Modulation recognition in underwater acoustic communication is challenging due to channel distortion and non-cooperative scenarios. This paper proposed a method based on Multi-Path ResNet to extract and fuse time-domain, frequency-domain, and multi-order spectral features of received signals for improving recognition performance. Moreover, for practical non-cooperative applications, a per-class distance thresholding method was proposed to reject unknown modulation types by leveraging the proximity of intra-class samples in the logits space and establishing decision thresholds individually for each class. To validate the proposed methods, this paper conducted both simulation and sea trials. Simulations using measured underwater acoustic channels confirmed the superior performance of both proposed methods in shallow water. Deep-water sea trials in the Western Pacific collected and established a dataset of 6776 samples, on which the proposed system achieved the recognition accuracy of 64.1%, while the rejection accuracy for unknown categories reached 74.4%. Furthermore, the study also revealed that recognition performance is directly correlated with the delay spread and dynamic variation of channels, offering physical insights into improving deep learning-based modulation recognition in underwater acoustic communication.
Zhang et al. (Fri,) studied this question.