Explainable multi-scale health index learning for non-destructive condition monitoring and wear severity assessment of aviation piston pumps

Abstract

Reliable condition monitoring of aviation piston pumps is essential for operational safety and predictive maintenance. This study proposes an Explainable Multi-Scale Health Index Learning (EMHI-L) framework integrating multimodal sensor fusion, multi-scale representation learning, fault classification, continuous health-index (HI) estimation, and explainability. Eleven pressure, vibration, flow, speed, and torque channels from a public dataset covering healthy operation and regulation valve, servo piston, and swashplate bearing bush wear yielded 3,850 synchronised analysis windows. Heterogeneous sensor streams were timestamp-aligned to a common 3-kHz grid before generating 4096-sample windows. Because measured wear-depth trajectories are unavailable, normalised HI targets of 1.00, 0.65, 0.45, and 0.25 were prescribed for the four conditions as an ordinal degradation-severity scale. To prevent leakage from overlapping windows, all windows from each recording were retained within a single partition using grouped/file-level cross-validation, complemented by leave-one-recording-out (LORO) evaluation. Grouped evaluation achieved 99.952 ± 0.067% classification accuracy, HI mean absolute error (MAE) of 0.0117 ± 0.0056, root mean square error (RMSE) of 0.0181 ± 0.0117, and coefficient of determination (R2) of 0.9942 ± 0.0061. LORO evaluation achieved 100% accuracy, HI MAE of 0.0093, RMSE of 0.0148, and R2 of 0.9967. Same-partition comparisons with deep-learning, conventional classification, and HI-regression baselines demonstrated competitive performance.