Motor Fault Detection with a Hybrid Physics-based and Data-Driven Method

Electric machine condition monitoring and fault detection using machine learning methods have been widely investigated in recent years. One main challenge for such data-driven approaches is the lack of real data, especially for machines under faulty conditions. In this paper, we propose a framework to address the data scarcity problem with a hybrid physics-based and data-driven method, and evaluate the effectiveness on induction motor eccentricity fault detection. We first use a simulation model to generate synthetic data for the motor under eccentricity fault; then we introduce a topological data analysis method to process the obtained data and extract fault-related features; next we apply domain adaptation technique to bridge the gap between the synthetic data and limited real data; finally with the adapted data, we train machine learning models to predict motor fault conditions. We show the prediction error is reduced from over 12% to about 5% compared with the same model trained without the domain adaptation process.