Robust Multi-Task Signal Classification via Multi-Scale Multi-Component Decomposition and Fusion Network

The increasing complexity of modern wireless communication systems calls for robust signal recognition techniques capable of capturing multidimensional characteristics under challenging channel conditions. Conventional single-label classification approaches often fall short in such scenarios, particularly exhibiting poor robustness in low signal-to-noise ratio (SNR) environments. To overcome these limitations, this letter proposes a Multi-Scale Multi-Component Decomposition and Fusion Network (MSCDFN) within a multi-task learning (MTL) framework for joint modulation classification and signal type recognition. The proposed MSCDFN employs a novel multi-scale decomposition module to extract both principal and residual components across multiple resolutions, followed by a hierarchical fusion strategy that effectively enhances discriminative feature representation. Experimental results on the RadComAWGN and RadComDynamic datasets demonstrate that MSCDFN consistently outperforms state-of-the-art baseline models across a wide range of SNR conditions, verifying its robustness and practical potential for real-world wireless signal recognition.