Xiaozhao Jin

Pursuit–evasion tasks involving multiple unmanned surface vehicles (USVs) in complex maritime environments present significant challenges in terms of adaptability, intelligence, and robustness. This paper proposes a novel multi-layer formation control method to significantly improve the maneuverability of USV formation in dynamic pursuit scenarios. The primary contributions of this work are threefold: (1) The development of a maneuverable pursuit–evasion decision layer that enables strategic encirclement of the evader is proposed, which can flexibly adjust the encirclement strategy for the evader and reduce the resource consumption of USVs based on the Apollonius circle-based multiple-angle encirclement strategy and the artificial potential field (APF). (2) A coordination layer based on the trajectory prediction of the evader is designed. This layer introduces a Head-Guidance (HG) operator into the APF to dynamically predict the evader trajectory with motion constraints and assist the Pursuit team to complete tasks. (3) A novel pursuit–evasion distributed Model Predictive Control (MPC) layer that ensures robust formation tracking and collision avoidance is proposed. This layer is specifically engineered to handle not only dynamic obstacle constraints but also lumped uncertainties. The overall performance of the proposed multi-layer framework is verified via abundant numerical simulations under complex marine disturbances and multiple obstacle environments, and the developed fixed-time extended state observer is validated with a real-water field experiment with a single self-developed USV prototype.

In this paper, an adaptive tracking control technique is proposed to stabilize the yaw angle of an unmanned surface vessel (USV) system subject to input saturation and external disturbances. Unlike existing saturation systems and disturbance rejection methods, the proposed control scheme simultaneously considers input saturation and unknown disturbances in a unified adaptive framework. First, an adaptive second-order system is presented to compensate for control input saturation. Such a saturated system can effectively reduce the amplitude of controller output. Then, an adaptive learning law is designed to estimate online the unknown upper bound of an external disturbance, removing the requirement for prior information on the disturbance. Based on Lyapunov stability theory, the convergence of the tracking error of closed-loop system and the uniform boundedness of all closed-loop signals are strictly proved. Furthermore, the transient performance of the real yaw tracking error is analyzed by considering the selected Lyapunov function, and an explicit upper bound of the transient error is derived for the first time in USV yaw control. Finally, simulations are carried out to illustrate the effectiveness of the proposed controller.

U-shaped Bolts & Nuts (UBNs) are essential for ensuring a smooth and continuous power supply to trains. UBN defects can lead to a mismatch between the pantograph and the overhead line, potentially causing severe damage to the pantograph and even halting train operations. Unfortunately, a specialized algorithm for detecting small-size UBN defects in high-resolution catenary inspection images under intricate real-world environments is notably absent. To remedy this, we propose a multiscale cross-attention convolutional neural network (CNN)-transformer two-branch fusion network (M2CTFNet). This network accurately detects small-size UBN defects across various natural scenes by integrating global and local features through a weight-based multiscale cross-attention fusion, and building a multiscale shuffling technique in the transformer branch. Extensive experiments demonstrate that M2CTFNet achieves a 2.7% higher precision and a 2.4% higher recall rate with lower computational cost and fewer weight parameters in UBN defect detection compared to state-of-the-art networks. Furthermore, M2CTFNet has been deployed on multiple inspection vehicles and has successfully detected several hundred UBN defects. These results verify that M2CTFNet is effective for addressing small object defects in challenging natural environments.