PaperlayerView exact revision on arXiv

Source-linked AI summary

Cycle-Aware Autoencoder with Cross-SignalConsistency for Railway Door Anomaly Detection

Ammar Bouketta, Smail Niar, Hamza Ouarnoughi, Eva Mutuzo Brindle

arXiv:2609.39035v1cs.LGcs.AI

TL;DR

Railway door monitoring must detect rare, diverse, often unlabeled abnormal behavior from real operations. TCAA-CS treats each complete door cycle as an unsupervised monitoring unit and combines dual-stream encoding with a hybrid anomaly score. On industrial data, it reports 93.8% recall, 97.3% precision, and a 0.5% false-alarm rate, with real-time onboard feasibility assessed on a Jetson AGX Xavier.

  • Problem

    Door faults are rare and often unlabeled, while heterogeneous physical and logical signals can become inconsistent across a complete operation cycle.

    Source: §I · p. 1

  • Method

    TCAA-CS trains on nominal cycles and combines separate physical/logical signal encoding with temporal attention and a score fusing reconstruction error, latent deviation, and phase-aware cross-signal consistency.

    Source: §I · p. 1, B. Model Overview · p. 3

  • Results

    93.8% recall, 97.3% precision, an F1-score of 0.955, and a false-alarm rate of 0.5%; versus GRU-AE, recall improves by 8.9 percentage points and F1-score by 8.5 percentage points.

    Source: C. Main Detection Results · p. 7

  • Takeaways & Limitations

    System-level evaluation on an NVIDIA Jetson AGX Xavier supports real-time onboard deployment feasibility with millisecond-level latency and limited memory usage.

    Source: §VI · p. 8

Abstract

from arXiv · show

Passenger access doors are safety-critical subsystems in railway vehicles, yet detecting abnormal door behavior in real operation is challenging because faults are rare, diverse, and often unlabeled. This paper addresses railway door condition monitoring as a cycle-level unsupervised anomaly detection problem, where each complete opening-dwell-closing cycle is treated as a single monitoring unit. We propose the Temporal Cycle-Aware Attention Autoencoder with Cross-Signal Consistency (TCAA-CS), trained exclusively on nominal cycles. It combines a dual-stream encoder that processes continuous physical measurements (position, current, voltage) and binary logical states (door-closed, door-locked) through separate 1D-CNN branches, an LSTM encoder with temporal attention pooling, and a triple hybrid anomaly score fusing reconstruction error, latent-space deviation, and phase-aware cross-signal consistency. The consistency term helps identify cases where individual signals appear plausible but their inter-signal relationships become physically or logically inconsistent. On real industrial data from a passenger train in commercial service, TCAA-CS achieves 93.8% recall, 97.3% precision, and a 0.5% false-alarm rate, outperforming representative unsupervised baselines. System-level evaluation on an NVIDIA Jetson AGX Xavier supports the feasibility of real-time onboard deployment.

I. INTRODUCTION

Railway door monitoring must detect rare, unlabeled faults across complete cycles and heterogeneous signals, while remaining a maintenance-support function rather than certified safety control. TCAA-CS addresses these challenges with cycle-level unsupervised detection, signal-type-aware encoding, and hybrid scoring.

I. INTRODUCTION

  • Each complete opening–dwell–closing operation is treated as one monitoring unit for unsupervised railway-door anomaly detection.The framework is intended for onboard maintenance support, while certified train systems retain safety-critical door control.Source: §I · p. 1
  • TCAA-CS combines separate physical/logical encoding streams, temporal attention pooling, and a hybrid score to detect signal deviations and cross-signal inconsistencies.The score combines reconstruction error, latent-space deviation, and phase-aware cross-signal consistency.Source: §I · p. 1
  • The study evaluates TCAA-CS on real industrial railway-door data, compares it with unsupervised baselines, and assesses embedded deployment feasibility.The stated deployment assessment covers latency, memory, and energy consumption.Source: §I · p. 1

A. Fault Detection in Railway Door Systems

  • Prior model-based studies estimate motor-current parameters or use Bond Graphs, while later data-driven work includes CNNs on current, acoustic-signal SVMs, and nominal-data autoencoders with one-class SVMs.Follow-up unsupervised work addresses transfer learning and GAN-based domain adaptation across door types.Source: A. Fault Detection in Railway Door… · p. 2
  • Existing approaches often use a single modality, short windows or samples, and one detection criterion without explicitly checking consistency between physical measurements and logical states.The paper identifies these limitations as gaps in prior railway-door monitoring methods.Source: A. Fault Detection in Railway Door… · p. 2

B. Anomaly Detection in Multivariate Time Series

Common multivariate time-series detectors generally use fixed windows, homogeneous input channels, and no explicit phase-wise consistency check between physical and logical signals. TCAA-CS instead aligns detection with complete door cycles and is evaluated on commercial-service data represented as synchronized, normalized cycle sequences.

