hybrid configurations, to enhance the performance of RTSHM systems. Figure 9 illustrates a comprehensive overview of these AI algorithms and their respective applications within the RTSHM domain.

Among the numerous applications of AI in RTSHM, a primary focus has been monitoring the track structure’s physical integrity. This domain is broadly bifurcated into two critical applications: (i) rail surface defect detection, encompassing anomalies such as plugs, switches, joints, cracks, squats, corrugation, head-checks, and shelling, and (ii) track geometry monitoring, which includes parameters like alignment, longitudinal level, cross-level, gauge, and irregularities. In rail surface defect detection, AI methodologies can be categorized by output precision. A prevalent approach is region-level detection, which employs frameworks like YOLOv5 and Faster R-CNN to place a bounding box around a detected anomaly. According to Minguell and Pandit [138], bounding boxes are manually annotated around anomalies (e.g., Defective Clip) in training images for automated learning. The model then localizes these objects, assigning class labels and confidence scores to quantify detection certainty. This makes it highly effective for identifying defects at track joints and components. While this method is effective for localization, a key finding is that its primary weakness is that precise boundary and shape information of defects is not accessible [139]. To address this limitation and achieve higher precision, pixel-level detection using semantic segmentation networks offers a superior alternative. An example is provided by Ye, et al. [139], who demonstrated a method that classifies each pixel into a predefined class (e.g., defective or regular). Their study utilized a U-Net architecture to generate these masks from 3D laser data, achieving a mean Intersection over Union of 87.9%. The significance of this high-precision mask is that it enables a complete 3D reconstruction of the defect, from which critical parameters like width, length, and depth can be measured, information that is not accessible from conventional 2D detection results and is essential for informing maintenance decisions. Hybrid models often enhance this approach. Du, et al. [140] combined Deep Convolutional Neural Networks (DCNN) with Support Vector Machines (SVM), while Niebling, et al. [141] paired DCNN with lightweight models like SqueezeNet to improve efficiency. In the domain of track geometry monitoring, the focus has shifted toward achieving high accuracy and predictive capabilities. A notable breakthrough was reported by Shen, et al. [142], whose Multi-scale Feature Dynamic Fusion Network (MFDF-Net) achieved an impressive accuracy of 99.333%. For predictive maintenance, sophisticated models such as the Attention Mechanism–Convolutional Neural Network–Gated Recurrent Unit model introduced by Hao, et al. [143] have shown excellent performance in estimating vertical irregularities. This trend is supported by other studies using ANN for prediction [144,145]. Further advancing the field, researchers are integrating AI with other analytical methods, such as combining ANN with Finite Element (FE) analysis for stress prediction, Brown, et al. [146], or using SVM with vibration data to diagnose irregularities [147,148].

In addition to monitoring the track structure’s physical integrity, AI models also detect faults in rail joints, connections, and auxiliary equipment (e.g., bolts, clips, and fishplates), alongside structural errors like crosstie deflection, track subsidence, and rail movements. For example, Gallo et al. Gallo, et al. [149] validated using a YOLO-Oriented Bounding Box (YOLO-OBB) model to measure the rotational displacement of rail clips directly. This is a significant advance, as standard YOLO architectures cannot analyze angled bounding boxes. The key contribution of this work was its metrological characterization, which confirmed the system’s measurement uncertainty at 0.42°, well within the 3° error margin specified by the manufacturer. Another standout innovation is the “BoltVision” study by Alif, et al. [150], which targeted the classification of missing bolts for deployment on resource-constrained edge devices. Their comparative analysis revealed that a pre-trained Vision Transformer (ViT) achieved a remarkable 93% accuracy, substantially outperforming traditional CNN. This superiority stems from the ViT’s self-attention mechanism, which enables it to capture global contextual relationships across an entire image, a task where conventional CNNs are less effective. Other methods, such as the YOLOv3 model for fastener defect detection, complement these advanced approaches [151], and the combination of Multi-Layer Perceptrons with wavelet transforms [152]. Beyond individual fasteners, the scope of AI applications extends to detecting wider structural and environmental risks. To mitigate scour dangers from flooding, Mammeri, et al. [153] employed a U-Net-based encoder–decoder network for the semantic segmentation of tracks and floodwater from UAV imagery, achieving high mIoU scores of 0.673 for water and 0.982 for rails. For structural deformation, ANN have been used to predict track geometry deterioration [145], while Structure-from-Motion has been applied for accurate 3D track reconstruction and analysis [154].

