面向铁路系统异常检测的深度学习：结构化综述

# Deep Learning for Anomaly Detection in Railway Systems: A Structured Survey
Source: [https://arxiv.org/html/2610.00363](https://arxiv.org/html/2610.00363)
Ammar BoukettaAffiliation:Université Polytechnique Hauts\-de\-France, LAMIH CNRS UMR 8201, Valenciennes, FranceAffiliation:Alstom, Crespin, FranceSmail NiarAffiliation:Université Polytechnique Hauts\-de\-France, LAMIH CNRS UMR 8201, INSA Hauts\-de\-France, Valenciennes, FranceHamza OuarnoughiAffiliation:Université Polytechnique Hauts\-de\-France, LAMIH CNRS UMR 8201, INSA Hauts\-de\-France, Valenciennes, FranceAffiliation:Computer Science Department, College of Computing and Informatics, University of Sharjah, Sharjah, UAE

###### Abstract

Ensuring safe and reliable operation of modern railway systems increasingly relies on data\-driven monitoring and intelligent fault detection solutions\. In recent years, deep learning has emerged as an effective paradigm for anomaly detection in railway applications, driven by the growing availability of heterogeneous sensor data collected from rolling stock and infrastructure\. This paper presents a comprehensive survey of deep learning–based anomaly detection approaches developed for railway systems\. The surveyed methods are systematically analyzed and organized using a unified taxonomy that captures the location of anomaly occurrence, data representation and manifestation, sensing modality, and temporal characteristics\. Existing approaches, spanning convolutional neural networks, recurrent and attention\-based architectures, autoencoders, generative adversarial networks, and transformer\-based models, are further structured into major learning paradigms, including classification\-based, prediction\-based, reconstruction\-based, and hybrid strategies\. Beyond algorithmic design, this survey discusses key data\-centric challenges and reviews commonly adopted evaluation practices and performance metrics\. Practical deployment aspects are also addressed, including edge–cloud architectures, computational constraints, and hardware\-aware optimization strategies for embedded railway applications\. Building on this analysis, the survey introduces a decision\-oriented framework that links anomaly characteristics, data properties, and operational constraints to appropriate detection paradigms and deployment configurations\. By jointly considering learning models, data characteristics, evaluation protocols, and deployment constraints, this work provides engineers and researchers with a structured and application\-oriented reference for selecting and deploying deep learning solutions for railway anomaly detection, while highlighting open challenges and future research directions toward reliable and scalable intelligent monitoring systems\.

###### Keywords:

Railway anomaly detection , deep learning , predictive maintenance , anomaly taxonomy , edge deployment , decision framework

## 1Introduction

Railway systems constitute a fundamental component of modern transportation infrastructures, supporting large\-scale passenger mobility and freight logistics worldwide\. With nearly 930,000 kilometers of railway networks in operation and trillions of passenger\- and tonne\-kilometers transported annually, railways play a central role in sustainable, energy\-efficient, and high\-capacity transportation systems\[[67](https://arxiv.org/html/2610.00363#bib.bib151),[10](https://arxiv.org/html/2610.00363#bib.bib7)\]\. As railway networks continue to expand and aging assets are operated under increasing performance and safety requirements, ensuring reliability, availability, and operational continuity has become a major engineering challenge\.

Anomaly detection has emerged as a core function in modern railway monitoring and maintenance systems, as undetected faults can rapidly propagate across interconnected subsystems and lead to service disruptions, costly downtime, or safety\-critical incidents\. Typical anomalies include track degradation, rolling\-stock malfunctions, electrical faults, and abnormal environmental interactions, all of which may impact system performance and safety if not identified at an early stage\[[151](https://arxiv.org/html/2610.00363#bib.bib54)\]\. To address these challenges, the railway industry is progressively adopting intelligent condition monitoring and predictive maintenance strategies based on embedded sensing technologies and artificial intelligence \(AI\)\. Such systems are now deployed across a wide range of subsystems, including traction units, passenger access doors, braking systems, HVAC \(Heating, Ventilation, and Air Conditioning\) units, pantograph\-catenary interfaces, and track infrastructure\[[31](https://arxiv.org/html/2610.00363#bib.bib53)\]\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/overview_railway.png)Figure 1:Overview of railway subsystems and anomaly detection domains addressed in this survey, organized into three levels: infrastructure, rolling stock, and environmental or behavioral contexts\.As illustrated in Figure[1](https://arxiv.org/html/2610.00363#S1.F1), railway anomaly detection spans multiple operational domains, including onboard rolling\-stock subsystems \(e\.g\., doors, HVAC, traction\), wayside and infrastructure components \(e\.g\., rails, fasteners, catenary\), as well as environmental and behavioral contexts \(e\.g\., foreign objects, wind effects, passenger\-related events\)\. These domains differ substantially in sensing modalities, temporal dynamics, fault mechanisms, and data availability, which prevents the use of a single unified detection strategy\. Railway monitoring data are often high\-dimensional, strongly imbalanced, and temporally structured, posing significant challenges to traditional rule\-based or threshold\-based detection approaches\[[119](https://arxiv.org/html/2610.00363#bib.bib65),[90](https://arxiv.org/html/2610.00363#bib.bib73)\]\. Such methods typically struggle to generalize across nonlinear operating regimes, multisensor interactions, and evolving environmental conditions, resulting in brittle performance when deployed in real operational settings\[[128](https://arxiv.org/html/2610.00363#bib.bib66)\]\.

Historically, anomaly detection in railway systems has relied on expert\-defined rules, statistical thresholds, and handcrafted features derived from classical signal processing techniques, such as spectral analysis or principal component analysis \(PCA\)\[[18](https://arxiv.org/html/2610.00363#bib.bib64)\]\. While these approaches remain effective in well\-understood and stationary scenarios, they exhibit important limitations in modern railway environments\. Fixed thresholds lack adaptability to variations in operating conditions, component aging, and environmental influences, often leading to high false\-alarm rates or missed detections\. Furthermore, manual feature engineering is labor\-intensive and frequently fails to capture complex nonlinear and cross\-modal relationships observed in contemporary railway subsystems, such as traction motors, door actuation mechanisms, or pantograph\-catenary interfaces\. These limitations are exacerbated by the rarity of safety\-critical fault events, which results in severe class imbalance and further degrades detection reliability\[[119](https://arxiv.org/html/2610.00363#bib.bib65),[6](https://arxiv.org/html/2610.00363#bib.bib72),[46](https://arxiv.org/html/2610.00363#bib.bib8)\]\.

Deep learning \(DL\) has emerged as a powerful alternative by enabling data\-driven modeling of complex spatial and temporal patterns directly from raw or minimally processed sensor data\[[131](https://arxiv.org/html/2610.00363#bib.bib115),[144](https://arxiv.org/html/2610.00363#bib.bib116)\]\. A wide range of architectures, including convolutional neural networks \(CNN\), recurrent and attention\-based models, autoencoders, and transformer\-based frameworks, have demonstrated strong potential for railway anomaly detection\[[64](https://arxiv.org/html/2610.00363#bib.bib117),[105](https://arxiv.org/html/2610.00363#bib.bib56),[79](https://arxiv.org/html/2610.00363#bib.bib9)\]\. Representative applications include wheel\-flat detection from vibration signals\[[76](https://arxiv.org/html/2610.00363#bib.bib118),[116](https://arxiv.org/html/2610.00363#bib.bib138)\], electrical arc detection\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\], door obstruction detection\[[60](https://arxiv.org/html/2610.00363#bib.bib85)\], and rail\-surface inspection using visual or Light Detection and Ranging \(LiDAR\) data\[[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. Despite this rapid methodological progress, the practical adoption of deep learning–based anomaly detection in operational railway systems remains limited\.

A key challenge lies in the lack of system\-level alignment between anomaly characteristics, sensing modalities, learning paradigms, evaluation practices, and deployment constraints\[[39](https://arxiv.org/html/2610.00363#bib.bib74),[11](https://arxiv.org/html/2610.00363#bib.bib76)\]\. Many reported approaches achieve high performance under controlled experimental conditions but fail to translate into robust, comparable, and deployable solutions when confronted with real\-world railway data and hardware limitations\[[17](https://arxiv.org/html/2610.00363#bib.bib87),[20](https://arxiv.org/html/2610.00363#bib.bib86)\]\. This gap highlights the need for a structured and engineering\-oriented analysis that jointly considers data characteristics, model design, evaluation criteria, and deployment feasibility\. Motivated by this observation, this survey provides a comprehensive system\-level review of deep learning–based anomaly detection approaches for railway systems, with particular emphasis on practical applicability, reproducibility, and deployment\-oriented decision support\.

### 1\.1Research Questions

This survey is structured around six research questions that collectively address the full pipeline from anomaly characterization to deployment in railway systems:

- •RQ1:How can railway anomalies be systematically characterized to support consistent method selection across heterogeneous subsystems and sensing modalities? \(Section[3](https://arxiv.org/html/2610.00363#S3)\)
- •RQ2:What data\-centric challenges recur across the surveyed literature, and which datasets are available for railway anomaly detection in terms of accessibility, modality, and subsystem coverage? \(Section[4](https://arxiv.org/html/2610.00363#S4)\)
- •RQ3:Which deep learning detection paradigms are applied to railway anomaly detection, and how do their properties align with the characteristics of railway data and safety requirements? \(Section[5](https://arxiv.org/html/2610.00363#S5)\)
- •RQ4:How is performance evaluated in the surveyed studies, and do existing evaluation practices adequately reflect the safety\-critical nature of railway anomaly detection? \(Section[6](https://arxiv.org/html/2610.00363#S6)\)
- •RQ5:How are deep learning models deployed in railway applications, and what hardware constraints, optimization strategies, and real\-time requirements govern these deployments? \(Section[7](https://arxiv.org/html/2610.00363#S7)\)
- •RQ6:What decision\-oriented framework can guide practitioners in selecting detection paradigms and deployment configurations for specific railway anomaly scenarios? \(Section[7\.6](https://arxiv.org/html/2610.00363#S7.SS6)\)

### 1\.2Survey Methodology

This survey follows a structured and reproducible literature review protocol inspired by the PRISMA guidelines\[[102](https://arxiv.org/html/2610.00363#bib.bib119)\], adapted to the requirements of an engineering\-oriented analysis\.

#### 1\.2\.1Scope and Time Window

The review focuses on peer\-reviewed journal articles and conference papers published between January 2019 and 2026\. This period corresponds to the widespread adoption of deep learning techniques in railway monitoring and condition\-based maintenance\.

#### 1\.2\.2Search Strategy

The literature search was conducted using IEEE Xplore and Google Scholar\. IEEE Xplore was used as the primary structured database to ensure reliable and reproducible retrieval of peer\-reviewed engineering studies\. Google Scholar was used as a complementary source to broaden the search and capture relevant publications across major publishers, including Elsevier, Springer, and ACM\.

The search query was applied to titles, abstracts, and keywords \(when supported\), and adapted to each database:

> \("railway" OR "rail transit" OR "train" OR "metro" OR "tram"\) AND \("anomaly detection" OR "fault detection" OR "fault diagnosis" OR "condition monitoring"\) AND \("deep learning" OR "neural network" OR "CNN" OR "LSTM" OR "autoencoder" OR "transformer"\)

Results were filtered by publication period \(2019–2026\), document type \(journal articles and conference papers\), and language \(English\)\.

#### 1\.2\.3Inclusion and Exclusion Criteria

Studies were included if they:

- •apply at least one deep learning method,
- •focus on railway systems,
- •address anomaly detection, fault detection, or condition monitoring,
- •report quantitative experimental results,
- •provide sufficient methodological detail, including model design, training procedure, and evaluation protocol,
- •are published in peer\-reviewed venues\.

To ensure the quality and relevance of the survey, additional quality\-oriented criteria were considered:

- •studies lacking methodological transparency or reproducibility were excluded,
- •studies published in peer\-reviewed and established venues were prioritized to improve the reliability of the reviewed evidence\.

Studies were excluded if they:

- •rely only on classical machine learning methods,
- •target non\-railway domains,
- •lack quantitative evaluation,
- •are survey or review papers,
- •are not accessible in full text\.

#### 1\.2\.4Selection Process

The study selection followed a PRISMA\-inspired workflow \(Figure[2](https://arxiv.org/html/2610.00363#S1.F2)\)\.

Records were first retrieved from the selected databases and filtered based on publication period, document type, and language, followed by duplicate removal\.

Titles and abstracts were then screened to retain only studies relevant to railway systems, deep learning, and anomaly or fault detection tasks\.

The remaining studies were assessed through full\-text analysis, where papers lacking sufficient experimental validation, methodological transparency, or relevance were excluded\.

Finally, additional studies were identified through snowballing to improve coverage\.

Overall, the process follows adepth\-over\-breadth strategy, retaining only studies with sufficient technical detail and experimental rigor for meaningful analysis\. This strategy was adopted to avoid a purely descriptive accumulation of loosely related studies and to ensure that each included work could be analyzed consistently in terms of model design, data characteristics, evaluation protocol, and relevance to deep learning\-based railway anomaly detection\.

The exclusion counts and primary reasons are summarized in Figure[2](https://arxiv.org/html/2610.00363#S1.F2): 100 records were excluded during screening because they were not related to railway systems, deep learning, anomaly detection, or peer\-reviewed research\. A further 44 records were excluded during full\-text eligibility assessment due to insufficient experimental validation \(n=14n=14\), limited methodological transparency \(n=21n=21\), or marginal relevance to deep learning\-based railway anomaly detection \(n=9n=9\), yielding a final corpus of 68 studies\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/slr_survey1-Page-1.drawio.png)Figure 2:PRISMA\-inspired flow diagram of the literature identification, screening, eligibility, and inclusion process adopted in this survey\.

## 2Related Work

Deep learning has rapidly emerged as a transformative approach for anomaly detection in railway systems, prompting several surveys that analyze progress and challenges in this field\. While these reviews have advanced understanding, they remain largely fragmented\[[103](https://arxiv.org/html/2610.00363#bib.bib166),[22](https://arxiv.org/html/2610.00363#bib.bib67),[158](https://arxiv.org/html/2610.00363#bib.bib21)\], often limited to specific subsystems, single data modalities, or algorithmic benchmarks, without addressing DL methods categorization and deployment considerations such as hardware constraints, latency, and system integration\. This section critically synthesizes existing surveys to contextualize and motivate the present work\.

Several reviews provide broad overviews of artificial intelligence \(AI\) applications in railways\. Oh et al\.\[[100](https://arxiv.org/html/2610.00363#bib.bib69)\]cover infrastructure, rolling stock, and station safety, but their analysis is primarily descriptive and lacks discussion of deployment feasibility or modality\-specific challenges\. Baldyga et al\.\[[6](https://arxiv.org/html/2610.00363#bib.bib72)\]benchmark classical and deep\-learning algorithms on time\-series sensor data, revealing performance differences in operational environments, yet they do not address multimodal fusion or embedded constraints\. Lourenço et al\.\[[90](https://arxiv.org/html/2610.00363#bib.bib73)\]focus on predictive maintenance for wheel–track systems, emphasizing data sparsity and non\-stationarity, but their scope is confined to a narrow subsystem\. Tang et al\.\[[137](https://arxiv.org/html/2610.00363#bib.bib70)\]survey AI applications across seven railway domains, including safety and autonomous driving, but provide only a high\-level overview without in\-depth analysis of learning strategies or deployment considerations\.

Other surveys concentrate on vision\-based or modality\-specific techniques\. Kumar et al\.\[[78](https://arxiv.org/html/2610.00363#bib.bib71)\]and Di Summa et al\.\[[41](https://arxiv.org/html/2610.00363#bib.bib75)\]review computer\-vision approaches for defect detection, reporting high accuracy in controlled environments but limited generalization to multimodal data or operational settings\. De Donato et al\.\[[39](https://arxiv.org/html/2610.00363#bib.bib74)\]extend the analysis to audio and video modalities for nonintrusive predictive maintenance, yet do not discuss system\-level integration or inference under hardware constraints\. Binder et al\.\[[11](https://arxiv.org/html/2610.00363#bib.bib76)\]provide a large\-scale review of predictive maintenance methods across twenty railway components, classifying them by failure mode and algorithm type, but offering limited insights into real\-time feasibility, interpretability, or embedded implementation\. More recently, Phusakulkajorn et al\.\[[17](https://arxiv.org/html/2610.00363#bib.bib87)\]provide a structured review of AI techniques for railway infrastructure maintenance, covering track geometry, vibration monitoring, and imaging\-based inspection across multiple subsystems\. While comprehensive in scope, their analysis focuses on maintenance strategies rather than anomaly detection paradigms, and does not address deployment constraints or hardware\-aware model design\. Similarly, Hadj\-Mabrouk et al\.\[[58](https://arxiv.org/html/2610.00363#bib.bib62)\]surveys AI and machine learning applications in railway operations and safety, but focuses primarily on classical AI techniques and does not address deep learning\-based anomaly detection or embedded deployment\.

Table[1](https://arxiv.org/html/2610.00363#S2.T1)summarizes the coverage of these surveys across major dimensions, including subsystem scope, deep\-learning strategies, data challenges, and deployment considerations\.

Table 1:Comparative Overview of Survey Coverage in Railway AI\. Infra: infrastructure, Env: environmental, DL Strat: DL Strategy, Data Chal: Data Challenges, Depl: Deployment\.SurveyYearInfra\.TrainEnv\.DL Strat\.Data Chal\.Depl\.Oh et al\.\[[100](https://arxiv.org/html/2610.00363#bib.bib69)\]2022✓✓✓✓✗NTLourenço et al\.\[[90](https://arxiv.org/html/2610.00363#bib.bib73)\]2024✓✓✗✓✓NTTang et al\.\[[137](https://arxiv.org/html/2610.00363#bib.bib70)\]2022✓✓✓✓✗NTKumar et al\.\[[78](https://arxiv.org/html/2610.00363#bib.bib71)\]2024✓✓✗✗✗NTDe Donato et al\.\[[39](https://arxiv.org/html/2610.00363#bib.bib74)\]2022✓✗✗✓✗TDi Summa et al\.\[[41](https://arxiv.org/html/2610.00363#bib.bib75)\]2023✓✗✗✗✗NTBałdyga et al\.\[[6](https://arxiv.org/html/2610.00363#bib.bib72)\]2024✓✓✗✗✓NTBinder et al\.\[[11](https://arxiv.org/html/2610.00363#bib.bib76)\]2023✓✓✓✗✗NTPhusakulkajorn et al\.\[[17](https://arxiv.org/html/2610.00363#bib.bib87)\]2025✓✗✗✓✓NTHadj\-Mabrouk\[[58](https://arxiv.org/html/2610.00363#bib.bib62)\]2025✓✓✗✗✗NTOur Survey2026✓✓✓✓✓T
Note: ✓ = covered; ✗ = not covered; T = treated; NT = not treated\.

Column descriptions:

- •Infra\.,Train\.,Env\.: Coverage of anomaly detection in infrastructure \(e\.g\., rails, joints\), train subsystems \(e\.g\., brakes, doors\), and environmental scenarios \(e\.g\., passenger behavior, environmental hazards\)\.
- •DL Strat\.: Whether the survey categorizes deep\-learning methods by detection paradigm \(classification, reconstruction, prediction, hybrid\)\.
- •Data Ch\.: Whether data\-centric challenges, such as class imbalance, noise, or drift, are analyzed\.
- •Deploy\.: Whether deployment aspects, including edge inference, hardware constraints, or model compression, are treated\.

Synthesis and Positioning:A close inspection of prior railway anomaly\-detection surveys reveals several recurring limitations that reduce their usefulness for system\-level engineering\. First, the literature remainsfragmented by subsystem and modality: many surveys focus on isolated domains \(e\.g\., vision\-based track inspection, wheel\-rail dynamics, or vibration monitoring\) and therefore do not provide a unified perspective spanning infrastructure, rolling stock, and operational or environmental anomalies\. Second, deep\-learning methods are frequently reported in acatalog\-stylemanner, listing architectures or applications without sufficient abstraction intoactionable detection paradigms\. As a result, readers receive limited guidance on which learning strategy \(classification, reconstruction, prediction, or hybrid\) is appropriate under specific anomaly characteristics and data conditions\. Third,deployment and decision constraintsare often treated as secondary aspects\. In practice, however, railway feasibility is governed by latency, memory and power budgets, reliability requirements, and operational risk, which are rarely integrated into the methodological discussion\.

Contributions and Novelty of This Survey:Rather than a descriptive aggregation of prior work, this survey provides a design\- and decision\-oriented synthesis of deep learning\-based railway anomaly detection\. Its main contributions are four complementary elements that support systematic method selection and deployment planning:

- •Four\-axis anomaly taxonomy:A unified taxonomy based on location, data manifestation, sensing modality, and temporal behavior, enabling consistent characterization of anomalies across heterogeneous railway subsystems and data sources\.
- •Paradigm\-level organization of deep learning:A structured classification of existing methods into classification\-, reconstruction\-, prediction\-, and hybrid\-based paradigms, with explicit links between paradigm choice, anomaly properties, and data regimes\.
- •Evaluation metrics and assessment practices:We review the evaluation metrics and experimental practices used in existing railway anomaly detection studies, and we propose an extended evaluation framework that explicitly defines how anomaly detection performance should be assessed for railway systems beyond standard accuracy reporting\.
- •Deployment analysis and decision\-oriented framework:We review how deep learning models are deployed in railway applications by analyzing deployment architectures, hardware platforms, and optimization strategies, and we propose a decision\-oriented framework that guides the selection of detection approaches and deployment configurations under real\-world railway constraints\.

## 3Railway Anomalies Categorization

Railway anomalies exhibit diverse characteristics depending on their physical origin, data manifestation, sensing modality, and temporal dynamics\[[6](https://arxiv.org/html/2610.00363#bib.bib72),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\. To support systematic analysis and model selection, we propose a four\-axis taxonomy that links anomaly types to appropriate detection strategies and deployment considerations\. This framework facilitates reproducibility, multimodal integration, and real\-world applicability\. Figure[3](https://arxiv.org/html/2610.00363#S3.F3)provides an overview of the four axes and their categories\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/taxonomy_railway_anomaly.png)Figure 3:Four\-axis taxonomy for railway anomaly characterization: location of occurrence, manifestation in data, sensing modality, and temporal behavior\.### 3\.1Categorization by Location of Occurrence

Anomalies can be first classified by the physical domain in which they arise\[[6](https://arxiv.org/html/2610.00363#bib.bib72),[39](https://arxiv.org/html/2610.00363#bib.bib74)\]:

- •Infrastructure\-Level: Includes faults in static assets such as rails \(e\.g\., cracks, corrugation\), insulated joints \(spark erosion\), ballast and fasteners \(fouling, loosening\), sleepers \(cracking, misalignment\), and track geometry \(e\.g\., curvature or super\-elevation errors\)\[[17](https://arxiv.org/html/2610.00363#bib.bib87),[90](https://arxiv.org/html/2610.00363#bib.bib73)\]\.
- •Rolling Stock\-Level: Covers dynamic subsystems such as train doors, traction and braking systems, wheels and bogies, pantographs, HVAC units, and structural components of the train body\[[31](https://arxiv.org/html/2610.00363#bib.bib53),[11](https://arxiv.org/html/2610.00363#bib.bib76)\]\.
- •Environmental and Behavioral\-Level: Encompasses anomalies arising from environmental interaction or user behavior, including trespassing, extreme weather \(e\.g\., crosswinds\), and unsafe passenger actions\[[85](https://arxiv.org/html/2610.00363#bib.bib104),[137](https://arxiv.org/html/2610.00363#bib.bib70),[45](https://arxiv.org/html/2610.00363#bib.bib22)\]\.

### 3\.2Categorization by Manifestation in Data

Anomalies can also be distinguished by how they deviate from expected data patterns\[[23](https://arxiv.org/html/2610.00363#bib.bib158),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]:

- •Point Anomalies: Single data points that significantly deviate from the norm, such as a sudden voltage drop or vibration spike\[[23](https://arxiv.org/html/2610.00363#bib.bib158)\]\.
- •Contextual Anomalies: Values that are anomalous only within a specific context \(temporal, spatial, or operational\)\. For example, elevated brake temperature may be normal during descent but anomalous on a level track\[[23](https://arxiv.org/html/2610.00363#bib.bib158),[92](https://arxiv.org/html/2610.00363#bib.bib159)\]\.
- •Collective Anomalies: Sequences of individually normal observations that form an anomalous pattern, such as gradual bearing degradation or pantograph instability\[[23](https://arxiv.org/html/2610.00363#bib.bib158),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.
- •Trend Anomalies: Gradual and persistent shifts in the statistical distribution of sensor data that deviate from long\-term baseline behavior\. Representative examples include progressive wheel wear, rail surface degradation, or HVAC filter clogging, where no single observation is anomalous but the cumulative drift signals a developing fault\[[161](https://arxiv.org/html/2610.00363#bib.bib61),[35](https://arxiv.org/html/2610.00363#bib.bib57)\]\.

