Deep Learning for Anomaly Detection in Railway Systems: A Structured Survey

1. Introduction

• Heterogeneous railway domains and imbalanced, temporally structured data challenge traditional detection, while missed faults can disrupt service or create safety-critical incidents.The survey identifies faults across rolling stock, infrastructure, and environmental interactions, and notes that rare safety-critical events further degrade detection reliability.Source: §1 · pp. 2–3

1.1. Research Questions

• The six research questions span anomaly characterization, data and datasets, detection paradigms, evaluation, deployment constraints, and decision support for selecting methods.They address the pipeline from characterization through deployment in railway systems.Source: §1.1 · p. 4

1.2.1. Scope and Time Window

• The review covers peer-reviewed journal and conference papers published from January 2019 through 2026.The period is associated in the paper with widespread adoption of deep learning in railway monitoring and condition-based maintenance.Source: §1.2.1 · p. 4

1.2.2. Search Strategy

• Searches of IEEE Xplore and Google Scholar used railway, anomaly or fault detection, and deep-learning terms across titles, abstracts, and keywords.Results were filtered to English journal articles and conference papers published during 2019–2026.Source: §1.2.2 · p. 5

1.2.3. Inclusion and Exclusion Criteria

• Eligible studies had to address railway anomaly or fault detection using deep learning, report quantitative results, and provide methodological detail in peer-reviewed venues.Studies without methodological transparency, quantitative evaluation, railway relevance, or accessible full text were excluded.Source: §1.2.3 · p. 5

1.2. Survey Methodology

• The survey uses a PRISMA-inspired, depth-over-breadth review protocol to retain technically detailed studies for consistent analysis.The review assesses model design, data characteristics, evaluation protocols, and relevance.Source: §1.2 · p. 4, §1.2.4 · p. 6

1.2.4. Selection Process

• The PRISMA-inspired screening excluded 100 records and 44 full-text studies, leaving a final corpus of 68 studies.Full-text exclusions comprised 14 studies with insufficient experimental validation, 21 with limited methodological transparency, and 9 with marginal relevance.Source: Figure 2 · p. 6, §1.2.4 · p. 6

Existing Railway AI Surveys

• Broad railway AI reviews cover multiple domains but vary in depth on deployment feasibility, modality-specific challenges, multimodal fusion, and learning strategies.Some focus on predictive maintenance for a narrow subsystem or provide high-level domain overviews.Source: §2 · p. 7

Modality-Specific Surveys

• Vision- and modality-focused reviews report controlled-setting results but offer limited coverage of multimodal generalization, system integration, real-time feasibility, or embedded implementation.Other reviews focus on predictive maintenance components, infrastructure maintenance, or classical AI rather than deep-learning anomaly detection.Source: §2 · p. 7

Synthesis and Positioning

• Prior railway anomaly surveys remain fragmented by subsystem and modality, often catalog methods without actionable paradigm guidance or integration of deployment constraints.Feasibility depends on latency, memory and power budgets, reliability requirements, and operational risk.Source: §2 · p. 8

Contributions and Novelty

• This survey combines a four-axis anomaly taxonomy, paradigm-level method organization, expanded evaluation analysis, and deployment-oriented decision support.Its framework links method selection and deployment configurations to anomaly properties, data regimes, and real-world railway constraints.Source: §2 · pp. 8–9

3. Railway Anomalies Categorization

• The four-axis taxonomy links anomaly characteristics to detection strategies and deployment considerations across diverse railway domains.Its axes are location, manifestation in data, sensing modality, and temporal behavior.Source: §3 · p. 9, Figure 3 · p. 9

3.1. Categorization by Location of Occurrence

• Location distinguishes infrastructure faults, rolling-stock failures, and environmental or behavioral anomalies.Examples include rail cracks, door or traction-system faults, trespassing, and extreme weather.Source: §3.1 · p. 10