In detecting obstacles on the track, AI algorithms have become crucial in augmenting safety systems. They plays a crucial role in identifying and monitoring individuals, detecting hazardous items, classifying unusual movements, and recognizing obstructions on railway lines [155]. An algorithm developed by Kapoor, et al. [156] improves detection in thermal images by first using 2D Singular Spectrum Analysis to decompose an image into its constituent components. This pre-processing step effectively isolates the most relevant information about an obstacle, which is then exclusively fed into DL detectors like Faster R-CNN, SSD, and YOLO for more robust recognition across various lighting conditions. The results of this study indicated that the combination of 2D-SSA with Faster R-CNN yielded the highest accuracy at 85.2%. In a comparative study using aerial UAV imagery, Rampriya, et al. [157] evaluated nine different DL models. A notable finding from their work was that the lightweight SSD MobileNet V2 model proved the most effective, achieving a significant accuracy of 96.75%, with Faster R-CNN also demonstrating strong performance. Fantini, et al. [158] designed a dedicated, self-powered system known as the AI Camera Prototype for the specific and acute threat of falling rocks. This system operates by performing change detection between consecutive frames within predefined Regions of Interest for “alert” (the rock wall) and “alarm” (the track). Utilizing a dynamically updated background model transmits an alarm when pixel differences from a stationary object on the track exceed a set threshold, proving its effectiveness even in nighttime operations with LED illumination. To improve the quality of input data for such models, Perić, et al. [159] developed an improved Laplacian-based edge detection algorithm to automatically identify and remove blurred images from training datasets. These diverse approaches have highlighted the capability of AI to enhance railway safety from multiple perspectives.

The analytical results presented are evidence of AI’s great potential in the railway sector, from detecting surface and geometric defects to solving structural and component integrity issues and enhancing obstacle detection, contributing to comprehensive assessments, and promoting a “predict and prevent” model. What is surprising is that these application models are also facing difficulties, such as high-quality data collection, as noted by Ghiasi, et al. [147], high implementation costs, as highlighted by Gallo, et al. [149], and limited generalizability, as pointed out by Aydin, et al. [160]. Another unexpected finding was the extent to which, although AI excels in detection, interpretation requires human expertise, necessitating hybrid systems, as Ye pointed out.

4.3. Predictive Maintenance and Degradation Modeling for Railway Track Infrastructure

A pivotal application of previously discussed data acquisition techniques lies in predictive maintenance and degradation modeling for railway track infrastructure. This predictive capability spans multiple temporal scales, from the instantaneous detection of surface defects and the short-term forecasting of operational failures to the long-term prediction of gradual structural degradation. This section, therefore, reviews the evolution of research across these three critical timescales, analyzing the key methodologies and existing challenges in enhancing railway safety and operational efficiency.