### 3\.3Categorization by Sensing Modality

The type of sensor and data captured directly influences the detection strategy and deployment feasibility\[[39](https://arxiv.org/html/2610.00363#bib.bib74),[6](https://arxiv.org/html/2610.00363#bib.bib72)\]:

- •Image and Video: Used for visual inspection of structural defects and dynamic events\. Widely applied to rail surface inspection, fastener detection, and foreign object recognition\[[78](https://arxiv.org/html/2610.00363#bib.bib71),[41](https://arxiv.org/html/2610.00363#bib.bib75)\]\.
- •Time Series: Includes vibration, temperature, current, and pressure signals acquired from onboard and wayside sensors\. These signals are often non\-stationary and exhibit strong temporal dependencies, requiring sequential modeling approaches\[[6](https://arxiv.org/html/2610.00363#bib.bib72),[90](https://arxiv.org/html/2610.00363#bib.bib73)\]\.
- •Acoustic Data: Captures contact or ambient sounds, often transformed into spectrograms for convolutional neural network \(CNN\) or attention\-based analysis\[[70](https://arxiv.org/html/2610.00363#bib.bib152),[122](https://arxiv.org/html/2610.00363#bib.bib92)\]\.
- •Structured and Tabular Data: Includes numerical logs, operational alarms, and event timestamps, commonly used for condition monitoring and fault diagnosis\[[11](https://arxiv.org/html/2610.00363#bib.bib76)\]\.
- •Textual and Log Data: Comprises maintenance records and incident reports, processed using Natural Language Processing \(NLP\) techniques such as transformers or embedding models\[[47](https://arxiv.org/html/2610.00363#bib.bib155)\]\.
- •Emerging Modalities: Light Detection and Ranging \(LiDAR\), radar, multispectral sensors, and Distributed Acoustic Sensing \(DAS\) are increasingly investigated for enhanced spatial perception and environmental awareness\[[96](https://arxiv.org/html/2610.00363#bib.bib154),[29](https://arxiv.org/html/2610.00363#bib.bib58),[160](https://arxiv.org/html/2610.00363#bib.bib59)\]\.

### 3\.4Categorization by Temporal Behavior

Temporal dynamics further differentiate anomaly types and directly guide model selection\[[53](https://arxiv.org/html/2610.00363#bib.bib160),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]:

- •Static Anomalies: Detectable from a single snapshot, such as rail cracks or foreign objects on tracks\. Typically addressed using CNN\-based visual inspection\[[33](https://arxiv.org/html/2610.00363#bib.bib165),[78](https://arxiv.org/html/2610.00363#bib.bib71)\]\.
- •Sequential Anomalies: Require temporal modeling to capture evolving patterns, such as progressive wheel wear or air leakage\. Recurrent Neural Networks \(RNNs\), Long Short\-Term Memory networks \(LSTMs\), Temporal Convolutional Networks \(TCNs\), and Transformer architectures are commonly employed\[[148](https://arxiv.org/html/2610.00363#bib.bib108),[4](https://arxiv.org/html/2610.00363#bib.bib60),[43](https://arxiv.org/html/2610.00363#bib.bib37)\]\.

Temporal categorization guides the selection of models with appropriate memory and sequence learning capabilities\.

## 4Data Challenges, Preprocessing, and Datasets for Railway Anomaly Detection

### 4\.1Data\-Centric Challenges and Preprocessing Strategies

The performance of deep learning\-based anomaly detection systems in railway applications is strongly influenced by the quality, structure, and variability of sensor data collected from onboard and wayside monitoring systems\[[9](https://arxiv.org/html/2610.00363#bib.bib109),[21](https://arxiv.org/html/2610.00363#bib.bib110)\]\. The following challenges are particularly critical and recur across the surveyed literature\.

- •Severe class imbalance:Safety\-critical faults such as wheel flats, pantograph arc events, and door obstructions occur far less frequently than normal operation, causing deep models to overfit the majority class\. Oversampling techniques such as the Synthetic Minority Oversampling Technique\[[26](https://arxiv.org/html/2610.00363#bib.bib77)\]and Adaptive Synthetic Sampling\[[63](https://arxiv.org/html/2610.00363#bib.bib78)\], and Generative Adversarial Network\-based augmentation\[[88](https://arxiv.org/html/2610.00363#bib.bib79)\]can partially mitigate this imbalance, but must preserve temporal integrity, particularly for cycle\-based systems such as door actuation sequences\.
- •Contextual and operational variability:Normal behavior in railway systems depends on train speed, weather conditions, vehicle type, passenger load, and track geometry\[[154](https://arxiv.org/html/2610.00363#bib.bib111)\]\. Models trained under specific operational conditions often fail to generalize across different fleets or routes\. Context\-aware normalization, route\-specific calibration, and domain adaptation methods\[[130](https://arxiv.org/html/2610.00363#bib.bib80)\]are essential to improve cross\-fleet and cross\-network generalization\.
- •Sensor noise and missing data:Vibration signals are affected by wheel–rail roughness; current measurements suffer from power\-electronic interference; and wireless sensor streams frequently contain missing or corrupted segments\[[114](https://arxiv.org/html/2610.00363#bib.bib112)\]\. Signal filtering, interpolation, and denoising autoencoders are commonly applied to improve signal quality for downstream deep learning models\[[118](https://arxiv.org/html/2610.00363#bib.bib81)\]\.
- •Multimodal misalignment:Integrating heterogeneous data streams from vibration sensors, cameras, acoustic sensors, and current logs is challenging due to different sampling rates and asynchronous time bases\. Cross\-modal representation learning\[[5](https://arxiv.org/html/2610.00363#bib.bib153)\], dynamic time\-warping alignment, and late\-fusion architectures enable more reliable multimodal anomaly detection\.
- •Limited availability of public datasets:Safety and confidentiality constraints severely restrict access to operational railway data\. A systematic review by Pappaterra et al\.\[[104](https://arxiv.org/html/2610.00363#bib.bib127)\]revealed that the number of publicly available railway datasets remains critically low relative to the volume of published research\. The present survey confirms that this situation has only partially improved between 2019 and 2025: of the 32 datasets identified, fewer than one third are fully publicly accessible\. Simulated datasets derived from digital twins\[[101](https://arxiv.org/html/2610.00363#bib.bib82)\]and transfer learning strategies\[[66](https://arxiv.org/html/2610.00363#bib.bib83)\]represent promising directions to address this barrier\.

### 4\.2Datasets for Railway Anomaly Detection

This section provides a structured overview of the datasets used across the surveyed literature\. A total of32 datasetswere identified, covering infrastructure, rolling stock, and environmental monitoring contexts\. These datasets exhibit substantial heterogeneity in terms of origin, accessibility, and sensing modality\[[53](https://arxiv.org/html/2610.00363#bib.bib160),[22](https://arxiv.org/html/2610.00363#bib.bib67)\], which directly affects model design choices, evaluation protocols, and reproducibility of reported results\.

Origin of the data:Four categories are distinguished:

- •Real\-worlddatasets, collected during in\-service operations using accelerometers, cameras, current sensors\[[85](https://arxiv.org/html/2610.00363#bib.bib104),[121](https://arxiv.org/html/2610.00363#bib.bib122)\]\.
- •Simulateddatasets, generated through digital twins or physics\-based simulators\[[87](https://arxiv.org/html/2610.00363#bib.bib89)\]\.
- •Syntheticdatasets, produced via data augmentation or generative models to address rare\-event scarcity\[[146](https://arxiv.org/html/2610.00363#bib.bib123)\]\.
- •Laboratorydatasets, recorded on test rigs or scaled\-down prototypes, offering clean annotations but limited operational representativeness\[[61](https://arxiv.org/html/2610.00363#bib.bib91)\]\.

Accessibility of the data:Accessibility describes the level of public availability of a dataset, independently of its origin\. Railway datasets range from fully public benchmarks to partially public datasets shared upon request, and to restricted datasets owned by railway operators or industrial partners\. Limited accessibility remains a major barrier to fair benchmarking and reproducibility\[[22](https://arxiv.org/html/2610.00363#bib.bib67),[53](https://arxiv.org/html/2610.00363#bib.bib160)\]\. Among the surveyed datasets, notable fully public examples include MetroPT\-3 for air production unit monitoring\[[140](https://arxiv.org/html/2610.00363#bib.bib126)\], Rail Surface Defect Datasets\[[99](https://arxiv.org/html/2610.00363#bib.bib113)\], Rail\-5k\[[165](https://arxiv.org/html/2610.00363#bib.bib114)\], FaultSeg for wheel defect detection\[[117](https://arxiv.org/html/2610.00363#bib.bib148)\], and the pantograph arc datasets collected from European railway networks\[[123](https://arxiv.org/html/2610.00363#bib.bib128),[93](https://arxiv.org/html/2610.00363#bib.bib129)\]\.

To ensure consistency with the anomaly taxonomy introduced in Section[3](https://arxiv.org/html/2610.00363#S3), the surveyed datasets are organized by their location of occurrence into three categories: infrastructure\-level, rolling\-stock\-level, and environmental or behavioral datasets\. Dataset scale is reported as the exact sample count when disclosed in the original study, and marked as NR \(not reported\) otherwise\.

TheLabeledcolumn indicates whether the dataset provides ground\-truth annotations:Yesmeans fully annotated samples;Partialmeans annotations are incomplete or derived from maintenance reports only; andNomeans no annotations are provided, restricting usage to unsupervised approaches\.

1\) Infrastructure\-Level Datasets:These datasets focus on track geometry, fastener condition, rail surface defects, weld integrity, and catenary inspection \(Table[2](https://arxiv.org/html/2610.00363#S4.T2)\)\.

Table 2:Infrastructure\-Level Railway Datasets\. Accessibility\(Acc\.\): Public \(P\), Restricted \(R\)\. NR = not reported\.Ref\.Dataset nameAcc\.SystemModalityScale \(samples\)LabeledFault Types\[[129](https://arxiv.org/html/2610.00363#bib.bib90)\]Track\-Geom\-DefectRRailway trackTrack geometry172 436YesRail, switch, crossing defects\[[122](https://arxiv.org/html/2610.00363#bib.bib92)\]Rail\-AcousticRTrack, fastenersAudio signals1625YesWheel burn, loose nut\-bolt\[[73](https://arxiv.org/html/2610.00363#bib.bib95)\]IRJ\-SparkVisionRRail joints \(IRJ\)RGB images∼\\sim28 150YesSpark erosion, rail damage\[[156](https://arxiv.org/html/2610.00363#bib.bib98)\]Drone FastenerRRail fastenersRGB images500YesMissing fastener\[[107](https://arxiv.org/html/2610.00363#bib.bib99)\]Track FaultPRails, fastenersRGB images293PartialCracked rails, rusted bolts\[[52](https://arxiv.org/html/2610.00363#bib.bib88)\]Fastener VisionRRail fastenersRGB \+ depth3 300YesElastic strip, Nut, Baffle miss\[[159](https://arxiv.org/html/2610.00363#bib.bib150)\]RailSquat\-ABAPRail surfaceVibration signalsNRPartialRail squat, surface defect\[[36](https://arxiv.org/html/2610.00363#bib.bib156)\]Fastener 3DRRail fasteners3D point cloud∼\\sim5 500YesSkewed spring bar\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\]6C SystemRPantographVideo \+ signalsNRPartialPantograph arc, catenary defects\[[99](https://arxiv.org/html/2610.00363#bib.bib113)\]RSDDs\-113PRail surfaceRGB images \+ Depth113YesRolling scar, corrosion\[[165](https://arxiv.org/html/2610.00363#bib.bib114)\]Rail\-5kPRail surfaceRGB images5000PartialCrack, spalling\[[2](https://arxiv.org/html/2610.00363#bib.bib121)\]RailSurface\-FaultsPRail surfaceRGB images5050YesCracks, squats\[[162](https://arxiv.org/html/2610.00363#bib.bib124)\]RailSem19PRailway environmentVideo frames8500YesObstacles\[[134](https://arxiv.org/html/2610.00363#bib.bib125)\]OSDaR23PTrack, obstaclesLiDAR \+RGB \+IR1534YesObstacles, foreign objects\[[123](https://arxiv.org/html/2610.00363#bib.bib128)\]Pantograph Arc DCPPantograph catenaryVoltage \+ currentNRNoElectric arc\[[93](https://arxiv.org/html/2610.00363#bib.bib129)\]Pantograph Arc ACPPantograph catenaryVoltage \+ current21 033NoHarmonic distortion, power disturbance\[[86](https://arxiv.org/html/2610.00363#bib.bib130)\]DR\-TrainPLight railAcceleration \+ GPSNRNoTrack condition changes2\) Rolling Stock\-Level Datasets:These datasets capture onboard subsystems including doors, traction systems, braking systems, and heating, ventilation, and air conditioning units \(Table[3](https://arxiv.org/html/2610.00363#S4.T3)\)\. They include vibration, acoustic, current, and multivariate signals, often recorded during real train operations\.

Table 3:Rolling Stock\-Level Railway Datasets\. Accessibility\(Acc\.\): Public \(P\), Restricted \(R\)\. NR = not reported\.Ref\.Dataset nameAcc\.SystemModalityScale \(samples\)LabeledFault Types\[[87](https://arxiv.org/html/2610.00363#bib.bib89)\]HSR\-SimRRailway electrification systemMultivariate signals103 200YesShort\-circuit fault\[[61](https://arxiv.org/html/2610.00363#bib.bib91)\]Train\-DoorRigRTrain doorsCurrent626YesBearing wear, obstruction\[[49](https://arxiv.org/html/2610.00363#bib.bib84)\]RailHVAC\-RULRHVAC systemMultivariate signalsNRYesAir filter clogging\[[133](https://arxiv.org/html/2610.00363#bib.bib93)\]PlugDoor SoundRSliding doorsAcoustic signals164YesDoor faults\[[141](https://arxiv.org/html/2610.00363#bib.bib97)\]Braking SystemRBraking systemMultivariate signals28996YesPressure, braking force deviation\[[120](https://arxiv.org/html/2610.00363#bib.bib101)\]Wheel FlatPWheel\-railSpectrogram images4977YesWheel flat detection\[[149](https://arxiv.org/html/2610.00363#bib.bib102)\]Rail Transit IoTRMultiple systemsMultivariate streamsNRNoDevice anomalies \(door, motor, HVAC\)\[[50](https://arxiv.org/html/2610.00363#bib.bib103)\]HVAC FilterRHVAC air filterMultivariate signalsNRYesAir filter degradation, clogging\[[140](https://arxiv.org/html/2610.00363#bib.bib126)\]MetroPT\-3PAir productionPressure, current10 979 547NoAir leak, oil leak\[[80](https://arxiv.org/html/2610.00363#bib.bib149)\]FRailTRI20\-DODRPassenger doorRGB video44 099YesPassenger stuck, bag stuck\[[121](https://arxiv.org/html/2610.00363#bib.bib122)\]Railway DoorRDoor actuationCurrent signalsNRNoDoor actuation anomaly\[[117](https://arxiv.org/html/2610.00363#bib.bib148)\]FaultSegPWheel defectsRGB images1872YesCracks/Scratches, Shelling, Discoloration3\) Environmental and Behavioral\-Level Datasets:These datasets address hazards arising from environmental interactions and human behavior, including foreign object detection on tracks, trespassing events, obstacle intrusion, and wind\-induced operational risks \(Table[4](https://arxiv.org/html/2610.00363#S4.T4)\)\.

Table 4:Environmental and Behavioral\-Level Railway Datasets for Anomaly Detection\. Accessibility\(Acc\.\): Public \(P\), Restricted \(R\)\. NR = not reported\.Ref\.Dataset nameAcc\.SystemModalityScale \(samples\)LabeledFault Types\[[146](https://arxiv.org/html/2610.00363#bib.bib123)\]RFOD TrackRTrack surroundingsRGB images7235PartialForeign objects, obstacles\[[98](https://arxiv.org/html/2610.00363#bib.bib96)\]Foreign ObjectRTrack areaRGB images2534YesPedestrians, falling rocks\[[85](https://arxiv.org/html/2610.00363#bib.bib104)\]HSR Strong\-WindRWind\-exposed trackMeteoro\-logical signalsNRYesWind\-induced speed restriction eventsCritical perspective on dataset comparability and reported performance:Despite the diversity of datasets identified in this survey, the absence of standardized benchmarks significantly limits the comparability and generalizability of reported results\[[104](https://arxiv.org/html/2610.00363#bib.bib127),[27](https://arxiv.org/html/2610.00363#bib.bib120)\]\. Three structural limitations recur across the literature\.

First, the majority of studies evaluate their models on proprietary or task\-specific datasets with distinct fault definitions, sensing configurations, and operating conditions, making cross\-study comparison unreliable\. Vision\-based infrastructure inspection datasets frequently report high accuracy or mean Average Precision scores under controlled conditions\[[107](https://arxiv.org/html/2610.00363#bib.bib99),[156](https://arxiv.org/html/2610.00363#bib.bib98)\], whereas rolling stock monitoring datasets typically yield lower performance due to operational noise and variability\[[149](https://arxiv.org/html/2610.00363#bib.bib102),[50](https://arxiv.org/html/2610.00363#bib.bib103)\]\.

Second, models validated on laboratory\-scale or simulated datasets frequently demonstrate optimistic results that do not transfer to in\-service conditions\[[87](https://arxiv.org/html/2610.00363#bib.bib89)\]\. These discrepancies indicate that reported performance improvements are often influenced as much by dataset characteristics as by model architecture\.

Third, documentation quality remains poor\. Among the 32 datasets surveyed, several do not report their exact size, collection protocol, or known limitations\-a shortcoming already identified by Pappaterra et al\.\[[104](https://arxiv.org/html/2610.00363#bib.bib127)\]in 2021 and not yet resolved in the recent literature\. The absence of unified anomaly definitions, consistent train\-test splits, and cross\-dataset evaluation protocols further weakens reproducibility and hinders fair assessment of algorithmic progress\. Establishing open, annotated, and subsystem\-specific benchmarks remains a critical open challenge for the railway anomaly detection community\.

## 5Deep Learning Paradigms for Railway Anomaly Detection

Railway anomaly detection encompasses a diverse range of tasks, including binary anomaly detection, multiclass fault classification, spatial localization, early fault prediction, and remaining useful life estimation\[[22](https://arxiv.org/html/2610.00363#bib.bib67),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. These tasks differ in their supervision requirements, temporal modeling complexity, and operational objectives\. Binary detection distinguishes normal from abnormal states, multiclass classification targets specific fault types such as broken clips or valve malfunctions, localization extends detection to spatial mapping via segmentation or object detection, and temporal tasks such as early fault prediction leverage sequence models for prognostics\[[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.

Regardless of the target task, the choice of deep learning strategy fundamentally determines how a model learns to distinguish normal from anomalous behavior\. Based on our analysis of the 68 surveyed studies, we organize existing approaches into four paradigms according to their core detection principle\. Each paradigm can support one or more of the above tasks depending on data availability, labeling conditions, and operational constraints:

- •Classification\-basedapproaches learn discriminative boundaries between normal and fault classes from labeled data\.
- •Reconstruction\-basedapproaches detect anomalies as deviations from a learned normal reconstruction\.
- •Prediction\-basedapproaches flag anomalies as significant departures from expected temporal patterns\.
- •Hybridapproaches combine two or more of the above paradigms to address complex scenarios where a single strategy is insufficient\.

This taxonomy serves as the structural backbone of the survey and guides the analysis of methods across railway subsystems, sensing modalities, and deployment contexts\.

### 5\.1Classification\-Based Approaches

Classification\-based anomaly detection assigns each input sample to a predefined class \(either a binary label \(normal or abnormal\) or a specific fault type through supervised learning\[[6](https://arxiv.org/html/2610.00363#bib.bib72),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. This paradigm is particularly effective when fault categories are defined and labeled data are available, covering scenarios such as broken fasteners, door malfunctions, wheel surface defects, or foreign object intrusion\. Unlike reconstruction\- or prediction\-based methods, classification models learn explicit decision boundaries between known classes, making them well\-suited for recurring and distinguishable anomalies\.

Figure[4](https://arxiv.org/html/2610.00363#S5.F4)illustrates the typical pipeline: raw sensor inputs are acquired and preprocessed, then fed into a representation learning module \(e\.g\., CNN\[[129](https://arxiv.org/html/2610.00363#bib.bib90)\], Vision Transformer\[[94](https://arxiv.org/html/2610.00363#bib.bib55)\], or PointNet\+\+\[[36](https://arxiv.org/html/2610.00363#bib.bib156)\]\) that extracts discriminative features\. A classifier head produces class scores, and a decision layer applies thresholding or a final softmax to yield the fault label\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/classification_based.png)Figure 4:Typical pipeline for classification\-based anomaly detection in railway systems\.Based on our analysis of the surveyed studies, classification\-based methods are organized into three groups according to their input modality\. For each group, a summary table lists the surveyed works with their targeted subsystem, architecture, sensing modality, and reported metric\. These tables are intended as a reference map to help identify which approaches address their specific railway component and data type and not as a direct performance comparison, since each study uses a different dataset and evaluation protocol\.

#### 5\.1\.1Classical Machine Learning \(ML\) Classifiers

Classical ML pipelines combining handcrafted features with shallow classifiers remain relevant when labeled data are scarce or computation is constrained\[[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. SVM \(Support Vector Machine\), Random Forest, and MLP \(Multilayer Perceptron\) have been applied to acoustic and vibration signals for door, HVAC, and track fault classification\[[113](https://arxiv.org/html/2610.00363#bib.bib134),[133](https://arxiv.org/html/2610.00363#bib.bib93),[49](https://arxiv.org/html/2610.00363#bib.bib84),[122](https://arxiv.org/html/2610.00363#bib.bib92),[61](https://arxiv.org/html/2610.00363#bib.bib91)\], achieving competitive accuracy on small datasets \(Table[5](https://arxiv.org/html/2610.00363#S5.T5)\)\. However, their reliance on manual feature design limits scalability to complex or high\-dimensional inputs, motivating the shift to deep learning\.

Table 5:Classical Machine Learning classifiers for railway anomaly detection\. All methods are supervised\. Acc\.: Accuracy\.Ref\.YearSystemArchitectureModalityMetricAD Task\[[113](https://arxiv.org/html/2610.00363#bib.bib134)\]2021Rail trackSVM, RF, MLPAcousticAcc\.: 97%Classify track faults \(wheel burn, loose bolt\) from acoustic signals\.\[[133](https://arxiv.org/html/2610.00363#bib.bib93)\]2020Train doorsSVM \(\+MFCC features\)AcousticAcc\.:\>\>95%Classify plug\-door fault types from acoustic measurements\.\[[49](https://arxiv.org/html/2610.00363#bib.bib84)\]2021HVAC systemGradient Boosted Trees \(GBT\)MultivariateAcc\.: 92\.2%Detect air filter faults in railway HVAC from sensor data\.\[[122](https://arxiv.org/html/2610.00363#bib.bib92)\]2022Rail trackMLP \+ MFCCAcousticAcc\.: 98\.4%Classify rail structural faults from IoT acoustic data\.\[[61](https://arxiv.org/html/2610.00363#bib.bib91)\]2019Train doorsKNN vs\. CNNCurrentAcc\.: 99\.5%Detect abnormal current patterns in train door motors\.\[[111](https://arxiv.org/html/2610.00363#bib.bib135)\]2025Rail trackSVM, GRU, LSTMAcousticAcc\.:\>\>96%Classify track faults using combined MFCC and CQT features\.
#### 5\.1\.2Vision\-Based Deep Classifiers

Vision\-based classification is the dominant paradigm in the surveyed literature, covering rail surface defects, fasteners, joints, and wheel anomalies from wayside and onboard imagery\.

Standard CNNs with transfer learning \(Residual Network\-ResNet\[[77](https://arxiv.org/html/2610.00363#bib.bib132)\], Visual Geometry Group\-VGG\[[135](https://arxiv.org/html/2610.00363#bib.bib133)\]\) effectively handle data scarcity in visual inspection\[[129](https://arxiv.org/html/2610.00363#bib.bib90),[107](https://arxiv.org/html/2610.00363#bib.bib99)\]\. Ensemble CNN approaches combining VGG\-19, MobileNetV3, and ResNet50 achieve 99% accuracy on railway defect images\[[13](https://arxiv.org/html/2610.00363#bib.bib136)\], while YOLO\-based architectures \(You Only Look Once\) consistently outperform static CNN classifiers for binary fault classification\[[109](https://arxiv.org/html/2610.00363#bib.bib140)\]\.

For real\-time defect localization, attention\-enhanced YOLO\[[71](https://arxiv.org/html/2610.00363#bib.bib147)\]variants achieve strong speed\-accuracy trade\-offs under operational constraints\[[52](https://arxiv.org/html/2610.00363#bib.bib88),[73](https://arxiv.org/html/2610.00363#bib.bib95),[16](https://arxiv.org/html/2610.00363#bib.bib141),[167](https://arxiv.org/html/2610.00363#bib.bib10)\]\. Lightweight transformer\-based detectors such as RFD\-DETR \(Railway Fastener Defect Detection Transformer\) further improve fastener defect detection under size\-constrained deployment\[[166](https://arxiv.org/html/2610.00363#bib.bib142)\]\. Mask R\-CNN achieves 98% precision for missing fastener detection from drone imagery\[[156](https://arxiv.org/html/2610.00363#bib.bib98)\], and hybrid segmentation\-detection pipelines extend coverage to foreign object localization\[[98](https://arxiv.org/html/2610.00363#bib.bib96)\]\.

