Evaluation of 2D Acoustic Signal Representations for Acoustic-Based Machine Condition Monitoring

Abstract

Acoustic-based machine condition monitoring (MCM) provides an improved alternative to conventional MCM approaches, including vibration analysis and lubrication monitoring, among others. Several challenges arise in anomalous machine operating sound classification, as it requires effective 2D acoustic signal representation. This paper explores this question. A baseline convolutional neural network (CNN) is implemented and trained with rolling element bearing acoustic fault data. Three representations are considered, such as log-spectrogram, short-time Fourier transform and log-Mel spectrogram. The results establish log-Mel spectrogram and log-spectrogram, as promising candidates for further exploration.