Channel Replay Aided Modulation Classification of Underwater Acoustic Communication Signals

Abstract

Modulation classification or recognition of underwater acoustic (UWA) communication signals faces great difficulties due to the unpredictable characteristics of underwater acoustic channels as well as lack of training data. As traditional feature-based methods work poorly in such scenarios, resurgent artificial neural networks are gaining more attentions in the area of automatic modulation classification. Deep learning, however, demands large amount of training data for good recognition performance. To address the problem of insufficient training data, we propose to adopt the channel replay technique. With this technique, channel realizations with the same statistical characteristics as the measured underwater acoustic channel, can be generated as much as one needs. Training and validation data sets can then be generated with the replayed channels. A residual network (ResNet) is adopted to achieve simultaneously feature extraction and modulation recognition. It is shown the proposed ResNet not only classifies artificial test data but also experimental data collected in an at-sea UWA communication trail.