Low probability of intercept (LPI) microwave waveforms are commonly utilized in multiple radar systems. Identifying precisely the modulation formats of LPI waveform is of importance in complex electromagnetic environment. We propose a framework for recognizing overlapping LPI radar waveforms. By utilizing two distinct encoders for text and image, the model effectively aligns radar image and context prompt embeddings into a unified feature space. This alignment enables the model to more deeply learn and capture the intrinsic characteristics of radar waveforms, enhancing its ability to discern subtle modulation features. On the premise of using only single modulation signal as training data, high-precision identification of highly overlapping waveforms at random frequency bands under different noise interference levels has been successfully achieved. Experimental results show that the proposed model not only implements zero-shot classification task, but also demonstrates robustness to the variation of signal frequency bands and environment interference via Gradient-weighted Class Activation Mapping (Grad-CAM), which is more practically meaningful.
Yang et al. (Wed,) studied this question.