Proposal of noninvasive failure diagnosis of electrical motor using googlenet

Abstract

Fault diagnosis is a useful tool that reduces system maintenance risks and costs. However, data related to the system's nominal and fault operating behavior is often not collected and stored adequately, it is difficult to identify and suggest automated fault detection methods. This study proposes a solution to apply deep learning technique on the convolutional neural network (CNN) to identify some common errors on induction motors based on operation sound. The opreration sound signal emitted from on a 0.37 kW two-pole induction motor is collected in some cases such as normal operation, phase loss, phase difference and bearing breakage. Their 2-D scalogram images are analyzed by continuous Wavelet transformation which is used to train and evaluate the deep learning CNN (i.e. GoogLeNet) to identify the above faults. Experimental results show that this method can diagnose induction motor faults with accuracy up to 98.8%.