Automatic LPI radar waveform recognition of overlapping signals based on vision language model

Area of Science:

• Electrical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Low probability of intercept (LPI) microwave waveforms are crucial in modern radar systems.
• Accurate identification of LPI waveform modulation is challenging in complex electromagnetic environments.

Purpose of the Study:

• To develop a robust framework for recognizing overlapping LPI radar waveforms.
• To enhance the precision of waveform identification under varying noise and frequency conditions.

Main Methods:

• Utilized dual encoders (text and image) to align radar image and context prompt embeddings.
• Created a unified feature space for deep learning of intrinsic radar waveform characteristics.
• Achieved zero-shot classification using single modulation signal training data.

Main Results:

• High-precision identification of highly overlapping LPI waveforms was successfully achieved.
• The model demonstrated robustness against random frequency bands and varying noise interference levels.
• Gradient-weighted Class Activation Mapping (Grad-CAM) confirmed the model's practical utility and robustness.

Conclusions:

• The proposed framework enables accurate recognition of overlapping LPI radar waveforms.
• The model exhibits practical significance due to its zero-shot classification capability and robustness.
• This approach offers a valuable tool for signal intelligence in complex electromagnetic scenarios.