Sound based fault classify diagnosis method using artificial neural network and autoencoder processing

Abstract

We achieved a fault diagnosis for a certain air pump using an artificial neural network. The operating sound of the pump is recorded by a single microphone, after processing by an unsupervised autoencoder, 108 groups of samples containing only 1-second audio data are inputted to the neural network classifier. The training rounds and the neurons of the autoencoder are tested. After training, the provided detection network can finally give the classifying accuracy of up to 99% according to 1-sec sound data.