We investigate whether early air leaks in railway compressor systems can be detected from the shape of their operating patterns rather than from simple signal thresholds. Using pressure and motor current data from the MetroPT-3 dataset, we rebuild the compressor behavior in a state space, measure how its geometry changes over time, and compare each window with a healthy baseline. The method captures the regular closed-loop pattern of normal operation and tracks when that pattern starts to break down. In tests on four confirmed leak events, the approach detected all cases before the low-pressure switch was triggered and produced no false alarms on the healthy test data. Lead times ranged from 97 minutes to nearly 16 hours, and two events were flagged more than two hours before shutdown. The results show that geometric changes in compressor behavior can provide an early and interpretable warning of air leaks, even when ordinary signal values still appear normal. This offers a practical way to improve maintenance planning and reduce unplanned train withdrawals.

Keywords

Topological Data Analysis, Air Leak Detection, Railway Compressors, Persistent Homology, Predictive Maintenance Railway.