Abstract:

Motor performances such as cogging torque and torque ripple are difficult to predict accurately with surrogate models. In this work, we propose text-transformer-based models to tackle the problem. We first formatted the motor design parameters into natural-language sentences describing the design, which are fed into a text transformer model, and fine-tuned on the motor design dataset to make predictions. Compared with conventional artificial neural network models, our approach is more flexible in the input parameter dimensions, and can be jointly trained on multiple motor types. Numerical tests are performed on two datasets with different types of interior permanent magnet motor designs, and initial results show substantial improvements on cogging torque prediction accuracy over artificial neural network models on both datasets.