3.2. Categorization by Manifestation in Data

• Data manifestation distinguishes point, contextual, collective, and trend anomalies, from isolated deviations to persistent distribution shifts.For example, a brake temperature may be normal during descent but anomalous on level track; gradual wheel wear can signal a trend anomaly.Source: §3.2 · p. 10

3.3. Categorization by Sensing Modality

• Sensing modalities include images and video, time series, acoustic, structured or tabular, textual or log data, and emerging modalities such as LiDAR and distributed acoustic sensing.Non-stationary time series have temporal dependencies, while acoustic data may be transformed into spectrograms for analysis.Source: §3.3 · pp. 10–11

3.4. Categorization by Temporal Behavior

• Temporal behavior separates snapshot-detectable static anomalies from sequential anomalies that require models capable of learning evolving patterns.CNNs are used for static visual inspection, while RNNs, LSTMs, TCNs, and transformers are commonly employed for sequential anomalies.Source: §3.4 · p. 11

4.2. Datasets for Railway Anomaly Detection

• Fewer than one third of the 32 surveyed datasets are fully publicly accessible, while safety and confidentiality restrict access to operational railway data.The datasets span infrastructure, rolling stock, and environmental monitoring, with differing origins, access, and sensing modalities.Source: §4.1 · p. 12, §4.2 · p. 12

4.1. Data-Centric Challenges and Preprocessing Strategies

• Fault rarity, operational variability, sensor noise or missing data, and multimodal misalignment challenge model training and cross-fleet generalization.The survey describes oversampling and generative augmentation, context-aware normalization and domain adaptation, denoising and interpolation, and cross-modal alignment as responses to these issues.Source: §4.1 · pp. 11–12

4.2. Datasets for Railway Anomaly Detection

• The survey classifies datasets by origin, access, anomaly location, and sensing modality, and records exact sample counts when disclosed or NR otherwise.Origin categories are real-world, simulated, synthetic, and laboratory; the surveyed tables cover infrastructure, rolling stock, and environmental or behavioral datasets.Source: §4.2 · pp. 12–15, Table 2 · p. 14
• Most studies use proprietary or task-specific datasets with distinct fault definitions, sensing configurations, and operating conditions, making cross-study comparison unreliable.The survey notes differing reported performance across vision-based infrastructure and rolling-stock monitoring datasets, which operate under different conditions.Source: §4.2 · p. 16
• Laboratory-scale or simulated datasets often yield optimistic results that do not transfer to in-service conditions.The survey identifies dataset characteristics as influencing reported performance alongside model architecture.Source: §4.2 · p. 16
• Poor dataset documentation and inconsistent anomaly definitions, train-test splits, and cross-dataset protocols weaken reproducibility and fair assessment.The survey identifies open, annotated, subsystem-specific benchmarks as an ongoing challenge.Source: §4.2 · p. 16

• Railway detection tasks range from binary or multiclass fault classification to spatial localization, early fault prediction, and remaining useful life estimation.These tasks differ in supervision requirements, temporal modeling complexity, and operational objectives.Source: §5 · p. 16
• The survey organizes 68 studies into classification-, reconstruction-, prediction-, and hybrid-based paradigms according to their core detection principle.Paradigms can support different tasks depending on data availability, labeling conditions, and operational constraints.Source: §5 · pp. 16–17
• The four-paradigm taxonomy provides the survey's structural backbone for analyzing methods across railway subsystems, sensing modalities, and deployment contexts.Source: §5 · p. 17

5.1. Classification-Based Approaches

• Classification learns explicit decision boundaries for predefined normal or fault classes, and is most applicable when fault categories are defined and labeled data are available.The typical pipeline preprocesses raw sensor inputs, extracts discriminative features, and applies a classifier and decision layer to produce a fault label.Source: §5.1 · p. 17, Figure 4 · p. 17

