Automatic Modulation Classification for Short Burst Underwater Acoustic Communication Signals Based on Hybrid Neural Networks

Abstract

Automatic modulation classification (AMC) is challenging for short burst underwater acoustic (UWA) communication signals. Difficulties include but are not limited to the poor UWA channels, impulsive noise, and data scarcity. To address these problems, a method based on hybrid neural networks (HNNs) is proposed in this paper. First, an impulsive noise preprocessor is adopted to mitigate the impulse in the received signals. Subsequently, an HNN consisting of an attention aided convolutional neural network (Att-CNN) and a sparse auto-encoder is built to extract features from the temporal waveforms and square spectra of the preprocessed signals after burst detection. Finally, a late fusion is made to combine the prediction results of the two sub-networks. To overcome the variable signal duration relative to the fixed input size of the Att-CNN, a data-reusing approach is proposed to perform dimension preprocessing on the waveforms. Moreover, a transfer learning strategy is introduced to resolve the issue of insufficient training data from the testing channel. The results of simulation experiments and practical signal tests both demonstrate that the proposed method is robust against UWA channels and ambient noise. Our approach significantly outperforms existing deep learning-based methods in dealing with short and weak signal bursts, while requiring less training data from the testing channel.