MFD-MoE: Multi-scale Fusion Denoising Mixture of Experts Network for Stable Recognition of Underwater Acoustic Targets

Abstract

Stable recognition of underwater acoustic targets is a key technology for ensuring continuous and efficient maritime defense, resource exploration, and environmental monitoring. However, in complex marine environments, background noise, propagation disturbances, and dynamic changes in target features often lead to unstable recognition results. To address this issue, this paper proposes the Multi-scale Fusion Denoising Mixture of Experts (MFD-MoE) Network, with a focus on improving recognition stability. We design the Denoised Time-Frequency Feature Fusion module (DTFF), in which first performs soft-threshold denoising, followed by a feature decomposition strategy that includes key steps such as channel splitting, multi-scale feature extraction, and upsampling. We then design a multi-scale feature fusion strategy that combines multi-modal and pooling results, enhancing the stable expression of features in the presence of interference and target state changes, resulting in a feature extraction architecture consisting of four DTFF layers. Furthermore, we introduce a mixture of experts model, and based on the characteristics of underwater acoustic information recognition, we design an expert mixture mechanism with residual connections to improve the model's stability under target motion or interference. Experimental results show that the MFD-MoE model significantly outperforms existing methods in multi-class underwater acoustic target recognition tasks, offering both high stability and high accuracy, thus providing a new solution for intelligent underwater target perception in complex environments.