ScholarWorks@성균관대학교

초록

This paper presents a computationally efficient, learning-based design framework employing Multi-Stage Deep Transfer Learning (DTL) for accurate efficiency prediction and optimal geometry optimization of rib-less Interior Permanent Magnet Synchronous Motors (IPMSMs), explicitly considering PWM-induced core losses. The framework adopts a stage-wise transfer learning process in two stages: the model is first fine-tuned on efficiency map data (first transfer stage) with all layers updated and is then freeze-tuned on PWM-induced harmonic loss data (second transfer stage) with the feature extractor fixed and only the regression head updated. Hyperparameters are optimized using Particle Swarm Optimization (PSO), and the model architecture is adaptively configured for each dataset. The trained model achieves high prediction accuracy for both sinusoidal and PWM excitation, as verified by Finite Element Analysis (FEA) and experimental results. Moreover, the proposed framework enables rapid and accurate generation of efficiency maps, thus supporting efficiency map-based optimal design and providing a practical and scalable solution for motor-inverter co-design in real-world applications.

키워드