Physics-Guided Variational Causal Intervention Network for Few-Shot Radar Jamming Recognition.

Area of Science:

• Artificial Intelligence
• Signal Processing
• Causal Inference

Background:

• Accurate radar active jamming recognition is crucial for cognitive electronic countermeasures.
• Deep learning models struggle with complex electromagnetic environments and scarce training data, leading to performance degradation due to spurious correlations.

Purpose of the Study:

• To develop a novel method for robust radar active jamming recognition under conditions of limited training samples and complex electromagnetic environments.
• To address the issue of causal confounding in deep learning models for electronic warfare applications.

Main Methods:

• Proposed a physics-guided variational causal intervention network (PG-VCIN) integrating physical mechanisms and causal inference.
• Reconstructed a structured causal model to decouple jamming signal observations into physical statistical features and time-frequency representations.
• Leveraged physical priors for precision-weighted modulation and employed an active inference framework with a variational information bottleneck for deconfounding.
• Imposed dual intervention constraints (intra-class invariance and confounder invariance) in the latent space to approximate causal effects.

Main Results:

• The PG-VCIN method achieved substantially higher recognition accuracy compared to representative few-shot baselines.
• Demonstrated significant performance improvements in extremely low-sample regimes.
• Validated the effectiveness of integrating physical mechanisms with causal inference for jamming recognition.

Conclusions:

• The proposed PG-VCIN effectively mitigates causal confounding in radar active jamming recognition.
• The integration of physics-guided principles and causal inference offers a promising direction for enhancing cognitive electronic countermeasures.
• The method shows strong potential for real-world applications requiring accurate signal recognition with limited data.