Source-linked AI summary
Anomaly Detection and Localization for the Pantograph-Catenary System
Francesco Vitale, Hangli Ge, Francesco Flammini
TL;DR
Video-based CNN monitoring of PCS height and stagger does not account for the train’s geographic location. The paper aligns measurements with nominal route GPS coordinates and applies collective anomaly detection; on an Italian railway track, localized LSTM-AE models showed robust detection across segments.
Problem
Existing video-based CNN solutions monitor PCS height and stagger without accounting for the train’s geographical localization.
Source: §I · pp. 1–2Method
The framework standardizes GPS and telemetry on a discretized geographical grid and trains localized LSTM-AE models to detect PCS anomalies.
Source: §III · p. 3Results
On real-world data from an Italian railway track with synthetically injected anomalies, localized LSTM-AE models showed robust collective detection with high AUC and precision across route segments.
Source: §III · p. 3Takeaways & Limitations
The framework is presented as a practical predictive-maintenance tool for detecting PCS degradation and pinpointing its physical location.
Source: §III · p. 3
Abstract
from arXiv · showhide
Monitoring the Pantograph-Catenary System (PCS) provides insight into the health conditions of the pantograph and the railway infrastructure. Recent industrial solutions trace the pantograph's contact wire height and stagger (PCS height/stagger) using video monitoring through convolutional neural networks. However, these solutions do not account for the train route's geographic location. Therefore, in this paper we propose a novel framework for 1) localization of the PCS height/stagger by alignment with the nominal GPS coordinates of the reference route, and 2) collective anomaly detection to evaluate the health conditions of the PCS. We apply and assess the localization and detection performance of the methodology to a case-study based on a real-world industrial dataset provided by a railway transportation company, which includes the PCS height/stagger of several train journeys across Italian railway routes.
I. INTRODUCTION
The paper addresses geographic context missing from video-based PCS monitoring by proposing route-based localization and collective anomaly detection. The work applies this framework to an industrial dataset of train journeys on Italian railway routes.
I. INTRODUCTION
- Abnormal contact-wire height or stagger may cause wire breakage, pantograph destruction, material damage, fire, or electric shock.The PCS connects onboard equipment to the overhead catenary through the pantograph’s contact wire.Source: §I · p. 1
- Prior video-based CNN work monitors catenary defects, pantograph wear, or PCS height and stagger, including checks against safe operational zones.The cited studies address these monitoring tasks, but the paper identifies missing geographic localization in existing CNN-based height/stagger solutions.Source: §I · pp. 1–2
- The proposed framework localizes PCS height and stagger using nominal reference-route GPS coordinates and performs collective anomaly detection on train-run measurements.The case study uses data from several train journeys across routes managed by an Italian railway company.Source: §I · p. 2
II. REFERENCE FRAMEWORK AND CURRENT RESULTS
The framework aligns train-run measurements to a spatial reference route, segments the route, and trains localized LSTM autoencoders for anomaly detection. Results vary by segment, with strong AUC and precision but lower recall in Zone 2.
II. REFERENCE FRAMEWORK AND CURRENT RESULTS
- Onboard video tracks contact-wire height and stagger, while GPS synchronizes the measurements with train location to produce run-level telemetry logs.The framework aggregates segment-level classification metrics to assess collective detection performance.Source: §II · pp. 2–3
- Raw telemetry is aligned to a nominal route grid resampled at 22.0-meter intervals, with GPS readings snapped to their nearest grid points using Haversine distance.Preprocessing imputes missing values, filters stationary windows, and aggregates redundant geographic readings.Source: §II · p. 2
- The route is divided into geographical segments with L = 4, and synthetic test anomalies model cantilever shifts and auto-tensioning failures over 30-point windows.The simulated failures apply a lateral stagger offset or a smooth sinusoidal penalty to height.Source: §II · pp. 2–3
- A separate bidirectional LSTM autoencoder is trained from scratch for each route segment to model location-specific healthy geometry.Training uses only healthy data; test sequences are flagged when reconstruction error exceeds the validation set’s 99th-percentile MSE threshold.Source: §II · p. 3
- AUC scores were 0.968 in Zone 1 and 0.967 in Zone 3; precision exceeded 0.820 across zones, while Zone 2 recall was 0.644 with FN = 53.The results are reported for the anomaly detector with L = 4.Source: Table I · p. 3, §II · p. 3
III. CONCLUSIONS
The framework standardizes data geographically and partitions routes to establish localized geometric baselines. Tests on real-world Italian railway data with synthetic anomalies showed robust collective detection across route segments, while the authors identify broader validation and real-time processing as future work.
III. CONCLUSIONS
- Standardizing telemetry onto a discretized geographic grid and partitioning routes established localized baselines for healthy geometric variation.The baselines reflect natural variation at specific geographic coordinates.Source: §III · p. 3
- Localized LSTM-AE models achieved high AUC and precision across route segments when evaluated against synthetically injected anomalies.The experiments used real-world data from an Italian railway track; the authors suggest location-specific training minimizes false positives caused by natural topological variance.Source: §III · p. 3
- The framework is presented as a practical predictive-maintenance tool for detecting PCS degradation and pinpointing its physical location.Future work includes real-time edge processing, evaluation on a wider dataset of naturally occurring faults, and Graph Neural Networks for segment-boundary transitions.Source: §III · pp. 3–4