5.1.1. Classical Machine Learning (ML) Classifiers

• Classical SVM, Random Forest, and MLP pipelines remain competitive on small acoustic and vibration datasets, especially when labels are scarce or computation is constrained.Their reliance on handcrafted features limits scalability to complex or high-dimensional inputs.Source: §5.1.1 · p. 18, Table 5 · p. 18

5.1.2. Vision-Based Deep Classifiers

• Vision classifiers cover infrastructure and wheel defects, using transfer learning, ensembles, YOLO detectors, vision transformers, and hybrid localization pipelines.The surveyed examples include 99% accuracy for an ensemble CNN on railway defect images, 98% precision for Mask R-CNN missing-fastener detection from drone imagery, and DeiT 95.0% vs VGG16 91.5% for fastener classification.Source: §5.1.2 · pp. 18–19, Table 6 · p. 19

5.1.3. Time-Series Signal Classifiers

• Time-series classifiers use spectrograms, time-frequency representations, or raw sensor signals for rolling-stock faults and passenger behavioral anomalies.Reported architectures include CNNs, GRUs, 1D-CNNs, LSTMs, and multimodal visual-acoustic classifiers.Source: §5.1.3 · p. 20, Table 7 · p. 20
• Classification can offer high accuracy, interpretable outputs, and embedded deployment, but requires curated fault labels and may silently miss out-of-distribution anomalies.Class imbalance requires mitigation, and heterogeneous metrics leave standardized benchmarking unresolved.Source: §5.1.3 · pp. 20–21
• Transformers consistently outperform CNN baselines across visual inspection tasks, but require higher computational resources.Vision classifiers dominate the surveyed literature, while time-series classifiers show strong results on rolling-stock subsystems but are often tested on simulated or single-operator data.Source: §5.1.3 · p. 20

5.2. Prediction-Based Approaches

• Prediction-based detection learns normal temporal dynamics and flags deviations, typically using normal operational data when fault examples are rare or unavailable.A sliding window predicts the next state; prediction error is thresholded to trigger a fault alert, enabling detection of unseen fault types without fault labels.Source: §5.2 · p. 21, Figure 5 · p. 21

5.2.1. Recurrent Sequence Predictors

• Recurrent predictors, especially LSTMs, model sequential sensor data; railway applications include HVAC, compressor, track-geometry, and suspension monitoring, including failure anticipation hours ahead.Standard RNNs have gradient instability over long sequences, while LSTM gated memory addresses that limitation.Source: §5.2.1 · p. 22, Table 8 · p. 23

5.2.2. Attention and Transformer Predictors

• Attention and transformer predictors weight informative channels or time intervals, with applications to wind-risk prediction and axle-bearing fault prediction from vibration signals.Transformers use self-attention for parallel computation and long-range temporal dependencies; the surveyed IEWT model improved performance over recurrent baselines on long sequences.Source: §5.2.2 · p. 22, Table 9 · p. 23

5.2.3. Convolutional Temporal Predictors

• Convolutional temporal predictors capture local sequence patterns at lower computational cost than recurrent models, supporting embedded and edge use.A 1D-CNN predicted rail corrugation profiles from vehicle acceleration with prediction errors below 5%; a hybrid attention-TCN-GRU model outperformed individual recurrent and convolutional baselines on a multi-domain feature set.Source: §5.2.3 · p. 24, Table 10 · p. 24
• Prediction errors are sensitive to sensor noise and operational drift, threshold calibration must balance sensitivity and false alarms, and models may degrade across speed, load, or environmental conditions.Source: §5.2.3 · p. 24
• LSTM predictors dominate surveyed railway studies, while attention and transformer models are emerging alternatives for long-range dependencies and convolutional predictors offer lower-cost resource-constrained deployment.Most studies use proprietary or single-operator data, limiting conclusions about cross-fleet generalizability.Source: §5.2.3 · pp. 24–25

5.3. Reconstruction-Based Approaches

• Training on normal data only lets reconstruction-based methods detect deviations without fault labels across railway signals, images, and vibrations.Models compare inputs with reconstructions and use the resulting error as an anomaly score.Source: §5.3 · p. 25

