Optimized LSTM-BiTCN parallel network model for anomalous sound detection in rotating machinery

Abstract In small sample scenarios where rotating machinery signals are heavily obscured by noise, complicating feature extraction and reducing model learn efficiency, this paper proposes an optimized LSTM-BiTCN parallel network for anomalous sound detection of rotating machinery. Firstly, fast Fourier transform is used to extract spectral characteristics, while VMD captures multi-scale temporal characteristics, mitigating key feature loss. Secondly, LSTM networks and BiTCN are utilized to construct and optimize a parallel network model, which simultaneously extracts the temporal and spatial information from the rotating machinery’s sound signals to enhance feature extraction efficiency. Then, a cross-attention mechanism is used to build temporal-spatial feature, enhancing the network model’s focus on important features; finally, the temporal-spatial features are reduced in dimensionality, and the detection results are obtained through the sigmoid function in the output layer. Experiments show that under strong noise and small sample conditions, the proposed method outperforms the DCASE2022 Task2 winning system, with an 16.2% increase in F 1-score and a 0.7% boost in AUC. It also surpasses other models, including the pre-trained model and transformer and dynamic graph convolution (Unsuper-TDGCN).