Intra-class and Inter-class Differences in Mel-spectrogram Images of DC Motor Sounds

Abstract

One of the most used approaches for application of deep learning on audio signals is to use spectrogram-based images as an input to a neural network. There are various spectrogram-based images including mel-spectrogram representing an option often used in practice. In such as case, it is worth knowing what are the intra-class and inter-class differences of the input images. These differences are studied here by analyzing the Pearson’s correlation coefficient. They are calculated from the mel-spectrograms extracted from the audio signals containing sounds of DC motors. The recorded signals are classified into 8 classes used separately for intra-class difference, while specific pairs of classes are grouped into 12 binary sets of classes used for inter-class difference analysis.