VIRTUAL VIBRATION SENSOR CALIBRATION BASED ON VAE-BASED TRANSFER LEARNING FOR PREDICTING LOW FREQUENCY VIBRATION ON PERMANENT MAGNET SYNCHRONOUS MOTOR

Abstract

Permanent magnet synchronous motors (PMSMs) are widely used in industrial applications for efficient power conversion. However, vibration generated by radial and tangential electromagnetic forces poses a significant challenge in ensuring mechanical integrity and operational reliability. These forces originate from the magnetomotive force (MMF) of the field and armature, and their interaction, and the resulting radial vibration is transmitted through the motor housing. Accurately predicting such vibrations typically requires high-cost 3-D finite element analysis (FEA), making it computationally inefficient for large-scale modeling. To overcome this, we propose a variational autoencoder (VAE)-based transfer learning framework that combines a large low-fidelity dataset generated from 2-D FEA with a small high-fidelity dataset from 3-D FEA. The low-fidelity VAE is pre-trained and its learned parameters are transferred to the high-fidelity model for fine-tuning. This method enables accurate vibration prediction across varying armature currents and rotational speeds with significantly reduced computational burden. The trained model can also serve as a virtual vibration sensor for PMSM condition monitoring.