LPI Radar Signal Recognition Method Based on Adaptive Wavelet Decomposition and Deep Residual Shrinkage Network

Low Probability of Intercept(LPI) radar signals utilize complex modulation techniques to achieve low intercept characteristics, making their accurate identification in complex electromagnetic environments a challenging task. To address this issue, we propose LW-BERSNet, an automatic modulation recognition method that combines adaptive wavelet decomposition with deep residual shrinkage networks. The method initially transforms the radar signal into a time-frequency image through time-frequency analysis. Thereafter, the Learnable Wavelet Band Enhancement (LWBE) module is applied to enhance the signal’s time-frequency feature representation during the automatic radar signal modulation recognition stage. In the final stage, leveraging the extracted time–frequency representations, a Deep Residual Shrinkage Network integrated with a Cross Entropy–Dynamic Center Loss (CE–DC Loss) function is designed for model training. Experimental results in environments with various complex multipath fading effects and colored noise demonstrate that the proposed method is highly adaptable to complex conditions, achieving a recognition accuracy of over 96% even at a signal-to-noise ratio of -10 dB.