5.3.1. Standard and Recurrent Autoencoders

• LSTM-AEs outperformed SAEs in F1-score and recall on MetroPT-3 by capturing temporal dependencies in pressure and current signals.Other studies applied autoencoders to analog and digital APU sensor streams, brake-unit current signals, and online multivariate monitoring.Source: §5.3.1 · p. 26
• A VAE detects anomalies through low reconstruction likelihood or deviation from its learned probabilistic latent distribution.Railway VAE applications remain less explored than recurrent autoencoders, despite reported sensitivity to subtle structural anomalies.Source: §5.3.1 · p. 26, §5.3.2 · p. 28
• A CVAE identified rail-surface squat defects as shallow as 0.03 mm from axle-box acceleration without fault labels.An Elliptic Envelope detector used latent features to identify deviations from the normal distribution.Source: §5.3.1 · p. 26

5.3.2. Generative Adversarial Reconstruction Models

• Reconstruction methods can miss anomalies that resemble normal data, rely on thresholds that may not transfer across conditions or fleets, and require careful GAN stabilization.Most surveyed reconstruction studies use single-operator or proprietary datasets, limiting conclusions about cross-fleet generalizability.Source: §5.3.2 · p. 28
• Transformer and memory-augmented reconstruction models are emerging for visual inspection, with reported robustness against the anomaly reconstruction problem.A reverse-distillation approach for pantograph inspection achieved image- and pixel-level reconstruction scores competitive with supervised baselines.Source: §5.3.2 · p. 28

5.4. Hybrid Approaches

• Hybrid pipelines combine reconstruction with prediction, or supervised detection with unsupervised reconstruction or rule-based reasoning.Parallel or sequential branches can fuse outputs into a unified score or decision.Source: §5.4 · pp. 29–30
• A door-hazard cGAN jointly reconstructs video frames and predicts optical flow, scoring deviations in appearance and motion.The dual objective detects obstructed doors, irregular passenger movements, and mechanical faults that either objective alone would miss.Source: §5.4 · p. 30
• A pantograph-catenary system combines classification, reconstruction, and prediction across RGB, infrared, ultraviolet, and vibration sensing.Its specialized modules jointly interpret object detections, signal deviations, and behavioral forecasts.Source: §5.4 · p. 30
• A GAN and structural-similarity meta-learning framework reduced inspection workload by over 99.7% across five high-speed railway lines in China.The framework detects anomalies in high-speed rail inspection from limited labeled examples.Source: §5.4 · p. 30
• Hybrid approaches combine detection paradigms to broaden anomaly coverage, improve robustness to label scarcity, and generalize better than single-paradigm methods.They are suited to safety-critical scenarios where both known and unknown faults must be handled.Source: §5.4 · pp. 29–30
• Hybrid deployment faces greater design, training, and calibration complexity, while fusion and multimodal processing can raise compute and memory costs.Under real-time constraints, one fusion branch may dominate the anomaly score, and resource demands may limit embedded deployment.Source: §5.4 · pp. 30–31

6.1. Metrics by Detection Task

• Metric choice should match model output and detection objective; Table 14 maps metrics to tasks and notes their suitability or limits in railway settings.The supplied text does not provide the table's individual metric mappings.Source: §6.1 · p. 32, Table 14 · p. 32

6.2. Why Standard Metrics Are Insufficient for Railway Systems

• With railway fault rates typically below 1%, a model that detects no faults can still report 99% accuracy, while false negatives may cost EUR 100,000 or more.PR-AUC focuses on the anomaly class; cost-sensitive evaluation assigns asymmetric penalties, and latency, memory, and throughput matter for deployment.Source: §6.2 · p. 33

