Small-Sample Demagnetization-Level Diagnosis of Permanent Magnet Synchronous Motors Based on Information Fusion

Abstract

Accurate demagnetization-grade diagnosis of permanent magnet synchronous motors is difficult when operating conditions change and only a small number of independent fault records are available. This paper presents a condition-calibrated dual-modal workflow that combines complex Morlet continuous wavelet transform (CWT) order features of external radial stray flux with angular-domain current order-amplitude ratios. A healthy D0 record at the same operating point is used for calibration; feature selection, support vector classifier tuning, probability calibration, and fusion-weight selection are performed only within the outer training conditions. The dataset contains 54 simulated records with six demagnetization grades, three speeds, and three loads, and nested leave-one-speed-load-condition-out validation is used. The fusion model gives 96.30% accuracy and 96.37% macro-F1, compared with 83.33% and 82.00% for the Bronly model; grade MAE decreases from 0.389 to 0.037. The Br branch provides the stronger stand-alone basis, while current probabilities correct complementary boundary errors; their fusion improves cross-condition diagnosis.