cs.LGSep 30, 2026
SaveDeep Learning for Anomaly Detection in Railway Systems: A Structured Survey
Organizations: Université Polytechnique Hauts-de-France, LAMIH CNRS UMR 8201, Valenciennes, France · Alstom, Crespin, France · Université Polytechnique Hauts-de-France, LAMIH CNRS UMR 8201, INSA Hauts-de-France, Valenciennes, France · 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. Deep learning has emerged as an effective paradigm for railway anomaly detection, driven by the growing availability of heterogeneous sensor data from rolling stock and infrastructure. This paper presents a structured survey of deep learning-based anomaly detection approaches for railway systems. The surveyed methods are organized using a unified taxonomy covering anomaly location, data representation and manifestation, sensing modality, and temporal characteristics. Existing approaches, including convolutional, recurrent and attention-based architectures, autoencoders, generative adversarial networks, and transformers, are structured into classification-based, prediction-based, reconstruction-based, and hybrid learning paradigms. The survey also examines data-centric challenges, evaluation practices, performance metrics, and practical deployment aspects, including edge-cloud architectures, computational constraints, and hardware-aware optimization. Finally, a decision-oriented framework links anomaly characteristics, data properties, and operational constraints to suitable detection paradigms and deployment configurations. This work provides a structured reference for selecting and deploying deep learning solutions for railway anomaly detection and highlights open challenges toward reliable and scalable intelligent monitoring systems.
Figures & tables
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.
Figure 2: PRISMA-inspired flow diagram of the literature identification, screening, eligibility, and inclusion process adopted in this survey.
| Survey | Year | Infra. | Train | Env. | DL Strat. | Data Chal. | Depl. |
| Oh et al. [ 100 ] | 2022 | ✓ | ✓ | ✓ | ✓ | ✗ | NT |
| Lourenço et al. [ 90 ] | 2024 | ✓ | ✓ | ✗ | ✓ | ✓ | NT |
| Tang et al. [ 137 ] | 2022 | ✓ | ✓ | ✓ | ✓ | ✗ | NT |
| Kumar et al. [ 78 ] | 2024 | ✓ | ✓ | ✗ | ✗ | ✗ | NT |
| De Donato et al. [ 39 ] | 2022 | ✓ | ✗ | ✗ | ✓ | ✗ | T |
| Di Summa et al. [ 41 ] | 2023 | ✓ | ✗ | ✗ | ✗ | ✗ | NT |
Table 1: Comparative Overview of Survey Coverage in Railway AI. Infra: infrastructure, Env: environmental, DL Strat: DL Strategy, Data Chal: Data Challenges, Depl: Deployment.
Figure 3: Four-axis taxonomy for railway anomaly characterization: location of occurrence, manifestation in data, sensing modality, and temporal behavior.
| Ref. | Dataset name | Acc. | System | Modality | Scale (samples) | Labeled | Fault Types |
| [ 129 ] | Track-Geom-Defect | R | Railway track | Track geometry | 172 436 | Yes | Rail, switch, crossing defects |
| [ 122 ] | Rail-Acoustic | R | Track, fasteners | Audio signals | 1625 | Yes | Wheel burn, loose nut-bolt |
| [ 73 ] | IRJ-SparkVision | R | Rail joints (IRJ) | RGB images | 28 150 | Yes | Spark erosion, rail damage |
| [ 156 ] | Drone Fastener | R | Rail fasteners | RGB images | 500 | Yes | Missing fastener |
| [ 107 ] | Track Fault | P | Rails, fasteners | RGB images | 293 | Partial | Cracked rails, rusted bolts |
| [ 52 ] | Fastener Vision | R | Rail fasteners | RGB + depth | 3 300 | Yes | Elastic strip, Nut, Baffle miss |
Table 2: Infrastructure-Level Railway Datasets. Accessibility(Acc.): Public (P), Restricted (R). NR = not reported.
