This paper designs a deep learning model MLSTM-FCN in combination with the advantages of fully convoluted neural network, recurrent neural network and compression and excitation module aimed at the problems that the existing air target recognition methods are not high enough in agility and reliability. The complex local features can be extracted from the air combat data by the fully convoluted network, and the long and short memory neural network can capture the temporal features of air combat intention data. The results of ablation experiments and comparative experiments show that the MLSTM-FCN model is superior to the existing air target intention recognition model in terms of intention recognition accuracy, reaction speed and anti-interference ability, and the results of sota are obtained, providing a more effective basis for commanders in making air combat decisions.