B. Anomaly Detection in Multivariate Time Series

  • Common multivariate anomaly detectors use fixed-length windows, homogeneous channels, and generally do not assess phase-wise consistency between continuous physical measurements and binary logical states.Examples include USAD, DAGMM, and the Anomaly Transformer, which add criteria such as adversarial reconstruction, latent density estimation, and association discrepancy.Source: B. Anomaly Detection in Multivariate… · p. 2

C. Positioning of This Work

  • TCAA-CS aligns detection to complete door cycles, separately encodes physical and logical signals, and combines reconstruction error, latent deviation, and phase-aware cross-signal consistency.Its evaluation also assesses onboard deployment feasibility in terms of latency, memory footprint, and energy consumption.Source: C. Positioning of This Work · p. 2

III. DATASET OVERVIEW

  • The dataset comes from a Regio 2N passenger train in regular commercial service, with 16 doors across five cars and 30 operating days.Fault events are rare and not systematically annotated; expert-identified operational issues do not provide exhaustive fault labels, and the dataset is proprietary.Source: §III · p. 2

A. Cycle-Based Data Representation

  • Each sample is a complete opening–dwell–closing cycle segmented from a door-day recording, using five synchronized signals acquired at 50 Hz.Cycles are the atomic anomaly-detection samples and capture mechanical, electrical, and logical behavior for one passenger access event.Source: A. Cycle-Based Data Representation · p. 2
  • Cycles are resampled to 256 time points; position, current, and voltage are scaled independently, while door-closed and door-locked states retain binary values.Each resulting sequence has three continuous variables and two binary states, with operating day, door identifier, and original duration retained only as metadata.Source: A. Cycle-Based Data Representation · pp. 2–3

B. Dataset Statistics

The supplied dataset-statistics passages identify Table I as a summary of characteristics after cycle segmentation and temporal normalization, but do not provide its cell values. The accompanying methodology passages define cycle-level scoring and describe nominal-only training with a dual-stream attention autoencoder and three-part anomaly score.

B. Dataset Statistics

  • Table I is identified as summarizing the dataset's main characteristics after cycle segmentation and temporal normalization.The supplied table passage is a caption only and contains no statistics to report.Source: B. Dataset Statistics · p. 3, Table I · p. 3

IV. METHODOLOGY

  • Each complete door operation is one monitoring instance that produces one anomaly score and one decision.The model is trained exclusively on nominal cycles; a test cycle is anomalous when its score exceeds a threshold calibrated on nominal validation data.Source: §IV · p. 3, A. Problem Setting · p. 3

A. Problem Setting

  • The problem-setting representation uses 256 time points and five synchronized signals: three continuous physical signals and two binary logical states.The excerpt identifies the representation dimensions but does not provide dataset-statistics cell values.Source: A. Problem Setting · p. 3

B. Model Overview

  • TCAA-CS trains its reconstruction model on nominal cycles and assigns one anomaly score per complete cycle at inference.Inference combines reconstruction fidelity, latent-space deviation, and cross-signal consistency into one decision against a calibrated threshold.Source: B. Model Overview · p. 3, Figure 2 · p. 3
  • The model separates continuous physical measurements from binary logical states, then uses attention-based encoding and sequence reconstruction.The architecture includes LSTM temporal modeling with attention pooling and an LSTM decoder.Source: B. Model Overview · p. 3

C. Dual-Stream Signal Encoding

TCAA-CS keeps continuous physical measurements separate from binary logical states during initial encoding, then uses temporal attention to form a cycle-level representation that emphasizes informative operation segments.

C. Dual-Stream Signal Encoding

  • Three physical measurements and two logical states enter separate temporal-convolution branches, preserving signal-type distinctions before joint modeling.The physical branch learns position, current, and voltage patterns; the logical branch learns state transitions, producing a unified 64-feature sequence.Source: C. Dual-Stream Signal Encoding · pp. 3–4

D. LSTM Encoder with Temporal Attention Pooling

  • A two-layer LSTM encodes the fused sequence, and additive temporal attention pools its hidden states into a full-cycle representation.The attention weights emphasize informative phases such as motion onset, phase transitions, closing impact, or lock engagement.Source: D. LSTM Encoder with Temporal… · p. 4
  • Attention pooling reduces the influence of less informative segments while emphasizing informative parts of the opening–dwell–closing operation.Source: D. LSTM Encoder with Temporal… · p. 4

E. Decoder and Training Objective

The latent cycle representation is decoded into a complete multivariate sequence, with training on nominal cycles only. At inference, reconstruction error measures signal-level deviations such as abnormal current peaks or distorted motion profiles.

