Can Anomaly Scores Provide Credible Early Warning of Compressor Failures? Evidence from a Chronological MetroPT-3 Evaluation

Anomaly detectors for predictive maintenance are often compared with observation-level ranking metrics, although maintenance action is triggered by temporally consolidated alarms. This study tested whether three unsupervised scoring rules translate from ranking to event-specific early warning on public MetroPT-3 railway-compressor telemetry. Models were fitted on February–March 2020, thresholds were fixed during April–May, and June–July—containing only two documented test failures—was held out chronologically. At the prespecified 0.995 validation quantile, 15-minute consolidation, and 24-hour window, every method placed an alarm before 2/2 test failures, while producing 3.67–7.17 unrelated episodes per day. Crucially, 1,000 circular shifts of each unchanged alarm sequence also preceded 2/2 failures in 86.3% (Robust-MAD), 93.1% (Isolation Forest), and 86.8% (PCA) of shifts. The observed two-of-two warning count was therefore unsurprising given alarm occupancy and timing. PCA still achieved the strongest pointwise ranking (AUROC 0.9550; AUPRC 0.3668), but ranking and two events cannot validate event-specific warning. The experiment demonstrates why chronological evaluation must include episode burden and a chance-alignment control before an early-warning claim is made.