Spectral Graph Convolutional Networks for Rotor Temperature Estimation in Permanent Magnet Synchronous Machines

Estimating rotor temperature is a challenging problem, yet an indispensable aspect of modern permanent magnet synchronous machines (PMSMs). This is critically important for health monitoring in PMSMs operating under various conditions. This paper presents a data-driven estimation framework by leveraging spectral graph convolutional networks (GCN) due to their low computational overhead and improved real-time performance. In particular, we construct several datasets using an experimental setup to evaluate the estimation performance under diverse operating conditions. Each dataset is initially processed using a novel graph-based correlation-aware mechanism to construct a graph with strongly connected nodes. This graph is then fed into the GCN module for regression. An enhanced GCN model using Chebyshev convolution is proposed to capture the temporal and overall data patterns within each dataset. Compared with a regular GCN, results show improved performance and superiority of the enhanced GCN for rotor temperature estimation.