Single-Channel Blind Separation Algorithm for Co-Frequency Co-Modulated Communication Signals

In modern electronic warfare, deceptive jamming closely imitates the frequency and modulation of communication signals, making it highly covert and a central challenge in anti-jamming research. We focus on single-channel anti-jamming with limited information, reflecting the trend toward low-cost, miniaturized communication systems. However, existing methods in this setting often face high computational complexity, low separation accuracy, and unstable performance. To overcome these limitations, we propose the Multi-Scale Feature Separation Network (MFSNet) with Minimized Average Error Probability (AEP). It employs a feature extractor to identify high-dimensional feature differences, a feature separator to isolate features in high-dimensional space, and a feature mapper to reconstruct separation results in the original signal domain. This approach separates signals with subtle differences that are difficult to observe in the time-frequency domain. By incorporating theoretical error rate analysis, the proposed method minimizes the AEP, thereby improving the reliability of the recovered signals. Simulation results demonstrate that MFSNet significantly outperforms existing algorithms, demonstrating superior separation performance and improved generalization capability.