A Study on Electric Motor Fault Diagnosis Using Time-Series Imaging and Unsupervised Feature Learning

Abstract

Electric motors can cause human, financial, and societal losses due to unexpected failures. Therefore, it is important to diagnose the electric motors to ensure the health of the electric motors. However, it is difficult to guarantee diagnostic performance due to signal noise or complex features. This study presents an electric motor health diagnosis method that combines time-series data imaging with deep learning. The overall procedure includes three major steps: (1) converting one-dimensional signals into two-dimensional images, (2) extracting features using convolutional neural networks, and (3) calculating a health index using Mahalanobis distance. The proposed time-series imaging is designed to be robust against noise and sensitive to weak signal changes. The trained convolutional neural networks are used to extract features in an unsupervised manner. The distance from the normal state features is measured using Mahalanobis distance. To evaluate the performance of the proposed method, simulation study and case study are conducted. In the simulation study, electric motor current signals of healthy and faulty states are collected from a stator winding inter-turn short circuit fault simulation model. In the case study, electric current signals are acquired from artificial fault seeded testbed with different levels of fault severities. This study can help improve the accuracy of fault diagnosis for electric motors.