6.3. Evaluation Practices in the Surveyed Literature

• Accuracy remains the most reported metric, PR-AUC appears in fewer than 8% of studies, and latency is reported in 7 studies versus memory or energy in 5.The survey recommends reporting F1-score, PR-AUC, false negative rate for safety-critical scenarios, target-platform latency, and cost-sensitive evaluation when costs can be estimated.Source: §6.3 · p. 33, Table 15 · p. 34

7.1. Deployment Evidence in the Surveyed Literature

• Only three studies reported power; measured latency ranged from 0.7 ms on an FPGA to approximately 1200 ms on a Jetson AGX Xavier, while embedded real-time inference was feasible for specific tasks.The authors say the evidence remains too sparse to generalize deployability across subsystems or operating conditions.Source: §7.1 · pp. 35–36

7.2. Hardware Platforms for Railway Deployment

• CPUs/GPUs primarily serve training and offline evaluation; Jetson edge platforms trade performance for deployability, FPGAs offer deterministic low latency at development cost, and NPUs remain underrepresented.Embedded CPUs are limited to lightweight or low-frequency inference; NPUs are projected as energy-efficient candidates for future onboard use.Source: §7.2 · p. 36

7.3. Deployment Architectures: Edge, Cloud, and Hybrid

• Edge suits predictable responses and local autonomy, cloud supports fleet analytics and retraining, while hybrid systems combine fast edge screening with centralized refinement.The survey identifies edge for safety-critical events, cloud for slow degradation and retrospective analysis, and hybrid for immediate alerts plus fleet context.Source: §7.3 · pp. 37–38

7.4. Real-Time and Non-Real-Time Anomalies

• Hard real-time events require bounded latency and favor deterministic FPGAs, whereas gradual degradations permit offline or batch analysis prioritizing robustness and interpretability.The survey distinguishes hard real-time tasks, where deadline misses are unacceptable, from soft real-time tasks, where occasional misses degrade performance.Source: §7.4 · p. 38

7.5.1. Latency-oriented techniques

• Lightweight models, input reduction, accelerated runtimes, and parallel pipelines target latency; YOLOv5nu achieved 20–50 ms mean latency on Jetson AGX Xavier across dataset complexities.FP16 inference reduced latency by 40–50% versus FP32 with negligible accuracy loss in most tested cases.Source: §7.5.1 · pp. 38–39

7.5.2. Size-oriented techniques

• Quantization is the most common compression technique: Fu et al. report 93.4% accuracy with 0.3% degradation using 1-bit weights and activations, while Li et al. report no accuracy loss.Tsiktsiris et al. report an approximately 0.5% accuracy drop with INT8 GPU-edge quantization; pruning is rare and distillation is essentially absent.Source: §7.5.2 · p. 39

7.5.3. Summary

• Railway deployments mainly use quantization and lightweight architectures; pruning is rare, distillation remains unexplored, and systematic multi-technique compression pipelines are absent.The survey identifies hardware-aware architecture search and hardware–software co-design as open directions tailored to railway constraints.Source: §7.5.3 · pp. 39–40

7.6.1. Stepwise Selection Procedure

• The seven-step procedure links anomaly characterization and label availability to constraints, paradigm and hardware selection, model optimization, and operationally relevant evaluation.It recommends evaluating across operational variability, including fleets, routes, seasons, speeds, and maintenance states.Source: §7.6.1 · pp. 41–42

7.6.2. Analytical Comparison of Detection Paradigms

• Table 18 rates four detection paradigms across ten engineering criteria; suitability depends on whether each paradigm typically meets a criterion or needs adaptation.The rating guide defines High, Medium, and Low suitability, with examples including classification's limited coverage of unseen anomalies.Source: §7.6.2 · p. 43, Table 18 · p. 44

