Comparison of acoustic signal–based fault detection of mechanical faults in induction motors using image classification models

Abstract

This study presents a novel deep neural network–based method for fault detection in induction motors. The focus was on identifying five types of mechanical cases: normal operation, shaft/load breakage, misalignment, mounting bolt looseness, and cooling fan problems. To increase the realism of the results, a laboratory-collected dataset of stereo microphone recordings was augmented with real factory noise. The audio data was transformed into image data using Mel-frequency cepstral coefficients as the feature extraction method and then processed with image-based classifiers. A comparison was made among 12 different networks in terms of accuracy and number of parameters, revealing that Mobilenet_v2, EfficientNetV2B0, and NASNetMobile had the best performance in terms of both network size and accuracy.