Radar Working Mode Fusion Recognition Based on DS-BiLSTM

To address the limited reconnaissance perspective and susceptibility to noise interference of single-platform approaches in multifunction radar (MFR) working mode recognition, this paper proposes a fusion recognition model based on bidirectional long short-term memory networks (BiLSTM) and Dempster–Shafer (DS) evidence theory. The model first employs a BiLSTM network to extract deep temporal features from radar phrase sequences observed by a single platform, thereby obtaining the state probability distribution for each platform. The probability distributions from multiple platforms are then treated as independent evidence and fused at the decision level using DS evidence theory to enable accurate recognition of radar working modes. Experiments construct a multi-platform co-simulation dataset based on the operational process of a typical “Mercury” radar and compare the proposed method with various algorithms, including the hidden Markov model–Viterbi (HMM-Viterbi), recurrent neural network (RNN), and long short-term memory (LSTM) algorithms. The results show that the proposed model achieves excellent recognition performance under varying noise and missing-data rates. Under the extreme condition of both a 25% noise rate and a 25% missing rate, the recognition accuracy still reaches 95.64%, which is 3.82 percentage points higher than the 91.82% achieved by the conventional HMM-Viterbi algorithm, validating the model’s high accuracy and strong robustness in complex electromagnetic environments.