Domain-Adversarial Adaptation Regression Model for IPMSM Permanent Magnet Temperature Estimation

Permanent magnet synchronous motors (PMSMs) are widely applied for industrial uses as well as electric vehicles (EVs) due to their remarkable electrical characteristics. However, PMSMs are vulnerable to damage from high temperatures, which can cause the demagnetization of permanent magnets (PMs). To prevent demagnetization, the accurate real-time estimation of the PM temperature (PMT) is crucial. This article proposes a simulation-driven unsupervised transfer learning approach that leverages the domain-adversarial adaptation regression (DAAR) model to enable online PMT estimation. First, Motor-CAD PMSM simulation provides a convenient way to obtain a large amount of simulated PMT data, which overcomes the difficulty in acquiring labeled PMT data. Second, this simulation dataset serves as the source-domain data to train the proposed modified transformer-based long sequence time-series regression (MT-LSTR) model, which generates coarse predicting labels (CP-labels). The CP-labels can be used to label the target-domain dataset collected from a real interior PMSM (IPMSM). Finally, the DAAR model is trained with both feature-level and domain-level adaptations to obtain a well-trained online PMT estimator. In experiments, the effectiveness of the proposed method is validated using the IPMSM temperature dataset provided by Paderborn University (PU).