IUDR‐Net: Improved U‐Lite Denoising and Recognition Network for Automatic Modulation Recognition of LPI Radar Signals

Low probability of intercept (LPI) radars are critical in electronic warfare due to their difficulty in being intercepted. Accurate automatic modulation recognition (AMR) of LPI radar signals is essential for electronic warfare decision‐making and offers critical insights for optimising algorithms in cognitive radio (CR). However, existing deep learning‐based AMR methods have exhibited significant performance degradation in low signal‐to‐noise ratio (SNR) conditions, making this task challenging. To address this issue, we propose a novel method for AMR of LPI radar signals in low SNR conditions, named the improved U‐Lite denoising and recognition network (IUDR‐Net). The method integrates an improved U‐Lite denoising network and a recognition network, combining a two‐stage training process with a joint loss function to achieve accurate modulation recognition. Experimental results demonstrate that the IUDR‐Net outperforms state‐of‐the‐art methods in low SNR conditions as well as in a lightweight design and strong robustness. In the additive white Gaussian noise environment, the IUDR‐Net achieves an average recognition accuracy of 92.5% at an SNR of −12 dB for 13 types of LPI radar modulation signals.