Machine sound anomaly detection based on dual-channel feature fusion variational auto-encoder

Abstract

With the increasing intelligence and automation of industrial equipment, the technology for detecting equipment anomalies has become increasingly important. Compared to image-based anomaly detection methods, sound-based anomaly detection methods have the advantages of being non-intrusive, real-time and having lower data collection costs. These advantages make them highly valuable for research. Currently, deep learning methods that focus on spectrogram reconstruction have become widely utilized in the field of machine sound anomaly detection research. However, previous methods only attempted to mitigate the impact of noise without enabling the model to fully learn the distribution of sound features during the reconstruction process. In this paper, a novel Dual-Channel Feature Fusion Variational Autoencoder (DCFF-VAE) is proposed to effectively improve its reconstruction ability and help it better learn the normal sound features. In this method, the deep features extracted from the convolution layer and bidirectional gated cycle unit in the encoder are integrated by means of concatenation to make full use of the important features in the sound. Subsequently, grouped deconvolution is applied in the decoder to reduce model complexity while enhancing its perceptual ability for features. Additionally, during the anomaly detection phase, anomaly scores are calculated based on the Mahalanobis distance to better capture the differences between normal and abnormal sounds. Anomaly detection experiments conducted on five types of machines demonstrate that DCFF-VAE not only achieves the best stability but also surpasses the best comparison method by 3.14% and 1.21% in AUC and pAUC metrics, respectively.