Cross-domain formations composed of Unmanned Aerial Vehicles (UAVs) and Unmanned Surface Vessels (USVs) are critical for maritime defense but face significant challenges in countering complex aerial threats and developing flexible, collaborative strategies. Addressing the limitations of traditional decision support systems in semantic understanding and dynamic adaptation, this paper proposes a novel Large Language Model (LLM)-driven decision support framework grounded in the Department of Defense Architecture Framework (DoDAF). By integrating Retrieval-Augmented Generation (RAG) with a domain-specific knowledge base, the framework enhances the LLM’s ability to align natural-language directives with standardized DoDAF view models, effectively mitigating hallucinations in tactical generation. The proposed framework coordinates a closed-loop process, using Petri net-based static logic verification to ensure structural consistency and Monte Carlo-based dynamic effectiveness evaluation to optimize the selection of kill chains. Experimental validations in a simulated UAV-USV maritime defense scenario demonstrate that the framework achieves 96.6% entity accuracy and 100% format compliance in model generation. In comparison, the generated cooperative kill chains significantly outperform non-cooperative methods by improving interception efficacy by approximately 26.08% under saturation attack conditions. This study develops an automated, interpretable workflow that transforms unstructured situational understanding into decision reporting, significantly enhancing the efficiency and reliability of cross-domain collaborative mission planning.

Air target intent recognition plays an important role in situation awareness and decision support in complex air-combat environments. However, existing methods often process motion and semantic information in the same manner, making it difficult to fully exploit their different characteristics. They also have difficulty modeling how target intent changes over different time periods. To address these limitations, this paper develops a deep learning framework that learns motion and semantic information separately and combines historical observations from multiple time ranges. A one-dimensional convolutional network extracts local movement patterns from continuous motion variables, while an embedding layer represents discrete semantic variables as dense feature vectors. These representations are synchronized in time and jointly analyzed by a Transformer. The model selectively uses recent, intermediate, distant, and overall historical information to capture instantaneous maneuvers, changes in behavioral stages, and long-term tactical trends. Experiments on an AFSIM-generated dataset show that the proposed method achieves higher recognition accuracy and better robustness than representative baseline models. Further analysis confirms that both the separate processing of motion and semantic information and the use of multiple historical time ranges contribute to the performance improvements.

The increasing popularity of civilian unmanned aerial vehicles (UAVs) poses serious safety challenges to low-altitude airspace near power transmission and distribution lines. Existing monitoring technologies rely heavily on expensive radar or radio frequency devices that are susceptible to interference, while mainstream vision schemes are typically limited to object detection and fail to analyze flight intentions. To address these problems, this paper proposes a UAV maneuver recognition method for transmission and distribution line protection, providing precise semantic support for action-based risk warnings through multi-dimensional visual feature fusion. First, a YOLO object detection model extracts real-time UAV bounding boxes and trajectories from video streams to construct raw motion state sequences. Second, to suppress detection noise while preserving critical maneuvering features, a feature-preserving trajectory reconstruction algorithm based on median absolute deviation (MAD) enhancement (FP-SGM) is proposed, which combines a Savitzky-Golay (S-G) filter with an adaptive statistical threshold. On this basis, a multi-dimensional feature space comprising trajectory and scale features is constructed to design a cascade decision strategy. By fusing these features, the lateral and radial motions of UAVs are effectively decoupled, achieving robust recognition of four typical maneuvers: ‘Hovering’, ‘Circling’, ‘Approaching’, and ‘Receding’. Simulation experiments on the Gazebo platform demonstrate a 95% accuracy rate in maneuver recognition. The results indicate that this low-cost method presents a viable potential for integration into existing surveillance terminals, laying a foundation for intelligent power line defense systems.