Infrasound signal classification based on attention-CNN-BiLSTM

Accurate classification of infrasound signals holds significant value for disaster early warning and military reconnaissance. To achieve precise recognition, this paper proposes an Attention-CNN-BiLSTM fusion method. The method first extracts energy aggregation features within local frequency bands of the infrasound signal power spectrum using a Convolutional Neural Network (CNN). Subsequently, it captures long-range contextual information through a Bidirectional Long Short-Term Memory (BiLSTM) network. Crucially, a Convolutional Block Attention Module (CBAM) is embedded to adaptively weight discriminative features, thereby enhancing focus on critical features. The experiments are conducted using the Comprehensive Nuclear-Test-Ban Treaty Organization historical infrasound data and publicly available infrasound datasets. Results demonstrate that the proposed Attention-CNN-BiLSTM classification model achieves an improvement of 6.97%–13.12% in classification accuracy on the proprietary dataset compared to four classical deep learning benchmarks: LeNet-5, AlexNet, ResNet, and LSTM. Furthermore, it attained a classification accuracy of 98.06% on the public dataset. These findings indicate that the method offers high classification accuracy and exhibits strong practical potential for infrasound signal classification.