Multifunction radar (MFR) systems dynamically switch among diverse working modes according to mission requirements and environmental conditions. Accurate, real-time recognition of these modes is essential for radar intelligence and electronic countermeasures. Existing pulse descriptor word (PDW)-based approaches depend on pulse-level parameter extraction, suffering from accumulated processing latency and performance degradation under low signal-to-noise ratio (SNR) and severe pulse-loss scenarios. To address these limitations, this article proposes an end-to-end signal-level framework for streaming MFR working mode recognition. The framework directly processes the time-frequency spectrogram of the received intermediate-frequency (IF) signal to generate temporally aligned working-mode labels at the input sampling resolution. To enable efficient online inference, a perception-knowledge-memory network (PKM-Net) was further proposed. It uniquely combines a Transformer-based perception module for local feature extraction with a GRU-based interparagraph memory module to capture long-range contextual dependencies while maintaining a constant memory cost. The experimental results demonstrate that PKM-Net achieves 99.88% accuracy, significantly outperforming recent state-of-the-art (SOTA) baselines in robustness tests. Furthermore, with a compact footprint (10.43 MiB) and low latency (1.64 ms), it offers a feasible solution for real-time online processing in complex electromagnetic scenarios.
Liu et al. (Thu,) studied this question.