E. Decoder and Training Objective

  • A symmetric two-layer LSTM decoder reconstructs the complete multivariate sequence from the latent representation.Source: E. Decoder and Training Objective · p. 4
  • Training uses nominal cycles only and minimizes mean squared reconstruction error across all timesteps and signal dimensions.The objective compares each nominal training cycle with its reconstruction.Source: E. Decoder and Training Objective · p. 4

F. Triple Hybrid Anomaly Scoring

1) Reconstruction Error:

  • Reconstruction error captures signal-level deviations that alter expected signal shape or amplitude, including abnormal current peaks or distorted motion profiles.Source: 1) Reconstruction Error · p. 5

2) Latent-Space Deviation (Cycle-Level Abnormality):

Latent-space deviation identifies cycles whose global temporal structure differs from nominal behavior, while phase-aware cross-signal consistency detects departures in relationships among signals during opening and closing.

2) Latent-Space Deviation (Cycle-Level Abnormality):

  • Latent-space deviation captures cycles whose global temporal structure differs from nominal behavior, including unusual timing, phase durations, or overall dynamics.The nominal latent center is computed from training cycles, and test-cycle deviation is assessed against nominal latent behavior.Source: 2) Latent-Space Deviation (Cycle-Level… · p. 5

3) Phase-Aware Cross-Signal Consistency:

  • Cross-signal consistency tests whether relationships among physical measurements and logical states remain coherent with nominal operation.A cycle can be abnormal even when individual signals appear locally plausible if their relationships are physically or logically inconsistent.Source: 3) Phase-Aware Cross-Signal Consistency · p. 5
  • The consistency measure compares pairwise signal correlations in the opening and closing phases with nominal reference matrices; high values indicate departures from nominal relationships.The phases are identified from the position trajectory after temporal normalization.Source: 3) Phase-Aware Cross-Signal Consistency · p. 5

4) Score Normalization and Fusion:

The method standardizes its three anomaly components using nominal-validation statistics, fuses them with nonnegative weights, and compares the resulting score with a threshold calibrated from nominal validation cycles. The evaluation section outlines comparisons with unsupervised baselines, ablations, deployment experiments, and qualitative analysis.

4) Score Normalization and Fusion:

  • Each anomaly component is standardized using statistics computed on nominal validation data before the components are fused.The normalization uses each component’s nominal-validation mean and standard deviation, with a small constant for numerical stability.Source: 4) Score Normalization and Fusion · p. 5
  • The final anomaly score is a weighted sum of normalized reconstruction error, latent-space deviation, and cross-signal consistency, with nonnegative weights summing to one.The weights are selected through sensitivity analysis.Source: 4) Score Normalization and Fusion · p. 5

G. Decision Rule

  • The decision threshold is calibrated exclusively on nominal validation data, and each complete door operation cycle receives an anomaly or nominal decision.The validation-score percentile controls the tolerated false-alarm level.Source: G. Decision Rule · p. 5

V. EXPERIMENTAL EVALUATION

  • The evaluation covers real-world railway door data, representative unsupervised baselines, ablations, deployment experiments, and qualitative analysis of abnormal cycles.Source: §V · p. 5

A. Experimental Protocol

The evaluation uses a day-wise split with nominal-only training and validation, then adds controlled anomalies only to the test set. Detection is assessed against unsupervised baselines using recall, precision, F1-score, and false-alarm rate.

1) Dataset Split:

  • The day-wise split keeps each operating day in one subset, with expert-identified anomalies excluded from training and validation and reserved for testing.Validation supports early stopping, score normalization, and threshold calibration.Source: 1) Dataset Split · pp. 5–6, Table II · p. 6

2) Test-Time Anomaly Construction:

  • Because only 12 real abnormal cycles were identified, 180 nominal test cycles were replaced with one injected anomaly each, yielding 192 anomalous and 969 nominal test cycles.Injection patterns were derived from documented maintenance records and expert observations, then validated by domain engineers for physical and operational plausibility.Source: 2) Test-Time Anomaly Construction · p. 6, Table III · p. 6

3) Evaluation Metrics:

  • Recall measures anomaly detection, while FAR counts nominal test cycles incorrectly flagged as anomalous and reflects maintenance workload.Recall is primary because missed anomalies are abnormal operations that remain undetected.Source: 3) Evaluation Metrics · p. 6

4) Compared Methods:

  • The comparison includes recurrent, convolutional, handcrafted-feature, and transformer-based unsupervised methods, all trained on nominal data and thresholded using the same validation rule.The Anomaly Transformer uses the same fixed-length cycle representation, with timestep scores aggregated to one score per cycle.Source: 4) Compared Methods · p. 6

B. TCAA-CS Configuration

TCAA-CS is configured as a dual-stream recurrent autoencoder and evaluated against unsupervised baselines on cycle-level test data. It achieves the strongest reported detection results, while most remaining misses are low-magnitude injected anomalies.