7.6.3. Illustrative Case Studies

• For labeled, recurring wheel-flat faults, the example selects supervised spectrogram-CNN classification and a GPU-edge platform, with FPGA as an option when deterministic latency is required.The scenario uses vibration data, a soft-real-time target below 50 ms per inference cycle, and a moderate 10–30 W power budget.Source: §7.6.3 · pp. 43–44
• The wheel-flat example calls for recall, F1-score, and false-negative-rate reporting, with validation across wheel wear, speeds, and seasonal temperatures.It also requires target-platform latency and power measurements.Source: §7.6.3 · p. 44
• For unlabeled sequential door-current signals, the example uses a normal-cycle-trained bidirectional LSTM autoencoder and reconstruction error to flag deviations.It specifies onboard soft-real-time detection below 200 ms per cycle, a power budget below 10 W, intermittent connectivity, and INT8 quantization.Source: §7.6.3 · p. 45
• The door example recommends PR-AUC, false-alarm rate, and reconstruction-error analysis, with validation across door types, passenger loads, and seasonal temperatures.Threshold calibration and drift monitoring are identified as essential for long-term deployment.Source: §7.6.3 · p. 45

7.6.4. Engineering Implications

• The framework turns the survey taxonomy and paradigm analysis into a structured, reproducible process for selecting and deploying railway anomaly-detection solutions.It connects anomaly characterization, supervision, deployment constraints, hardware, optimization, and evaluation.Source: §7.6.4 · p. 46

Multimodal fusion and multi-task railway monitoring

• Most studies address one task and sensing modality, motivating multimodal architectures that combine heterogeneous signals and support detection, localization, severity estimation, and condition assessment.The passage names vision, vibration, acoustic, current, voltage, log, and operational signals as potential inputs.Source: §8 · p. 46

Open datasets and standardized evaluation protocols

• Proprietary datasets, differing fault definitions, and incompatible experimental settings impede cross-study comparison and reproducibility, motivating open datasets and standardized evaluation protocols.Proposed protocols include clear train-test splits, cross-route and cross-fleet validation, cost-sensitive metrics, false-alarm analysis, and inference-latency reporting.Source: §8 · p. 46

Hardware-aware neural architecture search and hardware–software co-design

• Hardware-aware neural architecture search remains uncommon, while railway studies lack shared deployment benchmarks and safety-related searches require reproducibility, robustness, and auditability.Search can increase development complexity because candidate architectures must be evaluated under realistic hardware constraints; railway-grade platforms also impose real-time, resource, vibration, temperature, electromagnetic, and reliability requirements.Source: §8 · pp. 46–47

Domain adaptation and cross-fleet generalization

• Models trained on one fleet, route, season, or operating condition may not generalize across railway environments with different sensors, maintenance, infrastructure, weather, loads, or operating profiles.Domain adaptation, transfer learning, and federated learning are proposed to improve robustness and reduce each operator's need for large labeled datasets from scratch while preserving data confidentiality.Source: §8 · p. 47

Foundation models and large language models for railway monitoring

• Foundation models and language models could support few-shot image defect detection, maintenance-record analysis, and cross-modal reasoning, but their computational cost, uncertain latency, limited interpretability, and certification difficulty hinder safety-critical deployment.The survey frames efficient, explainable, and certifiable solutions as future research, with use in railway systems remaining a long-term direction.Source: §8 · p. 47

9. Conclusion

• The survey introduces a four-axis anomaly taxonomy and organizes deep learning approaches into classification-, prediction-, reconstruction-, and hybrid-based paradigms.The paradigms are analyzed with respect to data availability, supervision requirements, and deployment feasibility.Source: §9 · p. 48
• The survey finds conventional accuracy-centric metrics insufficient for safety-critical applications and proposes evaluation integrating cost-sensitive assessment, interpretability, and latency reporting.Source: §9 · p. 48
• A deployment-driven decision framework translates anomaly characteristics and operational constraints into paradigm and architecture choices, emphasizing embedded platforms, model compression, and hardware-aware design.The survey presents these contributions as a unified reference bridging academic performance and operational feasibility in safety-critical monitoring.Source: §9 · p. 48