A regularized leakage-free lightweight temporal sensor-fusion framework for permanent-magnet temperature estimation in PMSMs

Abstract

Accurate estimation of permanent-magnet (PM) temperature is important for supervisory thermal management of permanent magnet synchronous motors (PMSMs), but direct rotor-side measurement is usually impractical. This study presents a regularized, leakage-free, lightweight temporal sensor-fusion framework evaluated on a physically measured 50 kW PMSM benchmark sampled at 2 Hz. Complete operating profiles were assigned to training, validation, and test subsets (48/10/11 profiles) before any window construction; z-score parameters and hyperparameters were learned without test information. Twenty-one operational and thermal channels were transformed into 168 descriptors using 256-sample windows (128 s) updated every 32 samples (16 s), and a Ridge estimator was selected on validation profiles. On 11 completely unseen profiles, the full model achieved an mean absolute error of 4.16 °C, root mean squared error (RMSE) of 5.31 °C, R2 of 0.909, and 95th-percentile absolute error of 10.67 °C; 93.71% of predictions were within 10 °C. The regression stage processed all 5743 test windows in 0.010 s (approximately 1.75 microseconds per window) and required 169 fitted coefficients including the intercept. Post-review accuracy-complexity experiments identified 77- and 37-descriptor configurations that reduced the descriptor count by 54.2% and 78.0% while increasing RMSE by only 0.67% and 1.35%, respectively. The largest errors occurred on short, high-temperature, distribution-shifted profiles, so the point estimator is intended for thermal supervision rather than standalone protection. A validation-calibrated upper bound is additionally analyzed for conservative derating decisions. The results show that explicit leakage control, compact temporal representation, and quantitative sensor/complexity trade-offs can support transparent PMSM thermal monitoring without a high-capacity sequence model.