EARLY RADAR WORK MODE RECOGNITION USING HYBRID CNN-TRANSFORMER AND REINFORCEMENT LEARNING

In this paper, a novel approach is proposed for early recognition of Radar Work Mode, which integrates a hybrid CNN-Transformer architecture and a Reinforcement Learning strategy. The model processes Pulse Descriptor Word (PDW) sequences and adaptively determines when enough pulses have been collected to make a reliable classification decision. Our architecture replaces the traditional Recurrent Neural Network with a Transformer module to better capture global dependencies in the PDW sequence while retaining the CNN branch for local feature extraction. The model is evaluated on the Radar Mode Recognition task of a 64-pulse PDW sequence with 5 parameters (pulse width, bandwidth, repetition interval, amplitude, frequency). Experimental results show that the proposed method achieves excellent accuracy and earliness, making it suitable for application scenarios of early recognition of operating modes.