Investigation into LSTM Deep Learning for Induction Motor Fault Diagnosis

Abstract

As motor faults could lead to unwanted loss in industry, it is important to find out the motor faults in time. Currently, with the popularity and mature application of deep learning, researchers in the field of electrical machine health assessment have begun to focus on deep learning methods. It is hoped that motor fault detection can be achieved with the help of deep learning methods. This paper presents to adopt deep learning methods represented by LSTM neural network for motor fault diagnosis and evaluates on our own experimental platform. Considering two typical motor faults with two different degrees of severity, the results show that the proposed LSTM approach has a high accuracy (98.81%) on motor fault classification. The results also confirm that: (1) adequate effort of preprocessing, including sample length selection in the time domain and frequency band selection in the frequency domain, can significantly improve accuracy and computational efficiency; (2) different faults can be separated through the information in frequency band of 100–1000 Hz, which has not been fully modelled analytically before.