Run-Disjoint Few-Shot XGBoost Framework for Compound Fault Diagnosis of Induction Motors

Le résumé fourni par la source

Few-shot compound fault diagnosis of induction motors can be overestimated when correlated windows from the same continuous run are split across support and query sets. This study develops a run-disjoint few-shot framework in which each complete experimental run is treated as one shot and support and query sets are separated by run ID. Forty-eight multidomain features are extracted from synchronized triaxial vibration windows, classified using task-specific XGBoost, and aggregated to obtain run-level predictions; TreeSHAP provides post hoc feature attribution. In a matched comparison with identical query runs and windows, window-mixed partitioning increased the task-level mean run-level Macro-F1 from 0.9212 to 0.9934. After repeated predictions were aggregated over 108 unique query runs, the corresponding difference was 0.0093 with a 95% paired-bootstrap confidence interval of [0.0000, 0.0282], showing that the estimated magnitude depends on the statistical unit. Under the predefined strict 3-shot protocol, XGBoost achieved a Macro-F1 of 0.9263 and run-level accuracy of 0.9292. Additional sensitivity and controlled comparisons showed that performance depends on within-run sampling, representation, and classifier design, while strict cross-speed tests revealed the limitation of fixed-frequency features under rotational-speed shifts. The framework provides a leakage-aware evaluation procedure for few-shot compound-fault diagnosis using independently labeled runs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.