Evaluating Unsupervised Anomaly Detection Models to Detect Faults in Heavy Haul Railway Operations

摘要

Tuning a fault detector to balance false positive and false negative rates is fundamental to optimize maintenance operations. Unbalanced detectors can either lead to a high demand rate on the maintenance team (biased to false positives) or let failures happen with no preventive action (biased to false negatives), causing stoppages and accidents. In the context of railway operations, the use of sensors and their maintenance history generates rich data sources that can be explored to detect, identify, predict and treat faults before a possible incident. Machine Learning has been applied to this fault management context, in which supervised and semi-supervised models are extensively used to discriminate faulty observations. Supervised and semi-supervised models are effective for wellknown cases, but they are limited in novel cases, since faults can happen in unpredictable ways. Thus, unsupervised models are an alternative approach to deal with this limitation. This paper aims to evaluate the metrics and effectiveness of two unsupervised anomaly detection models – Isolation Forest and Autoencoders – to detect faults on rail cars. These models were applied to real measurements obtained from thermal, acoustic and impact sensors installed in a heavy haul railway line in Brazil. The results were compared to maintenance rules that guide general decisions for field inspections from railway operators. As main outcomes, Autoencoders produced balanced results in different scenarios, showing that these models can autonomously detect faults with great robustness. Therefore, they can compose predictive methods, improving the efficiency of maintenance tasks and railway operations.