The air battlefield situation provides a general description of the actions and states of all participants， while target combat intention recognition serves as a critical foundation for air battlefield posture assessment. To address the challenge of acquiring a large amount of labeled air target battlefield data amidst intense confrontation and rapidly evolving scenarios， a semi-supervised air targets combat intention recognition model is proposed based on metric learning. The model offers a method for uncovering potential patterns from unlabeled samples， thus reducing the reliance on extensive labeled data. In this model， the target time-series data encoder reduces the dimensionality of target data and produces an embedded representation. Based on this， loss values are calculated by measuring the similarity between labeled target sequences and intention types， as well as between labeled and unlabeled target sequences. The experimental results demonstrate that the model achieves combat intention recognition accuracy rates of 86%， 89%， and 91% with labeled sample rates of 30%， 40%， and 50%， respectively.

Blind source separation of complex signals composed of reconnaissance， interference， detection， and communication signals is the first step in the processing of received signals in the integrated reconnaissance， interference， detection， and communication system， which requires higher accuracy of blind source separation. The traditional blind source separation method based on independent component analysis has such defects as the easy acquisition of local optimums and poor separation performance. To solve these problems， this paper proposes an improved pigeon-inspired optimization algorithm for blind source separation of complex signals. A location factor is added to the map and compass operators of the pigeon-inspired optimization algorithm， and a compression factor is added to the landmark operator. In this way， the global exploration ability in the early stage and the local search accuracy in the later stage of the algorithm are balanced， and the problems of the easy fall into local optimums and premature convergence are solved. The optimization ability and convergence speed of the algorithm are improved. Simulations show that the algorithm proposed in this paper can better separate complex signals under low and strong noise. Compared with the traditional independent component analysis method， the algorithm has better separation performance and convergence speed.