Multi-Domain Fusion for High-Sensitivity Interference Identification: A Lightweight Model Using Multi-Scale Extraction and Compression

Multi-domain fusion (MDF) has become a new approach in wireless interference identification (WII) especially at low interference-to-signal ratios (ISR), thanks to the complementarity of diversified domain features. As the development of WII at miniaturized edge devices, the demand for low-complexity models is becoming increasingly urgent. However, lightweighting the MDF model faces the significant contradiction between efficiency and accuracy, due to the severe imbalance, redundancy, and the larger model size from multiple independent domains. To address these challenges, we propose a lightweight multi-domain cross-fusion model which integrates the information bottleneck with enhanced feature extraction and fusion modules to achieve high-sensitivity WII with ultra-low complexity. We consider the time-frequency and fractional Fourier domains and propose a novel lightweight dual-branch multi-scale feature extraction module to capture global and local features. We then propose an efficient fusion module consisting of a cross-attention mechanism and depthwise separable convolution to significantly decrease complexity. To enhance WII performance at low ISRs, the proposed model introduces a multi-domain information bottleneck to reduce redundancy by maximizing the mutual information lower bound. Finally, we design an innovative adaptive weighted cascaded loss to adjust the importance of features from different domains and scales, while proving the optimal loss weights that maintain multi-domain balance. Experimental results show that the proposed model reduces the parameter requirements for feature extraction by 91.7% compared to the advanced lightweight model, and improves the WII accuracy by 11% at the low ISR compared to the SOTA MDF model.