MSANet: Multi-Stage attention network for anomalous sound detection in machine condition monitoring

Abstract

Anomalous Sound Detection (ASD) system identifies sound waves from sensors to detect the anomaly of industrial machines. However, recent methods have failed to sufficiently focus on partial details and long-term dependence information in acoustic features, resulting in poor performance on certain machine types. To address this challenge, we propose a novel ASD model based on the multi-stage attention network (MSANet). The spectral-temporal concatenated spectrogram of the audio samples is used as the MSANet input, and serially modeled by the network. The fusion spectrogram attention network (FSAN) enhances inter-spectrogram correlation via directional pooling and attention weighting. Convolutional Block Attention Module (CBAM) is used in the local attention network to focus on the channel and spatial information in acoustic vectors, hence improving capacity of ASD system for modeling local information. In global attention network, the gated recurrent unit (GRU) is applied to improve the feed-forward layer of transformer, enhancing the model to capture global correlation feature and contextual information. Extensive experiments are conducted out on the DCASE 2020 Challenge Task 2 dataset to evaluate the proposed model. Experimental results demonstrate that MSANet achieves an average AUC of 94.89 % and an average pAUC of 89.11 %, both of which surpass the performance of previously methods.