Acoustic noise prediction for synchronous motors with different topologies using MFCC-based datasets

Unpleasant acoustic noise from electric machines not only affects human perception but also serves as a potential indicator of underlying mechanical or electromagnetic issues. From an electromagnetic perspective, this noise primarily originates from radial electromagnetic forces acting on the stator and rotor structures. This paper presents a novel framework that integrates deep learning with multiphysics system modeling to predict and classify the conditions of permanent-magnet motors based on Mel-frequency cepstral coefficient (MFCC)-derived acoustic datasets. Measured acoustic data from a target motor are compared with simulated results obtained through finite element analysis (FEA), with waterfall diagrams demonstrating high correlation between experiment and simulation. Literature shows limited coverage of integrated methods that couple multiphysics-based dataset generation with neural-network training for acoustic-noise prediction. The FEA-based simulation process enables efficient generation of 2,640 labeled datasets within 470 hours, representing four motor conditions: reference, adjusted magnet size, adjusted air gap, and combined magnet-air-gap adjustment. Using these datasets, a trained neural network achieves a classification accuracy of 95.68%, successfully identifying the causes of motor design deviations through acoustic-noise features for further design improvements.