Bearing fault diagnosis is vital for maintaining the reliability and security of industrial machinery. Although deep learning approaches are widely utilized, the scarcity of labeled fault data significantly hinders the effective training of large-parameter fault diagnosis models rooted in deep neural networks. Moreover, such models often exhibit a lack of interpretability. To tackle these challenges, this study proposes a novel hybrid-domain fault diagnosis framework for bearings, synergistically enhanced by adaptive feature optimization and Ensemble Learning (EL). The core methodology involves three principal innovations: (1) the extraction and fusion of condition-indicative features from different domains to construct a comprehensive representation, overcoming the insufficiency of single-domain analysis; (2) the development of an adaptive feature optimization mechanism that automatically selects the most discriminative features from the hybrid domain, enhancing diagnostic relevance and reducing dimensionality while providing physical insights into the diagnostic process; (3) the formulation of a robust diagnostic model that integrates the optimized hybrid-domain features with an ensemble learning strategy to significantly boost diagnostic accuracy and stability. Comparative experiments were conducted, comparing models utilizing single-domain features, simple hybrid-domain fusion, hybrid-domain with adaptive optimization, and the proposed framework. Experimental evidence indicates that the proposed hybrid-domain model possesses superior fault diagnosis performance, validating that the adaptive optimization within the hybrid domain effectively retains critical information while EL ensures reliable generalization. This work presents a data-efficient, interpretable, and high-performance methodology for intelligent bearing health management.
Trajectory planning for a multi-UAV system requires jointly considering obstacle avoidance, inter-UAV conflict avoidance, and target reaching. To address this problem, this paper proposes a graph attention network-based method for multi-UAV trajectory planning. The multi-UAV system is represented as an interaction graph, where UAVs are modeled as nodes and communication-based inter-UAV relationships are modeled as edges. For each UAV, local perception, target-related direction information, previous motion direction, and neighborhood information are integrated into the node representation, while the relative geometric relationship between neighboring UAVs is used as the edge feature. The constructed graph is fed into a multi-head graph attention network to extract interaction-aware features and output an action score vector over discrete flight direction labels for each UAV. During online execution, candidate flight actions are generated according to the action scores, and the final action is selected using the geodesic cost-to-go map. The trajectories of all UAVs are then generated step by step through the online decision process. By combining local perception, target guidance, motion history, and inter-UAV interaction information, the proposed method can learn cooperative action preferences for multi-UAV trajectory generation. Experiments are conducted on different flight maps and swarm sizes using multiple performance metrics. The results show that the proposed method achieves effective performance in mission success, flight efficiency, and safety-related metrics, and it also demonstrates generalization ability on unseen maps. Compared with a CBF-based collision avoidance method, the proposed method achieves better performance in task completion, inter-UAV collision avoidance, and trajectory efficiency.
Action chunking strategies in robot imitation learning struggle to dynamically balance between long-range motion efficiency and short-range operational precision due to their fixed planning horizon. This paper presents an Adaptive Action Chunking framework that enables robots to dynamically predict the optimal action chunk length based on real-time visual context. We design an end-to-end dual-branch network comprising a shared visual encoder, a parallel action prediction head, and a chunk-size prediction head. Experiments on two real-world bimanual robot manipulation tasks (transport-and-place and flip-and-handover) demonstrate that the method autonomously derives two distinct intelligent strategy patterns—phase-aware switching and sustained high-frequency adjustment—in response to task uncertainty. It significantly outperforms fixed-chunk baselines in both success rate and efficiency. Ablation studies confirm that the performance gain stems from the adaptive decision-making mechanism itself.