Dual-Path Transformer For Machine Condition Monitoring

Anomalous sound detection (ASD) aims to detect anomaly sounds for earlier warning. Recently, with the development of machine learning methods for automatically detecting anomalous situations, ASD has attracted much attention. Previous work mainly focuses on finding acoustic patterns by deep neural networks, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs). CNNs can catch local characteristics, but are difficult to model long sequences. RNNs are able to model long-time dependencies, but are timeconsuming for training. In this paper, we propose DPTrans, a novel dual-path Transformer-based neural network for machine condition monitoring in ASD. DPTrans consists of several encoders, in which acoustic features are modeled on time-dimension and frequency-dimension by Transformer encoders. DPTrans can learn temporal and frequency dependencies and model interactive information effectively by taking advantages of self-attention module. Finally, we evaluated DPTrans on the dataset of DCASE2021 task2, the averaged AUC score was improved by 12 % compared with official baseline systems.