A Hybrid Physical and Data-Driven Framework for Temperature Evaluation of Permanent Magnets in PMSMs

Abstract

To ensure the reliable operation of permanent magnet synchronous machines (PMSMs), accurate temperature evaluation and monitoring of permanent magnets (PM) are essential, as excessive heat can lead to performance degradation and irreversible damage. A hybrid physical and data-driven framework is proposed to address this challenge. The framework combines a physical model, which captures the thermal dynamics of PM, with advanced machine learning techniques. The physical model is translated into a graph structure, enabling the use of a graph attention network (GAT) to update node features and extract complex thermal interactions. A convolutional neural network (CNN) is subsequently applied for precise temperature evaluation. Experimental results validate the robustness and generalization capability of the proposed method, providing a reliable approach for assessing the thermal performance of PM in PMSMs. This work establishes a foundation for exploring temperature-dependent magnet degradation and implementing preventive measures in PMSMs. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.