The real-time detection of rail surface defects, such as squats, rail corrugation, and insulated rail joints, is a core focus of RTSHM, aimed at ensuring operational safety and optimizing maintenance costs. The evolution of this field reveals a clear trajectory toward enhanced accuracy and reliability. The most fundamental approach has been using Axle Box Acceleration (ABA) measurements, pioneered by studies such as Molodova, et al. [161], though this method was initially limited by environmental noise. Subsequent refinements, such as the use of longitudinal ABA, have significantly increased detection sensitivity and accuracy [162]. In parallel, advanced signal processing algorithms like Variational Mode Decomposition and Empirical Mode Decomposition have become a dominant trend for isolating defect signatures from complex background noise [163,164]. More recent optimized versions of these algorithms continue to improve separation efficiency and reduce errors [165,166]. In this context, novel sensing technologies and AI have ushered in a new era of monitoring effectiveness. Research has explored alternative physical principles, such as acoustic analysis [167], and particularly Distributed Acoustic Sensing, which achieved a breakthrough accuracy of 97.3% by leveraging existing fiber-optic cables [168]. Concurrently, AI and ML have become indispensable tools for automating identification. Noteworthy examples include unsupervised learning models like the convolutional variational autoencoder-elliptic envelope, which achieved 96.7% accuracy [169], and other integrated methods such as AI-Principal Component Analysis [170] and the Dynamic Differential Evolution algorithm [171], all of which enhance system reliability. In summary, the field has evolved from single-sensor methods (ABA) to enhancement through complex signal processing (Variational Mode Decomposition and Empirical Mode Decomposition), and finally to the emergence of breakthrough sensing technologies and intelligent AI models. This combination promises to advance monitoring efficacy significantly. However, environmental noise and algorithm optimization challenges still necessitate extensive field trials for validation and refinement.

Moving beyond detecting existing surface defects, another critical research frontier in RTSHM involves the short-term prediction of operational and environmental failures, marking a paradigm shift from detection to prediction. In wheel slide forecasting, pioneering studies by Hubbard, et al. [172] utilized on-board monitoring to estimate risk. This approach was later significantly improved by Namoano, et al. [173], who combined wavelet analysis with a Long Short-Term Memory (LSTM) network to achieve an F1-score of 99.94% and the ability to predict incidents seconds in advance, albeit with a dependency on expert-annotated data. In parallel, the prediction of railway track degradation due to failures, such as slurry pumping, has made significant progress with the development of an explainable AI model [174,175]. By integrating multi-source data (GIS, weather, and on-board sensors), this model achieved a high balanced accuracy of 90.84% and identified rainfall as a key risk factor. However, it still faces challenges with high false alarm rates. These studies demonstrate the successful application of real-time sensors and AI models (especially LSTM) for fast and accurate forecasting. However, a standard limitation remains the sensitivity to noise and a lack of contextual data, a problem that multi-source integration methods like that of Zeng, et al. [174] are beginning to address. Therefore, while initial results are auspicious, extensive field testing is imperative to ensure the robustness and reliability of these predictive models under real-world operational conditions.

Extending the predictive horizon from short-term incidents to long-term structural deterioration, a final cornerstone of RTSHM is the long-term forecasting of track degradation, including wear and Rolling Contact Fatigue. Research in this area has progressed along two main, complementary paths: physics-based simulation and data-driven modeling. The physics-based simulation approach has evolved from early empirical studies on friction reduction in the works of Tumanishvili, et al. [176] to the widespread use of FE models to understand mechanisms and predict the degradation of track components under various loading conditions [177,178,179,180]. The culmination of this approach is the integration of Multi-Body Simulation to forecast the service life of the entire system and prioritize maintenance based on fundamental operational factors [18,181]. Concurrently, data-driven approaches have gained prominence as a robust toolset, spanning from statistical models like logistic regression, utilized by Soleimanmeigouni, et al. [182] to pinpoint risk factors, to sophisticated AI frameworks such as BiTCN-BiGRU, leveraged by Liu, et al. [183] for precise degradation trend forecasting. These two approaches are complementary: simulation provides mechanistic insights, while data-driven models offer powerful predictive capabilities. Nevertheless, common challenges persist, including high computational complexity, sensitivity to input data quality, and especially the scarcity of long-term field data for model validation. Consequently, verifying these methods under diverse operational conditions is critical to ensuring their practical applicability.

4.4. Dynamic Response Analysis and Performance Optimization of Railway Tracks

Analyzing dynamic response and optimizing track performance are core tenets of RTSHM, aimed at ensuring safety, enhancing passenger comfort, and reducing life-cycle costs. This research theme encompasses the study of the physical behavior of track structures under dynamic loads, load transfer mechanisms, real-time response monitoring, and performance optimization through advanced methodologies such as simulation, field testing, and AI. These efforts provide innovative solutions to enhance track durability and efficiency under diverse operational conditions. The following analysis, therefore, is structured around four key aspects of these research efforts: (1) the dynamic response of tracks under loading, (2) load transfer mechanisms and structural stability, (3) real-time structural response monitoring, and (4) performance optimization strategies.