Vision Transformers \(ViT\)\[[74](https://arxiv.org/html/2610.00363#bib.bib131)\]overcome the locality bias of CNNs by capturing global structural dependencies\. ViT and DeiT\(Data\-efficient Image Transformer\) outperform CNN baselines for fastener \(DeiT: 95\.0% vs VGG16: 91\.5%\)\[[95](https://arxiv.org/html/2610.00363#bib.bib26),[94](https://arxiv.org/html/2610.00363#bib.bib55),[105](https://arxiv.org/html/2610.00363#bib.bib56)\]\. Non\-destructive testing pipelines further extend classification to subsurface defects by fusing camera and ultrasonic signals\[[153](https://arxiv.org/html/2610.00363#bib.bib100)\]\.In parallel, Wang et al\.\[[145](https://arxiv.org/html/2610.00363#bib.bib2)\]proposed a hierarchical transfer\-generative framework that automates multiple analytical tasks in rail surface defect inspection simultaneously\.

Table[6](https://arxiv.org/html/2610.00363#S5.T6)summarizes the surveyed vision\-based classification works\.

Table 6:Vision\-based deep classifiers for railway anomaly detection\. All methods are supervised\. Acc\.: Accuracy; mAP: mean Average Precision; mIoU: mean Intersection over Union; Prec\.: Precision\.Ref\.YearSystemArchitectureModalityMetricAD Task\[[129](https://arxiv.org/html/2610.00363#bib.bib90)\]2022Track geometryCNN, DNNTrack geometry signalsAcc\.: 94\.3%Classify rail geometry defects\.\[[107](https://arxiv.org/html/2610.00363#bib.bib99)\]2022Rail surfaceResNet\-50, VGG\-16RGB imagesAcc\.: 83\.8%Detect surface defects \(cracks, corrosion\)\.\[[13](https://arxiv.org/html/2610.00363#bib.bib136)\]2024Rail and wheelEnsemble CNNRGB imagesAcc\.: 99%Classify rail and wheel surface defects\.\[[109](https://arxiv.org/html/2610.00363#bib.bib140)\]2024Rail trackYOLOv11 \+ TL\(Transfer Learning\)RGB imagesAcc\.: 92\.5%Binary classification of track images: defective vs\. normal\.\[[52](https://arxiv.org/html/2610.00363#bib.bib88)\]2023Rail fastenersCSP\-Darknet53RGB \+ DepthmAP: 86\.6%Detect and localize missing and broken fasteners\.\[[73](https://arxiv.org/html/2610.00363#bib.bib95)\]2025Rail jointsSqueezeNetRGB imagesAcc\.: 98\.2%Detect and segment spark erosion on insulated rail joints\.\[[116](https://arxiv.org/html/2610.00363#bib.bib138)\]2025Wheel surfaceYOLOv5\-segRGB imagesmAP: 92%Detect and segment wheel defects \(flats, shelling, cracks, discoloration\)\.\[[16](https://arxiv.org/html/2610.00363#bib.bib141)\]2024Fasteners \+ trackFOD\-YOLONetRGB imagesAcc: 98\.14%Simultaneously detect fastener faults and foreign objects on tracks\.\[[166](https://arxiv.org/html/2610.00363#bib.bib142)\]2025Rail fastenersRFD\-DETRRGB imagesmAP: 98\.27%Detect and localize fastener defects\.\[[167](https://arxiv.org/html/2610.00363#bib.bib10)\]2024Rail surfaceYOLOv5/v8 \+ CBAMRGB imagesAcc\.: 80\.3%Detect and classify rail defects\.\[[156](https://arxiv.org/html/2610.00363#bib.bib98)\]2022Rail fastenersMask R\-CNNRGB \(drone\)Prec\.: 98%Detect and segment missing fasteners from drone imagery\.\[[98](https://arxiv.org/html/2610.00363#bib.bib96)\]2024Track obstaclesU\-Net \+ YOLOv5RGB imagesmAP@0\.5: 77\.9%Detect foreign objects on tracks and assess intrusion risk\.\[[153](https://arxiv.org/html/2610.00363#bib.bib100)\]2023Rail weldsCNN \+ SVMRGB imagesQualitativeDetect internal and surface weld flaws\.\[[94](https://arxiv.org/html/2610.00363#bib.bib55)\]2025Rail surfaceCNN vs\. ViTRGB imagesAcc\.:∼\\sim95%Classify rail surface and fastener defects comparing CNN and ViT\.\[[95](https://arxiv.org/html/2610.00363#bib.bib26)\]2025Rail fastenersViT, DeiTRGB imagesAcc\.: 95\.0%Detect fastener defects\.\[[105](https://arxiv.org/html/2610.00363#bib.bib56)\]2024Multi\-componentRailTrack\-DaViTRGB imagesAcc\.:\>\>98%Classify defects across rail, fastener, fishplate, and multi\-fault categories\.
#### 5\.1\.3Time\-Series Signal Classifiers

Rolling stock fault diagnosis relies on temporal signals from onboard sensors rather than visual inspection\. CNN\-based classifiers applied to spectrograms and time\-frequency representations effectively detect wheel flats and braking system faults\[[120](https://arxiv.org/html/2610.00363#bib.bib101),[141](https://arxiv.org/html/2610.00363#bib.bib97)\]\. GRU\-based hierarchical models enable ultra\-low\-latency multiclass fault classification in traction systems\[[87](https://arxiv.org/html/2610.00363#bib.bib89)\], while 1D\-CNN and LSTM architectures classify wheelset and suspension faults directly from raw acceleration signals\[[112](https://arxiv.org/html/2610.00363#bib.bib137)\]\. Multimodal classifiers combining visual and acoustic streams extend this paradigm to passenger behavioral anomaly detection\[[136](https://arxiv.org/html/2610.00363#bib.bib63)\]\.Table[7](https://arxiv.org/html/2610.00363#S5.T7)summarizes the surveyed time\-series classifiers\.

Table 7:Time\-series signal classifiers for railway anomaly detection\. All methods are supervised\. Acc\.: Accuracy; AUC: Area Under the ROC Curve; NR: Not Reported\.Ref\.YearSystemArchitectureModalityMetricAD Task\[[120](https://arxiv.org/html/2610.00363#bib.bib101)\]2023Wheel\-railLeNet\-5 CNNVibration signalsAcc\.: 97%, AUC: 0\.97Detect and classify wheel flat defects from vibration spectrograms\.\[[141](https://arxiv.org/html/2610.00363#bib.bib97)\]2021Braking systemMultivariate CNNPressure \+ forceAcc\.: 93%Classify braking faults from multivariate signals\.\[[87](https://arxiv.org/html/2610.00363#bib.bib89)\]2020Traction systemGRU hierarchicalMultivariateAcc\.:\>\>93%Classify 29 traction fault types from high\-frequency sensor streams\.\[[112](https://arxiv.org/html/2610.00363#bib.bib137)\]2023Wheelset1D\-CNN, LSTMAccelerationNRDetect and classify wheelset faults from raw acceleration time series\.\[[136](https://arxiv.org/html/2610.00363#bib.bib63)\]2023Passenger cabinSlowFast CNNVideo \+ acousticAcc\.: 85\.1%Classify abnormal passenger behavioral events\.Summary and limitations:Classification\-based approaches offer high accuracy, interpretable outputs, and efficient deployment on embedded hardware\[[73](https://arxiv.org/html/2610.00363#bib.bib95),[141](https://arxiv.org/html/2610.00363#bib.bib97)\]\. However, three structural limitations constrain their applicability in open railway environments\. First, they require curated labeled datasets per fault type, which is costly for rare or novel anomalies\[[22](https://arxiv.org/html/2610.00363#bib.bib67),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. Second, models trained on a closed fault set fail silently on out\-of\-distribution anomalies \(a critical safety concern in safety\-critical railway contexts\)\[[107](https://arxiv.org/html/2610.00363#bib.bib99)\]\. Third, class imbalance between normal and fault samples requires active mitigation through loss reweighting or data augmentation\[[52](https://arxiv.org/html/2610.00363#bib.bib88),[26](https://arxiv.org/html/2610.00363#bib.bib77)\]\.

Cross\-category observations:Across the three groups summarized in Tables[5](https://arxiv.org/html/2610.00363#S5.T5)\-[7](https://arxiv.org/html/2610.00363#S5.T7), several trends emerge\. Vision\-based classifiers dominate the surveyed literature, reflecting the widespread deployment of wayside cameras and the availability of large annotated image datasets for infrastructure inspection\. Classical ML remains competitive on small acoustic and vibration datasets but is systematically outperformed by deep models on raw high\-dimensional inputs\[[61](https://arxiv.org/html/2610.00363#bib.bib91),[129](https://arxiv.org/html/2610.00363#bib.bib90)\]\. Time\-series classifiers achieve strong results on rolling stock subsystems but are predominantly validated on simulated or single\-operator datasets, raising concerns about operational generalizability\[[87](https://arxiv.org/html/2610.00363#bib.bib89),[112](https://arxiv.org/html/2610.00363#bib.bib137)\]\. Transformer\-based architectures \(ViT, DeiT, DETR\) consistently outperform CNN baselines across visual inspection tasks at the cost of higher computational demands\[[94](https://arxiv.org/html/2610.00363#bib.bib55),[95](https://arxiv.org/html/2610.00363#bib.bib26),[166](https://arxiv.org/html/2610.00363#bib.bib142)\]\. Finally, the heterogeneity of reported metrics confirms that standardized benchmarking remains an open challenge for the railway anomaly detection community\[[103](https://arxiv.org/html/2610.00363#bib.bib166),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. These limitations motivate the complementary reconstruction\- and prediction\-based paradigms reviewed in subsequent sections\.

### 5\.2Prediction\-Based Approaches

Prediction\-based anomaly detection learns the normal temporal dynamics of a system and flags observations that deviate significantly from predicted values\[[161](https://arxiv.org/html/2610.00363#bib.bib61),[35](https://arxiv.org/html/2610.00363#bib.bib57),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. Unlike classification\-based approaches, this paradigm is typically trained on normal operational data, reducing dependence on fault labels, making it particularly suitable for railway subsystems where fault examples are rare or unavailable such as traction motors, braking units, Heating Ventilation and Air Conditioning \(HVAC\) systems, and track geometry monitoring\[[68](https://arxiv.org/html/2610.00363#bib.bib161),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\.

As illustrated in Figure[5](https://arxiv.org/html/2610.00363#S5.F5), a deep predictive model learns to predict the next system statex^t\+1\\hat\{x\}\_\{t\+1\}from a sliding window of past observationsXt=\[xt−n,…,xt\]X\_\{t\}=\[x\_\{t\-n\},\\ldots,x\_\{t\}\]\. An anomaly score is computed as the prediction errore=\|xt\+1−x^t\+1\|e=\|x\_\{t\+1\}\-\\hat\{x\}\_\{t\+1\}\|, and a static or adaptive threshold triggers a fault alert wheneeexceeds a predefined limit\. The model is trained exclusively on normal data, enabling detection of previously unseen fault types without any fault labels\[[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/prediction_based.png)Figure 5:Typical pipeline for prediction\-based anomaly detection in railway systems\.Based on our analysis of the surveyed studies, we organize prediction\-based railway studies into three architecture\-oriented groups \( recurrent sequence predictors, attention and transformer predictors, and convolutional temporal predictors\)based on the dominant architectural family used in each work\. For each group, a summary table lists the surveyed works\.

#### 5\.2\.1Recurrent Sequence Predictors

Recurrent Neural Networks \(RNNs\) and their gated variants constitute the dominant family for prediction\-based anomaly detection in railway systems\[[161](https://arxiv.org/html/2610.00363#bib.bib61),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. These architectures maintain a hidden state that evolves over time, enabling them to learn normal operational dynamics from sequential sensor data and detect anomalies as deviations between predicted and observed values\.

Standard RNNs have been applied to anticipate unsafe operational conditions such as abnormal running behavior and derailment risk from multivariate onboard sensor streams\[[81](https://arxiv.org/html/2610.00363#bib.bib105)\]\. However, standard recurrent networks suffer from gradient instability over long sequences, which limits their use in complex railway monitoring scenarios\.

LSTM networks overcome this limitation through gated memory cells that selectively retain or discard information across time steps\. In railway applications, LSTM predictors trained exclusively on normal operational data have been applied across multiple subsystems\. Wang et al\.\[[149](https://arxiv.org/html/2610.00363#bib.bib102)\]trained a modified Long Short\-Term Memory on normal data from heating ventilation and air conditioning units, and compressors to flag anomalies based on prediction error thresholds\. Yokouchi and Kondo\[[157](https://arxiv.org/html/2610.00363#bib.bib11)\]applied LSTM to condition monitoring data from railway vehicle air conditioning units, demonstrating that prediction\-based detection can identify abnormal behavior without any fault labels\. Hao et al\.\[[62](https://arxiv.org/html/2610.00363#bib.bib106)\]applied LSTM to predict future track geometry parameters from inertial measurements, enabling early detection of misalignment before safety thresholds are exceeded\. Ribeiro et al\.\[[108](https://arxiv.org/html/2610.00363#bib.bib12)\]further demonstrated that LSTM predictors can anticipate rolling stock equipment failures hours in advance using the MetroPT\-3 dataset of compressor monitoring signals\. Chen et al\.\[[30](https://arxiv.org/html/2610.00363#bib.bib139)\]introduced a deep LSTM method for suspension fault detection, modeling normal vibration responses and flagging deviations from the learned baseline \(Table[8](https://arxiv.org/html/2610.00363#S5.T8)\)\.

Table 8:Recurrent sequence predictors for railway anomaly detection\. Models are trained on normal data only unless otherwise noted\. F1: F1\-score; Acc\.: Accuracy; AUC: Area Under the ROC Curve; NR: not reported\.Ref\.YearSystemArchitectureModalityAD Task\[[81](https://arxiv.org/html/2610.00363#bib.bib105)\]2021Rolling stockLSTM, RNN\+ DNNMultivariate signalsPredict derailment risk and abnormal running behavior from onboard sensors\.\[[149](https://arxiv.org/html/2610.00363#bib.bib102)\]2022HVAC, compressorsModified Long Short\-Term MemoryMultivariate signalsPredict normal system state and flag anomalies in multiple rolling stock subsystems\.\[[157](https://arxiv.org/html/2610.00363#bib.bib11)\]2021Air conditioning unitLSTM, RNNTemperature \+ currentDetect anomalies in vehicle air conditioning units from monitoring data deviations\.\[[121](https://arxiv.org/html/2610.00363#bib.bib122)\]2022Air conditioning unitLSTMCurrent signalsDetect abnormal air conditioning units recorded in vehicule on operation\.\[[62](https://arxiv.org/html/2610.00363#bib.bib106)\]2023Track geometryLSTMInertial signalsForecast track geometry parameters to detect misalignment before safety limits are exceeded\.\[[108](https://arxiv.org/html/2610.00363#bib.bib12)\]2023Train traction converter coolingLSTMMultivariate signalsPredict rolling stock equipment failures hours in advance from sensor streams\.\[[30](https://arxiv.org/html/2610.00363#bib.bib139)\]2024SuspensionDeep LSTMVibration signalsDetect suspension faults by measuring deviations from learned normal vibration\.
#### 5\.2\.2Attention and Transformer Predictors

Attention mechanisms and Transformer architectures extend recurrent predictors by explicitly weighting informative sensor channels and critical time intervals, improving both accuracy and interpretability\[[35](https://arxiv.org/html/2610.00363#bib.bib57),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.

Dual Attention Recurrent Neural Networks integrate input\-level and temporal attention within an encoder\-decoder structure\. Liu et al\.\[[85](https://arxiv.org/html/2610.00363#bib.bib104)\]applied this architecture to real\-time wind risk prediction along high\-speed rail corridors, where the attention mechanism captures spatial\-temporal correlations across distributed weather sensor networks under weak supervision\. Transformer\-based architectures replace recurrence entirely with self\-attention, enabling parallel computation and superior modeling of long\-range temporal dependencies\. Ding et al\.\[[43](https://arxiv.org/html/2610.00363#bib.bib37)\]proposed Improved EmpiricalWavelet Transform \(IEWT\), a Transformer\-based model for railway axle bearing fault prediction from vibration signals, demonstrating improved performance over recurrent baselines on long temporal sequences \(Table[9](https://arxiv.org/html/2610.00363#S5.T9)\)\.

Table 9:Attention and Transformer predictors for railway anomaly detection\. AUC: Area Under the ROC Curve; NR: not reported\.
#### 5\.2\.3Convolutional Temporal Predictors

One\-dimensional CNNs and Temporal Convolutional Networks \(TCN\) apply convolutional filters along the temporal dimension to capture local temporal patterns with significantly lower computational cost than recurrent architectures\[[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. This makes them well\-suited for embedded and edge deployment in resource\-constrained railway environments\.

Galvez et al\.\[[50](https://arxiv.org/html/2610.00363#bib.bib103)\]combined physics\-based degradation models with a one\-dimensional CNN predictor to estimate the remaining useful life of railway heating ventilation and air conditioning air filters, detecting anomalies as deviations from the learned degradation trajectory\. Haghbin et al\.\[[59](https://arxiv.org/html/2610.00363#bib.bib36)\]applied 1D\-CNN to predict rail corrugation profiles directly from vehicle acceleration measurements, achieving prediction errors below 5% and demonstrating the feasibility of real\-time infrastructure monitoring from onboard sensors\. Zhang et al\.\[[163](https://arxiv.org/html/2610.00363#bib.bib38)\]proposed a hybrid architecture combining self\-attention, bidirectional TCNs, and bidirectional GRUs for predicting the evolution trend of rail corrugation, outperforming individual recurrent and convolutional baselines on a multi\-domain feature set extracted from acceleration signals \(Table[10](https://arxiv.org/html/2610.00363#S5.T10)\)\.

Table 10:Convolutional temporal predictors for railway anomaly detection\. Acc\.: Accuracy; NR: not reported; TCN: Temporal Convolutional Network; GRU: Gated Recurrent Unit\.Ref\.YearSystemArchitectureModalityAD Task\[[50](https://arxiv.org/html/2610.00363#bib.bib103)\]2022Heating ventilation air conditioning filter1D CNN \+ physics modelMultivariate signalsEstimate remaining useful life of air filter and detect degradation anomalies\.\[[59](https://arxiv.org/html/2610.00363#bib.bib36)\]2024Rail corrugation1D CNNAcceleration signalsPredict rail corrugation profile from vehicle acceleration to detect surface anomalies\.\[[163](https://arxiv.org/html/2610.00363#bib.bib38)\]2024Rail corrugationSelf\-attention \+ TCN \+ GRUAcceleration signalsPredict corrugation evolution trend and detect anomalous degradation progression\.Summary and limitations:Prediction\-based approaches offer flexibility and generalizability for railway monitoring, requiring only normal operational data and enabling the detection of previously unseen fault types\[[103](https://arxiv.org/html/2610.00363#bib.bib166),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. However, three limitations constrain their applicability\. First, prediction errors are sensitive to sensor noise and operational drift, which may generate false alarms unrelated to genuine faults\[[35](https://arxiv.org/html/2610.00363#bib.bib57)\]\. Second, threshold calibration is non\-trivial and must balance detection sensitivity against false alarm rates\[[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\. Third, models trained on normal data from one operational context often degrade significantly when deployed under different speed, load, or environmental conditions\[[149](https://arxiv.org/html/2610.00363#bib.bib102),[62](https://arxiv.org/html/2610.00363#bib.bib106)\]\.

Cross\-category observations:Across the three groups summarized in Tables[8](https://arxiv.org/html/2610.00363#S5.T8)\-[10](https://arxiv.org/html/2610.00363#S5.T10), LSTM architectures dominate the surveyed railway literature, reflecting their established stability over long sensor sequences and their suitability for multivariate onboard monitoring\[[149](https://arxiv.org/html/2610.00363#bib.bib102),[157](https://arxiv.org/html/2610.00363#bib.bib11),[108](https://arxiv.org/html/2610.00363#bib.bib12)\]\.Attention and Transformer\-based predictors are emerging as strong alternatives for both rolling stock and infrastructure monitoring, have demonstrated competitive or superior performance compared to recurrent baselines in several railway studies when long\-range temporal dependencies are present\[[43](https://arxiv.org/html/2610.00363#bib.bib37),[85](https://arxiv.org/html/2610.00363#bib.bib104)\]\. Convolutional temporal predictors represent a practical option for resource\-constrained deployments due to their lower computational cost where inference latency and memory footprint are critical, particularly for edge\-side track monitoring applications\[[50](https://arxiv.org/html/2610.00363#bib.bib103),[59](https://arxiv.org/html/2610.00363#bib.bib36)\]\. Finally, the majority of surveyed studies validate their approaches on proprietary or single\-operator datasets, limiting conclusions about cross\-fleet generalizability \(a gap that remains open for future work\)\[[37](https://arxiv.org/html/2610.00363#bib.bib107),[163](https://arxiv.org/html/2610.00363#bib.bib38)\]\.

### 5\.3Reconstruction\-Based Approaches

Reconstruction\-based anomaly detection trains models exclusively on normal data to learn compact representations of healthy system behavior\. At inference, inputs that deviate significantly from learned normal patterns produce elevated reconstruction errors, which serve as anomaly scores\[[8](https://arxiv.org/html/2610.00363#bib.bib164),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. Anomalies are flagged when the reconstruction errore=∥x−x^∥e=\\lVert x\-\\hat\{x\}\\rVertexceeds a static or adaptive threshold\. This paradigm requires no fault labels, making it particularly suited to railway contexts where annotated fault data are scarce or unavailable\[[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\.

As illustrated in Figure[6](https://arxiv.org/html/2610.00363#S5.F6), the pipeline consists of an encoder that compresses the inputx∈ℝnx\\in\\mathbb\{R\}^\{n\}into a latent representationzz, a decoder that reconstructsx^=D⁡\(E⁡\(x\)\)\\hat\{x\}=D\(E\(x\)\), and an anomaly scoring module that compares the input and its reconstruction\. The model is trained exclusively on normal data by minimizing the reconstruction lossℒ=∥x−x^∥2\\mathcal\{L\}=\\lVert x\-\\hat\{x\}\\rVert^\{2\}\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/reconstruction_based.png)Figure 6:Typical pipeline for reconstruction\-based anomaly detection in railway systems\.In this survey, reconstruction\-based railway studies are organized into two groups according to their core architectural family: standard and recurrent autoencoders, and generative adversarial models\. For each group, a summary table lists the surveyed works\.

#### 5\.3\.1Standard and Recurrent Autoencoders

Standard Autoencoders \(AEs\) and their recurrent variants constitute the dominant family for reconstruction\-based anomaly detection in railway systems\. These architectures compress input data into a bottleneck latent space and reconstruct it, flagging anomalies through elevated reconstruction error on unseen inputs\.

Sparse Autoencoders \(SAEs\) extend standard AEs by regularizing hidden activations to promote sparse and interpretable latent representations\[[24](https://arxiv.org/html/2610.00363#bib.bib48)\]\. Davari et al\.\[[37](https://arxiv.org/html/2610.00363#bib.bib107)\]applied separate SAE models to analog and digital sensor streamsfrom metro train Air Production Units \(APUs\), demonstrating that reconstruction\-based detection can identify operational anomalies in heterogeneous sensor environments without fault labels\. A subsequent study on the same MetroPT\-3 dataset\[[38](https://arxiv.org/html/2610.00363#bib.bib14)\]compared Long Short\-Term Memory Autoencoders \(LSTM\-AEs\) with SAEs, finding that LSTM\-AEs achieve superior F1\-score and recall by capturing temporal dependencies in pressure and current signals\. Kang et al\.\[[72](https://arxiv.org/html/2610.00363#bib.bib13)\]applied a one\-class LSTM\-AE to detect anomalies in the brake operating unit of metro vehicles, training exclusively on healthy current signals and flagging deviations with a reconstruction error threshold\.

Bidirectional LSTM Autoencoders \(Bi\-LSTM\-AEs\) extend recurrent AEs by processing sequences in both forward and backward directions, improving sensitivity to subtle temporal anomalies\. Correia et al\.\[[35](https://arxiv.org/html/2610.00363#bib.bib57)\]extended LSTM\-AE\-based reconstruction to online anomaly detection in multivariate railway sensor streams, demonstrating that reconstruction errors evaluated with the Mahalanobis distance improve detection robustness compared to mean squared error thresholding\. Seo et al\.\[[110](https://arxiv.org/html/2610.00363#bib.bib15)\]applied autoencoder\-based reconstruction to vibration signals from Korean high\-speed trains for unsupervised earthquake anomaly detection, training only on normal operational vibration data and outperforming the conventional Short\-Time Average over Long\-Time Average detector for peak ground acceleration events above the derailment threshold\.

Variational Autoencoders \(VAEs\) extend standard AEs by learning a probabilistic latent distributionq⁡\(z\|x\)q\(z\|x\)rather than a fixed encoding\. The decoder samples from this distribution to reconstruct the input, and anomalies are detected when the reconstruction likelihood or the latent deviation from the learned normal distribution is low\[[8](https://arxiv.org/html/2610.00363#bib.bib164),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.

Yuan et al\.\[[159](https://arxiv.org/html/2610.00363#bib.bib150)\]applied a Convolutional VAE \(CVAE\) to detect squat defects in rail surfaces from Axle Box Acceleration \(ABA\) signals, using an Elliptic Envelope detector on latent features to identify deviations from the normal distribution\. The model successfully identified defects as shallow as 0\.03 mm without any fault labels, demonstrating the sensitivity of probabilistic reconstruction to subtle infrastructure anomalies \(Table[11](https://arxiv.org/html/2610.00363#S5.T11)\)\.

