Abstract:

To enhance the accuracy of radar jamming recognition and quantify the confidence of model judgement results, a radar jamming recognition method based on multi-domain refined feature extraction and subjective logic is proposed. Firstly, an attention long short-term memory module is designed to leverage global temporal dependencies in signals based on one-dimensional convolution features and dynamically quantify the contribution of different time segments to the classification task, achieving refined feature extraction in the time domain. Secondly, based on the shape prior of interference signal time-frequency diagram, a multi-angle strip pooling module is designed, which uses strip pooling kernels at different angles to aggregate contextual information and generate attention masks, thereby enhancing model key region localization ability and suppressing background noise and relalizing time-frequency domain fine feature extraction. Finally, the concatenated and fused feature vectors are mapped to the Dirichlet distribution space through a fully connected layer, and the network’s judgment results on the input signals and their reliability measures are modeled based on the subjective logic theory. Experimental results demonstrate that the proposed method achieves a 1.89% improvement in accuracy over the next-best approach and provides accurate measurement of prediction confidence, showing dual advantages in recognition performance and reliability measurement.

Key words: radar jamming recognition, attention long short-term memory (A-LSTM) module, multi-angle strip pooling (MASP) module, subjective logic theory