An Intelligent Inference Method for Combat Simulation Based on the Semi-Supervised Two-Stage Domain Feature Alignment Model

With the rapid development of information technology and artificial intelligence, combat simulation has become an indispensable tool in military assessment. The complexity of combat environments presents significant challenges for simulation systems, particularly in heterogeneous scenarios where disparities in feature distributions undermine the effectiveness of traditional inference methods. Cross-domain transfer inference has therefore emerged as a promising solution. However, existing approaches often fail to adequately balance the preservation of intra-domain features with the alignment of inter-domain representations. In addition, feature alignment loss is introduced at early training stages before source domain feature learning has sufficiently converged, which can degrade transfer performance and reduce inference accuracy. To address these limitations, this study proposes a novel domain adaptation (DA) framework that incorporates a dual feature extraction mechanism and a staged training strategy. This framework facilitates effective inter-domain alignment while preserving domain-specific information, and avoiding the detrimental effects of premature feature alignment. Experimental results demonstrate that the proposed model outperforms in various cross-domain scenarios, particularly when labeled target domain samples are scarce, significantly improving prediction accuracy while exhibiting strong generalization and robustness. This model offers an innovative solution and theoretical support for addressing cross-domain feature alignment and transfer inference challenges in combat simulation systems.