Table 11:Standard and recurrent autoencoder approaches for railway anomaly detection\.Ref\.YearSubsystemArchitectureModalityAD Task\[[37](https://arxiv.org/html/2610.00363#bib.bib107)\]2021Air production unitSAEPressure \+ currentDetect anomalies in APU analog and digital sensor streams without fault labels\.\[[38](https://arxiv.org/html/2610.00363#bib.bib14)\]2024Air production unitLSTM\-AE vs\. SAEPressure \+ currentDetect APU anomalies using temporal reconstruction on MetroPT\-3 dataset\.\[[72](https://arxiv.org/html/2610.00363#bib.bib13)\]2021Braking systemOne\-class LSTM\-AECurrent signalsDetect anomalies in metro brake operating unit from current signal reconstruction\.\[[35](https://arxiv.org/html/2610.00363#bib.bib57)\]2024Rolling stockLSTM\-AEMultivariate signalsOnline reconstruction\-based anomaly detection in multivariate railway sensor streams\.\[[110](https://arxiv.org/html/2610.00363#bib.bib15)\]2024High\-speed railConvolutional AEVibration signalsDetect seismic anomalies from train vibration using normal\-only autoencoder reconstruction\.\[[159](https://arxiv.org/html/2610.00363#bib.bib150)\]2021Rail surfaceCVAEABA signalsDetect rail squat defects as shallow as 0\.03 mm from normal ABA signal reconstruction\.
#### 5\.3\.2Generative Adversarial Reconstruction Models

Generative Adversarial Networks \(GANs\) for anomaly detection train a generatorGGto reconstruct normal inputs and a discriminatorDDto distinguish real from generated samples\. Anomalies are flagged when reconstruction errors \(either pixel\-wise or perceptual\)exceed a threshold\. This family offers superior expressiveness for high\-dimensional visual data but requires careful training stabilization\[[40](https://arxiv.org/html/2610.00363#bib.bib162)\]\.

Wang et al\.\[[146](https://arxiv.org/html/2610.00363#bib.bib123)\]proposed a GAN\-based framework to detect foreign objects on railway tracks using only normal track images, reconstructing clean scenes and flagging deviations via error maps\. Wang et al\.\[[142](https://arxiv.org/html/2610.00363#bib.bib17)\]applied an adversarial memory\-enhanced reconstruction network to metro vehicle undercarriage images, combining a memory module with adversarial training to prevent the model from reconstructing anomalous patterns \(a known failure mode of standard GANs in reconstruction\)\[[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\.

For visual track inspection, transformer\-based reconstruction architectures have recently emerged as an alternative to convolutional autoencoders\. Wang et al\.\[[143](https://arxiv.org/html/2610.00363#bib.bib16)\]proposed MemFormer, a memory\-augmented Convolutional Neural Network \(CNN\) and Transformer model that reconstructs normal metro track images and flags anomalies via reconstruction error\. The memory module constrains the model to reconstruct only normal patterns, improving robustness to the anomaly reconstruction problem\. A similar approach applied to railway turnout inspection, Rail\-PatchCore\[[164](https://arxiv.org/html/2610.00363#bib.bib146)\], combines unsupervised feature\-level reconstruction with a dual\-dimensional channel attention module to reduce background noise interference in complex turnout environments\. Wang et al\.\[[147](https://arxiv.org/html/2610.00363#bib.bib1)\]introduced an anomaly\-free representation learning approach for railway foreign object detection that learns discriminative features exclusively from normal track images\. For pantograph inspection, an unsupervised reverse distillation network\[[155](https://arxiv.org/html/2610.00363#bib.bib143)\]reconstructs normal pantograph image features through an encoder\-decoder architecture trained on defect\-free images, achieving image\-level and pixel\-level reconstruction scores competitive with supervised baselines \(Table[12](https://arxiv.org/html/2610.00363#S5.T12)\)\.

Table 12:Generative adversarial and transformer reconstruction models for railway anomaly detection\. All methods are trained on normal data only\. GAN: Generative Adversarial Network; CNN: Convolutional Neural Network\.Ref\.YearSubsystemArchitectureModalityAD Task\[[146](https://arxiv.org/html/2610.00363#bib.bib123)\]2022Track obstaclesGAN \(adversarial AE\)RGB imagesDetect foreign objects on tracks by reconstructing normal track scenes from onboard cameras\.\[[142](https://arxiv.org/html/2610.00363#bib.bib17)\]2024Metro undercarriageGAN \+ memory moduleRGB imagesDetect undercarriage anomalies via adversarial memory\-enhanced image reconstruction\.\[[143](https://arxiv.org/html/2610.00363#bib.bib16)\]2024Metro tracksCNN \+ TransformerRGB imagesDetect track surface anomalies by reconstructing normal track images with memory augmentation\.\[[164](https://arxiv.org/html/2610.00363#bib.bib146)\]2025Railway turnoutPatchCore \+ attentionRGB imagesDetect visual anomalies in turnout environments using unsupervised feature reconstruction\.\[[155](https://arxiv.org/html/2610.00363#bib.bib143)\]2024PantographReverse distillation networkRGB imagesDetect pantograph anomalies via unsupervised reverse distillation image reconstruction\.Summary and limitations:Reconstruction\-based approaches offer label\-free anomaly detection across diverse railway modalities \(signals, images, and vibrations\)and generalize to fault types not seen during training\[[103](https://arxiv.org/html/2610.00363#bib.bib166),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. However, three limitations constrain their applicability in real railway environments\. First, standard autoencoders may reconstruct anomalous inputs with low error if the anomaly shares low\-level features with normal data \(a phenomenon known as the anomaly reconstruction problem\)\[[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\. Second, reconstruction thresholds are typically calibrated heuristically and may not generalize across different operational conditions or fleets\[[35](https://arxiv.org/html/2610.00363#bib.bib57)\]\. Third, generative models such as GANs require careful training stabilization and are sensitive to hyperparameter choices\[[40](https://arxiv.org/html/2610.00363#bib.bib162)\]\.

Cross\-category observations:Across the three groups summarized in Tables[11](https://arxiv.org/html/2610.00363#S5.T11)\-[12](https://arxiv.org/html/2610.00363#S5.T12), recurrent autoencoders\(particularly LSTM\-AE and Bi\-LSTM\-AE variants\)dominate the surveyed railway literature for signal\-based monitoring, reflecting their ability to capture temporal dependencies in multivariate sensor streams\[[72](https://arxiv.org/html/2610.00363#bib.bib13),[35](https://arxiv.org/html/2610.00363#bib.bib57)\]\. For visual inspection tasks, transformer and memory\-augmented reconstruction models are emerging as alternatives to convolutional autoencoders, demonstrating improved robustness against the anomaly reconstruction problem\[[143](https://arxiv.org/html/2610.00363#bib.bib16),[164](https://arxiv.org/html/2610.00363#bib.bib146)\]\. Probabilistic models such as VAEs remain less explored in the railway domain despite their demonstrated sensitivity to subtle structural anomalies\[[159](https://arxiv.org/html/2610.00363#bib.bib150)\]\. Finally, as with prediction\-based approaches, the majority of surveyed reconstruction studies are validated on single\-operator or proprietary datasets, limiting conclusions about cross\-fleet generalizability\[[37](https://arxiv.org/html/2610.00363#bib.bib107),[142](https://arxiv.org/html/2610.00363#bib.bib17)\]\.

### 5\.4Hybrid Approaches

None of the three paradigms reviewed in the preceding sections \- classification, reconstruction, and prediction \- is universally sufficient for railway anomaly detection\. Classification\-based models achieve high accuracy on known fault types but fail on unseen anomalies\. Reconstruction\-based models detect novel deviations without labels but produce elevated false alarm rates in noisy environments\. Prediction\-based models generalize well to temporal anomalies but require stable operating conditions\. Hybrid approaches address these limitations by combining two or more paradigms within a unified framework, exploiting their complementary strengths\[[161](https://arxiv.org/html/2610.00363#bib.bib61),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\.

As illustrated in Figure[7](https://arxiv.org/html/2610.00363#S5.F7), hybrid frameworks process the input through parallel or sequential branches \(each implementing a different detection principle\) and fuse their outputs into a unified anomaly score or decision\. Fusion can be rule\-based, score\-weighted, or learned end\-to\-end depending on the availability of labeled data and the complexity of the target scenario\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/hybrid_based.png)Figure 7:Typical pipeline for hybrid anomaly detection in railway systems\. Two or more detection paradigms process the input in parallel\. Their outputs are fused into a unified anomaly score for the final decision\.In the surveyed railway literature, hybrid approaches appear in two main forms\. The first form combines reconstruction and prediction within a single generative model, jointly learning to reconstruct the current state and predict future states, with anomalies scored as combined deviations from both objectives\. The second form integrates supervised detection with unsupervised reconstruction or rule\-based reasoning, enabling simultaneous identification of known and unknown anomalies within the same pipeline \(Table[13](https://arxiv.org/html/2610.00363#S5.T13)\)\.

Laurendin et al\.\[[80](https://arxiv.org/html/2610.00363#bib.bib149)\]proposed a hybrid framework for detecting hazardous events near train doors, combining reconstruction and prediction within a conditional Generative Adversarial Network \(cGAN\) with a U\-Net generator\. The model jointly reconstructs the current video frame and predicts optical flow, scoring anomalies as deviations in both image appearance and motion patterns\. This dual\-objective design enables detection of obstructed doors, irregular passenger movements, and mechanical faults that would be missed by either objective alone\.

Gao et al\.\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\]presented a multimodal hybrid system for pantograph\-catenary inspection that integrates classification, reconstruction, and prediction across four sensing modalities: Red Green Blue \(RGB\) imaging, infrared thermography, ultraviolet discharge detection, and vibration signals\. Each modality is processed by a specialized module \(object detectors for known defect localization, autoencoders for signal deviation detection, and time\-series models for behavioral forecasting\) with outputs interpreted jointly\. This modular architecture demonstrates that combining complementary detection principles and sensing modalities can achieve broader anomaly coverage than any single approach\.

Chen et al\.\[[28](https://arxiv.org/html/2610.00363#bib.bib18)\]proposed a meta\-learning framework combining Generative Adversarial Networks with structural similarity\-based loss functions for anomaly detection in high\-speed rail inspection\. The framework learns to detect anomalies from limited labeled examples by combining generative reconstruction with a supervised meta\-learner, reducing inspection workload by over 99\.7% when deployed on five high\-speed railway lines in China\.

Sun et al\.\[[132](https://arxiv.org/html/2610.00363#bib.bib144)\]introduced RailFDNet, a hybrid supervision model for railway anomalous object detection that combines an unsupervised normalizing flow\-based backbone with a supervised transformer\-based anomaly decoder\. The unsupervised component enhances the feature discrepancy between normal and abnormal data, while the supervised decoder uses artificially generated anomaly images to improve precise localization\. This dual\-stage training strategy enables the model to handle both the absence of real anomaly labels and the need for accurate spatial localization in open railway environments\.

Wu et al\.\[[152](https://arxiv.org/html/2610.00363#bib.bib35)\]proposed a hybrid deep learning framework for automatic railroad track component inspection that combines convolutional feature extraction with rule\-based geometric reasoning\. The supervised classification component identifies track component types, while the rule\-based module applies domain\-specific geometric constraints to flag anomalous spatial configurations that the neural network alone cannot capture\.

Table 13:Hybrid approaches for railway anomaly detection\. IR: Infrared; UV: Ultraviolet\.Ref\.YearSubsystemArchitectureModalityAD Task\[[80](https://arxiv.org/html/2610.00363#bib.bib149)\]2021Train doorsConditional GAN \+ U\-Net \(reconstruction \+ prediction\)VideoDetect hazardous door events by jointly reconstructing frames and predicting optical flow\.\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\]2020Pantograph\-catenaryYOLO \+ AE \+ LSTM \(classification \+ reconstruction \+ prediction\)RGB \+ IR \+ UV \+ vibrationDetect pantograph anomalies across four modalities using parallel specialized detection modules\.\[[28](https://arxiv.org/html/2610.00363#bib.bib18)\]2022Track infrastructureGAN \+ meta\-learner \(reconstruction \+ supervised\)RGB imagesDetect track anomalies from limited labeled examples using generative reconstruction and meta\-learning\.\[[132](https://arxiv.org/html/2610.00363#bib.bib144)\]2025Track obstaclesNormalizing flow \+ Transformer \(unsupervised \+ supervised\)RGB imagesDetect and localize anomalous objects using hybrid unsupervised backbone and supervised decoder\.\[[152](https://arxiv.org/html/2610.00363#bib.bib35)\]2023Track componentsCNN \+ rule\-based module \(classification \+ rules\)RGB imagesInspect track components by combining learned classification with geometric domain rules\.Summary and limitations:Hybrid approaches offer broader anomaly coverage, greater robustness to label scarcity, and improved generalization compared to single\-paradigm methods\[[103](https://arxiv.org/html/2610.00363#bib.bib166),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\. They are particularly suited to safety\-critical railway scenarios where both known and unknown fault types must be handled simultaneously\. However, three limitations constrain their deployment\. First, multi\-branch architectures are more complex to design, train, and calibrate than single\-paradigm models, particularly under real\-time inference constraints\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\]\. Second, fusion strategies \(whether rule\-based or learned\) require careful design to avoid one branch dominating the anomaly score\[[80](https://arxiv.org/html/2610.00363#bib.bib149)\]\. Third, hybrid models that process multiple modalities simultaneously impose higher computational and memory costs, which may limit their applicability on resource\-constrained embedded platforms\[[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\.

General Interpretation and Comparative Analysis:

The surveyed approaches reveal complementary strengths across four paradigms \(classification, prediction, reconstruction, and hybrid\) each suited to specific railway contexts\.

Classification\-based methods excel when labeled datasets are available and fault types are well\-defined\. CNN\-based models achieve high accuracy in visual inspection tasks such as fastener detection\[[52](https://arxiv.org/html/2610.00363#bib.bib88)\]and pantograph monitoring\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\], but their reliance on supervised learning limits adaptability to unseen anomalies\.

Prediction\-based methods \(e\.g\., LSTM, GRU\) are effective for continuous monitoring of multivariate time series, enabling early fault detection without anomaly labels\[[149](https://arxiv.org/html/2610.00363#bib.bib102)\]\. They capture temporal dynamics but require retraining under non\-stationary conditions\.

Reconstruction\-based methods \(AEs, VAEs, GANs\) address scenarios with scarce or unlabeled fault data by modeling normal behavior and detecting deviations\[[37](https://arxiv.org/html/2610.00363#bib.bib107),[159](https://arxiv.org/html/2610.00363#bib.bib150),[146](https://arxiv.org/html/2610.00363#bib.bib123)\]\. While robust for rare anomalies, defining interpretable thresholds remains challenging\.

Hybrid approaches combine multiple paradigms for complex, multimodal environments\. Examples include appearance\-motion fusion for door hazard detection\[[80](https://arxiv.org/html/2610.00363#bib.bib149)\]and integrated classification\-prediction pipelines for pantograph monitoring\[[51](https://arxiv.org/html/2610.00363#bib.bib157)\], offering flexibility at the cost of higher computational complexity\.

In practice, deployment strategies often mix paradigms: classification for frequent visual defects, prediction for real\-time sensor forecasting, reconstruction for rare failures, and hybrid models for safety\-critical multimodal monitoring\.

## 6Performance Evaluation for Railway Anomaly Detection

Selecting the right evaluation metric is as important as selecting the right model\. In railway anomaly detection, faults are rare, error costs are asymmetric, and models must often run on embedded hardware under strict latency constraints\. Standard metrics designed for balanced classification tasks frequently misrepresent performance in this context\. This section reviews which metrics are appropriate for each detection task, identifies the limitations of common reporting practices observed across the 68 primary studies of this survey, and proposes concrete evaluation criteria better aligned with railway safety and deployment requirements\[[103](https://arxiv.org/html/2610.00363#bib.bib166),[22](https://arxiv.org/html/2610.00363#bib.bib67)\]\.

### 6\.1Metrics by Detection Task

The appropriate metric depends on the model output and the detection objective\. Table[14](https://arxiv.org/html/2610.00363#S6.T14)maps each metric to its detection task, explains what it measures, and highlights when it is appropriate or insufficient in railway contexts\.

Table 14:Evaluation metrics for railway anomaly detection\.
### 6\.2Why Standard Metrics Are Insufficient for Railway Systems

Three structural properties of railway anomaly detection make standard accuracy\-oriented metrics unreliable\.

- •Class imbalance:Operational fault rates in railway systems are typically below 1%\. Under such imbalance, accuracy and AUROC can both be misleading: a model that never detects any fault can report 99% accuracy and a non\-trivial AUROC\[[127](https://arxiv.org/html/2610.00363#bib.bib163),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. PR\-AUC is more informative because it focuses exclusively on the anomaly class and penalizes models that miss rare faults\[[92](https://arxiv.org/html/2610.00363#bib.bib159)\]\.
- •Asymmetric error costs:False negatives and false positives carry fundamentally different consequences\. Missing a brake fault \(FN\) may cause a safety incident costing EUR 100,000 or more, while a false alarm \(FP\) typically triggers an unnecessary inspection costing a few hundred euros\. Standard metrics treat both errors as equally costly, which is inappropriate for safety\-critical monitoring\[[115](https://arxiv.org/html/2610.00363#bib.bib46),[91](https://arxiv.org/html/2610.00363#bib.bib52)\]\. Cost\-sensitive evaluation assigns asymmetric penaltiescF​N≫cF​Pc\_\{FN\}\\gg c\_\{FP\}and reports expected operational cost: 𝒞¯=1N​\(cT​P​T​P\+cF​P​F​P\+cF​N​F​N\+cT​N​T​N\)\\bar\{\\mathcal\{C\}\}=\\frac\{1\}\{N\}\\bigl\(c\_\{TP\}\\,TP\+c\_\{FP\}\\,FP\+c\_\{FN\}\\,FN\+c\_\{TN\}\\,TN\\bigr\)This formulation makes model selection directly accountable to operational and safety objectives\[[91](https://arxiv.org/html/2610.00363#bib.bib52),[12](https://arxiv.org/html/2610.00363#bib.bib50)\]\.
- •Deployment gap:High offline accuracy does not guarantee deployability\. A model achieving 99% F1\-score on a GPU server may fail to meet a 100 ms latency budget on an embedded platform\. For axle temperature monitored at 10 Hz, a 2 s detection latency implies a delay of 20 samples \(even 98% AUROC cannot compensate for this operational gap\)\[[7](https://arxiv.org/html/2610.00363#bib.bib68),[139](https://arxiv.org/html/2610.00363#bib.bib94)\]\. Inference latency, memory footprint, and throughput must be reported alongside accuracy metrics for any deployment\-oriented study\.

### 6\.3Evaluation Practices in the Surveyed Literature

Table[15](https://arxiv.org/html/2610.00363#S6.T15)summarizes how evaluation metrics are used across the 68 primary studies included in this survey\.

Table 15:Metric adoption across the 68 surveyed studies\. Inference latency is counted when at least one quantitative inference\-time indicator is explicitly reported \(execution time, frames per second, or end\-to\-end latency\)\.Three key observations emerge from this analysis\. First, accuracy remains the most frequently reported metric, despite being poorly suited to imbalanced railway fault detection scenarios\. Second, PR\-AUC, which is more appropriate for rare\-event detection, is reported in fewer than 8% of studies\. Third, deployment\-related metrics are largely underreported: inference latency is reported in only 7 studies, and memory or energy footprint in only 5 studies\. This confirms that practical deployment considerations remain secondary in the current literature\[[7](https://arxiv.org/html/2610.00363#bib.bib68)\]\.

Minimum reporting standard:Based on these findings, future studies should report at minimum: \(i\) F1\-score and PR\-AUC for imbalanced detection; \(ii\) false negative rate for safety\-critical scenarios; \(iii\) inference latency measured on the target platform; and \(iv\) cost\-sensitive evaluation when asymmetric error costs can be estimated\. These criteria are consistent with the deployment\-oriented decision framework presented in Section[7\.6](https://arxiv.org/html/2610.00363#S7.SS6)\.

## 7Deployment of Deep Learning Models in Railway Systems

Deploying deep learning models in railway systems requires balancing predictive performance with strict operational constraints, including latency, memory footprint, energy consumption, and safety compliance\[[75](https://arxiv.org/html/2610.00363#bib.bib49)\]\. Railway systems operate on resource\-constrained embedded platforms under harsh conditions \(vibration, thermal stress, electromagnetic interference\) where long\-term reliability is mandatory\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\]\. As a result, model design cannot be decoupled from hardware and deployment constraints\.

Railway applications must comply with safety standards such as EN 50155 \(environmental requirements for onboard electronics\) and EN 50128 \(software safety integrity levels SIL1–SIL4\)\[[44](https://arxiv.org/html/2610.00363#bib.bib29)\]\. However, no standardized certification pathway currently exists for deep learning models, due to their non\-deterministic behavior and limited interpretability\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\]\. This lack of certification remains a major barrier to deploying neural networks in safety\-critical railway subsystems\.

Despite the importance of these constraints, deployment remains the least documented aspect of the railway anomaly detection literature\. While recent work has begun to address this gap for specific subsystems such as track defect detection\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\], no prior survey provides a deployment\-oriented analysis spanning infrastructure, rolling stock, and environmental anomaly detection across heterogeneous hardware platforms\. Among the studies surveyed in this work, few explicitly report at least one quantitative deployment metric such as inference latency, power consumption, or accuracy degradation under model compression\. The following analysis combines this limited railway\-specific evidence with established practices from the broader embedded deep learning literature, adapted to railway constraints\. Each subsection distinguishes between findings from surveyed railway studies, requirements from railway engineering standards, and guidance from the general embedded AI literature\.

### 7\.1Deployment Evidence in the Surveyed Literature

To assess the current state of deployment in railway anomaly detection, we examined all surveyed studies for quantitative deployment metrics\. Table[16](https://arxiv.org/html/2610.00363#S7.T16)summarizes the studies that report at least one hardware\-level measurement\. The remaining studies evaluate their models exclusively on desktop or server\-grade GPUs without reporting inference latency on a target embedded platform, power consumption, or accuracy impact of model compression\.

Table 16:Deployment metrics reported across the surveyed studies\. Memory values refer to platform RAM capacity \(GB\) or FPGA on\-chip resource usage \(BRAM\), depending on reporting conventions in each study\. NR: not reported\.Ref\.YearSub\-systemPlatformClassLatency \(ms\)Power \(W\)AccuracyMemory\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]2025Track faults \(vision\)Zynq 7Z020 / KV260FPGA0\.7–102\.3–3\.993\.4%4\.5–11\.6 Mb BRAM\[[87](https://arxiv.org/html/2610.00363#bib.bib89)\]2020Traction faultsXilinx VCU128FPGA<<1NR\>\>93%FPGA resources\[[82](https://arxiv.org/html/2610.00363#bib.bib24)\]2024Track faults \(vision\)Xilinx ZCU104FPGA2\.16\.988\.9%230\.5 BRAM\[[36](https://arxiv.org/html/2610.00363#bib.bib156)\]2020Fastener \(3D\)Jetson TX2 \+ AtomGPU \+ CPU∼\\sim15NR99\.7%8 GB\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]2024In\-cabin eventsJetson AGX XavierGPU edge∼\\sim120010–3085\.1%32 GB\[[42](https://arxiv.org/html/2610.00363#bib.bib25)\]2025Passenger countingJetson NanoGPU edge80–100∼\\sim5–1096\.85%4 GB\[[25](https://arxiv.org/html/2610.00363#bib.bib31)\]2025Track faults \(IoT\)Jetson Xavier \+ ESP32GPU \+ IoT105–160NR90–96%∼\\sim6 GBFour observations emerge from this table\.

Platform diversity is narrow:Three studies deploy on Field\-Programmable Gate Array \(FPGA\) platforms\[[48](https://arxiv.org/html/2610.00363#bib.bib23),[87](https://arxiv.org/html/2610.00363#bib.bib89),[82](https://arxiv.org/html/2610.00363#bib.bib24)\], three use NVIDIA Jetson\-family GPU edge devices\[[36](https://arxiv.org/html/2610.00363#bib.bib156),[139](https://arxiv.org/html/2610.00363#bib.bib94),[42](https://arxiv.org/html/2610.00363#bib.bib25)\], and one combines a Jetson platform with IoT modules\[[25](https://arxiv.org/html/2610.00363#bib.bib31)\]\.

Latency spans three orders of magnitude:Inference times range from 0\.7 ms for binary classification on an FPGA\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]to approximately 1200 ms for a full multimodal pipeline on a Jetson AGX Xavier\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]\. This range reflects differences in task complexity and input resolution rather than platform capability alone, which underlines the need for standardized deployment benchmarks in the railway domain\.

Power consumption is rarely measured:Only three studies report power measurements\[[48](https://arxiv.org/html/2610.00363#bib.bib23),[82](https://arxiv.org/html/2610.00363#bib.bib24),[139](https://arxiv.org/html/2610.00363#bib.bib94)\]\. FPGA\-based systems operate at 2\.3–6\.9 W\[[48](https://arxiv.org/html/2610.00363#bib.bib23),[82](https://arxiv.org/html/2610.00363#bib.bib24)\], an order of magnitude below the 10–30 W reported for GPU edge platforms\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]\. Fu et al\.\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]introduce a cost\-energy efficiency metric \(CEES = throughput / power×\\timescost\), demonstrating 101\.9×\\timesand 33\.8×\\timesimprovements over CPU and GPU platforms respectively\. The remaining studies do not report power consumption, making it impossible to assess their suitability for onboard deployment where power budgets are typically constrained to 500 W or less for the full monitoring system\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\]\.

