Unsupervised anomalous sound detection for industrial monitoring based on ArcFace classifier and gaussian mixture model

The operation state of machine can be monitored by performing anomalous sound detection (ASD). Unsupervised-ASD is a detection task in which the model detects unknown anomalous sounds without the use of anomalous sounds to train it. However, when detecting completely unknown anomalous samples, it is challenging to classify samples with high similarities and determine decision boundaries, leading to the poor detection performance. In light of the above deficiencies, we propose the ArcFace classifier and Gaussian mixture model (GMM) based unsupervised-ASD method. The Arcface loss-based classifier is proposed to aggregate the hidden features of different classes into the corresponding arc space, which increases the separability of samples. The GMM-based anomaly score calculation is presented to determine the more complex decision boundary. Experiments are carried out on the datasets provided by DCASE 2020 Task 2 and CRWU. As demonstrated by the areas under the receiver operating characteristic curve (AUC) and the partial AUC (pAUC), the proposed method has better performance compared with other methods in comparison.