Physics-Guided Neural Temperature Estimator for End-to-End Interpretable PMSM Thermal State Monitoring

Accurate thermal monitoring is essential to avoid winding short-circuit and demagnetization in permanent magnet synchronous motors (PMSMs) fed by high-power voltagesource inverters (VSIs). Conventional physical model-based estimation methods demand in-depth knowledge (e.g., heat dissipation characteristics, VSI nonlinearity) and exhibit sensitivity to specific motor designs. While supervised machine learning (ML) approaches alternatively learn a black-box mapping from readily measurable quantities to temperature using empirical data without expert knowledge, their inherent lack of physical interpretability limits reliability, hindering industrial application. To bridge this gap, this paper proposes a physics-guided neural temperature estimator (PG-NTE) tailored for traction drive PMSMs fed by VSIs, built upon an encoder-decoder architecture for end-to-end estimation. Key physics, namely the dqaxis voltage equations, are embedded into the loss function to guide the encoder's latent space toward estimating temperaturesensitive electrical parameters, assisting the decoder in predicting temperatures. Furthermore, the physics-aware framework adopts a dynamic weighting scheme to balance composite loss constraints. Experimental validation on a PMSM demonstrates desirable estimation accuracy and strong physical consistency, proving its effectiveness and interpretability.