Abstract: Radar signal classification is a key aspect of radar electronic reconnaissance in modern warfare. However, as the electromagnetic environment becomes increasingly complex, the signals intercepted by reconnaissance receivers often suffer from a large number of missing pulses, false pulses, and pulse jitter. The complex electromagnetic environment is characterized by long-range signal correlations, diverse data distributions, and complementary information across different time steps. Existing Bi-LSTM-based radar signal sorting methods perform inadequately, primarily due to their insufficient ability to model global dependencies, risks of gradient vanishing and overfitting, and limited feature interaction. This paper proposes an integrated attention Bi-LSTM method that addresses these issues through the incorporation of multi-head self-attention mechanisms, layer normalization, and residual connections. The main improvements include: 1. Data Preprocessing: Converting the original Time of Arrival data into Time Difference of Arrival data to enhance the robustness and accuracy of the input data. 2. Feature Extraction and Fusion: Utilizing Bi-LSTM layers to extract both forward and backward temporal information, while introducing a multi-head self-attention mechanism for deep feature fusion, balancing the capture of long-range dependencies and local detail extraction. 3. Optimizing Training Stability: Employing residual connections and dropout regularization to ensure smooth gradient propagation. Experimental results indicate that this method achieves significantly improved accuracy compared to traditional Bi-LSTM approaches, and the structural improvements effectively enhance the model's robustness against missing pulses, false pulses, and pulse jitter.

Key words: radar signal sorting, Bi-LSTM, multi-head self-attention, missing, false pulses and pulse jitter