Abstract:

Vision Transformer (ViT) models have recently demonstrated strong performance in predicting electric motor cogging torque from visual representations of motor geometry. However, the high computational cost and memory footprint of ViT architectures hinder their deployment in resource- constrained environments such as embedded motor drives and real-time digital twins. This paper investigates post-training model compression techniques for a ViT-based cogging torque prediction model. Specifically, we evaluate some activation-aware pruning methods, which leverage activation statistics to identify redundant weights, as well as low-rank factorization applied to attention and feed-forward layers. Experimental results demonstrate that substantial reductions in parameter count and inference cost can be achieved with negligible degradation in prediction performance. These findings highlight the feasibility of deploying compressed ViT models for efficient and scalable electric motor analysis.