Model optimization is inconsistent:Model compression techniques \(reviewed in detail in Section[7\.5](https://arxiv.org/html/2610.00363#S7.SS5)\) are applied unevenly\. Fu et al\.\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]provide the most complete pipeline, combining extreme quantization with custom hardware resource balancing\. Li et al\.\[[82](https://arxiv.org/html/2610.00363#bib.bib24)\]apply reduced\-precision arithmetic with no accuracy loss\. Tsiktsiris et al\.\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]use hardware\-accelerated inference optimization with approximately 0\.5% accuracy drop\. The remaining studies apply minimal or no optimization\.

This analysis confirms that real\-time inference on embedded platforms is technically feasible for specific railway tasks\. However, the evidence base remains too sparse to support general conclusions about deployability across subsystems or operational conditions\. Establishing deployment reporting as a minimum standard \(including inference latency, power consumption, and accuracy impact of compression\) would significantly improve the comparability of future studies\.

### 7\.2Hardware Platforms for Railway Deployment

Based on the surveyed literature and the broader embedded AI domain, embedded platforms for railway anomaly detection fall into five classes\.

- •General\-purpose CPU/GPU systemsare primarily used for training and offline evaluation\. They provide high computational throughput and large memory capacity\. However, their high power consumption, non\-deterministic latency, and large physical footprint make them unsuitable for onboard railway deployment\[[52](https://arxiv.org/html/2610.00363#bib.bib88),[73](https://arxiv.org/html/2610.00363#bib.bib95),[34](https://arxiv.org/html/2610.00363#bib.bib145)\]\.
- •GPU\-accelerated edge platforms, such as the NVIDIA Jetson family \(TX2, Xavier NX, Orin\), offer a practical compromise between performance and deployability\[[139](https://arxiv.org/html/2610.00363#bib.bib94),[36](https://arxiv.org/html/2610.00363#bib.bib156)\]\. Their software ecosystem supports rapid deployment, although latency can vary under concurrent workloads\[[34](https://arxiv.org/html/2610.00363#bib.bib145)\]\.
- •Field\-Programmable Gate Array \(FPGA\)\-based systemsprovide deterministic execution and ultra\-low latency, making them well\-suited for safety\-critical subsystems such as braking monitoring\[[87](https://arxiv.org/html/2610.00363#bib.bib89),[97](https://arxiv.org/html/2610.00363#bib.bib51),[82](https://arxiv.org/html/2610.00363#bib.bib24)\]\. Their energy efficiency is high due to custom dataflow and fixed\-point computation, at the cost of high development complexity and limited flexibility\.
- •Embedded Central Processing Unit platforms, including ARM Cortex\-A and Intel Atom processors, are used for preprocessing, lightweight inference, and Train Control and Monitoring System \(TCMS\) integration\. Their limited parallelism restricts them to lightweight models or low\-frequency monitoring tasks\[[122](https://arxiv.org/html/2610.00363#bib.bib92),[149](https://arxiv.org/html/2610.00363#bib.bib102),[7](https://arxiv.org/html/2610.00363#bib.bib68)\]\.
- •Neural Processing Units \(NPUs\) and AI accelerators\(e\.g\., ARM Ethos, Hailo\-8\) are purpose\-built for neural network inference and achieve high throughput at low power consumption through architectures optimized for matrix operations and quantized computation\[[34](https://arxiv.org/html/2610.00363#bib.bib145)\]\. Although currently underrepresented in the railway anomaly detection literature, NPUs represent a strong candidate for future onboard deployment due to their energy efficiency and compact form factor\.

Table[17](https://arxiv.org/html/2610.00363#S7.T17)summarizes their trade\-offs and typical use cases\.

Table 17:Hardware platforms for deep learning deployment in railway anomaly detection systems\. TheRailway evidencecolumn indicates whether the platform has been validated in a surveyed railway study \(Yes\) or is projected from the general embedded AI literature \(Projected\)\.PlatformDevicesWhen to ChooseKey LimitationRailway evidenceCPU/GPU serverIntel Xeon, RTXOffline training with high compute requirements, no onboard constraints\[[52](https://arxiv.org/html/2610.00363#bib.bib88),[73](https://arxiv.org/html/2610.00363#bib.bib95)\]\.High power \(\>\>200 W\); non\-deterministic latency incompatible with real\-time safety\-critical systems\[[7](https://arxiv.org/html/2610.00363#bib.bib68),[34](https://arxiv.org/html/2610.00363#bib.bib145)\]\.Yes \(training only\)GPU edgeJetson TX2, Xavier, OrinReal\-time vision and multimodal tasks under moderate power constraints\[[139](https://arxiv.org/html/2610.00363#bib.bib94),[36](https://arxiv.org/html/2610.00363#bib.bib156)\]\.Latency variability under concurrent workloads; limited for real\-time constraints\[[34](https://arxiv.org/html/2610.00363#bib.bib145)\]\.YesFPGAXilinx Zynq, ZCU, VCUSafety\-critical subsystems requiring deterministic and low latency\[[87](https://arxiv.org/html/2610.00363#bib.bib89),[97](https://arxiv.org/html/2610.00363#bib.bib51)\]\.High development complexity; long design cycles; limited flexibility for model updates\[[97](https://arxiv.org/html/2610.00363#bib.bib51)\]\.YesEmbedded CPUARM Cortex\-A, Intel AtomPreprocessing, lightweight inference\[[122](https://arxiv.org/html/2610.00363#bib.bib92),[149](https://arxiv.org/html/2610.00363#bib.bib102)\]\.Low parallelism; unsuitable for complex deep learning models or high\-frequency inference\[[7](https://arxiv.org/html/2610.00363#bib.bib68)\]\.PartialNPU / AI accel\.Hailo\-8, ARM EthosLow\-power onboard inference with dedicated neural acceleration; strong candidate for future embedded deployment\[[34](https://arxiv.org/html/2610.00363#bib.bib145)\]\.Limited validation in railway applications; integration and certification challenges remain open\.Projected
### 7\.3Deployment Architectures: Edge, Cloud, and Hybrid

Deployment architecture directly impacts end\-to\-end latency, availability under degraded connectivity, and integration with maintenance workflows\[[56](https://arxiv.org/html/2610.00363#bib.bib6),[3](https://arxiv.org/html/2610.00363#bib.bib45),[1](https://arxiv.org/html/2610.00363#bib.bib44)\]\.

- •Edge deploymentexecutes inference close to sensors and is suited to anomalies requiring predictable response times or local autonomy\. European railway architectures emphasize edge and wayside processing for continuous detection\[[56](https://arxiv.org/html/2610.00363#bib.bib6)\]\. Edge offers low latency and resilience to connectivity outages, but is constrained by compute, memory, and power budgets\. It is best suited for safety\-critical events such as track intrusion and door obstruction\.
- •Cloud deploymentcentralizes storage and compute for fleet\-level analytics, long\-horizon trend analysis, and model retraining\[[1](https://arxiv.org/html/2610.00363#bib.bib44)\]\. It provides elastic compute and cross\-fleet aggregation, but network\-induced latency variability makes it unsuitable for real\-time detection\. It is best suited for slow degradation analytics and retrospective root\-cause analysis\.
- •Hybrid edge–cloud deploymentcombines fast edge screening with centralized refinement: edge nodes triage events and forward only informative content, while centralized services correlate anomalies across assets and update models\[[3](https://arxiv.org/html/2610.00363#bib.bib45)\]\. This architecture balances responsiveness and scalability at the cost of higher orchestration complexity\. It is best suited for multimodal, network\-wide deployments requiring both immediate alerts and fleet\-level context\.

In practice, deployment selection is driven by latency determinism and connectivity constraints: safety\-critical tasks with strict deadlines require edge deployment, while non\-critical analytics can tolerate cloud\-based processing\. Hybrid architectures are increasingly necessary for multimodal railway monitoring systems that must simultaneously provide immediate local alerts and support fleet\-level decision making\.

### 7\.4Real\-Time and Non\-Real\-Time Anomalies

Railway anomalies differ in urgency, which directly influences hardware selection and algorithm design\. In this context, real\-time refers to predictable \(deterministic\) response time rather than simply low average latency\[[75](https://arxiv.org/html/2610.00363#bib.bib49)\]\.

- •Real\-time anomaliesare safety\-critical events \(sudden motor failures, braking faults, track intrusions\) requiring bounded inference latency\. These are further divided intohard real\-time\(e\.g\., braking protection, where deadline misses are unacceptable\) andsoft real\-time\(e\.g\., door monitoring, where occasional misses degrade performance without immediate safety impact\)\. Hard real\-time constraints favor deterministic platforms such as FPGAs\[[87](https://arxiv.org/html/2610.00363#bib.bib89),[48](https://arxiv.org/html/2610.00363#bib.bib23)\], while soft real\-time tasks are typically handled by GPU edge devices\[[139](https://arxiv.org/html/2610.00363#bib.bib94),[42](https://arxiv.org/html/2610.00363#bib.bib25)\]\. For video\-based inspection at operational speeds below 160 km/h, a frame rate of approximately 30 fps at 1504\-pixel resolution is required for continuous coverage\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\], corresponding to a per\-frame budget of approximately 33 ms\. For 60 fps video streams, this budget tightens to 16\.7 ms\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]\.
- •Non\-real\-time anomaliesare gradual degradations \(HVAC inefficiency, bearing wear, insulation aging\) that allow offline or batch analysis integrated into condition\-based maintenance\[[37](https://arxiv.org/html/2610.00363#bib.bib107),[49](https://arxiv.org/html/2610.00363#bib.bib84)\]\. These tasks prioritize robustness and interpretability over latency, and can be deployed on cloud or server platforms\.

This distinction is operationally important because it directly constrains feasible model architectures, optimization strategies, and hardware platforms, as formalized in the decision framework of Section[7\.6](https://arxiv.org/html/2610.00363#S7.SS6)\.

### 7\.5Model Optimization for Embedded Railway Deployment

Unlike general embedded AI applications, railway deployment imposes additional constraints on determinism, long\-term reliability, and safety certification that limit which optimization strategies are applicable in practice\[[106](https://arxiv.org/html/2610.00363#bib.bib30),[75](https://arxiv.org/html/2610.00363#bib.bib49)\]\. Onboard and wayside platforms further constrain memory, power, and thermal budgets, limiting the deployability of large deep learning models\[[7](https://arxiv.org/html/2610.00363#bib.bib68),[83](https://arxiv.org/html/2610.00363#bib.bib41),[32](https://arxiv.org/html/2610.00363#bib.bib19)\]\. Six optimization strategies are identified in the surveyed literature, grouped into two categories: techniques that primarily reduce latency, and techniques that primarily reduce model size\. Figure[8](https://arxiv.org/html/2610.00363#S7.F8)illustrates the compression workflow\.

#### 7\.5\.1Latency\-oriented techniques

- •Lightweight architecturessuch as MobileNetV3, SqueezeNet, and Efficient Net\-Lite use depthwise separable convolutions to reduce computational cost on edge hardware\[[7](https://arxiv.org/html/2610.00363#bib.bib68)\]\. In the railway domain, MobileNetV3 combined with Transformer encoders achieved latency below 100 ms for foreign object detection\[[98](https://arxiv.org/html/2610.00363#bib.bib96)\]\. Radosavljevic et al\.\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\]systematically evaluated six YOLO variants \(v5 through v12\) on two Jetson platforms, finding that YOLOv5nu consistently achieved the best accuracy–latency trade\-off across both FP32 and FP16 precision levels, with mean latency of 20–50 ms on Jetson AGX Xavier depending on dataset complexity\.
- •Input reductionlowers computational load through frame downsampling or dimensionality reduction\. Laurendin et al\.\[[80](https://arxiv.org/html/2610.00363#bib.bib149)\]applied optical flow on reduced frame windows for efficient real\-time door surveillance\.
- •Hardware\-accelerated inferenceuses platform\-specific runtimes \(TensorRT, ONNX Runtime, TFLite\) to optimize execution\. Tsiktsiris et al\.\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]achieved 2–3×\\timesspeedup using INT8 quantization with TensorRT on Jetson platforms\. Radosavljevic et al\.\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\]confirmed that TensorRT export provides the best inference performance across all tested YOLO variants on both Jetson AGX Xavier and Orin Nano, with FP16 inference reducing latency by 40–50% compared to FP32 with negligible accuracy loss \(<<0\.3% mAP drop in most cases\)\.
- •Parallel pipelinesenable asynchronous processing across CPU threads and GPU streams, reducing end\-to\-end latency in multimodal systems\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]\.

#### 7\.5\.2Size\-oriented techniques

- •Quantizationreduces numerical precision from FP32 to FP16, INT8, or lower\. It is the most commonly applied compression technique in railway deployments\. Fu et al\.\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]demonstrate the most extreme case: binary neural networks \(1\-bit weights and activations\) deployed on FPGA, achieving 93\.4% accuracy with only 0\.3% degradation from full precision and power consumption of 2\.3–3\.9 W\. Li et al\.\[[82](https://arxiv.org/html/2610.00363#bib.bib24)\]apply mixed fixed\-point quantization \(12\-bit and 22\-bit\) on FPGA with no accuracy loss\. Tsiktsiris et al\.\[[139](https://arxiv.org/html/2610.00363#bib.bib94)\]report INT8 quantization on GPU edge with∼\\sim0\.5% accuracy drop\. These results confirm that quantization is effective for railway deployment, with accuracy degradation typically below 1% when calibration is performed carefully\[[69](https://arxiv.org/html/2610.00363#bib.bib33)\]\.
- •Pruningremoves redundant weights or filters to reduce model complexity\[[32](https://arxiv.org/html/2610.00363#bib.bib19)\]\. Fu et al\.\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]apply entropy\-based pruning to reduce network depth before FPGA deployment\. Beyond this, explicit pruning strategies are rare in the railway literature; most works rely on lightweight architectures rather than formal pruning pipelines\[[150](https://arxiv.org/html/2610.00363#bib.bib39)\]\.
- •Knowledge distillationtransfers representations from a large teacher model to a compact student model\[[65](https://arxiv.org/html/2610.00363#bib.bib34)\]\. This technique remains essentially absent from railway anomaly detection, though it has shown promise for fault diagnosis under hardware constraints in adjacent domains\[[55](https://arxiv.org/html/2610.00363#bib.bib43)\]\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/compression_workflow.png)Figure 8:Model compression workflow: quantization, pruning, and knowledge distillation transform a baseline model into a compact version for embedded railway deployment\.
#### 7\.5\.3Summary

Existing railway deployments primarily exploit quantization and lightweight architecture selection\. Pruning is applied in one study\[[48](https://arxiv.org/html/2610.00363#bib.bib23)\]\. Knowledge distillation and early\-exit strategies\[[124](https://arxiv.org/html/2610.00363#bib.bib47),[15](https://arxiv.org/html/2610.00363#bib.bib40)\]remain unexplored in this domain\. No study applies a systematic end\-to\-end compression pipeline combining multiple techniques\. This pattern suggests that most railway anomaly detection models are optimized for accuracy first and adapted for deployment post hoc, rather than designed with hardware constraints as a primary objective\. Establishing systematic compression pipelines tailored to railway\-specific constraints \(deterministic latency, memory limits, and safety certification\) remains an open challenge\. More advanced approaches such as hardware\-aware neural architecture search and hardware–software co\-design, which integrate deployment constraints directly into model design, are discussed as future research directions in Section[8](https://arxiv.org/html/2610.00363#S8)\.

### 7\.6Deployment\-Driven Decision Framework

The preceding subsections established three complementary views: \(i\) the empirical deployment evidence from the surveyed literature \(Section[7\.1](https://arxiv.org/html/2610.00363#S7.SS1)\), \(ii\) the hardware platforms and deployment architectures available for railway systems \(Sections[7\.2](https://arxiv.org/html/2610.00363#S7.SS2)–[7\.3](https://arxiv.org/html/2610.00363#S7.SS3)\), and \(iii\) the model optimization techniques applied in practice \(Section[7\.5](https://arxiv.org/html/2610.00363#S7.SS5)\)\. Combined with the four\-axis anomaly taxonomy \(Section[3](https://arxiv.org/html/2610.00363#S3)\) and the detection paradigms \(Section[5](https://arxiv.org/html/2610.00363#S5)\), these elements provide the building blocks for systematic deployment planning\. However, they do not directly answer the practitioner’s question:which detection paradigm, model family, and deployment configuration should be selected for a given railway scenario?

This subsection introduces a deployment\-driven decision framework that maps \(A\) anomaly and data characteristics and \(B\) operational constraints to \(C\) a justified selection of detection paradigm, hardware platform, optimization strategy, and evaluation protocol\. The framework is operationalized in Figure[9](https://arxiv.org/html/2610.00363#S7.F9), which summarizes the complete decision path from anomaly characterization to deployment validation\. The framework synthesizes the surveyed evidence with railway engineering requirements and is intended as design guidance rather than a replacement for domain\-specific risk analysis or safety certification procedures such as EN 50128\[[75](https://arxiv.org/html/2610.00363#bib.bib49)\]\.

![Refer to caption](https://arxiv.org/html/2610.00363v1/img/decision_framework.png)Figure 9:Deployment decision framework for deep learning\-based anomaly detection in railway systems\.Scope:The framework targets safety\-relevant railway monitoring applications, including infrastructure inspection, onboard rolling\-stock subsystems, and environmental risk scenarios\.

#### 7\.6\.1Stepwise Selection Procedure

For a given railway anomaly detection scenario, the following seven steps should be applied sequentially\.

1. 1\.Characterize the anomaly:Describe the target anomaly using the four\-axis taxonomy introduced in Section[3](https://arxiv.org/html/2610.00363#S3): location \(infrastructure, rolling stock, environmental\), manifestation \(point, contextual, collective, trend\), sensing modality \(visual, vibration, acoustic, multivariate\), and temporal behavior \(static, sequential, long\-horizon\)\. This step determines whether spatial localization, temporal modeling, or multimodal fusion is required\.
2. 2\.Assess label availability and data regime:Determine whether annotated fault examples are available at scale, partially available, or absent\. In railway systems, fault rarity, data confidentiality, and class imbalance frequently limit supervision\[[22](https://arxiv.org/html/2610.00363#bib.bib67),[103](https://arxiv.org/html/2610.00363#bib.bib166)\]\. Normal\-only learning, reconstruction, and prediction strategies are often more realistic than purely supervised classification\.
3. 3\.Specify deployment constraints:Define the operational constraints that affect both model and platform selection: - •Latency:hard real\-time \(<<10 ms, e\.g\., braking protection\), soft real\-time \(<<100 ms, e\.g\., door monitoring\), or offline/batch\. - •Compute and memory budget:tight embedded \(<<8 W,<<1 GB\), moderate edge \(10–30 W, 4–32 GB\), or unconstrained\. - •Connectivity:offline\-capable, connected, or hybrid edge\-cloud\. - •Safety requirement:safety\-critical \(SIL2\+\) or condition\-monitoring\.
4. 4\.Select the detection paradigm:Based on Steps 1–3, the following rules guide paradigm selection: - •Labels availableand stable recurring fault types→\\rightarrowclassification\-based\. - •No labelsand temporal or gradual degradation→\\rightarrowprediction\-basedorreconstruction\-based\. - •No labelsand rare, unknown, or weakly characterized faults→\\rightarrowreconstruction\-based\. - •Multimodal sensingor safety redundancy required→\\rightarrowhybrid\. Table[18](https://arxiv.org/html/2610.00363#S7.T18)provides an analytical comparison supporting these rules\.
5. 5\.Select the hardware platform:Platform choice is driven by the latency and power constraints specified in Step 3: - •Hard real\-time and deterministic latency→\\rightarrowFPGA\[[87](https://arxiv.org/html/2610.00363#bib.bib89),[48](https://arxiv.org/html/2610.00363#bib.bib23),[82](https://arxiv.org/html/2610.00363#bib.bib24)\]\. - •Soft real\-time vision or multimodal inference→\\rightarrowGPU edge\(e\.g\., Jetson TX2, Xavier, Orin\)\[[139](https://arxiv.org/html/2610.00363#bib.bib94),[36](https://arxiv.org/html/2610.00363#bib.bib156)\]\. - •Low\-power onboard inference→\\rightarrowembedded CPUorNPU\. - •Offline training or fleet\-level analytics→\\rightarrowCPU/GPU server\.
6. 6\.Optimize the model for the target platform:The selected model must be adapted to the target hardware budget: - •If latency is limiting→\\rightarrowlightweight architectures or hardware\-accelerated inference\[[139](https://arxiv.org/html/2610.00363#bib.bib94),[98](https://arxiv.org/html/2610.00363#bib.bib96),[106](https://arxiv.org/html/2610.00363#bib.bib30)\]\. - •If memory is limiting→\\rightarrowquantization, pruning, or knowledge distillation\[[69](https://arxiv.org/html/2610.00363#bib.bib33),[65](https://arxiv.org/html/2610.00363#bib.bib34),[32](https://arxiv.org/html/2610.00363#bib.bib19)\]\. - •If both are limiting→\\rightarrowconsider hardware\-aware neural architecture search or hardware–software co\-design\[[7](https://arxiv.org/html/2610.00363#bib.bib68),[19](https://arxiv.org/html/2610.00363#bib.bib167),[14](https://arxiv.org/html/2610.00363#bib.bib20)\]\.
7. 7\.Define the evaluation protocol and validate under operational variability:Evaluation must reflect the detection task and safety requirements: - •Classification: F1\-score, minority\-class recall, and AUROC\. - •Detection and localization: mAP and mIoU\. - •Reconstruction and prediction: PR\-AUC, anomaly\-score AUROC, and false alarm rate\. - •Safety\-critical deployment: false negative rate, conservative thresholds, and uncertainty monitoring\. - •All deployed systems: inference latency on target platform and power consumption\. Validation should include operational variability across fleets, routes, seasons, speeds, and maintenance states\. Drift monitoring and adaptive thresholding should be considered for long\-term deployment\[[35](https://arxiv.org/html/2610.00363#bib.bib57),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.

#### 7\.6\.2Analytical Comparison of Detection Paradigms

Table[18](https://arxiv.org/html/2610.00363#S7.T18)complements the rule\-based framework by comparing the four paradigms across ten engineering criteria relevant to railway deployment\. It is intended as a structured summary supporting the selection rules in Step 4, not as a standalone decision tool\.

Rating scale and criteria:Each cell is rated as High \(H\), Medium \(M\), or Low \(L\) based on the following definitions:

- •High \(H\):The paradigm satisfies the criterion under typical railway conditions without requiring additional mechanisms\.Example:Classification achieves high interpretability because it produces explicit fault labels that maintenance engineers can act on directly\.
- •Medium \(M\):Suitability depends on implementation choices, data quality, or operational conditions, and may require moderate adaptation\.Example:Prediction handles temporal dependencies well but requires threshold recalibration when operating conditions change \(e\.g\., seasonal variation, fleet differences\)\.
- •Low \(L\):The paradigm does not naturally satisfy the criterion and typically requires substantial additional mechanisms such as data augmentation, domain adaptation, or hybridization\.Example:Classification has low coverage of unseen anomalies because it assumes a closed set of known fault types and fails silently on out\-of\-distribution inputs\.

Table 18:Analytical comparison of detection paradigms for deployment\-oriented railway anomaly detection\. H: High suitability; M: Medium; L: Low\.CriterionClassificationPredictionReconstructionHybridSuitability when fault labels are availableHMMHSuitability with normal data onlyLHHMCoverage of unseen anomaliesLMHHRobustness to class imbalanceLMHMTemporal dependency handlingMHM–HHTolerance to operational variabilityL–MMMHInterpretability for maintenanceHML–MMEmbedded feasibility \(tight budget\)HMML–MMultimodal extensibilityMMMHSafety redundancy potentialLMMH
#### 7\.6\.3Illustrative Case Studies

Two representative scenarios are used to illustrate how the stepwise procedure maps anomaly characteristics, label availability, deployment constraints, paradigm selection, hardware choice, model optimization, and evaluation metrics to concrete deployment recommendations\.