Understanding the dynamic response is fundamental to comprehending how track components such as rails, sleepers, and ballast behave under train-induced forces. Research methodologies have evolved from inverse analysis to complex physics-based simulations and hybrid approaches. Initially, inverse analysis methods, typically the Inverse Analysis Method (IAM) by Lee and Chiu [184], were used to estimate vertical wheel impact force magnitudes based on track acceleration response signals obtained in field trials. IAM was proven to be a reliable tool for reconstructing impact force waveforms, overcoming the limitations of the traditional shear-bridge method. Subsequently, FE simulation emerged as the dominant tool for detailed mechanical analysis. The 3D Finite Element Method (FEM) is used to study nonlinear impact forces from wheel onto rail and sleepers [185,186], and the extended finite element method is used to simulate rolling contact fatigue crack growth by adding realistic residual stresses to a frictional contact model [187]. In addition to traditional dynamic models, complex dynamic coupling models have been applied to analyze system performance comprehensively when combined with advanced simulation methods. Rhylane and Ajdour [188] focused on the influence of the overall route geometry, using FEM to model the entire layout (including curves) and determine the critical speed of the system. In contrast, Ren, et al. [189] focused on the vibration damping mechanism, using the vehicle track coupled dynamics model to show how the floating slab track absorbs and redistributes energy from periodic track irregularities. Similarly, Liu, et al. [190] focused on design performance comparison, using a refined joint model to evaluate the safety and comfort indices of three types of ballastless rails on bridges at very high speeds (up to 400 km/h). Complementing these simulation efforts are experimental studies and data-driven methods. For instance, full-scale tests on floating slab tracks have provided empirical data on vibration mitigation and time-dependent stiffness degradation [191], while ML techniques have been applied to predict structural properties, such as the stiffness of non-ballasted slabs, from real-time vibration data, achieving more than 98% accuracy [192].

The load transfer mechanism from the rail through sleepers and ballast to the subgrade is decisive for the long-term stability of the track structure. Consequently, research has focused on improving each component within this system. To enhance the stability of the ballast layer and subgrade, solutions such as reinforcement with polyurethane can reduce settlement by up to 99% [22,193], and the application of Cement Asphalt Mortar to reduce ballast pressure [194] has been proposed. In parallel, optimizing sleeper design has been another critical research avenue. Novel designs like Y steel sleepers by Szabó [195] or Fiber-Reinforced Foamed Urethane sleepers with bottom grooves optimized via Discrete Element Method simulation have demonstrated the ability to increase lateral resistance by 19–80.2% [196,197]. A breakthrough in this area is the development of savvy sleepers integrated with Carbon Nanotubes that exhibit self-sensing capabilities, enabling the effective detection of hanging sleepers [198]. Furthermore, to address abrupt stiffness changes at transitional zones, the integration of Under Sleeper Pads and Under Ballast Mats has proven effective in promoting uniform load transfer [199].

The ongoing monitoring of structural responses is pivotal in validating predictive models and facilitating the early identification of anomalies. Monitoring technologies have advanced to address diverse spatial scales. At the local level, sensors are integrated directly into the track structure, including magnetostrictive sensors employed by Cui, et al. [200] to measure the lateral displacement of Continuously Welded Rail under challenging thermal conditions, and FBG sensors are utilized by Li, et al. [201] for high-resolution assessment of rail joint reaction forces. At the regional scale, Wireless Sensor Networks, implemented by Cañete, et al. [202], enable continuous, cost-effective collection of vibration and displacement data, with ML serving as a critical tool for data analysis, as demonstrated by Yong and Lee [203]. For large-scale network monitoring, remote sensing technologies offer significant advantages. Techniques such as Persistent Scatterer Interferometric Synthetic Aperture Radar, leveraging satellite data from TerraSAR-X, as explored Kim, et al. [204], facilitate deformation monitoring across extensive areas (e.g., 30 km × 50 km), while the integration of GPR and InSAR with ML, as investigated by Koohmishi, et al. [205], enhances the detection of network-wide deformations.

