Intention Reasoning for Unmanned Swarm Based on Cluster-Feature-Enhanced Attention Mechanism

In the context of unmanned swarm confrontation, intention reasoning for swarm targets will contribute to a deep understanding of the battlefield situation and facilitate decision-making. In order to accurately uncover the operational intentions of targets from diverse, highly dynamic, and strongly concealed intelligence information, we propose a cluster feature-enhanced Long Short-Term Memory (LSTM) network. On one hand, considering the interactions among multiple targets, a cluster spatial feature library is designed to capture the spatial patterns of the swarm, in which a series of cluster features that reflect operational intent are selected to enhance the model’s recognition capabilities. Among them, Fourier descriptor is used to encode the formations into spectrum vectors. On the other hand, we incorporate temporal attention mechanism to give corresponding attention to different historical information, highlighting actions associated with specific intents. Through model ablation study conducted on a combat simulation dataset, the model demonstrates a high level of accuracy (approximately 97.8%, a 6.7% improvement compared to the baseline model) in inferring enemy target intentions, along with a certain degree of generalization performance.