Unsupervised detection of anomalous sounds for machine condition monitoring

Abstract

In an era of explosive machine applications , abnormal sound detection is gaining increasing attention from machine learning engineers. This report presents a novel solution to monitoring abnormal machine sounds by ensemble of models. Dense autoencoder and convolutional autoencoder were ensembled with self-supervised classifiers which output the confidence for machine-type predictions. In the datastet, six types of machines, containing around 20,000 pieces of normally operating recordings were used in the project. Time-series recordings were processed as mel spectrograms to be fed in the models. Competitive results were achieved by the ensembled system of dense autoencoder and self-supervised model using ResNet50V2 as the backbone. On average, the self-supervised model achieved a classification accuracy of 99 percent, and the ensemble system reached a prediction accuracy of 80 percent. In conclusion, this paper presents designs of dense autoencoder , self-supervised model using transfer learning and an ensemble method between dense autoencoder and self supervised classifiers. Further exploratory attempts on Generative Adversarial Network (GAN) and different feature extraction techniques could be researched for better generalization and performance