An Approach for Infrasound Event Classification Based on DenseNet-BiLSTM Fusion and Self-attention Mechanism

To solve the impact of environmental interference on the accuracy of identifying infrasound signals during propagation, a novel infrasound events classification model is proposed, which integrates dense connection convolution networks (DenseNet) and bidirectional long short-term memory networks (BiLSTM) with self-attention mechanism. The model adopts an end-to-end approach. Firstly, the original signals are processed using wide-kernel convolution and DenseNet to suppress noise while extracting detailed features, thereby enhancing the flow of information between layers of the model. Then, BiLSTM is employed to learn long-term dependencies and explore temporal sequence information from the infrasound signals. Finally, the self-attention mechanism is incorporated to focus on the temporal variations of the feature, improving the accuracy of infrasound events classification. Experimental results demonstrate that compared to traditional single-network and other fusion networks, the proposed method achieves higher classification accuracy and noise robustness.