Variational bayesian clustering algorithm for unsupervised anomalous sound detection incorporating VH-BCL+

Zhongqin Bi, Huanfeng Li, Weina Zhang, Zhen Dong

Abstract: In industrial environments, acoustic detection of equipment can often significantly change the acoustic characteristics between training and testing data due to the absence of anomalous data guidance, as well as changes in machine operating conditions or environmental noise. In addition, during model learning, there are cases where the target data is inaccessible, but an accurate model is still needed for the invisible target domain. To address these challenges, we propose an unsupervised approach for anomalous sound detection applied to unlabeled anomalous sound detection scenarios, which is an optimization algorithm based on joint deep learning and variational Gaussian mixture models, by two jointly training neural networks for feature extraction and using the variational Gaussian mixture model to analyze the obtained embeddings for clustering. We propose a new hybrid example data enhancement method, to generate examples in multiple ways, combining various methods to align the distribution between different domains. The improved sub-cluster AdaCos loss function is used to exclude potential anomalies by learning fewer finite distributions than the standard AdaCos. Experimental results show that the proposed method achieves an average area under curve of 82.70\(\%\) and an average F1 score of 69.42\(\%\) on the datasets of seven industrial equipment machine types. The average area under curve for the source domain is 85.49\(\%\) and for the target domain is 75.63\(\%\).