Modulation Recognition of Underwater Acoustic Signals Using Dynamic Dilated Convolutional Neural Network and Transformer

This study proposes a parallel two-branch Dynamic Dilated Convolutional Neural Network (DDCNN) Transformer model for identifying modulation types of underwater acoustic signals. In the signal preprocessing stage, wavelet transform and instantaneous autocorrelation processing are used to extract the signal's time-frequency and instantaneous autocorrelation characteristics, constructing a joint feature plane for model training. To enable the model to extract more refined feature information, a dynamic dilated convolution (DDConv) module was proposed and combined with traditional convolutional neural networks to form a DDCNN network, enhancing feature extraction capability over the original CNN. By simultaneously combining the transformer model, a parallel dual-channel network model is formed. The DDCNN branch extracts local information, while the transformer branch extracts global information. A weighted sum is used for classification and recognition. Simulation results show that the proposed model achieves a classification accuracy of 95.6% for eight types of modulation signals, demonstrating good classification and recognition ability.