An open-set recognition method for ship radiated noise signals based on hybrid supervision

Abstract

Ship radiated noise signals are one of the most crucial sources of information for ship perception. With the intensification of human marine activities, the underwater acoustic environment has become increasingly complex, subjecting underwater acoustic perception systems to various interferences. In this paper, we construct a recognition scenario involving multiple target ship radiated noise signals and various types of interference. In this scenario, target signals comprise various ship radiated noise, whereas interference signals consist of both decoy and non-target ship noise. The objective of this study is to accurately recognize multiple target signals while mitigating the effects of interference. We formulate this as an open-set ship radiated noise recognition problem. We design an encoder–decoder architecture for processing time-domain ship radiated noise. Pre-trained on a closed-set dataset, this model effectively captures the distribution of closed-set data. Furthermore, we design a classifier integrating supervised and self-supervised learning, augmented with an attention mechanism to enhance its representation learning capability. We emulate decoy signals, and all experimental data are collected from real sea trials. Experimental results demonstrate that our method can accurately recognize multiple target ship radiated signals while remaining robust to interference.