Transfer Learning Enhanced Model for Performance Prediction of Electrical Machines

Deep learning-based image-to-image translation models (I2I) have shown strong potential in accelerating magnetic performance prediction of interior permanent magnet synchronous machines (IPMSMs). However, most existing approaches are limited to a single machine topology, restricting their generalisation capability. To address this limitation, this paper proposes a transfer learning enhanced conditional generative adversarial network (cGAN) for cross-topology magnetic flux density prediction across multiple IPMSM designs, including V-shaped, U-shaped and parallel-magnet configurations. The proposed method leverages pre-trained feature representations and adapts them to new topologies using a significantly reduced dataset, thereby lowering sampling requirements while maintaining high prediction accuracy. Various transfer learning strategies are systematically investigated to balance generalisation and efficiency. Experimental results demonstrate that the proposed method achieves accurate predictions with substantially reduced computational cost, highlighting its effectiveness for fast performance evaluation and design optimization.