1. 1\.Case 1: Real\-time wheel flat detection from vibration signals Step 1 Anomaly characterization:The anomaly is a rolling\-stock\-level fault \(wheel flat\) that manifests as periodic abnormal vibration patterns, with sequential temporal behavior as the defect persists or worsens over operating cycles\. Step 2 Label availability:Labeled vibration datasets with annotated wheel flat events are available from test rig experiments\[[120](https://arxiv.org/html/2610.00363#bib.bib101)\]\. The data regime supports supervised learning\. Step 3 Deployment constraints:Detection must occur onboard during operation\. The latency requirement is soft real\-time \(<<50 ms per inference cycle\)\. Power budget is moderate \(10–30 W\)\. The system must operate offline \(no continuous connectivity\)\. Safety classification: condition\-monitoring with escalation to safety\-critical if severity thresholds are exceeded\. Step 4 Paradigm selection:Labels are available and the fault type is recurring and well\-characterized→\\rightarrowclassification\-basedapproach\. A spectrogram\-based CNN converts vibration windows into time\-frequency images for supervised classification\[[120](https://arxiv.org/html/2610.00363#bib.bib101)\]\. Step 5 Hardware platform:Soft real\-time with moderate power budget→\\rightarrowGPU edge platform\(e\.g\., Jetson Xavier NX\)\. If deterministic latency is required for integration with braking systems, anFPGAshould be considered instead\. Step 6 Model optimization:The spectrogram CNN is lightweight \(<<5M parameters\)\. FP16 quantization via TensorRT can be used to reduce inference time and memory usage, but the final accuracy–latency trade\-off must be validated on the target platform\[[106](https://arxiv.org/html/2610.00363#bib.bib30)\]\. Step 7 Evaluation:Primary metrics: recall \(to minimize missed flats\), F1\-score, and false negative rate\. Validation must cover multiple wheel wear levels, train speeds \(80–300 km/h\), and seasonal temperature variations\. Inference latency and power consumption must be reported on the target platform\.
2. 2\.Case 2: Door anomaly detection from current signals Step 1 Anomaly characterization:The anomaly is a rolling\-stock\-level fault \(door obstruction or mechanical degradation\) that manifests as a collective or contextual deviation in sequential current signals recorded during each door opening/closing cycle\. Step 2 Label availability:Fault labels are unavailable\. Only normal door cycles are recorded during routine operation\[[121](https://arxiv.org/html/2610.00363#bib.bib122)\]\. The data regime requires unsupervised or normal\-only learning\. Step 3 Deployment constraints:Detection should occur onboard in near\-real\-time \(soft real\-time,<<200 ms per cycle\)\. Power budget is tight \(<<10 W\)\. Connectivity is intermittent\. Safety classification: condition\-monitoring\. Step 4 Paradigm selection:No labels, sequential temporal behavior, collective anomaly pattern→\\rightarrowreconstruction\-basedapproach\. A bidirectional LSTM autoencoder trained on normal cycles flags deviations via reconstruction error\[[121](https://arxiv.org/html/2610.00363#bib.bib122)\]\. Step 5 Hardware platform:Tight power budget with soft real\-time requirement→\\rightarrowembedded CPUfor lightweight LSTM models, orGPU edge\(Jetson Nano at∼\\sim5–10 W\) if model complexity requires GPU acceleration\. Step 6 Model optimization:INT8 quantization reduces the LSTM autoencoder memory footprint\. No pruning or distillation is needed given the small model size \(<<500K parameters\)\. Step 7 Evaluation:Primary metrics: PR\-AUC \(due to severe class imbalance\), false alarm rate, and reconstruction error distribution analysis\. Validation must include multiple door types, passenger load conditions, and seasonal temperature effects on mechanical resistance\. Threshold calibration and drift monitoring are essential for long\-term deployment\[[35](https://arxiv.org/html/2610.00363#bib.bib57)\]\. Inference latency and power consumption must be reported\.

Although the two detailed examples focus on onboard time\-series monitoring, the same seven\-step logic also applies to vision\-based and offline inspection scenarios, such as image\-based rail fastener or rail\-surface defect detection\. These scenarios are discussed within the vision\-based detection paradigm in Section 5\.1\.2, where image\-based classification and localization approaches are reviewed\.

#### 7\.6\.4Engineering Implications

For deployment\-oriented railway anomaly detection studies, the following information should be reported at minimum: \(i\) target hardware platform and latency constraints, \(ii\) label availability and class imbalance level, \(iii\) optimization strategy applied \(quantization level, pruning ratio, or distillation setup\), \(iv\) decision policy and threshold calibration method, and \(v\) robustness evaluation under noise, drift, and fleet variability\.

The proposed framework transforms the survey taxonomy and paradigm analysis into an end\-to\-end decision methodology\. It connects anomaly characterization, supervision level, deployment constraints, hardware choice, model optimization, and evaluation protocol, thereby providing practitioners with a structured and reproducible process for selecting and deploying deep learning solutions for railway anomaly detection\.

## 8Future Research Directions

Several research directions emerge from the limitations identified in this survey\. These directions concern not only detection accuracy, but also generalization, deployment readiness, reproducibility, and safety assurance in real railway environments\.

Multimodal fusion and multi\-task railway monitoring:Most existing studies focus on a single sensing modality and address one specific task, such as fault classification, defect detection, or condition monitoring\. However, railway systems naturally generate heterogeneous data from vision, vibration, acoustic, current, voltage, log, and operational signals\. Future work should therefore move toward unified multimodal architectures able to combine complementary information across subsystems\. Such models should not only detect anomalies, but also support fault localization, severity estimation, and condition assessment within the same framework\[[89](https://arxiv.org/html/2610.00363#bib.bib28),[84](https://arxiv.org/html/2610.00363#bib.bib27)\]\.

Open datasets and standardized evaluation protocols:A major obstacle to progress remains the lack of public railway anomaly detection benchmarks\. Most studies rely on proprietary datasets, use different fault definitions, and report results under incompatible experimental settings\. This makes cross\-study comparison difficult and limits reproducibility\. Future efforts should prioritize the creation of open, well\-documented datasets covering representative railway subsystems, fault types, operating conditions, and sensor modalities\. In addition, standardized evaluation protocols are needed, including clear train\-test split strategies, cross\-route and cross\-fleet validation, cost\-sensitive metrics, false\-alarm analysis, and inference latency reporting\[[104](https://arxiv.org/html/2610.00363#bib.bib127)\]\.

Hardware\-aware neural architecture search and hardware–software co\-design:Current railway anomaly detection studies generally address deployment constraints after training, using compression techniques such as quantization, pruning, or knowledge distillation\. Although these methods reduce model size and inference cost, they cannot fully compensate for an architecture that was not designed for the target hardware from the beginning\. Future research should therefore explore hardware\-aware neural architecture search \(HW\-NAS\) as a more principled approach for railway applications\. Instead of optimizing only detection performance, HW\-NAS can jointly consider accuracy, latency, memory footprint, and energy consumption during the architecture search process\[[7](https://arxiv.org/html/2610.00363#bib.bib68),[19](https://arxiv.org/html/2610.00363#bib.bib167)\]\.

However, despite its potential, HW\-NAS remains largely underexplored in railway anomaly detection for three main reasons\. First, although the search is performed offline and does not increase inference cost after deployment, it increases development complexity because many candidate architectures must be trained, evaluated, or estimated under realistic hardware constraints\[[126](https://arxiv.org/html/2610.00363#bib.bib5),[125](https://arxiv.org/html/2610.00363#bib.bib4)\]\. Second, the railway domain lacks standardized deployment\-oriented benchmarks: most studies rely on task\-specific or proprietary datasets and rarely report common hardware platforms, latency budgets, memory limits, or energy measurements\. This prevents fair comparison of hardware\-aware search strategies across railway subsystems\. Third, when anomaly detection models support safety\-related decision chains, automatically searched architectures require careful validation in terms of reproducibility, robustness, and auditability\. More generally, current railway anomaly detection research still focuses mainly on detection accuracy and proof\-of\-concept validation, while systematic architecture search under hardware constraints remains uncommon\.

This direction is especially relevant for railway\-grade embedded platforms, where models must operate under strict constraints related to real\-time response, limited resources, vibration, temperature variation, electromagnetic compatibility, and long\-term reliability\. Future studies should define railway\-specific hardware benchmarks and latency budgets for different subsystems\. Beyond HW\-NAS, hardware–software co\-design offers an even broader perspective by jointly optimizing the neural architecture and the deployment platform\[[57](https://arxiv.org/html/2610.00363#bib.bib42),[14](https://arxiv.org/html/2610.00363#bib.bib20),[54](https://arxiv.org/html/2610.00363#bib.bib32)\]\. Adjacent railway control studies, such as optimization\-based control for maglev suspension, also highlight the importance of robustness in railway\-grade systems, although they remain outside the core corpus of deep learning\-based anomaly detection\[[138](https://arxiv.org/html/2610.00363#bib.bib3)\]\.

Domain adaptation and cross\-fleet generalization:Another important limitation is the lack of evaluation across fleets, routes, seasons, and operating conditions\. Models trained on one vehicle type, operator, or geographic region may not generalize to another due to differences in sensors, maintenance practices, infrastructure quality, weather, loading conditions, and operational profiles\. Future work should investigate domain adaptation, transfer learning, and federated learning to improve robustness across heterogeneous railway environments\. These approaches could reduce the need for each operator to build large labeled datasets from scratch, while enabling knowledge transfer between fleets and preserving data confidentiality\[[107](https://arxiv.org/html/2610.00363#bib.bib99),[62](https://arxiv.org/html/2610.00363#bib.bib106),[161](https://arxiv.org/html/2610.00363#bib.bib61)\]\.

Foundation models and large language models for railway monitoring:Foundation models and large language models open new possibilities for railway monitoring, particularly when heterogeneous data sources are available\. Vision foundation models could support few\-shot defect detection in inspection images, while language models could help analyze maintenance reports, event logs, and operator notes\. More importantly, multimodal foundation models may enable cross\-modal reasoning between sensor signals, visual observations, and textual maintenance records\[[27](https://arxiv.org/html/2610.00363#bib.bib120)\]\. However, their use in railway systems remains a long\-term research direction\. Their high computational cost, uncertain latency, limited interpretability, and difficult certification raise important barriers for safety\-critical deployment\. Future research should therefore focus on efficient, explainable, and certifiable foundation\-model\-based solutions rather than directly transferring large generic models to railway applications\.

## 9Conclusion

This survey presented a structured and deployment\-oriented review of deep learning\-based anomaly detection in railway systems, covering infrastructure, rolling stock, and operational contexts\. Four complementary contributions were developed to support systematic method selection and deployment planning\.

A four\-axis anomaly taxonomy was introduced to characterize railway anomalies by location, manifestation, sensing modality, and temporal behavior\. Deep learning approaches were organized into four detection paradigms\-classification\-, prediction\-, reconstruction\-, and hybrid\-based\-and analyzed with respect to data availability, supervision requirements, and deployment feasibility\. Evaluation practices were critically reviewed, showing that conventional accuracy\-centric metrics are insufficient for safety\-critical applications; an expanded framework integrating cost\-sensitive assessment, interpretability, and latency reporting was proposed\. Finally, a deployment\-driven decision framework was introduced to translate anomaly characteristics and operational constraints into justified paradigm and architecture choices, with particular emphasis on embedded platforms, model compression, and hardware\-aware design\.

Taken together, these contributions establish a unified reference for researchers and engineers at the intersection of deep learning and railway systems, bridging the gap between academic performance and operational feasibility in safety\-critical monitoring applications\.