| Ref. | Dataset name | Acc. | System | Modality | Scale (samples) | Labeled | Fault Types |
| [ 87 ] | HSR-Sim | R | Railway electrification system | Multivariate signals | 103 200 | Yes | Short-circuit fault |
| [ 61 ] | Train-DoorRig | R | Train doors | Current | 626 | Yes | Bearing wear, obstruction |
| [ 49 ] | RailHVAC-RUL | R | HVAC system | Multivariate signals | NR | Yes | Air filter clogging |
| [ 133 ] | PlugDoor Sound | R | Sliding doors | Acoustic signals | 164 | Yes | Door faults |
| [ 141 ] | Braking System | R | Braking system | Multivariate signals | 28996 | Yes | Pressure, braking force deviation |
| [ 120 ] | Wheel Flat | P | Wheel-rail | Spectrogram images | 4977 | Yes | Wheel flat detection |
Table 3: Rolling Stock-Level Railway Datasets. Accessibility(Acc.): Public (P), Restricted (R). NR = not reported.
| Ref. | Dataset name | Acc. | System | Modality | Scale (samples) | Labeled | Fault Types |
| [ 146 ] | RFOD Track | R | Track surroundings | RGB images | 7235 | Partial | Foreign objects, obstacles |
| [ 98 ] | Foreign Object | R | Track area | RGB images | 2534 | Yes | Pedestrians, falling rocks |
| [ 85 ] | HSR Strong-Wind | R | Wind-exposed track | Meteoro-logical signals | NR | Yes | Wind-induced speed restriction events |
Table 4: Environmental and Behavioral-Level Railway Datasets for Anomaly Detection. Accessibility(Acc.): Public (P), Restricted (R). NR = not reported.
Figure 4: Typical pipeline for classification-based anomaly detection in railway systems.
| Ref. | Year | System | Architecture | Modality | Metric | AD Task |
| [ 113 ] | 2021 | Rail track | SVM, RF, MLP | Acoustic | Acc.: 97% | Classify track faults (wheel burn, loose bolt) from acoustic signals. |
| [ 133 ] | 2020 | Train doors | SVM (+MFCC features) | Acoustic | Acc.: 95% | Classify plug-door fault types from acoustic measurements. |
| [ 49 ] | 2021 | HVAC system | Gradient Boosted Trees (GBT) | Multivariate | Acc.: 92.2% | Detect air filter faults in railway HVAC from sensor data. |
| [ 122 ] | 2022 | Rail track | MLP + MFCC | Acoustic | Acc.: 98.4% | Classify rail structural faults from IoT acoustic data. |
| [ 61 ] | 2019 | Train doors | KNN vs. CNN | Current | Acc.: 99.5% | Detect abnormal current patterns in train door motors. |
| [ 111 ] | 2025 | Rail track | SVM, GRU, LSTM | Acoustic | Acc.: 96% | Classify track faults using combined MFCC and CQT features. |
Table 5: Classical Machine Learning classifiers for railway anomaly detection. All methods are supervised. Acc.: Accuracy.
