Multilabel Recognition Method for Ship-Radiated Noise Signals Based on Multidomain Information Fusion With Deep Equilibrium Models

The recognition of ship-radiated noise signals is currently the crucial means of perceiving ships. Numerous methods for the recognition of ship-radiated noise signals have been developed based on deep learning techniques. Prior studies on ship-radiated noise signal recognition have assumed a single-target scenario. In this article, we establish a multilabel recognition scenario for ship-radiated noise signals. We propose a multilabel recognition framework for ship-radiated noise signals with multidomain information fusion. Initially, we adopt two fundamental backbone network structures to extract preliminary features from both time-domain and time–frequency domain signal data. Subsequently, we construct a feature enhancement and fusion module based on the theory of deep balanced networks. This module enables information interaction from low-level to high-level between the time-domain and time-frequency domain information of ship-radiated noise signals. We introduce a transformer-based feature enhancement approach and a gated fusion feature update network structure. We also design a fusion strategy for secondary fusion updates and interdomain representations to obtain stable enhanced fusion feature representations. Finally, a linear classifier is employed to determine the categories of the mixed signals. We simulate multilabel data for ship-radiated noise signals using the publicly available Deepship data set. Experimental results demonstrate satisfactory recognition performance.