Automatic modulation identification for underwater acoustic signals based on the space-time neural network

In general, CNN gives the same weight to all position information, which will limit the expression ability of the model. Distinguishing modulation types that are significantly affected by the underwater environment becomes nearly impossible. The transformer attention mechanism is used for the feature aggregation, which can adaptively adjust the weight of feature aggregation according to the relationship between the underwater acoustic signal sequence and the location information. In this paper, a novel aggregation network is designed for the task of automatic modulation identification (AMI) in underwater acoustic communication. It is feasible to integrate the advantages of both CNN and transformer into a single streamlined network, which is productive and fast for signal feature extraction. The transformer overcomes the constraints of sequential signal input, establishing parallel connections between different modulations. Its attention mechanism enhances the modulation recognition by prioritizing the key information. Within the transformer network, the proposed network is strategically incorporated to form a spatial-temporal structure. This structure contributes to improved classification results, and it can obtain more deep features of underwater acoustic signals, particularly at lower signal-to-noise ratios (SNRs). The experiment results achieve an average of 89.4% at -4 dB <= SNR <= 0 dB, which exceeds other state-of-the-art neural networks.