The global transition to clean energy has intensified demand for intelligent wind power systems, yet operational challenges persist due to fragmented knowledge and limited integration across business processes and maintenance management. While large language models (LLMs) exhibit strong reasoning and orchestration capabilities, their application in wind power systems remains constrained by inefficient retrieval-augmented generation (RAG) over heterogeneous knowledge bases, insufficient multimodal data integration, and limited edge resources. To address these challenges, we propose a MetaGPT-based LLM agent framework that dynamically integrates tools, lightweight models, and domain-specific databases to enable seamless knowledge orchestration. The framework introduces mix-augmented generation (MAG), which combines cache-augmented generation (CAG) with an optimized RAG pipeline to enhance retrieval efficiency and semantic alignment, outperforming both RAG and CAG through adaptive knowledge integration and dynamic query routing. In addition, the framework employs an activation-aware weight quantization strategy to reduce model size and memory usage while maintaining minimal accuracy loss. Evaluated on the EleQA wind energy dataset, the proposed framework demonstrates coordinated reasoning and task execution capabilities within the multi-agent system, while significantly reducing retrieval latency and accelerating inference. These results highlight its effectiveness and efficiency for real-time edge deployment in wind power system operations.
The orchestration of heterogeneous swarm systems, such as Unmanned Aerial Vehicle (UAV) and Unmanned Ground Vehicle (UGV) teams, in unknown, GPS-denied environments requires coordination among active perception, collision avoidance, and topological flexibility. Traditional Artificial Potential Field (APF) methods can suffer from local-minimum deadlocks, while optimization-based Control Barrier Function Quadratic Programs (CBF-QPs) may produce high-frequency control variation in narrow corridors. To address these limitations, this paper proposes the Riemannian Information-geometric Swarm Motion Orchestration (RISMO) framework. RISMO integrates objective-driven navigation, perception-aware motion generation, and obstacle avoidance within a common non-Euclidean metric representation. We construct a Riemannian metric tensor that incorporates the Fisher Information Matrix (FIM) to encourage the UGVs to preserve line-of-sight tracking of the UAVs. To accommodate formation deformation in narrow passages, the swarm topology is modeled as a Semi-Markov Jump System (S-MJS). Under the enforced positive per-agent inter-event bound and the stated Average Dwell Time (ADT) condition, the modeled switching sequence has a bounded transition frequency and cannot exhibit Zeno accumulation. Furthermore, a goal-aligned tangential flux supplies a deterministic local directional bias near saddle-like stagnation. The numerical and ablation results show that this mechanism facilitates escape from the tested non-convex traps. In the planar simulations, RISMO reduces the observed gradient stagnation of APF and the high-frequency control variation of CBF-QP while preserving positive obstacle clearance and cooperative perception.
Abstract LiDAR-based Simultaneous Localization and Mapping (SLAM) systems suffer from environment-induced degeneracy in geometrically self-similar scenes such as tunnels and cylindrical structures, where the point cloud registration problem becomes ill-conditioned. Existing degeneracy detection methods predominantly rely on eigenvalue analysis of the Hessian matrix, which assesses geometric conditioning at a fixed linearization point. However, in semi-degenerate environments containing partial but sufficient geometric constraints, these methods tend to produce false positive detections because measurement noise inflates eigenvalues and obscures the contribution of minor structural features. In this paper, we propose the Optimization Convergence Efficiency (OCE) metric, which evaluates the relative contraction of the directional optimizer gradient across Levenberg-Marquardt evaluations and directly characterizes the convergence response of LiDAR registration along the monitored state direction. The proposed method is evaluated through controlled simulation with ground-truth-based quantitative assessment and real-world UAV experiments in a wind-turbine tower and a long tunnel. In the nominal simulation, OCE achieves 98.21% recall with a 2.33% false-positive rate, compared with 100% recall and a 49.42% false-positive rate for the conventional eigenvalue-ratio detector. Controlled measurement-noise experiments further demonstrate that OCE maintains substantially lower false-alarm rates as spectral discrimination becomes less reliable. These results show that OCE can retain usable LiDAR pose estimates in semi-degenerate environments while reliably identifying actual registration failure.
Abstract Rapid and reliable localization of survivors is critical in post-disaster search and rescue operations. However, conventional sensing methods such as visual imaging are often ineffective due to smoke, dust, or debris, and traditional RSSI-based localization suffers from severe multipath interference and noise. This paper proposes a subspace-driven multi-UAV system for dynamic crowd monitoring and rescue prioritization. The system employs Boustrophedon scanning to acquire spatially continuous RSSI sequences from survivors’ Wi-Fi devices. A spatial resampling technique is first applied to regularize the irregular measurements, followed by Hankel matrix construction and Singular Value Decomposition (SVD) to separate signal from noise, yielding robust coarse device locations. A multi-pass scanning strategy tracks devices over time, detects mobility as an indicator of vitality, and maintains device records with expiration mechanisms. Detected devices are mapped to predefined rescue zones, and a weighted priority score is computed for each zone based on device count, mobility, and temporal trends. Simulation results under low and high noise conditions demonstrate that the proposed method significantly outperforms Kalman-filtered trilateration and SVD without resampling. At 6 dB noise, it achieves a 46.7% reduction in RMSE and a 60.0% reduction in outlier ratio. Furthermore, spatial resampling reduces computational time by 17% per device-line, with greater gains expected in high-frequency scenarios. The proposed framework transforms raw RSSI measurements into actionable intelligence for rescue command, enabling efficient resource allocation.
In autonomous driving trajectory prediction, it is important to generate multi-modal trajectories. As a generative method, diffusion model has been increasingly adopted in the field of trajectory prediction. In this paper, we propose CDJMP (Conditional Diffusion model-based Joint Motion Prediction), an adaptive conditional diffusion framework designed for multi-agent trajectory forecasting in autonomous driving. However, when conventional diffusion models generate multi-modal trajectories, particularly for multi-agent joint prediction, they often fail to balance the diversity of generated trajectories and the accuracy of prediction. At the same time, diffusion models also suffer from time-consuming inference. To address these issues, CDJMP utilizes a two-stage framework to generate diverse and highly accurate trajectories. In the first stage, a probabilistic initializer predicts proposal trajectories together with adaptive denoising steps. In the second stage, a group-aware conditional encoder captures dynamic multi-agent interactions and guides the diffusion process to produce coherent multimodal outcomes. Experiments on the INTERACTION and Argoverse datasets demonstrate that CDJMP achieves state-of-the-art performance, reducing minADE and minFDE by up to 9