Ultimately, the insights gained from dynamic analysis and real-time monitoring reduce the evaluation process time while maintaining high accuracy of predictions, directed toward tracking performance optimization [206]. This effort focuses on enhancing both maintenance procedures and scheduling. Regarding procedures, studies by Shi, et al. [207] have employed simulation techniques to optimize tamping machine parameters, leading to a 29.83% reduction in contact forces, and Przybyłowicz, et al. [208] have used field experiments to compare different tamping methods to improve ballast compaction. Concerning scheduling, AI by Amiel, et al. [209] has been applied to optimize maintenance plans to reduce operational costs. Studies by Merheb, et al. [210], identifying environmental influences, such as seasonal effects that reduce track modulus, also provide critical input for this planning process. However, common limitations persist across these research areas, including the scarcity of long-term field data, susceptibility to environmental noise, and high computational complexity. Therefore, the multi-conditional validation of these methods and integration with multi-source data are essential next steps for enhancing their reliability and practical applicability.

4.5. Integrating RTSHM for Operational Safety Assessment and Risk-Based Decision-Making

The integration of RTSHM is pivotal in transforming operational safety assessment from reactive to proactive. Its primary function is to continuously surveil critical railway assets, including rails, sleepers, bridges, and tunnels, to identify systemic risks proactively. The evolution of RTSHM for safety assessment demonstrates a clear trajectory toward higher accuracy and broader scope. Pioneering work by Chen, et al. [211] introduced a multi-sensor system for high-speed turnouts, achieving high reliability through integrating FBG, optical imaging, and Lamb waves. However, this system was constrained by high deployment costs and environmental noise. Subsequent research addresses this data integration gap by utilizing FBG in system-wide measurements, enabling the detection of hazards like landslides, although sensor durability remains a challenge [212]. A significant paradigm shift has been demonstrated by Zeng, et al. [213], who leveraged data from in-service trains to predict high-risk events such as track failures up to three months in advance, giving the railway industry enough time to plan maintenance. Furthermore, their study links the predicted state to the enabling modes using methods such as SHAP for causal analysis, identifying why failures are predicted and thus helping to plan maintenance accordingly. These advanced models also incorporate techniques such as TimeGAN to mitigate the severe data imbalances commonly found in real-life failure data, thereby improving prediction accuracy.

In addition to monitoring, data streams from RTSHM play an important role in powering advanced analytics platforms for risk-based decision-making. By feeding real-time data into platforms such as Digital Twins and BIM-GIS, RTSHM enables a significant shift in maintenance paradigms: moving from localized fault detection to integrated risk management. Early efforts, such as using Faster R-CNN [214], or deploying an IoT multi-robot system with a CNN [215], have made significant progress in feeding data into Digital Twins for emergency repair decisions. This approach effectively addresses the signal integrity issues of previous vision-based approaches. Instead of just detecting faults, the transition from predicted track states to actual operating modes has also been investigated. For example, time series models such as LSTM are currently being applied in practice to predictive maintenance before the failure of tracks by estimating remaining useful life of tracks and their components [216]. Despite their superior power, these platforms still face limitations, including the lack of diverse training data, noted by Zhang, et al. [217], and the high computational demands highlighted by Li, et al. [218], as well as the significant challenge of working with highly imbalanced datasets common in real-world failure prediction, which requires further system integration and optimization research.

