Intelligent recognition of radar emitters with agile waveform based on deep reinforcement learning

Abstract In the traditional radar emitter recognition task, the number of emitter category tags is fixed and relatively small. We input the pulse characteristic parameters into the machine learning model to model the radar emitter. The radar emitters with agile waveform involves four different waveform categories. The number of specific emitter categories contained in different waveform categories is extremely unbalanced. Their parameters vary widely, so only using the same model or strategy to model each waveform category cannot effectively extract the core features of the radar emitters with agile waveform. The above is the main reason for the low recognition accuracy of the recognition task of radar emitters with agile waveform. In this paper, a deep reinforcement learning framework is designed to be used in the intelligent recognition task of radar emitters with agile waveform. The task is regarded as a two‐step decision game. We use CNN and Bi‐LSTM to model the radar emitter, and calculate the initial state and transition state of the game. At the same time, we design the penalty function in reinforcement learning and increase the penalty for the wrong decision in the first step to deal with the imbalance of the number of emitters between different waveform categories. Finally, the Q ‐Learning algorithm with approximate values is used to learn the control strategy of the game, that is, the modeling and recognition strategy adopted for emitters of different waveform categories. The simulation experiment results show that the deep reinforcement learning framework constructed in the article can improve the recognition accuracy of 1.2% compared with the state‐of‐the‐art method in the intelligent recognition task of radar emitters with agile waveform.