B. TCAA-CS Configuration

  • The dual-stream encoder uses separate 1D-CNN branches, followed by a two-layer LSTM encoder and decoder; training minimizes reconstruction error with AdamW.The score weights are λ1 = 0.6, λ2 = 0.1, and λ3 = 0.3, with the threshold set at the 95th percentile of nominal validation scores.Source: B. TCAA-CS Configuration · p. 6

C. Main Detection Results

  • 93.8% recall, 97.3% precision, an F1-score of 0.955, and a false-alarm rate of 0.5% make TCAA-CS the best overall method.Against the strongest baseline, GRU-AE, recall improves by 8.9 percentage points and F1-score by 8.5 percentage points under the same threshold calibration protocol.Source: Table IV · p. 6, C. Main Detection Results · p. 7
  • GRU-AE is the strongest baseline, with 84.9% recall and an F1-score of 0.870; the Anomaly Transformer follows with 82.3% recall and an F1-score of 0.859.TCN-AE has high precision but lower recall, while OC-SVM is less competitive than learned temporal representations.Source: C. Main Detection Results · pp. 6–7
  • All 12 expert-confirmed real anomalies are detected, while the 12 misses among 192 anomalous cycles are low-magnitude controlled perturbations close to nominal variability.Five false alarms on 969 nominal cycles mainly arise from rare valid events, including passenger-triggered reopening or atypical non-faulty timing.Source: C. Main Detection Results · p. 7

D. Ablation Study

Ablations show that attention, dual-stream encoding, and the complementary hybrid score each contribute to detection performance. The final model also meets the reported cycle-by-cycle latency, memory, and energy requirements on the embedded platform.

D. Ablation Study

1) Component Ablation:

  • Removing temporal attention drops recall from 93.8% to 88.0%; removing cross-signal consistency yields 89.1% recall, and removing dual-stream encoding yields 90.6%.The consistency-term loss is concentrated on logical anomalies whose individual signals appear plausible but whose relationships are inconsistent.Source: Table V · p. 7, 1) Component Ablation · p. 7

2) Score Weight Sensitivity:

  • Reconstruction error alone reduces recall to 78.1%, while the selected hybrid weights achieve the best F1-score of 0.955.Nearby weight settings remain competitive, but single-component settings have lower F1-scores: 0.847 for reconstruction, 0.474 for latent deviation, and 0.456 for cross-signal consistency.Source: 1) Component Ablation · p. 7, 2) Score Weight Sensitivity · p. 7

E. System-Level Evaluation

  • On Jetson AGX Xavier, mean inference latency is 4.98 ms per cycle, p95 latency is 6.52 ms, maximum latency is 8.92 ms, and inference energy use is approximately 8–12 W.GPU-based platforms use below 34 MB peak memory; the model has approximately 560k trainable parameters.Source: Table VI · p. 7, E. System-Level Evaluation · pp. 7–8

F. Qualitative Analysis on Real Anomalies

The qualitative examples show distinct load-related, kinematic, and logical fault patterns across full door cycles. The conclusion presents TCAA-CS as combining cycle-level temporal modeling with signal-level and inter-signal anomaly detection.

F. Qualitative Analysis on Real Anomalies

  • A load-related anomaly appears as abnormally slow opening and stalled position, with mismatch from the normal reconstructed trajectory also visible in current and voltage.The nominal example shows close agreement between observed signals and reconstruction.Source: Figure 4 · p. 8, F. Qualitative Analysis on Real… · p. 8
  • In the kinematic example, the door reverses before fully opening, while reconstruction follows the expected trajectory and exposes the missing motion.Figure 4 shows observed signals in blue and TCAA-CS reconstruction in red.Source: Figure 4 · p. 8, F. Qualitative Analysis on Real… · p. 8
  • The logical anomaly activates door-closed and door-locked states while position still indicates motion; the cross-signal consistency component primarily captures this fault.The individual signals appear plausible, but their relationship is physically inconsistent.Source: F. Qualitative Analysis on Real… · p. 8, Figure 4 · p. 8
  • The qualitative analysis finds the three scoring components complementary across failure modes, supporting a unified framework rather than separate handcrafted checks.The conclusion describes detection of both signal-level deviations and inter-signal inconsistencies across complete door cycles.Source: F. Qualitative Analysis on Real… · p. 8, §VI · p. 8

VI. CONCLUSION

  • 93.8% recall, 97.3% precision, and a 0.5% false-alarm rate are reported on real industrial data, with all expert-confirmed anomalies detected.The conclusion also reports real-time onboard deployment feasibility on Jetson AGX Xavier with millisecond-level latency and limited memory use.Source: §VI · p. 8
Loading 2609.39035v1…