Unsupervised Radar Signal Sorting Method Based on Co-Optimization of Feature Fusion and Data Augmentation

In electronic countermeasure scenarios such as radar jamming or spectrum congestion, traditional radar emitter signal-sorting methods face significant challenges due to limited prior knowledge and scarce intercepted samples, especially under low signal-to-noise ratio (SNR) conditions. To address these issues, this article proposes an unsupervised sorting framework [fusion-augmented collaborative generative adversarial network (FACGAN) + unsupervised contrastive multitask meta-learning (UCMTML)] that integrates ambiguity function (AF)-based feature extraction, generative modeling, and self-supervised contrastive learning (SSCL). The method extracts key signal features from the orthogonal and main ridge slices of the AF to represent essential signal properties. To improve feature fusion and sample diversity, we introduce the FACGAN. Furthermore, to improve representation learning under unlabeled conditions, the UCMTML model is constructed through the incorporation of meta-learning and multitasking strategies. At an SNR of −4 dB, the proposed method improves sorting accuracy by 65.48% compared to SIMCLR. On real-world radar emitter signal data, it achieves an accuracy of 92.86%, demonstrating strong robustness and excellent performance in complex, low-SNR environments.