## References

- \[1\]\(2025\)Integrating ai and edge computing for advanced safety at railroad grade crossings\.Journal of Rail Transport Planning & Management33,pp\.100501\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S7.I2.i2.p1.1),[§7\.3](https://arxiv.org/html/2610.00363#S7.SS3.p1.1)\.
- \[2\]A\. Arain, S\. Mehran, M\. Z\. Shaikh, D\. Kumar, B\. S\. Chowdhry, and T\. Hussain\(2024\)Railway track surface faults dataset\.Data in brief52,pp\.110050\.Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.13.1.1.1)\.
- \[3\]A\. Armijo and D\. Zamora\-Sánchez\(2024\)Integration of railway bridge structural health monitoring into the internet of things with a digital twin: a case study\.Sensors24\(7\),pp\.2115\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I2.i3.p1.1),[§7\.3](https://arxiv.org/html/2610.00363#S7.SS3.p1.1)\.
- \[4\]S\. Bai, J\. Z\. Kolter, and V\. Koltun\(2018\)An empirical evaluation of generic convolutional and recurrent networks for sequence modeling\.arXiv preprint arXiv:1803\.01271\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S3.I4.i2.p1.1)\.
- \[5\]T\. Baltrušaitis, C\. Ahuja, and L\. Morency\(2018\)Multimodal machine learning: a survey and taxonomy\.IEEE transactions on pattern analysis and machine intelligence41\(2\),pp\.423–443\.Cited by:[4th item](https://arxiv.org/html/2610.00363#S4.I1.i4.p1.1)\.
- \[6\]M\. Bałdyga, K\. Barański, J\. Belter, M\. Kalinowski, and P\. Weichbroth\(2024\)Anomaly detection in railway sensor data environments: state\-of\-the\-art methods and empirical performance evaluation\.Sensors24\(8\),pp\.2633\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p4.1),[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.8.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p2.1),[2nd item](https://arxiv.org/html/2610.00363#S3.I3.i2.p1.1),[§3\.1](https://arxiv.org/html/2610.00363#S3.SS1.p1.1),[§3\.3](https://arxiv.org/html/2610.00363#S3.SS3.p1.1),[§3](https://arxiv.org/html/2610.00363#S3.p1.1),[§5\.1](https://arxiv.org/html/2610.00363#S5.SS1.p1.1)\.
- \[7\]H\. Benmeziane, K\. E\. Maghraoui, M\. Wistuba, and N\. Wang\(2021\)A comprehensive survey on hardware\-aware neural architecture search\.arXiv preprint arXiv:2101\.09336\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S6.I1.i3.p1.1),[§6\.3](https://arxiv.org/html/2610.00363#S6.SS3.p2.1),[4th item](https://arxiv.org/html/2610.00363#S7.I1.i4.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I4.i1.p1.1),[3rd item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i3.p1.1),[§7\.5](https://arxiv.org/html/2610.00363#S7.SS5.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.2.4.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.5.4.1.1),[§8](https://arxiv.org/html/2610.00363#S8.p4.1)\.
- \[8\]K\. Berahmand, F\. Daneshfar, E\. S\. Salehi, Y\. Li, and Y\. Xu\(2024\)Autoencoders and their applications in machine learning: a survey\.Artificial intelligence review57\(2\),pp\.28\.Cited by:[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p4.1),[§5\.3](https://arxiv.org/html/2610.00363#S5.SS3.p1.1)\.
- \[9\]E\. Bernal, M\. Spiryagin, and C\. Cole\(2018\)Onboard condition monitoring sensors, systems and techniques for freight railway vehicles: a review\.IEEE sensors journal19\(1\),pp\.4–24\.Cited by:[§4\.1](https://arxiv.org/html/2610.00363#S4.SS1.p1.1)\.
- \[10\]A\. S\. Bhadouria, R\. P\. Mishra, A\. K\. Jalan, and A\. R\. Andrade\(2025\)Advances in railway condition monitoring and fault diagnosis: transition from industry 4\.0 automation to industry 5\.0 human\-centric maintenance\.Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit,pp\.09544097261417814\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p1.1)\.
- \[11\]M\. Binder, V\. Mezhuyev, and M\. Tschandl\(2023\)Predictive maintenance for railway domain: a systematic literature review\.IEEE Engineering Management Review51\(2\),pp\.120–140\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p6.1),[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.9.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p3.1),[2nd item](https://arxiv.org/html/2610.00363#S3.I1.i2.p1.1),[4th item](https://arxiv.org/html/2610.00363#S3.I3.i4.p1.1)\.
- \[12\]R\. Bodo, M\. Bertocco, and A\. Bianchi\(2021\)Cost\-sensitive fault identification in predictive maintenance applications: a case study\.In2021 IEEE International Instrumentation and Measurement Technology Conference \(I2MTC\),pp\.1–6\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S6.I1.i2.p1.2)\.
- \[13\]A\. Bouguettaya, H\. Zarzour, A\. Kechida, and A\. M\. Taberkit\(2024\)A reusable AI\-enabled defect detection system for railway using ensembled CNN\.Applied Intelligence\.External Links:[Document](https://dx.doi.org/10.1007/s10489-024-05676-8)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p2.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.4.1.1.1)\.
- \[14\]H\. Bouzidi, M\. Odema, H\. Ouarnoughi, M\. A\. Al Faruque, and S\. Niar\(2023\)HADAS: hardware\-aware dynamic neural architecture search for edge performance scaling\.InDesign, Automation and Test in Europe Conference \(DATE\),External Links:[Link](https://arxiv.org/abs/2212.03354)Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i3.p1.1),[§8](https://arxiv.org/html/2610.00363#S8.p6.1)\.
- \[15\]H\. Bouzidi\(2024\)Efficient deployment of deep neural networks on hardware devices for edge ai\.Ph\.D\. Thesis,Université Polytechnique Hauts\-de\-France\.Cited by:[§7\.5\.3](https://arxiv.org/html/2610.00363#S7.SS5.SSS3.p1.1)\.
- \[16\]K\. Brintha and S\. J\. Jawhar\(2024\)FOD\-YOLO net: fasteners fault and object detection in railway tracks using deep YOLO network\.Journal of Intelligent and Fuzzy Systems46\.External Links:[Document](https://dx.doi.org/10.3233/JIFS-236445)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.9.1.1.1)\.
- \[17\]F\. J\. Bris\-Peñalver, R\. Verdecia\-Peña, and J\. I\. Alonso\(2026\)A survey of ai\-enabled predictive maintenance for railway infrastructure: models, data sources, and research challenges\.Sensors26\(3\),pp\.906\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p6.1),[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.10.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p3.1),[1st item](https://arxiv.org/html/2610.00363#S3.I1.i1.p1.1)\.
- \[18\]M\. A\. Butakova, A\. V\. Chernov, P\. S\. Shevchuk, and V\. D\. Vereskun\(2017\)Complex event processing for network anomaly detection in digital railway communication services\.In2017 25th Telecommunication Forum \(TELFOR\),pp\.1–4\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p4.1)\.
- \[19\]H\. Cai, L\. Zhu, and S\. Han\(2018\)Proxylessnas: direct neural architecture search on target task and hardware\.arXiv preprint arXiv:1812\.00332\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i3.p1.1),[§8](https://arxiv.org/html/2610.00363#S8.p4.1)\.
- \[20\]Z\. Cao, Y\. Qin, L\. Jia, Z\. Xie, Y\. Gao, Y\. Wang, P\. Li, and Z\. Yu\(2024\)Railway intrusion detection based on machine vision: a survey, challenges, and perspectives\.IEEE Transactions on Intelligent Transportation Systems25\(7\),pp\.6427–6448\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p6.1)\.
- \[21\]J\. M\. Castillo\-Mingorance, M\. Sol\-Sánchez, F\. Moreno\-Navarro, and M\. C\. Rubio\-Gámez\(2020\)A critical review of sensors for the continuous monitoring of smart and sustainable railway infrastructures\.Sustainability12\(22\),pp\.9428\.Cited by:[§4\.1](https://arxiv.org/html/2610.00363#S4.SS1.p1.1)\.
- \[22\]R\. Chalapathy and S\. Chawla\(2019\)Deep learning for anomaly detection: a survey\.arXiv preprint arXiv:1901\.03407\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1),[§2](https://arxiv.org/html/2610.00363#S2.p1.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p1.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[§5\.1\.1](https://arxiv.org/html/2610.00363#S5.SS1.SSS1.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p1.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p3.1),[§5\.2](https://arxiv.org/html/2610.00363#S5.SS2.p1.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p2.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p4.1),[§5\.3](https://arxiv.org/html/2610.00363#S5.SS3.p1.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p1.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p9.1),[§5](https://arxiv.org/html/2610.00363#S5.p1.1),[§6](https://arxiv.org/html/2610.00363#S6.p1.1),[item 2](https://arxiv.org/html/2610.00363#S7.I6.i2.p1.1)\.
- \[23\]V\. Chandola, A\. Banerjee, and V\. Kumar\(2009\)Anomaly detection: a survey\.ACM computing surveys \(CSUR\)41\(3\),pp\.1–58\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S3.I2.i1.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S3.I2.i2.p1.1),[3rd item](https://arxiv.org/html/2610.00363#S3.I2.i3.p1.1),[§3\.2](https://arxiv.org/html/2610.00363#S3.SS2.p1.1)\.
- \[24\]S\. Chang, B\. Du, and L\. Zhang\(2019\)A sparse autoencoder based hyperspectral anomaly detection algorihtm using residual of reconstruction error\.InIGARSS 2019\-2019 IEEE International Geoscience and Remote Sensing Symposium,pp\.5488–5491\.Cited by:[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p2.1)\.
- \[25\]S\. Chavhan, D\. Gupta, and J\. J\. Rodrigues\(2025\)Bi\-lstm\-based train fault detection system using edge computing and iot for enhanced railway safety\.IEEE Transactions on Intelligent Transportation Systems\.Cited by:[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.8.1.1.1)\.
- \[26\]N\. V\. Chawla, K\. W\. Bowyer, L\. O\. Hall, and W\. P\. Kegelmeyer\(2002\)SMOTE: synthetic minority over\-sampling technique\.Journal of artificial intelligence research16,pp\.321–357\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S4.I1.i1.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1)\.
- \[27\]X\. Chen, Y\. Lei, Y\. Li, S\. Parkinson, X\. Li, J\. Liu, F\. Lu, H\. Wang, Z\. Wang, B\. Yang,et al\.\(2025\)Large models for machine monitoring and fault diagnostics: opportunities, challenges, and future direction\.Journal of Dynamics, Monitoring and Diagnostics4\(2\),pp\.76–90\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p9.1),[§8](https://arxiv.org/html/2610.00363#S8.p8.1)\.
- \[28\]Y\. Chenet al\.\(2022\)Meta\-learning with GANs for anomaly detection, with deployment in high\-speed rail inspection system\.arXiv preprint arXiv:2202\.05795\.Cited by:[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p6.1),[Table 13](https://arxiv.org/html/2610.00363#S5.T13.2.4.1.1.1)\.
- \[29\]Y\. Chenet al\.\(2024\)A real\-time method for railway track detection and 3D fitting based on camera and LiDAR fusion sensing\.Remote Sensing16\(8\),pp\.1441\.Cited by:[6th item](https://arxiv.org/html/2610.00363#S3.I3.i6.p1.1)\.
- \[30\]Y\. Chen, X\. Liu, W\. Fan, N\. Duan, and K\. Zhou\(2024\)A Deep\-LSTM\-based fault detection method for railway vehicle suspensions\.Machines12\(2\),pp\.116\.External Links:[Document](https://dx.doi.org/10.3390/machines12020116)Cited by:[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p3.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.8.1.1.1)\.
- \[31\]C\. Cheng, J\. Wang, H\. Chen, Z\. Chen, H\. Luo, and P\. Xie\(2020\)A review of intelligent fault diagnosis for high\-speed trains: qualitative approaches\.Entropy23\(1\),pp\.1\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p2.1),[2nd item](https://arxiv.org/html/2610.00363#S3.I1.i2.p1.1)\.
- \[32\]Y\. Cheng, D\. Wang, P\. Zhou, and T\. Zhang\(2017\)A survey of model compression and acceleration for deep neural networks\.arXiv preprint arXiv:1710\.09282\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S7.I5.i2.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i2.p1.1),[§7\.5](https://arxiv.org/html/2610.00363#S7.SS5.p1.1)\.
- \[33\]J\. Choi and J\. Han\(2024\)Deep learning \(fast r\-cnn\)\-based evaluation of rail surface defects\.Applied Sciences14\(5\),pp\.1874\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S3.I4.i1.p1.1)\.
- \[34\]R\. Cordova\-Cardenas, D\. Amor, and Á\. Gutiérrez\(2025\)Edge ai in practice: a survey and deployment framework for neural networks on embedded systems\.Electronics14\(24\),pp\.4877\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S7.I1.i1.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I1.i2.p1.1),[5th item](https://arxiv.org/html/2610.00363#S7.I1.i5.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.2.4.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.3.4.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.6.3.1.1)\.
- \[35\]L\. Correiaet al\.\(2024\)Online model\-based anomaly detection in multivariate time series: taxonomy, survey, research challenges and future directions\.Engineering Applications of Artificial Intelligence138,pp\.109323\.Cited by:[4th item](https://arxiv.org/html/2610.00363#S3.I2.i4.p1.1),[§5\.2\.2](https://arxiv.org/html/2610.00363#S5.SS2.SSS2.p1.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p3.1),[§5\.2](https://arxiv.org/html/2610.00363#S5.SS2.p1.1),[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p3.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p4.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[Table 11](https://arxiv.org/html/2610.00363#S5.T11.2.5.1.1.1),[item 7](https://arxiv.org/html/2610.00363#S7.I6.i7.p1.2),[item 2](https://arxiv.org/html/2610.00363#S7.I8.i2.p8.1)\.
- \[36\]H\. Cui, J\. Li, Q\. Hu, and Q\. Mao\(2019\)Real\-time inspection system for ballast railway fasteners based on point cloud deep learning\.IEEE Access8,pp\.61604–61614\.Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.9.1.1.1),[§5\.1](https://arxiv.org/html/2610.00363#S5.SS1.p2.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I1.i2.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I6.i5.I1.i2.p1.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.5.1.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.3.3.1.1)\.
- \[37\]N\. Davari, B\. Veloso, R\. P\. Ribeiro, P\. M\. Pereira, and J\. Gama\(2021\)Predictive maintenance based on anomaly detection using deep learning for air production unit in the railway industry\.In2021 IEEE 8th International Conference on Data Science and Advanced Analytics \(DSAA\),pp\.1–10\.Cited by:[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p2.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p14.1),[Table 11](https://arxiv.org/html/2610.00363#S5.T11.2.2.1.1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I3.i2.p1.1)\.
- \[38\]N\. Davari, B\. Veloso, R\. P\. Ribeiro, and J\. Gama\(2024\)Detecting and explaining anomalies in the air production unit of a train\.InProceedings of the 39th ACM/SIGAPP Symposium on Applied Computing,pp\.1–8\.External Links:[Document](https://dx.doi.org/10.1145/3605098.3635906)Cited by:[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p2.1),[Table 11](https://arxiv.org/html/2610.00363#S5.T11.2.3.1.1.1)\.
- \[39\]L\. De Donato, F\. Flammini, S\. Marrone, C\. Mazzariello, R\. Nardone, C\. Sansone, and V\. Vittorini\(2022\)A survey on audio\-video based defect detection through deep learning in railway maintenance\.IEEE Access10,pp\.65376–65400\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p6.1),[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.6.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p3.1),[§3\.1](https://arxiv.org/html/2610.00363#S3.SS1.p1.1),[§3\.3](https://arxiv.org/html/2610.00363#S3.SS3.p1.1)\.
- \[40\]F\. Di Mattia, P\. Galeone, M\. De Simoni, and E\. Ghelfi\(2019\)A survey on gans for anomaly detection\.arXiv preprint arXiv:1906\.11632\.Cited by:[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p1.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p4.1)\.
- \[41\]M\. Di Summa, M\. E\. Griseta, N\. Mosca, C\. Patruno, M\. Nitti, V\. Renò, and E\. Stella\(2023\)A review on deep learning techniques for railway infrastructure monitoring\.IEEE Access11,pp\.114638–114661\.Cited by:[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.7.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p3.1),[1st item](https://arxiv.org/html/2610.00363#S3.I3.i1.p1.1)\.
- \[42\]S\. Diaz\-Santos, P\. Caballero\-Gil, and C\. Caballero\-Gil\(2025\)Real\-time passenger flow analysis in tram stations using yolo\-based computer vision and edge ai on jetson nano\.Computers14\(11\),pp\.476\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S7.I3.i1.p1.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.7.1.1.1)\.
- \[43\]X\. Ding, Z\. Li, M\. Tang, X\. Shen, and L\. Zhou\(2026\)An integrated iewt and cnn–transformer deep architecture for intelligent fault diagnosis of bogie axle\-box bearings\.Electronics15\(4\),pp\.804\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S3.I4.i2.p1.1),[§5\.2\.2](https://arxiv.org/html/2610.00363#S5.SS2.SSS2.p2.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 9](https://arxiv.org/html/2610.00363#S5.T9.2.3.1.1.1)\.
- \[44\]C\. Donnarumma, P\. Fara, G\. Serra, S\. Di Leonardi, and M\. Marinoni\(2019\)EN\-50128 certification\-oriented design of a safety\-critical hard real\-time kernel\.In2019 IEEE International Symposium on Software Reliability Engineering Workshops \(ISSREW\),pp\.314–317\.Cited by:[§7](https://arxiv.org/html/2610.00363#S7.p2.1)\.
- \[45\]G\. Du, H\. Zhang, H\. Yu, P\. Hou, J\. He, S\. Cao, G\. Wang, and L\. Ma\(2024\)Study on automatic tracking system of microwave deicing device for railway contact wire\.IEEE Transactions on Instrumentation and Measurement73,pp\.1–11\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S3.I1.i3.p1.1)\.
- \[46\]H\. A\. Fernández\-Bobadilla, Y\. Bouchikhi, and U\. Martin\(2025\)GAN\-based data augmentation of time series for fault diagnosis in railway track: ha fernández\-bobadilla et al\.\.Railway Engineering Science33\(4\),pp\.642–683\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p4.1)\.
- \[47\]A\. Ferrari, G\. Gori, B\. Rosadini, I\. Trotta, S\. Bacherini, A\. Fantechi, and S\. Gnesi\(2018\)Detecting requirements defects with nlp patterns: an industrial experience in the railway domain\.Empirical Software Engineering23\(6\),pp\.3684–3733\.Cited by:[5th item](https://arxiv.org/html/2610.00363#S3.I3.i5.p1.1)\.
- \[48\]Y\. Fu, D\. Yan, J\. Li, S\. L\. Ma, C\. Sham, and H\. Chou\(2025\)An efficient fpga\-based edge ai system for railway fault detection\.IEEE Consumer Electronics Magazine14\(4\),pp\.76–86\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S7.I3.i1.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I5.i1.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I5.i2.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I6.i5.I1.i1.p1.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p4.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p5.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p6.1),[§7\.5\.3](https://arxiv.org/html/2610.00363#S7.SS5.SSS3.p1.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.2.1.1.1)\.
- \[49\]A\. Gálvez, A\. Diez\-Olivan, D\. Seneviratne, and D\. Galar\(2021\)Fault detection and rul estimation for railway hvac systems using a hybrid model\-based approach\.Sustainability13\(12\),pp\.6828\.Cited by:[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.4.1.1.1),[§5\.1\.1](https://arxiv.org/html/2610.00363#S5.SS1.SSS1.p1.1),[Table 5](https://arxiv.org/html/2610.00363#S5.T5.2.4.1.1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I3.i2.p1.1)\.
- \[50\]A\. Gálvez, D\. Galar, and D\. Seneviratne\(2022\)A hybrid model\-based approach on prognostics for railway hvac\.IEEE Access10,pp\.108117–108127\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p10.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.9.1.1.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p2.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 10](https://arxiv.org/html/2610.00363#S5.T10.2.2.1.1.1)\.
- \[51\]S\. Gao\(2020\)Automatic detection and monitoring system of pantograph–catenary in china’s high\-speed railways\.IEEE Transactions on Instrumentation and Measurement70,pp\.1–12\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.10.1.1.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p12.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p15.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p5.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p9.1),[Table 13](https://arxiv.org/html/2610.00363#S5.T13.2.3.1.1.1)\.
- \[52\]Y\. Gao, Z\. Cao, Y\. Qin, X\. Ge, L\. Lian, J\. Bai, and H\. Yu\(2023\)Railway fastener anomaly detection via multisensor fusion and self\-driven loss reweighting\.IEEE Sensors Journal24\(2\),pp\.1812–1825\.Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.7.1.1.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p12.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.6.1.1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I1.i1.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.2.3.1.1)\.
- \[53\]R\. Gasparini, A\. D’Eusanio, G\. Borghi, S\. Pini, G\. Scaglione, S\. Calderara, E\. Fedeli, and R\. Cucchiara\(2021\)Anomaly detection, localization and classification for railway inspection\.In2020 25th International Conference on Pattern Recognition \(ICPR\),pp\.3419–3426\.Cited by:[§3\.4](https://arxiv.org/html/2610.00363#S3.SS4.p1.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p1.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1)\.
- \[54\]M\. I\. E\. Ghebriout, H\. Bouzidi, S\. Niar, and H\. Ouarnoughi\(2023\)Harmonic\-NAS: hardware\-aware multimodal neural architecture search on resource\-constrained devices\.InAsian Conference on Machine Learning \(ACML\),External Links:[Link](https://api.semanticscholar.org/%0ACorpusID:261706155)Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p6.1)\.
- \[55\]R\. Gong, C\. Wang, J\. Li, and Y\. Xu\(2023\)Lightweight fault diagnosis method in embedded system based on knowledge distillation\.Journal of Mechanical Science and Technology37\(11\),pp\.5649–5660\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I5.i3.p1.1)\.
- \[56\]T\. Gong, L\. Zhu, F\. R\. Yu, and T\. Tang\(2024\)Train\-to\-edge cooperative intelligence for obstacle intrusion detection in rail transit\.IEEE Transactions on Vehicular Technology73\(6\),pp\.7669–7680\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S7.I2.i1.p1.1),[§7\.3](https://arxiv.org/html/2610.00363#S7.SS3.p1.1)\.
- \[57\]C\. Guo, F\. Cheng, Z\. Du, J\. Kiessling, J\. Ku, S\. Li, Z\. Li, M\. Ma, T\. Molom\-Ochir, B\. Morris,et al\.\(2025\)A survey: collaborative hardware and software design in the era of large language models\.IEEE Circuits and Systems Magazine25\(1\),pp\.35–57\.Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p6.1)\.
- \[58\]H\. Hadj\-Mabrouk\(2026\)Contributions and limitations of ai and machine learning in railway operations, maintenance, and safety: a literature review\.Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability240\(1\),pp\.3–36\.Cited by:[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.11.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p3.1)\.
- \[59\]M\. Haghbin J\. Chiachióet al\.\(2024\)Predicting rail corrugation based on convolutional neural networks using Vehicle’s acceleration measurements\.Sensors24\(14\),pp\.4627\.External Links:[Document](https://dx.doi.org/10.3390/s24144627)Cited by:[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p2.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 10](https://arxiv.org/html/2610.00363#S5.T10.2.3.1.1.1)\.
- \[60\]S\. Ham, S\. Han, S\. Kim, H\. J\. Park, K\. Park, and J\. Choi\(2019\)A comparative study of fault diagnosis for train door system: traditional versus deep learning approaches\.Sensors19\(23\),pp\.5160\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1)\.
- \[61\]S\. Ham, S\. Han, S\. Kim, H\. J\. Park, K\. Park, and J\. Choi\(2019\)A comparative study of fault diagnosis for train door system: traditional versus deep learning approaches\.Sensors19\(23\),pp\.5160\.Cited by:[4th item](https://arxiv.org/html/2610.00363#S4.I2.i4.p1.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.3.1.1.1),[§5\.1\.1](https://arxiv.org/html/2610.00363#S5.SS1.SSS1.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[Table 5](https://arxiv.org/html/2610.00363#S5.T5.2.6.1.1.1)\.
- \[62\]X\. Hao, J\. Yang, F\. Yang, X\. Sun, Y\. Hou, and J\. Wang\(2023\)Track geometry estimation from vehicle–body acceleration for high\-speed railway using deep learning technique\.Vehicle system dynamics61\(1\),pp\.239–259\.Cited by:[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p3.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p3.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.6.1.1.1),[§8](https://arxiv.org/html/2610.00363#S8.p7.1)\.
- \[63\]H\. He, Y\. Bai, E\. A\. Garcia, and S\. Li\(2008\)ADASYN: adaptive synthetic sampling approach for imbalanced learning\.In2008 IEEE international joint conference on neural networks \(IEEE world congress on computational intelligence\),pp\.1322–1328\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S4.I1.i1.p1.1)\.
- \[64\]J\. He, J\. He, Z\. Gou, and F\. Lv\(2025\)Defect detection in railway catenary systems: a survey\.Intelligent Transportation Infrastructure,pp\.liaf027\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1)\.
- \[65\]G\. Hinton, O\. Vinyals, and J\. Dean\(2015\)Distilling the knowledge in a neural network\.arXiv preprint arXiv:1503\.02531\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I5.i3.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i2.p1.1)\.
- \[66\]X\. Huang\(2025\)AI\-enabled knowledge transfer and learning for nondestructive evaluation toward intelligent and adaptive systems\.Ph\.D\. Thesis,Michigan State University\.Cited by:[5th item](https://arxiv.org/html/2610.00363#S4.I1.i5.p1.1)\.
- \[67\]International Union of Railways \(UIC\)\(2025\)Railway Statistics Synopsis 2025 Edition\.Note:Available: https://uic\-stats\.uic\.org[https://uic\-stats\.uic\.org](https://uic-stats.uic.org/)Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p1.1)\.
- \[68\]A\. Iqbal, R\. Amin, F\. S\. Alsubaei, and A\. Alzahrani\(2024\)Anomaly detection in multivariate time series data using deep ensemble models\.Plos one19\(6\),pp\.e0303890\.Cited by:[§5\.2](https://arxiv.org/html/2610.00363#S5.SS2.p1.1)\.
- \[69\]B\. Jacob, S\. Kligys, B\. Chen, M\. Zhu, M\. Tang, A\. Howard, H\. Adam, and D\. Kalenichenko\(2018\)Quantization and training of neural networks for efficient integer\-arithmetic\-only inference\.InProceedings of the IEEE conference on computer vision and pattern recognition,pp\.2704–2713\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S7.I5.i1.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i2.p1.1)\.
- \[70\]A\. R\. Javed, M\. Usman, S\. U\. Rehman, M\. U\. Khan, and M\. S\. Haghighi\(2020\)Anomaly detection in automated vehicles using multistage attention\-based convolutional neural network\.IEEE Transactions on Intelligent Transportation Systems22\(7\),pp\.4291–4300\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S3.I3.i3.p1.1)\.
- \[71\]P\. Jiang, D\. Ergu, F\. Liu, Y\. Cai, and B\. Ma\(2022\)A review of yolo algorithm developments\.Procedia computer science199,pp\.1066–1073\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1)\.
- \[72\]J\. Kang, C\. Kim, J\. Kang, and J\. Gwak\(2021\)Anomaly detection of the brake operating unit on metro vehicles using a one\-class LSTM autoencoder\.Applied Sciences11\(19\),pp\.9290\.External Links:[Document](https://dx.doi.org/10.3390/app11199290)Cited by:[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p2.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[Table 11](https://arxiv.org/html/2610.00363#S5.T11.2.4.1.1.1)\.
- \[73\]U\. Kaya\(2025\)Detection of spark erosion on insulated rail joints by deep learning\.IEEE Access13\(\),pp\.79291–79303\.External Links:[Document](https://dx.doi.org/10.1109/ACCESS.2025.3567193)Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.4.1.1.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.7.1.1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I1.i1.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.2.3.1.1)\.
- \[74\]S\. Khan, M\. Naseer, M\. Hayat, S\. W\. Zamir, F\. S\. Khan, and M\. Shah\(2022\)Transformers in vision: a survey\.ACM computing surveys \(CSUR\)54\(10s\),pp\.1–41\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p4.1)\.
- \[75\]Z\. Kljaić, D\. Pavković, M\. Cipek, M\. Trstenjak, T\. J\. Mlinarić, and M\. Nikšić\(2023\)An overview of current challenges and emerging technologies to facilitate increased energy efficiency, safety, and sustainability of railway transport\.Future Internet15\(11\),pp\.347\.Cited by:[§7\.4](https://arxiv.org/html/2610.00363#S7.SS4.p1.1),[§7\.5](https://arxiv.org/html/2610.00363#S7.SS5.p1.1),[§7\.6](https://arxiv.org/html/2610.00363#S7.SS6.p2.1),[§7](https://arxiv.org/html/2610.00363#S7.p1.1)\.
- \[76\]P\. Komorski, G\. M\. Szymanski, T\. Nowakowski, and M\. Orczyk\(2021\)Advanced acoustic signal analysis used for wheel\-flat detection\.Latin American Journal of Solids and Structures18\(01\),pp\.e338\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1)\.
- \[77\]B\. Koonce\(2021\)ResNet 50\.InConvolutional neural networks with swift for tensorflow: image recognition and dataset categorization,pp\.63–72\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p2.1)\.
- \[78\]A\. Kumar and S\. Harsha\(2024\)A systematic literature review of defect detection in railways using machine vision\-based inspection methods\.International Journal of Transportation Science and Technology\.Cited by:[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.5.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p3.1),[1st item](https://arxiv.org/html/2610.00363#S3.I3.i1.p1.1),[1st item](https://arxiv.org/html/2610.00363#S3.I4.i1.p1.1)\.
- \[79\]D\. C\. Larese, A\. B\. Cerrada, G\. D\. Tomei, A\. Guerrero\-López, P\. M\. Olmos, and M\. J\. G\. García\(2025\)Transformer vibration forecasting for advancing rail safety and maintenance 4\.0\.IEEE Open Journal of Intelligent Transportation Systems6,pp\.1612–1624\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1)\.
- \[80\]O\. Laurendin, S\. Ambellouis, A\. Fleury, A\. Mahtani, S\. Chafik, and C\. Strauss\(2021\)Hazardous events detection in automatic train doors vicinity using deep neural networks\.In2021 17th IEEE International Conference on Advanced Video and Signal Based Surveillance \(AVSS\),pp\.1–7\.Cited by:[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.11.1.1.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p15.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p4.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p9.1),[Table 13](https://arxiv.org/html/2610.00363#S5.T13.2.2.1.1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I4.i2.p1.1)\.
- \[81\]H\. Lee, S\. Han, K\. Park, H\. Lee, and T\. Kwon\(2021\)Real\-time hybrid deep learning\-based train running safety prediction framework of railway vehicle\.Machines9\(7\),pp\.130\.Cited by:[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p2.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.2.1.1.1)\.
- \[82\]J\. Li, Y\. Fu, D\. Yan, S\. L\. Ma, and C\. Sham\(2024\)An edge ai system based on fpga platform for railway fault detection\.pp\.1387–1389\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I1.i3.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I5.i1.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I6.i5.I1.i1.p1.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p5.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p6.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.4.1.1.1)\.
- \[83\]Z\. Li, H\. Li, and L\. Meng\(2023\)Model compression for deep neural networks: a survey\.Computers12\(3\),pp\.60\.Cited by:[§7\.5](https://arxiv.org/html/2610.00363#S7.SS5.p1.1)\.
- \[84\]H\. Liu, Y\. Qin, and D\. Tu\(2026\)Bearing anomaly detection method based on multimodal fusion and self\-adversarial learning\.Sensors26\(2\),pp\.629\.Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p2.1)\.
- \[85\]H\. Liu, C\. Liu, S\. He, and J\. Chen\(2021\)Short\-term strong wind risk prediction for high\-speed railway\.IEEE Transactions on Intelligent Transportation Systems22\(7\),pp\.4243–4255\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S3.I1.i3.p1.1),[1st item](https://arxiv.org/html/2610.00363#S4.I2.i1.p1.1),[Table 4](https://arxiv.org/html/2610.00363#S4.T4.2.4.1.1.1),[§5\.2\.2](https://arxiv.org/html/2610.00363#S5.SS2.SSS2.p2.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 9](https://arxiv.org/html/2610.00363#S5.T9.2.2.1.1.1)\.