| Ref. | Year | System | Architecture | Modality | Metric | AD Task |
| [ 129 ] | 2022 | Track geometry | CNN, DNN | Track geometry signals | Acc.: 94.3% | Classify rail geometry defects. |
| [ 107 ] | 2022 | Rail surface | ResNet-50, VGG-16 | RGB images | Acc.: 83.8% | Detect surface defects (cracks, corrosion). |
| [ 13 ] | 2024 | Rail and wheel | Ensemble CNN | RGB images | Acc.: 99% | Classify rail and wheel surface defects. |
| [ 109 ] | 2024 | Rail track | YOLOv11 + TL(Transfer Learning) | RGB images | Acc.: 92.5% | Binary classification of track images: defective vs. normal. |
| [ 52 ] | 2023 | Rail fasteners | CSP-Darknet53 | RGB + Depth | mAP: 86.6% | Detect and localize missing and broken fasteners. |
| [ 73 ] | 2025 | Rail joints | SqueezeNet | RGB images | Acc.: 98.2% | Detect and segment spark erosion on insulated rail joints. |
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. | Year | System | Architecture | Modality | Metric | AD Task |
| [ 120 ] | 2023 | Wheel-rail | LeNet-5 CNN | Vibration signals | Acc.: 97%, AUC: 0.97 | Detect and classify wheel flat defects from vibration spectrograms. |
| [ 141 ] | 2021 | Braking system | Multivariate CNN | Pressure + force | Acc.: 93% | Classify braking faults from multivariate signals. |
| [ 87 ] | 2020 | Traction system | GRU hierarchical | Multivariate | Acc.: 93% | Classify 29 traction fault types from high-frequency sensor streams. |
| [ 112 ] | 2023 | Wheelset | 1D-CNN, LSTM | Acceleration | NR | Detect and classify wheelset faults from raw acceleration time series. |
| [ 136 ] | 2023 | Passenger cabin | SlowFast CNN | Video + acoustic | Acc.: 85.1% | Classify abnormal passenger behavioral events. |
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.
Figure 5: Typical pipeline for prediction-based anomaly detection in railway systems.
| Ref. | Year | System | Architecture | Modality | AD Task |
| [ 81 ] | 2021 | Rolling stock | LSTM, RNN+ DNN | Multivariate signals | Predict derailment risk and abnormal running behavior from onboard sensors. |
| [ 149 ] | 2022 | HVAC, compressors | Modified Long Short-Term Memory | Multivariate signals | Predict normal system state and flag anomalies in multiple rolling stock subsystems. |
| [ 157 ] | 2021 | Air conditioning unit | LSTM, RNN | Temperature + current | Detect anomalies in vehicle air conditioning units from monitoring data deviations. |
| [ 121 ] | 2022 | Air conditioning unit | LSTM | Current signals | Detect abnormal air conditioning units recorded in vehicule on operation. |
| [ 62 ] | 2023 | Track geometry | LSTM | Inertial signals | Forecast track geometry parameters to detect misalignment before safety limits are exceeded. |
| [ 108 ] | 2023 | Train traction converter cooling | LSTM | Multivariate signals | Predict rolling stock equipment failures hours in advance from sensor streams. |
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. | Year | System | Architecture | Modality | AD Task |
|---|---|---|---|---|---|
| [ 85 ] | 2021 | High-speed rail wind monitoring | Multiple Attention Layer based Multi-Instance Learning (MAL-MIL) | Meteorological signals | Predict wind hazard probability in real-time along high-speed rail corridors. |
| [ 43 ] | 2025 | Axle bearing | Improved Empirical Wavelet Transform (IEWT)+CNN | Vibration signals | Predict bearing degradation and detect faults from vibration time series. |
Table 9: Attention and Transformer predictors for railway anomaly detection. AUC: Area Under the ROC Curve; NR: not reported.
| Ref. | Year | System | Architecture | Modality | AD Task |
| [ 50 ] | 2022 | Heating ventilation air conditioning filter | 1D CNN + physics model | Multivariate signals | Estimate remaining useful life of air filter and detect degradation anomalies. |
| [ 59 ] | 2024 | Rail corrugation | 1D CNN | Acceleration signals | Predict rail corrugation profile from vehicle acceleration to detect surface anomalies. |
| [ 163 ] | 2024 | Rail corrugation | Self-attention + TCN + GRU | Acceleration signals | Predict corrugation evolution trend and detect anomalous degradation progression. |
Table 10: Convolutional temporal predictors for railway anomaly detection. Acc.: Accuracy; NR: not reported; TCN: Temporal Convolutional Network; GRU: Gated Recurrent Unit.