Overall, the integration of RTSHM has yielded considerable achievements in railway safety management. Technologically, advancements are marked by significant improvements in detection reliability; for instance, systems developed by Yang, et al. [219] have reduced false alarm rates to 2% using laser-based technologies, while innovations introduced by Iyer, et al. [215] with IoT-enabled robotics achieve F1-scores exceeding 90%. Analytically, applying sophisticated AI, ranging from Bayesian ELM proposed by Li, et al. [218] to probabilistic DL explored by Zhang, et al. [217], has enabled precise risk quantification and standardized decision support. This analytical power is now being used to enable maintenance to transition from reactive to predictive scheduling, guided by data-driven forecasts rather than fixed timescales. Furthermore, validation of these AI models against real-world data from track inspection records confirms their practical utility, achieving high recall scores for operationally significant defects. However, several cross-cutting limitations persist. Technological and operational challenges include the high cost of sensors and environmental vulnerabilities [212], alongside unstable data connectivity in remote regions, highlighted by Iyer, et al. [215]. Data-related challenges remain a significant bottleneck, including image noise affecting visual inspection models by Wei, et al. [214], and a lack of diverse datasets for training generalizable models, emphasized by Zhang, et al. [217]. Furthermore, the computational complexity of advanced AI models limits their feasibility for on-board, real-time implementation on embedded devices [218,220].

5. Discussion

5.1. Identified Gaps in RTSHM

Through an integrated approach using the PRISMA protocol, bibliometric analysis, and PRODLDA topic modeling, this study identifies substantial progress across five key themes: (1) Sensing and Data Acquisition Techniques, (2) AI Approaches for Fault Detection and Prognosis, (3) Predictive Maintenance and Degradation Modeling, (4) Dynamic Response Analysis and Performance Optimization, and (5) Integrating RTSHM for Operational Safety Assessment and Risk-Based Decision-Making. Together, these themes outline a comprehensive pathway for managing railway track structural health. Still, notable gaps within and between these areas persist, highlighting critical challenges that hinder the development of fully integrated, holistic monitoring systems. Specifically, a gap exists between predictive maintenance and operational safety. Current studies often focus only on detecting individual defects while overlooking their integration into a holistic safety strategy, leaving them vulnerable to systemic risks [221]. Similarly, dynamic response analyses are often detached from the rich, high-fidelity data streams provided by advanced Sensing and Data Acquisition technologies, such as sophisticated LiDAR-camera fusion systems [222], which undermines the reliability of predictive models. Additionally, the potential collaboration between AI approaches and Operational Safety remains underexploited, especially in forecasting and responding to harsh weather conditions [223]. Even the most pioneering integrated platforms exhibit this inherent weakness by failing to link their outputs with dynamic analyses or formalized risk frameworks [215,224]. This systemic fragmentation underscores the urgent need for a unified platform that seamlessly synthesizes sensor data, analytics, and safety management into a cohesive, intelligent system to forge resilient and sustainable smart railway solutions.

Beyond integration issues, several foundational limitations persist within the topics themselves. A consistent theme in the literature is the data-related challenge, representing a critical bottleneck. This limitation is apparent across most RTSHM applications, where a lack of diverse, well-annotated datasets often constrains the efficacy of advanced AI models. This has been shown to limit the generalizability of safety assessments and fault detection systems [217,225]. Furthermore, the scarcity of long-term field data fundamentally restricts the reliability of predictive maintenance and dynamic response models [214,226]. Another critical area of concern is the environmental and operational robustness of sensing systems. For instance, the performance of vision-based models appears to be affected by environmental factors such as dust and humidity [222], while rapid weather changes can impact laser systems [219]. This suggests that such sensitivity directly affects downstream applications and limits the effectiveness of predictive models [227,228]. These difficulties are compounded by the need for intricate sensor calibration and auxiliary components, which introduce operational complexity and additional points of failure [142,212]. Finally, several practical barriers may impede the transition from research to large-scale deployment. A limitation of many advanced models is their computational demand, which often hinders real-time implementation on embedded devices [229,230]. This is exacerbated by the high implementation cost of advanced sensor hardware, which remains a substantial economic burden [212,231]. These findings suggest that the validation-to-practice gap, coupled with a continued dependency on human expertise for interpreting AI-driven results, presents multifaceted challenges that must be addressed to unlock the full potential of RTSHM.