- \[86\]J\. Liu, S\. Chen, G\. Lederman,et al\.\(2019\)The DR\-Train dataset: dynamic responses, GPS positions and environmental conditions of two light rail vehicles in Pittsburgh\.Scientific Data6,pp\.146\.External Links:[Document](https://dx.doi.org/10.1038/s41597-019-0148-8)Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.18.1.1.1)\.
- \[87\]Q\. Liu, T\. Liang, and V\. Dinavahi\(2020\)Real\-time hierarchical neural network based fault detection and isolation for high\-speed railway system under hybrid ac/dc grid\.IEEE Transactions on Power Delivery35\(6\),pp\.2853–2864\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S4.I2.i2.p1.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p11.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.2.1.1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[Table 7](https://arxiv.org/html/2610.00363#S5.T7.2.4.1.1.1),[3rd item](https://arxiv.org/html/2610.00363#S7.I1.i3.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I3.i1.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I6.i5.I1.i1.p1.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.3.1.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.4.3.1.1)\.
- \[88\]R\. Liu, W\. Liu, Z\. Zheng, L\. Wang, L\. Mao, Q\. Qiu, and G\. Ling\(2023\)Anomaly\-gan: a data augmentation method for train surface anomaly detection\.Expert Systems with Applications228,pp\.120284\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S4.I1.i1.p1.1)\.
- \[89\]Y\. Liyuan, G\. Osman, S\. A\. Rahman, and M\. F\. Mustapha\(2024\)Multi\-modal fusion for multi\-task fuzzy detection of rail anomalies\.IEEE Access12,pp\.73925–73935\.Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p2.1)\.
- \[90\]A\. Lourenço, D\. Ribeiro, M\. Fernandes, and G\. Marreiros\(2024\)Time series data mining for railway wheel and track monitoring: a survey\.Neural Computing and Applications36\(27\),pp\.16707–16725\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p3.1),[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.3.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p2.1),[1st item](https://arxiv.org/html/2610.00363#S3.I1.i1.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S3.I3.i2.p1.1)\.
- \[91\]C\. Ma, J\. Mi, W\. Gao, and S\. Tao\(2024\)Cost\-sensitive dual\-stream residual networks for imbalanced classification\.\.Computers, Materials & Continua80\(3\)\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S6.I1.i2.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S6.I1.i2.p1.2)\.
- \[92\]P\. Malhotra, L\. Vig, G\. Shroff, P\. Agarwal,et al\.\(2015\)Long short\-term memory networks for anomaly detection in time series\.InProceedings,Vol\.89,pp\.94\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S3.I2.i2.p1.1),[1st item](https://arxiv.org/html/2610.00363#S6.I1.i1.p1.1)\.
- \[93\]A\. Mariscotti\(2020\)Data sets of measured pantograph voltage and current of European AC railways\.Data in Brief30,pp\.105477\.External Links:[Document](https://dx.doi.org/10.1016/j.dib.2020.105477)Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.17.1.1.1)\.
- \[94\]S\. Mohammadi, S\. S\. Karganroudi, M\. Adda, and H\. Ibrahim\(2025\)Rail defect classification with deep learning method\.Green Energy and Intelligent Transportation,pp\.100332\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p4.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[§5\.1](https://arxiv.org/html/2610.00363#S5.SS1.p2.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.15.1.1.1)\.
- \[95\]S\. Mohammadi, S\. S\. Karganroudi, M\. Adda, and H\. Ibrahim\(2026\)Toward smart railway maintenance: ai\-enhanced non\-destructive evaluation using vision transformers and cnns for fastener defect detection\.Green Energy and Intelligent Transportation5\(3\),pp\.100332\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p4.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.16.1.1.1)\.
- \[96\]E\. Nateghinia and L\. F\. Miranda\-Moreno\(2025\)Development of an unsupervised 3d lidar\-based methodology for automated safety monitoring of railway facilities\.International Journal of Transportation Science and Technology\.Cited by:[6th item](https://arxiv.org/html/2610.00363#S3.I3.i6.p1.1)\.
- \[97\]A\. Nechi, L\. Groth, S\. Mulhem, F\. Merchant, R\. Buchty, and M\. Berekovic\(2023\)Fpga\-based deep learning inference accelerators: where are we standing?\.ACM Transactions on Reconfigurable Technology and Systems16\(4\),pp\.1–32\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S7.I1.i3.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.4.3.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.4.4.1.1)\.
- \[98\]S\. Ning, F\. Ding, and B\. Chen\(2024\)Research on the method of foreign object detection for railway tracks based on deep learning\.Sensors24\(14\),pp\.4483\.Cited by:[Table 4](https://arxiv.org/html/2610.00363#S4.T4.2.3.1.1.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.13.1.1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I4.i1.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i1.p1.1)\.
- \[99\]M\. Niu, K\. Song, L\. Huang, Q\. Wang, Y\. Yan, and Q\. Meng\(2020\)Unsupervised saliency detection of rail surface defects using stereoscopic images\.IEEE Transactions on Industrial Informatics17\(3\),pp\.2271–2281\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.11.1.1.1)\.
- \[100\]K\. Oh, M\. Yoo, N\. Jin, J\. Ko, J\. Seo, H\. Joo, and M\. Ko\(2022\)A review of deep learning applications for railway safety\.Applied Sciences12\(20\),pp\.10572\.Cited by:[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.2.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p2.1)\.
- \[101\]OpenRailLab\(2025\)OpenRailLab: simulation platform for railway systems\.Note:Accessed: 2025\-07\-03External Links:[Link](https://www.openraillab.org/)Cited by:[5th item](https://arxiv.org/html/2610.00363#S4.I1.i5.p1.1)\.
- \[102\]M\. J\. Page, J\. E\. McKenzie, P\. M\. Bossuyt, I\. Boutron, T\. C\. Hoffmann, C\. D\. Mulrow, L\. Shamseer, J\. M\. Tetzlaff, E\. A\. Akl, S\. E\. Brennan,et al\.\(2021\)The prisma 2020 statement: an updated guideline for reporting systematic reviews\.bmj372\.Cited by:[§1\.2](https://arxiv.org/html/2610.00363#S1.SS2.p1.1)\.
- \[103\]G\. Pang, C\. Shen, L\. Cao, and A\. V\. D\. Hengel\(2021\)Deep learning for anomaly detection: a review\.ACM computing surveys \(CSUR\)54\(2\),pp\.1–38\.Cited by:[§2](https://arxiv.org/html/2610.00363#S2.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[§5\.1](https://arxiv.org/html/2610.00363#S5.SS1.p1.1),[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p1.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p3.1),[§5\.2](https://arxiv.org/html/2610.00363#S5.SS2.p1.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p4.1),[§5\.3](https://arxiv.org/html/2610.00363#S5.SS3.p1.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p9.1),[§5](https://arxiv.org/html/2610.00363#S5.p1.1),[1st item](https://arxiv.org/html/2610.00363#S6.I1.i1.p1.1),[Table 14](https://arxiv.org/html/2610.00363#S6.T14.2.8.4.1.1),[§6](https://arxiv.org/html/2610.00363#S6.p1.1),[item 2](https://arxiv.org/html/2610.00363#S7.I6.i2.p1.1)\.
- \[104\]M\. J\. Pappaterra, F\. Flammini, V\. Vittorini, and N\. Bešinović\(2021\)A systematic review of artificial intelligence public datasets for railway applications\.Infrastructures6\(10\),pp\.136\.External Links:[Document](https://dx.doi.org/10.3390/infrastructures6100136)Cited by:[5th item](https://arxiv.org/html/2610.00363#S4.I1.i5.p1.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p12.1),[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p9.1),[§8](https://arxiv.org/html/2610.00363#S8.p3.1)\.
- \[105\]A\. Phaphuangwittayakul, N\. Harnpornchai, F\. Ying, and J\. Zhang\(2024\)RailTrack\-davit: a vision transformer\-based approach for automated railway track defect detection\.Journal of Imaging10\(8\),pp\.192\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p4.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.17.1.1.1)\.
- \[106\]S\. Radosavljevic, A\. Rivero, A\. El Ouardi, and S\. R\. Flórez\(2025\)Railway track defect detection: from a comprehensive review of methods to new embedded system modeling perspectives\.IEEE Transactions on Instrumentation and Measurement75,pp\.1–25\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S7.I3.i1.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I4.i1.p1.1),[3rd item](https://arxiv.org/html/2610.00363#S7.I4.i3.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i1.p1.1),[item 1](https://arxiv.org/html/2610.00363#S7.I8.i1.p7.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p5.1),[§7\.5](https://arxiv.org/html/2610.00363#S7.SS5.p1.1),[§7](https://arxiv.org/html/2610.00363#S7.p1.1),[§7](https://arxiv.org/html/2610.00363#S7.p2.1),[§7](https://arxiv.org/html/2610.00363#S7.p3.1)\.
- \[107\]S\. Rakshit B\. Sandeepet al\.\(2022\)Railway track fault detection using deep neural networks\.In2022 IEEE 6th Conference on Information and Communication Technology \(CICT\),pp\.1–5\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p10.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.6.1.1.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p2.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.3.1.1.1),[§8](https://arxiv.org/html/2610.00363#S8.p7.1)\.
- \[108\]R\. P\. Ribeiroet al\.\(2023\)LSTM\-based failure prediction for railway rolling stock equipment\.Expert Systems with Applications218,pp\.119420\.External Links:[Document](https://dx.doi.org/10.1016/j.eswa.2023.119420)Cited by:[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p3.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.7.1.1.1)\.
- \[109\]O\. Rodríguez\-Alvaradoet al\.\(2024\)Automatic detection of railway faults using neural networks: a comparative study of transfer learning models and YOLOv11\.Infrastructures10\(1\),pp\.3\.External Links:[Document](https://dx.doi.org/10.3390/infrastructures10010003)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p2.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.5.1.1.1)\.
- \[110\]J\. Seo, Y\. Kim, J\. Ha,et al\.\(2024\)Unsupervised anomaly detection for earthquake detection on Korea high\-speed trains using autoencoder\-based deep learning models\.Scientific Reports14,pp\.639\.External Links:[Document](https://dx.doi.org/10.1038/s41598-024-51354-7)Cited by:[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p3.1),[Table 11](https://arxiv.org/html/2610.00363#S5.T11.2.6.1.1.1)\.
- \[111\]R\. Shafique, K\. Kanwal, and G\. S\. Choi\(2025\)Improved railway track faults detection using Mel\-frequency cepstral coefficient and constant\-Q transform features\.Scientific Reports15\.External Links:[Document](https://dx.doi.org/10.1038/s41598-025-14763-w)Cited by:[Table 5](https://arxiv.org/html/2610.00363#S5.T5.2.7.1.1.1)\.
- \[112\]R\. Shafiqueet al\.\(2023\)Deep learning\-based fault detection in railway wheelsets using time series analysis\.Mehran University Research Journal of Engineering and Technology42\(3\)\.Cited by:[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[Table 7](https://arxiv.org/html/2610.00363#S5.T7.2.5.1.1.1)\.
- \[113\]R\. Shafique, H\. U\. R\. Siddiqui, F\. Rustam, S\. Ullah, M\. A\. Siddique, E\. Lee, I\. Ashraf, and S\. Dudley\(2021\)A novel approach to railway track faults detection using acoustic analysis\.Sensors21\(18\),pp\.6221\.External Links:[Document](https://dx.doi.org/10.3390/s21186221)Cited by:[§5\.1\.1](https://arxiv.org/html/2610.00363#S5.SS1.SSS1.p1.1),[Table 5](https://arxiv.org/html/2610.00363#S5.T5.2.2.1.1.1)\.
- \[114\]R\. Shah, V\. Mittal, and M\. Lotwin\(2025\)Recent advances in vibration analysis for predictive maintenance of modern automotive powertrains\.Vibration8\(4\),pp\.68\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S4.I1.i3.p1.1)\.
- \[115\]M\. Shahin, F\. F\. Chen, A\. Hosseinzadeh, and N\. Zand\(2023\)Using machine learning and deep learning algorithms for downtime minimization in manufacturing systems: an early failure detection diagnostic service\.The international journal of advanced manufacturing technology128\(9\),pp\.3857–3883\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S6.I1.i2.p1.1),[Table 14](https://arxiv.org/html/2610.00363#S6.T14.2.3.4.1.1)\.
- \[116\]M\. Z\. Shaikh, E\. N\. Baro, and S\. Jatoi\(2025\)A contribution to reliable rail transport: ai\-powered real\-time wheel defect detection\.Scientific Reports\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.8.1.1.1)\.
- \[117\]M\. Z\. Shaikh, S\. Jatoi, E\. N\. Baro, B\. Das, S\. Hussain, and B\. S\. Chowdhry\(2025\)FaultSeg: a dataset for train wheel defect detection\.Scientific Data12\(1\),pp\.309\.External Links:[Document](https://dx.doi.org/10.1038/s41597-025-04557-0)Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.13.1.1.1)\.
- \[118\]Z\. Shang, L\. Sun, Y\. Xia, and W\. Zhang\(2021\)Vibration\-based damage detection for bridges by deep convolutional denoising autoencoder\.Structural Health Monitoring20\(4\),pp\.1880–1903\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S4.I1.i3.p1.1)\.
- \[119\]J\. Shim, J\. Koo, Y\. Park, and J\. Kim\(2022\)Anomaly detection method in railway using signal processing and deep learning\.Applied Sciences12\(24\),pp\.12901\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p3.1),[§1](https://arxiv.org/html/2610.00363#S1.p4.1)\.
- \[120\]J\. Shim, J\. Koo, and Y\. Park\(2023\)A methodology of condition monitoring system utilizing supervised and semi\-supervised learning in railway\.Sensors23\(22\),pp\.9075\.Cited by:[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.7.1.1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p1.1),[Table 7](https://arxiv.org/html/2610.00363#S5.T7.2.2.1.1.1),[item 1](https://arxiv.org/html/2610.00363#S7.I8.i1.p3.1),[item 1](https://arxiv.org/html/2610.00363#S7.I8.i1.p5.1)\.
- \[121\]M\. Shimizu, S\. Perinpanayagam, and B\. Namoano\(2022\)A real\-time fault detection framework based on unsupervised deep learning for prognostics and health management of railway assets\.IEEE Access10,pp\.96442–96458\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S4.I2.i1.p1.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.12.1.1.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.5.1.1.1),[item 2](https://arxiv.org/html/2610.00363#S7.I8.i2.p3.1),[item 2](https://arxiv.org/html/2610.00363#S7.I8.i2.p5.1)\.
- \[122\]H\. U\. R\. Siddiqui, A\. A\. Saleem, M\. A\. Raza, K\. Zafar, K\. Munir, and S\. Dudley\(2022\)IoT based railway track faults detection and localization using acoustic analysis\.IEEE Access10,pp\.106520–106533\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S3.I3.i3.p1.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.3.1.1.1),[§5\.1\.1](https://arxiv.org/html/2610.00363#S5.SS1.SSS1.p1.1),[Table 5](https://arxiv.org/html/2610.00363#S5.T5.2.5.1.1.1),[4th item](https://arxiv.org/html/2610.00363#S7.I1.i4.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.5.3.1.1)\.
- \[123\]D\. Signorino, D\. Giordano, A\. Mariscotti,et al\.\(2020\)Dataset of measured and commented pantograph electric arcs in DC railways\.Data in Brief31,pp\.105978\.External Links:[Document](https://dx.doi.org/10.1016/j.dib.2020.105978)Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.16.1.1.1)\.
- \[124\]J\. Simioni, E\. K\. Viegas, A\. Santin, and P\. Horchulhack\(2025\)An early exit deep neural network for fast inference intrusion detection\.InProceedings of the 40th ACM/SIGAPP Symposium on Applied Computing,pp\.730–737\.Cited by:[§7\.5\.3](https://arxiv.org/html/2610.00363#S7.SS5.SSS3.p1.1)\.
- \[125\]N\. Sinha, A\. El Rahman Shabayek, A\. Kacem, P\. Rostami, C\. Shneider, and D\. Aouada\(2024\)Hardware aware evolutionary neural architecture search using representation similarity metric\.InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision,pp\.2628–2637\.Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p5.1)\.
- \[126\]N\. Sinha, P\. Rostami, A\. El Rahman Shabayek, A\. Kacem, and D\. Aouada\(2024\)Multi\-objective hardware aware neural architecture search using hardware cost diversity\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\.8032–8039\.Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p5.1)\.
- \[127\]V\. Škvára, T\. Pevnỳ, and V\. Šmídl\(2023\)Is auc the best measure for practical comparison of anomaly detectors?\.arXiv preprint arXiv:2305\.04754\.Cited by:[1st item](https://arxiv.org/html/2610.00363#S6.I1.i1.p1.1),[Table 14](https://arxiv.org/html/2610.00363#S6.T14.2.7.4.1.1)\.
- \[128\]J\. Song, P\. Kumar, Y\. Kim, and H\. S\. Kim\(2024\)A fault detection system for wiring harness manufacturing using artificial intelligence\.Mathematics12\(4\),pp\.537\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p3.1)\.
- \[129\]J\. Sresakoolchai and S\. Kaewunruen\(2022\)Railway defect detection based on track geometry using supervised and unsupervised machine learning\.Structural health monitoring21\(4\),pp\.1757–1767\.Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.2.1.1.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p2.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[§5\.1](https://arxiv.org/html/2610.00363#S5.SS1.p2.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.2.1.1.1)\.
- \[130\]P\. Su, K\. Wang, X\. Zeng, S\. Tang, D\. Chen, D\. Qiu, and X\. Wang\(2020\)Adapting object detectors with conditional domain normalization\.InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16,pp\.403–419\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S4.I1.i2.p1.1)\.
- \[131\]Q\. Sun and Z\. Ge\(2021\)A survey on deep learning for data\-driven soft sensors\.IEEE Transactions on Industrial Informatics17\(9\),pp\.5853–5866\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1)\.
- \[132\]T\. Sun, B\. Guo, T\. Ruan, X\. Zhou, D\. Bai, H\. Yu, and Y\. Wang\(2025\)RailFDNet: a hybrid supervision and feature discrepancy enhancement model for railway anomalous object detection\.Expert Systems with Applications275,pp\.127005\.Cited by:[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p7.1),[Table 13](https://arxiv.org/html/2610.00363#S5.T13.2.5.1.1.1)\.
- \[133\]Y\. Sun, Y\. Cao, and L\. Ma\(2020\)A fault diagnosis method for train plug doors via sound signals\.IEEE Intelligent Transportation Systems Magazine13\(3\),pp\.107–117\.Cited by:[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.5.1.1.1),[§5\.1\.1](https://arxiv.org/html/2610.00363#S5.SS1.SSS1.p1.1),[Table 5](https://arxiv.org/html/2610.00363#S5.T5.2.3.1.1.1)\.
- \[134\]R\. Tagiew, M\. Köppel, K\. Schwalbe,et al\.\(2023\)OSDaR23: open sensor data for rail 2023\.InInternational Conference on Robotics and Automation Engineering,pp\.270–276\.Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.15.1.1.1)\.
- \[135\]S\. Tamminaet al\.\(2019\)Transfer learning using vgg\-16 with deep convolutional neural network for classifying images\.International Journal of Scientific and Research Publications \(IJSRP\)9\(10\),pp\.143–150\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p2.1)\.
- \[136\]Q\. Tang, J\. Liang, and F\. Zhu\(2023\)A comparative review on multi\-modal sensors fusion based on deep learning\.Signal Processing213,pp\.109165\.Cited by:[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p1.1),[Table 7](https://arxiv.org/html/2610.00363#S5.T7.2.6.1.1.1)\.
- \[137\]R\. Tang, L\. De Donato, N\. Besinović, F\. Flammini, R\. M\. Goverde, Z\. Lin, R\. Liu, T\. Tang, V\. Vittorini, and Z\. Wang\(2022\)A literature review of artificial intelligence applications in railway systems\.Transportation Research Part C: Emerging Technologies140,pp\.103679\.Cited by:[Table 1](https://arxiv.org/html/2610.00363#S2.T1.2.4.1.1.1),[§2](https://arxiv.org/html/2610.00363#S2.p2.1),[3rd item](https://arxiv.org/html/2610.00363#S3.I1.i3.p1.1)\.
- \[138\]E\. A\. Teklu and C\. M\. Abdissa\(2023\)Genetic algorithm tuned super twisting sliding mode controller for suspension of maglev train with flexible track\.IEEE Access11\(\),pp\.30955–30969\.External Links:[Document](https://dx.doi.org/10.1109/ACCESS.2023.3262416)Cited by:[§8](https://arxiv.org/html/2610.00363#S8.p6.1)\.
- \[139\]D\. Tsiktsiris, A\. Lalas, M\. Dasygenis, and K\. Votis\(2024\)Multimodal abnormal event detection in public transportation\.IEEE Access\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S6.I1.i3.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I1.i2.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I3.i1.p1.1),[3rd item](https://arxiv.org/html/2610.00363#S7.I4.i3.p1.1),[4th item](https://arxiv.org/html/2610.00363#S7.I4.i4.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I5.i1.p1.1),[2nd item](https://arxiv.org/html/2610.00363#S7.I6.i5.I1.i2.p1.1),[1st item](https://arxiv.org/html/2610.00363#S7.I6.i6.I1.i1.p1.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p3.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p4.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p5.1),[§7\.1](https://arxiv.org/html/2610.00363#S7.SS1.p6.1),[Table 16](https://arxiv.org/html/2610.00363#S7.T16.2.6.1.1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.3.3.1.1)\.
- \[140\]B\. Veloso, J\. Gama, R\. P\. Ribeiro, and P\. M\. Pereira\(2022\)The MetroPT dataset for predictive maintenance\.Scientific Data9\(1\),pp\.764\.External Links:[Document](https://dx.doi.org/10.1038/s41597-022-01877-3)Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.10.1.1.1)\.
- \[141\]C\. Wang and J\. Liu\(2021\)An efficient anomaly detection for high\-speed train braking system using broad learning system\.IEEE Access9,pp\.63825–63832\.Cited by:[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.6.1.1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p1.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p2.1),[Table 7](https://arxiv.org/html/2610.00363#S5.T7.2.3.1.1.1)\.
- \[142\]L\. Wang K\. Zhanget al\.\(2024\)An undercarriage image driven anomaly detection method for metro vehicle based on adversarial memory enhancement\.Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit\.External Links:[Document](https://dx.doi.org/10.1177/09544097231201519)Cited by:[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p2.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[Table 12](https://arxiv.org/html/2610.00363#S5.T12.2.3.1.1.1)\.
- \[143\]L\. Wang, K\. Zhang, Q\. Zheng, G\. Ding, W\. Zhang, D\. Chen, and B\. Liu\(2024\)MemFormer: a memory based unified model for anomaly detection on metro railway tracks\.Expert Systems with Applications238,pp\.121824\.External Links:[Document](https://dx.doi.org/10.1016/j.eswa.2023.121824)Cited by:[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p3.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[Table 12](https://arxiv.org/html/2610.00363#S5.T12.2.4.1.1.1)\.
- \[144\]S\. Wang, J\. Cao, and S\. Y\. Philip\(2020\)Deep learning for spatio\-temporal data mining: a survey\.IEEE transactions on knowledge and data engineering34\(8\),pp\.3681–3700\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p5.1)\.
- \[145\]T\. Wang, Z\. Zhang, and K\. L\. Tsui\(2024\)A hierarchical transfer\-generative framework for automating multianalytical tasks in rail surface defect inspection\.IEEE Internet of Things Journal\.External Links:[Document](https://dx.doi.org/10.1109/JIOT.2024.3374751)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p4.1)\.
- \[146\]T\. Wang, Z\. Zhang, and K\. Tsui\(2022\)A deep generative approach for rail foreign object detections via semisupervised learning\.IEEE Transactions on Industrial Informatics19\(1\),pp\.459–468\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S4.I2.i3.p1.1),[Table 4](https://arxiv.org/html/2610.00363#S4.T4.2.2.1.1.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p2.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p14.1),[Table 12](https://arxiv.org/html/2610.00363#S5.T12.2.2.1.1.1)\.
- \[147\]T\. Wang, Z\. Zhang, and M\. Xie\(2024\)An anomaly\-free representation learning approach for efficient railway foreign object detection\.IEEE Transactions on Industrial Informatics20\(10\),pp\.12125–12135\.Cited by:[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p3.1)\.
- \[148\]X\. Wang, J\. Dai, and X\. Liu\(2025\)A spatial\-temporal neural network based on resnet\-transformer for predicting railroad broken rails\.Advanced Engineering Informatics65,pp\.103126\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S3.I4.i2.p1.1)\.
- \[149\]Y\. Wang, X\. Du, Z\. Lu, Q\. Duan, and J\. Wu\(2022\)Improved lstm\-based time\-series anomaly detection in rail transit operation environments\.IEEE Transactions on Industrial Informatics18\(12\),pp\.9027–9036\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p10.1),[Table 3](https://arxiv.org/html/2610.00363#S4.T3.2.8.1.1.1),[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p3.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p3.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p13.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.3.1.1.1),[4th item](https://arxiv.org/html/2610.00363#S7.I1.i4.p1.1),[Table 17](https://arxiv.org/html/2610.00363#S7.T17.4.5.3.1.1)\.
- \[150\]Y\. Weng, J\. Yang, C\. Zhang, J\. He, C\. Peng, L\. Jia, and H\. Xiang\(2025\)An improved deeplabv3\+ railway track extraction algorithm based on densely connected and attention mechanisms\.Scientific Reports15\(1\),pp\.2556\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S7.I5.i2.p1.1)\.
- \[151\]J\. Wróbel, P\. Bury, M\. Zając, A\. Kierzkowski, S\. Tubek, and J\. Blaut\(2024\)Fault detection and diagnostic methods for railway systems–a literature survey\.Advances in Science and Technology\. Research Journal18\(6\)\.Cited by:[§1](https://arxiv.org/html/2610.00363#S1.p2.1)\.
- \[152\]Y\. Wu, P\. Chen, Y\. Qin, Y\. Qian, F\. Xu, and L\. Jia\(2023\)Automatic railroad track components inspection using hybrid deep learning framework\.IEEE Transactions on Instrumentation and Measurement72,pp\.1–15\.External Links:[Document](https://dx.doi.org/10.1109/TIM.2022.3225031)Cited by:[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p8.1),[Table 13](https://arxiv.org/html/2610.00363#S5.T13.2.6.1.1.1)\.
- \[153\]L\. Xiong, G\. Jing, J\. Wang, X\. Liu, and Y\. Zhang\(2023\)Detection of rail defects using ndt methods\.Sensors23\(10\),pp\.4627\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p4.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.14.1.1.1)\.
- \[154\]L\. Xu and W\. Zhai\(2020\)Train–track coupled dynamics analysis: system spatial variation on geometry, physics and mechanics\.Railway Engineering Science28\(1\),pp\.36–53\.Cited by:[2nd item](https://arxiv.org/html/2610.00363#S4.I1.i2.p1.1)\.
- \[155\]H\. Yan, C\. Lin, Z\. Xu, N\. Guo, and A\. Qing\(2024\)Pantograph unsupervised anomaly detection network based on reverse distillation\.InProceedings of the 2024 8th International Conference on Advances in Image Processing \(ICAIP\),pp\.55–60\.Cited by:[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p3.1),[Table 12](https://arxiv.org/html/2610.00363#S5.T12.2.6.1.1.1)\.
- \[156\]M\. Yilmazer and M\. Karakose\(2022\)Mask r\-cnn architecture based railway fastener fault detection approach\.In2022 International Conference on Decision Aid Sciences and Applications \(DASA\),pp\.1363–1366\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p10.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.5.1.1.1),[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.12.1.1.1)\.
- \[157\]T\. Yokouchi and M\. Kondo\(2021\)LSTM\-based anomaly detection for railway vehicle air\-conditioning unit using monitoring data\.InIECON 2021 – 47th Annual Conference of the IEEE Industrial Electronics Society,pp\.1–6\.External Links:[Document](https://dx.doi.org/10.1109/IECON48115.2021.9589473)Cited by:[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p3.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 8](https://arxiv.org/html/2610.00363#S5.T8.2.4.1.1.1)\.
- \[158\]P\. You, P\. Chen, X\. Li, and A\. Bian\(2026\)Gated memory\-guided multi\-scale spatio–temporal–spectral feature fusion network for unsupervised internet of things time series anomaly detection\.Engineering Applications of Artificial Intelligence169,pp\.114104\.Cited by:[§2](https://arxiv.org/html/2610.00363#S2.p1.1)\.
- \[159\]Z\. Yuan, S\. Zhu, C\. Chang, X\. Yuan, Q\. Zhang, and W\. Zhai\(2021\)An unsupervised method based on convolutional variational auto\-encoder and anomaly detection algorithms for light rail squat localization\.Construction and Building Materials313,pp\.125563\.Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.8.1.1.1),[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p5.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p14.1),[Table 11](https://arxiv.org/html/2610.00363#S5.T11.2.7.1.1.1)\.
- \[160\]Y\. Yürekli, C\. Özarpa, and İ\. Avcı\(2025\)Real\-time railway hazard detection using distributed acoustic sensing and hybrid ensemble learning\.Sensors25\(13\),pp\.3992\.Cited by:[6th item](https://arxiv.org/html/2610.00363#S3.I3.i6.p1.1)\.
- \[161\]Z\. Zamanzadeh Darban, G\. I\. Webb, S\. Pan, C\. Aggarwal, and M\. Salehi\(2024\)Deep learning for time series anomaly detection: a survey\.ACM Computing Surveys57\(1\),pp\.1–42\.Cited by:[3rd item](https://arxiv.org/html/2610.00363#S3.I2.i3.p1.1),[4th item](https://arxiv.org/html/2610.00363#S3.I2.i4.p1.1),[§3\.2](https://arxiv.org/html/2610.00363#S3.SS2.p1.1),[§3\.4](https://arxiv.org/html/2610.00363#S3.SS4.p1.1),[§3](https://arxiv.org/html/2610.00363#S3.p1.1),[§5\.2\.1](https://arxiv.org/html/2610.00363#S5.SS2.SSS1.p1.1),[§5\.2\.2](https://arxiv.org/html/2610.00363#S5.SS2.SSS2.p1.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p3.1),[§5\.2](https://arxiv.org/html/2610.00363#S5.SS2.p1.1),[§5\.2](https://arxiv.org/html/2610.00363#S5.SS2.p2.1),[§5\.3\.1](https://arxiv.org/html/2610.00363#S5.SS3.SSS1.p4.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p1.1),[§5\.4](https://arxiv.org/html/2610.00363#S5.SS4.p9.1),[§5](https://arxiv.org/html/2610.00363#S5.p1.1),[item 7](https://arxiv.org/html/2610.00363#S7.I6.i7.p1.2),[§8](https://arxiv.org/html/2610.00363#S8.p7.1)\.
- \[162\]O\. Zendel, M\. Murschitz, M\. Zeilinger, D\. Steininger, S\. Abbasi, and C\. Beleznai\(2019\)RailSem19: a dataset for semantic rail scene understanding\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops,Cited by:[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.14.1.1.1)\.
- \[163\]C\. Zhanget al\.\(2024\)Intelligent prediction of rail corrugation evolution trend based on self\-attention bidirectional TCN and GRU\.Intelligent Robotics4\(4\),pp\.548–565\.External Links:[Document](https://dx.doi.org/10.20517/ir.2024.20)Cited by:[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p2.1),[§5\.2\.3](https://arxiv.org/html/2610.00363#S5.SS2.SSS3.p4.1),[Table 10](https://arxiv.org/html/2610.00363#S5.T10.2.4.1.1.1)\.
- \[164\]Y\. Zhang, Z\. Yu, L\. Zhu, B\. Guo, and Y\. Wang\(2025\)Rail\-patchcore: unsupervised learning\-based detection of visual anomalies in the railway\-turnout environment: y\. zhang et al\.\.Applied Intelligence55\(6\),pp\.408\.Cited by:[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p3.1),[§5\.3\.2](https://arxiv.org/html/2610.00363#S5.SS3.SSS2.p5.1),[Table 12](https://arxiv.org/html/2610.00363#S5.T12.2.5.1.1.1)\.
- \[165\]Z\. Zhang, S\. Yu, S\. Yang, Y\. Zhou, and B\. Zhao\(2021\)Rail\-5k: a real\-world dataset for rail surface defects detection\.arXiv preprint arXiv:2106\.14366\.Cited by:[§4\.2](https://arxiv.org/html/2610.00363#S4.SS2.p3.1),[Table 2](https://arxiv.org/html/2610.00363#S4.T2.2.12.1.1.1)\.
- \[166\]H\. Zhou, Y\. Liu, J\. Wang, A\. Fang, and A\. Mohammadzadeh\(2025\)Railway fastener defect detection using RFD\-DETR: a lightweight real\-time transformer\-based approach\.PLOS ONE\.External Links:[Document](https://dx.doi.org/10.1371/journal.pone.0331513)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[§5\.1\.3](https://arxiv.org/html/2610.00363#S5.SS1.SSS3.p3.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.10.1.1.1)\.
- \[167\]A\. Zhukov, A\. Rivero, J\. Benois\-Pineau, A\. Zemmari, and M\. Mosbah\(2024\)A hybrid system for defect detection on rail lines through the fusion of object and context information\.Sensors24\(4\),pp\.1171\.Cited by:[§5\.1\.2](https://arxiv.org/html/2610.00363#S5.SS1.SSS2.p3.1),[Table 6](https://arxiv.org/html/2610.00363#S5.T6.2.11.1.1.1)\.