Automatic Radar Modulation Recognition Using CRLNet with Rotary Encoding and LCA Mechanism

In modern electronic countermeasures, accurately identifying the modulation patterns of enemy radar signals is of vital importance for electronic reconnaissance and intelligence acquisition. A hybrid network model named CRLNet based on the convolutional neural network (CNN) and residual modules is proposed, which innovatively introduces the rotary position embedding (RoPE) and the linear-convolutional attention (LCA) mechanism. RoPE enhances the model's ability to locate key signal features in the environment with low signal-to-noise ratios (SNR) by encoding positional information. Meanwhile, the LCA combines the advantages of the linear attention and the convolutional block attention module (CBAM), improving the processing efficiency of long-sequence signals and the ability to extract features of weak signals. Through experimental evaluation, for 11 kinds of extremely complex radar signals, the network proposed in this paper can achieve a recognition accuracy of over 80% at −8 dB and over 90% at 0 dB. Moreover, as the noise weakens, the accuracy gradually approaches 99%, demonstrating great identification performance.