Radar Signal Pulse Train Recognition With Dual-Branch LSTM-Transformer Networks

Radar pulse train recognition plays a crucial role in modern electronic warfare (EW), serving as the foundation for further analysis of radar signal sources and operating modes. Existing studies primarily focus on the extraction and analysis of pulse repetition interval (PRI) modulation features, while the utilization of other key parameters remains insufficient. To address this issue, this paper proposes a radar pulse train recognition method based on a dual-branch LSTM-Transformer network, with optimizations in both input design and network architecture. In terms of input, the traditional complex manual feature extraction process is bypassed, and only normalized PRI sequences, frequency sequences, and their first-order difference sequences are used, significantly simplifying the input data. Regarding network architecture, the method combines the temporal modeling ability of LSTM with the global representation strength of Transformer, while also constructing a dual-branch framework and introducing a gated adaptive feature fusion mechanism to enhance feature representation and recognition performance. Experimental results demonstrate that the proposed method outperforms existing approaches in recognition accuracy, fine-grained classification, and noise robustness, highlighting its robustness and practical application value in complex electromagnetic environments.