IIoT-BASED FAULT DETECTION IN A RAIL VEHICLE AIR COMPRESSOR

Abstract

Scheduled maintenance does not always reflect the actual condition of equipment, and the built-in threshold signals fire only after a failure has begun. We assess, at event level, whether anomaly detection models applied to IIoT telemetry detect a failure earlier than the built-in threshold method or with fewer false alarms. The data are the open MetroPT-3 dataset from the air production unit of a metro train (February–August 2020, 1,516,948 records, 15 signals, CC BY 4.0 licence, 4 failures from company reports) and the MetroPT2 dataset (2022, 2 failures) for a transfer test. The methods comprised 38 features in 10-minute windows, a strictly chronological split, Isolation Forest, One-Class SVM, a fully connected autoencoder, an EWMA control chart and the built-in LPS rule; thresholds were chosen on validation data for a target number of false alarms per day, and each model was run with 5 random seeds. At the pre-set operating point of 0.1 false alarms per day the models detected on average 0–0.6 of 4 failures, EWMA 1, and no method alarmed before the reported start. With thresholds calibrated to 0.5 false alarms per day, the models and EWMA detected on average 3.8–4.0 failures at 0.15–0.19 false alarms per day, the built-in rule 3 failures at 0.66; median lead times were negative. The models ranked windows near failures better than EWMA (PR-AUC 0.40–0.45 versus 0.24). Models trained on 2020 data kept this ability on 2022 data (PR-AUC 0.69–0.73), but the 2020 threshold was too high. The novelty is an event-level comparison with the built-in rule on real metro compressor data, with a chronological protocol and a transfer test. In practice, the threshold can be set from the false-alarm rate a maintenance team accepts. Early warning was not demonstrated; thresholds must be recalibrated for each asset.