Figure 6: Typical pipeline for reconstruction-based anomaly detection in railway systems.
| Ref. | Year | Subsystem | Architecture | Modality | AD Task |
| [ 37 ] | 2021 | Air production unit | SAE | Pressure + current | Detect anomalies in APU analog and digital sensor streams without fault labels. |
| [ 38 ] | 2024 | Air production unit | LSTM-AE vs. SAE | Pressure + current | Detect APU anomalies using temporal reconstruction on MetroPT-3 dataset. |
| [ 72 ] | 2021 | Braking system | One-class LSTM-AE | Current signals | Detect anomalies in metro brake operating unit from current signal reconstruction. |
| [ 35 ] | 2024 | Rolling stock | LSTM-AE | Multivariate signals | Online reconstruction-based anomaly detection in multivariate railway sensor streams. |
| [ 110 ] | 2024 | High-speed rail | Convolutional AE | Vibration signals | Detect seismic anomalies from train vibration using normal-only autoencoder reconstruction. |
| [ 159 ] | 2021 | Rail surface | CVAE | ABA signals | Detect rail squat defects as shallow as 0.03 mm from normal ABA signal reconstruction. |
Table 11: Standard and recurrent autoencoder approaches for railway anomaly detection.
| Ref. | Year | Subsystem | Architecture | Modality | AD Task |
| [ 146 ] | 2022 | Track obstacles | GAN (adversarial AE) | RGB images | Detect foreign objects on tracks by reconstructing normal track scenes from onboard cameras. |
| [ 142 ] | 2024 | Metro undercarriage | GAN + memory module | RGB images | Detect undercarriage anomalies via adversarial memory-enhanced image reconstruction. |
| [ 143 ] | 2024 | Metro tracks | CNN + Transformer | RGB images | Detect track surface anomalies by reconstructing normal track images with memory augmentation. |
| [ 164 ] | 2025 | Railway turnout | PatchCore + attention | RGB images | Detect visual anomalies in turnout environments using unsupervised feature reconstruction. |
| [ 155 ] | 2024 | Pantograph | Reverse distillation network | RGB images | Detect pantograph anomalies via unsupervised reverse distillation image reconstruction. |
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.
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.
| Ref. | Year | Subsystem | Architecture | Modality | AD Task |
| [ 80 ] | 2021 | Train doors | Conditional GAN + U-Net (reconstruction + prediction) | Video | Detect hazardous door events by jointly reconstructing frames and predicting optical flow. |
| [ 51 ] | 2020 | Pantograph-catenary | YOLO + AE + LSTM (classification + reconstruction + prediction) | RGB + IR + UV + vibration | Detect pantograph anomalies across four modalities using parallel specialized detection modules. |
| [ 28 ] | 2022 | Track infrastructure | GAN + meta-learner (reconstruction + supervised) | RGB images | Detect track anomalies from limited labeled examples using generative reconstruction and meta-learning. |
| [ 132 ] | 2025 | Track obstacles | Normalizing flow + Transformer (unsupervised + supervised) | RGB images | Detect and localize anomalous objects using hybrid unsupervised backbone and supervised decoder. |
| [ 152 ] | 2023 | Track components | CNN + rule-based module (classification + rules) | RGB images | Inspect track components by combining learned classification with geometric domain rules. |
Table 13: Hybrid approaches for railway anomaly detection. IR: Infrared; UV: Ultraviolet.
| Task | Metric | What it measures | Use in railway / Limitation |
|---|---|---|---|
| Classification | Accuracy | Overall fraction of correct predictions. | Avoid: unreliable when faults represent 1% of data (a model predicting "normal" for every sample achieves 99% accuracy.) |
| Recall | Fraction of actual faults correctly detected. | Prioritize in safety-critical scenarios (braking): a missed fault is more dangerous than a false alarm [ 115 ] . | |
| Precision | Fraction of detected anomalies that are genuine faults. | Use when false alarms are operationally costly (unnecessary maintenance stops). | |
| F1-score | Harmonic mean of precision and recall. | Preferred for imbalanced binary detection. | |
| Specificity | Fraction of normal samples correctly identified. | Use to control false alarm rates in continuous monitoring; complements recall. | |
| Anomaly scoring (unsupervised) | AUROC | Ranking quality of anomaly scores across all thresholds. | Use for threshold-independent comparison; avoid when anomaly prevalence is very low (PR-AUC is preferable) [ 127 ] . |
Table 14: Evaluation metrics for railway anomaly detection.
