Efficient Design and Torque Prediction for Similar PMaSynRMs with a Generalized Attention Convolutional Network

Reducing computational load in the design optimization of similarly structured electric motors is crucial for industrial production. Recent advancements, particularly in data-driven surrogate models like neural networks, have shown significant potential in this area. This paper introduces a generalized attention convolutional network (GACN) for optimizing permanent magnet-assisted synchronous reluctance motors (PMaSynRMs) that vary in the number and shape of flux barrier layers. The network effectively identifies commonalities in the electromagnetics and flux barrier structures of similar PMaSynRMs within a high-dimensional feature space. Specifically, it utilizes ResNet34 as the backbone, enhanced with spatial and channel attention blocks for feature extraction, to discern motor structures independent of geometric interference. Additionally, the network incorporates a local maximum mean discrepancy block to measure data distribution similarities and adjust the nonlinear functions that correlate cross-sectional images with torque performance in similar motors. Experimental results demonstrate that, after training on a dataset of four-layer flux barrier PMaSynRM images, the proposed method accurately predicts torque and assesses geometric validity in PMaSynRMs with two- and three-layer barriers.