State-Aware Anomaly Detection for Railway Predictive Maintenance: A Reproducible Study on the MetroPT-3 Dataset

Authors

  • Ying Shi Guangzhou Railway Polytechnic, Guangzhou 511300, China.

DOI:

https://doi.org/10.54691/nnfh2g87

Keywords:

Railway Predictive Maintenance; Anomaly Detection; Operating-State Conditioning; Event-Level Evaluation.

Abstract

Railway air-production units operate under changing compressor and train-motion conditions, which can make a single global anomaly model sensitive to normal operating transitions. This paper presents a reproducible state-aware anomaly-detection study on the MetroPT dataset. The downloaded dataset_train.csv contains 10,773,588 one-second records and 21 columns covering pressure, temperature, current, digital control signals, and GPS variables. We aggregate the stream into 60-second windows, identify compressor and motion states, and compare a global robust Z-score baseline, a global Isolation Forest, an RBF kernel baseline, a state-conditioned Isolation Forest, and a temporally smoothed event detector. Time-ordered splits use healthy data for training, later healthy windows for validation, and three documented maintenance failures for testing. The state-conditioned model reaches window-level precision 1.000 and recall 0.0328, while the smoothed detector reaches precision 0.9259 and recall 0.0282. Both methods detect two of the three failure events at the documented failure intervals; smoothing changes the distribution of alarms and recovers the short second event, but it does not improve aggregate window-level F1 in this initial configuration. These results show why event-level and window-level evaluation should be reported together and provide a reproducible baseline for future railway predictive-maintenance studies.

Downloads

Download data is not yet available.

References

[1] Veloso, B., Ribeiro, R. P., Gama, J., & Pereira, P. M. (2022). The MetroPT dataset for predictive maintenance. Scientific Data, 9, 764. https://doi.org/10.1038/s41597-022-01877-3

[2] Veloso, B., Gama, J., Ribeiro, R. P., & Pereira, P. M. (2022). MetroPT: A benchmark dataset for predictive maintenance. Zenodo. https://doi.org/10.5281/zenodo.6854240

[3] Liu, F. T., Ting, K. M., & Zhou, Z.-H. (2008). Isolation forest. In Proceedings of the 8th IEEE International Conference on Data Mining (pp. 413–422). https://doi.org/10.1109/ICDM.2008.17

[4] Schölkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J., & Williamson, R. C. (2001). Estimating the support of a high-dimensional distribution. Neural Computation, 13(7), 1443–1471. https://doi.org/10.1162/089976601750264965

[5] Jardine, A. K. S., Lin, D., & Banjevic, D. (2006). A review on machinery diagnostics and prognostics implementing condition-based maintenance. Mechanical Systems and Signal Processing, 20(7), 1483–1510. https://doi.org/10.1016/j.ymssp.2005.09.012

Downloads

Published

2026-10-23

Issue

Section

Articles