| Metric | Studies | % |
|---|---|---|
| Accuracy | 42 | 61.8% |
| Precision / Recall | 38 | 55.9% |
| F1-score | 36 | 52.9% |
| mAP / mIoU | 14 | 20.6% |
| AUROC | 18 | 26.5% |
| RMSE / MAE | 11 | 16.2% |
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).
| Ref. | Year | Sub-system | Platform | Class | Latency (ms) | Power (W) | Accuracy | Memory |
| [ 48 ] | 2025 | Track faults (vision) | Zynq 7Z020 / KV260 | FPGA | 0.7–10 | 2.3–3.9 | 93.4% | 4.5–11.6 Mb BRAM |
| [ 87 ] | 2020 | Traction faults | Xilinx VCU128 | FPGA | 1 | NR | 93% | FPGA resources |
| [ 82 ] | 2024 | Track faults (vision) | Xilinx ZCU104 | FPGA | 2.1 | 6.9 | 88.9% | 230.5 BRAM |
| [ 36 ] | 2020 | Fastener (3D) | Jetson TX2 + Atom | GPU + CPU | 15 | NR | 99.7% | 8 GB |
| [ 139 ] | 2024 | In-cabin events | Jetson AGX Xavier | GPU edge | 1200 | 10–30 | 85.1% | 32 GB |
| [ 42 ] | 2025 | Passenger counting | Jetson Nano | GPU edge | 80–100 | 5–10 | 96.85% | 4 GB |
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.
| Platform | Devices | When to Choose | Key Limitation | Railway evidence |
| CPU/GPU server | Intel Xeon, RTX | Offline training with high compute requirements, no onboard constraints [ 52 , 73 ] . | High power ( 200 W); non-deterministic latency incompatible with real-time safety-critical systems [ 7 , 34 ] . | Yes (training only) |
| GPU edge | Jetson TX2, Xavier, Orin | Real-time vision and multimodal tasks under moderate power constraints [ 139 , 36 ] . | Latency variability under concurrent workloads; limited for real-time constraints [ 34 ] . | Yes |
| FPGA | Xilinx Zynq, ZCU, VCU | Safety-critical subsystems requiring deterministic and low latency [ 87 , 97 ] . | High development complexity; long design cycles; limited flexibility for model updates [ 97 ] . | Yes |
| Embedded CPU | ARM Cortex-A, Intel Atom | Preprocessing, lightweight inference [ 122 , 149 ] . | Low parallelism; unsuitable for complex deep learning models or high-frequency inference [ 7 ] . | Partial |
| NPU / AI accel. | Hailo-8, ARM Ethos | Low-power onboard inference with dedicated neural acceleration; strong candidate for future embedded deployment [ 34 ] . | Limited validation in railway applications; integration and certification challenges remain open. | Projected |
Table 17: Hardware platforms for deep learning deployment in railway anomaly detection systems. The Railway evidence column indicates whether the platform has been validated in a surveyed railway study (Yes) or is projected from the general embedded AI literature (Projected).
Figure 8: Model compression workflow: quantization, pruning, and knowledge distillation transform a baseline model into a compact version for embedded railway deployment.
Figure 9: Deployment decision framework for deep learning-based anomaly detection in railway systems.
| Criterion | Classification | Prediction | Reconstruction | Hybrid |
| Suitability when fault labels are available | H | M | M | H |
| Suitability with normal data only | L | H | H | M |
| Coverage of unseen anomalies | L | M | H | H |
| Robustness to class imbalance | L | M | H | M |
| Temporal dependency handling | M | H | M–H | H |
| Tolerance to operational variability | L–M | M | M | H |
Table 18: Analytical comparison of detection paradigms for deployment-oriented railway anomaly detection. H: High suitability; M: Medium; L: Low.