5.2. Future Research Agenda

To address the knowledge gaps identified, this study puts forth a research agenda as illustrated in Figure 10. The first direction focuses on evolving conventional tracks into intelligent and resilient infrastructure. The developmental pathway commences with material innovation, exploring advanced composites like fiber-reinforced foam urethane to enhance durability and mitigate settlement. A logical next step is to employ DL models to rapidly optimize the performance of these novel materials under diverse loading conditions, thereby reducing the dependency on extensive physical prototyping. Subsequently, long-term field tests are essential to confirm real-world durability and address the current scarcity of performance data. The final intelligence layer would involve embedding self-sensing capabilities directly into track components, for instance, by leveraging emerging technologies like carbon nanotube-infused concrete to monitor internal stress in real-time. While this integrated roadmap is promising, it is essential to note that challenges such as computational complexity and model generalizability will require dedicated investigation, potentially through the integration of interpretable AI.

However, resilient physical infrastructure is only part of the solution. Our findings on data fragmentation and siloed analytics underscore the need for a centralized, intelligent software ecosystem. Consequently, the second direction proposes developing an integrated platform system to unify AI applications and predictive maintenance. This system’s core should be a Digital Twin, creating a dynamic virtual replica of the physical railway network. This would facilitate sophisticated safety simulations and monitoring, thereby reducing the high costs and logistical burdens associated with extensive field trials. To power this virtual replica, a physics-informed AI engine, leveraging models analogous to the YOLOv5-RTO concept [232], could be developed to correlate subtle vibrational data with physical risks. To ensure the platform’s models are generalizable and robust, large-scale field validation over 3–5 years is necessary, as this will provide the diverse training data currently lacking. Finally, integrating multi-modal IoT data streams would provide the rich, holistic data required.

A successful predictive platform fundamentally depends on the quality, richness, and timeliness of the data it receives. Therefore, the third essential direction is the development of intelligent, low-cost, and autonomous data collection devices. This line of research directly tackles the observed limitations of environmental interference, deployment complexity, and low data fidelity. This progression begins with enhancing sensor durability through protective designs, such as heat-resistant coatings on FBG sensors, to ensure reliable performance across wide operational ranges. The next step towards autonomy is incorporating self-calibration features, similar to adaptive LiDAR systems that adjust for ambient conditions, thereby improving data integrity without manual intervention. Furthermore, these devices can locally create multi-source predictive maintenance insights by embedding AI-driven data fusion models directly at the source. The final step towards full autonomy is embedding edge computing capabilities, inspired by the success of distributed acoustic sensing in on-site processing, which enables real-time analysis and reduces the demand for constant connectivity and human expertise.

6. Conclusions

The mounting of research on RTSHM motivated this systematic review to identify key advancements and persistent challenges in this fast-evolving field. Synthesizing 474 studies through a hybrid review approach, this study found that sensing technologies have evolved from discrete sensors to integrated, multi-sensor systems, enabling more comprehensive data acquisition for track condition assessment. Concurrently, AI-driven, particularly DL-based, methods have advanced defect detection across multiple components, enhancing the accuracy and efficiency of predictive maintenance. These developments have expanded predictive modeling to include real-time detection, short-term forecasting, and long-term degradation analysis, supported by improvements in dynamic response modeling and physical optimization for track stability.

Despite these advances, RTSHM research remains fragmented, with predictive maintenance and dynamic analysis often developed in isolation from broader safety and operational frameworks. Persistent challenges include limited access to diverse, annotated datasets, sensor fragility, high implementation costs, and computational demands, which together hinder large-scale deployment. Future research should prioritize the integration of AI, IoT, and advanced materials into unified, real-time monitoring platforms. Developing cost-effective, durable, and multifunctional smart sensors will be critical to bridging the gap between laboratory innovation and field application. Such advancements can support railway operators in optimizing maintenance strategies, improving safety, and enhancing the overall efficiency of modern railway systems.

This study also has limitations. Reliance on the Web of Science, IEEE, and Scopus as the primary data source may have excluded relevant work from other repositories, while omitting industrial materials such as technical reports and national standards limits the ability to compare academic advances with real-world implementation. These constraints may affect the generalizability of recommendations, especially for operators balancing innovation with financial and infrastructural realities.