Machine Anomalous Sound Detection Based on Self-Supervised Classification

The Machine Anomalous Sound Detection task aims to design a system to detect unknown anomalous sounds given only the sounds of machines working normally. The sounds emitted by different types of machines often have different characteristics, and the environments in which the machines work (such as temperature, noise, etc.) are constantly changing, which also affects the acoustic characteristics of the machine sound, so this is a challenging task. To this end, we propose a method for anomalous sound detection based on self-supervised classification. First, we obtain an effective feature representation of the sound by extracting frequency domain and time domain features from the raw wave and extracting pre-trained features based on the pre-trained model. Then, we design an auxiliary loss based on the attribute information of the audio, which helps the model to distinguish different operating conditions of the machine. Finally, we extract latent representations from the trained model, and calculate the anomaly score of the machine based on the distance metric. Experimental results on the DCASE 2022 Challenge Task 2 dataset demonstrate the effectiveness of our method. Moreover, we analyze the complementarity between different feature representations, which proves that the feature representations used in our method are effective.