Parameter-invariant feature learning for radar jamming signal recognition

With the rapid development of electronic countermeasures in modern warfare, accurate Radar jamming signal recognition has become a critical challenge due to dynamic parameter variations (e.g., jamming-to-noise ratio (JNR) and pulse repetition interval (PRI) fluctuations). Existing methods suffer severe performance degradation in such environments, mainly due to parameter overfitting and limited generalization. To address this issue, we propose the Parameter-Invariant Feature Extraction LSTM (PIFE-LSTM) model, which integrates a high-dimensional feature expansion layer (mapping real and imaginary signal components to 80 latent dimensions), bidirectional LSTM networks, and a gradient-constrained attention mechanism. Under JNR (-10 to 10 dB) and PRI (±50%) variations, PIFE-LSTM achieves 78% cross-parameter recognition accuracy, outperforming BiLSTM-Att (68%) and demonstrating its robustness in capturing intrinsic jamming features, mitigating parameter sensitivity, and outperforming conventional approaches.