Gaoliang Peng

Vehicle-mounted optical platforms are susceptible to imaging instability under broadband vibration environments. Semi-active control strategies, particularly those employing magnetorheological (MR) dampers, have been widely adopted to mitigate such effects. While MR dampers with semi-active algorithm are effective in suppressing large-amplitude excitations, frequent variations in damping force during control may excite high-frequency structural modes, thereby deteriorating imaging performance. To overcome this issue, a fuzzy skyhook control strategy with integrated threshold limitation and slope transition algorithm (TAS), is proposed. An auxiliary oscillator is introduced to emulate the real modal characteristics of the optical platform, and both forward and inverse models of the MR damper are developed for control implementation and comparative analysis. Results demonstrate that, relative to existing control strategies, the proposed TAS method not only effectively attenuates low-frequency, large-amplitude vibrations but also significantly reduces acceleration transmissibility around the high-frequency structural mode (45.1 Hz). Specifically, under road excitation across a range of vehicle speeds, TAS achieves at least a 40% reduction in transmissibility compared with conventional skyhook control.

Mobile precision equipment requires efficient and accurate road surface perception for reliable operation. However, existing road classification methods, designed mainly for single-type surfaces, perform poorly on mixed surfaces, limiting system adaptability in complex environments. To address this gap, we constructed a dedicated dataset comprising images of four single-type and two mixed road surfaces, acquired using a custom-built intelligent vehicle equipped with multiple sensors. Each image was meticulously annotated to facilitate model training and evaluation. Leveraging this dataset, we propose a deep learning model, termed GS-ResNet, which integrates squeeze-and-excitation (SE) modules, a gating mechanism, and a region of interest (ROI) extraction inspired by the patch paradigm of Vision Transformers. The SE modules enhance feature discriminability through dynamic channel-wise recalibration, while the gating mechanism refines feature extraction to improve texture perception. The ROI strategy focuses computational resources on the most informative image regions. Our GS-ResNet achieves a classification accuracy of 97.23% with an average response time of 0.5 ms, outperforming state-of-the-art methods in both accuracy and efficiency. This study addresses a critical gap in mixed road surface classification and provides an effective solution for robust road recognition in complex scenarios, offering support for the stable operation of mobile precision equipment across diverse road conditions.

In quasi-distributed monitoring systems using Fiber Bragg Grating sensors, unfavorable sensor placement and fiber routing may lead to insufficient coverage, unnecessary data redundancy, or infeasible cabling. To address these issues, this study proposes an integrated optimization framework that jointly considers sensor configurations and fiber routing. The framework adopts a two-stage strategy. First, an improved multi-objective genetic algorithm, leveraging dynamic chromosome encoding and customized genetic operators, is employed to optimize sensor layouts along the Pareto front, regarding modal orthogonality, strain energy coverage, and the amount of sensors. Second, a tailored traveling salesman problem is formulated and solved to obtain the shortest feasible fiber length under the minimum bending radius constraint. Furthermore, benefiting from large language models, a novel parameter tuning mechanism is introduced to reduce reliance on human decisions and enhance global search capability. More broadly, the proposed approach provides an efficient, scalable, and engineering-practical solution framework for the automated design of quasi-distributed sensing networks.

Gas–solid two-phase flow continuous filling faces challenges in achieving high precision, efficiency, and robustness under complex and dynamic working conditions. This article presents emotion-aware reinforcement learning (ERL)-Fill framework with staged pretraining, designed to address these challenges. Built upon the soft actor-critic (SAC) architecture, ERL-Fill integrates a Transformer-based emotion adapter and a gradient-guided reward mechanism to enhance adaptive control under diverse constraints and operational demands. To comprehensively evaluate the performance of ERL-Fill, four experimental groups are designed: emotion mechanism ablation, different algorithms comparison, multicondition testing, and staged pretraining validation. The staged training involves pretraining in virtual simulation and on-board communication-based control, followed by final training and validation on real-world systems using polypropylene, based on a 5000-1000-200-episode curriculum. Results demonstrate that the ERL-Fill effectively achieves intended control objectives by adaptively regulating filling processes across varying working conditions.

The widespread application of Unmanned Aerial Vehicles (UAVs) presents substantial challenges to public safety, thereby necessitating accurate visual tracking for effective counter-UAV operations. However, most existing transformer-based trackers rely on increasingly large and deep networks to improve accuracy in complex backgrounds, which constrains their practicality in real-world deployments. To address this limitation, we propose a lightweight anti‑UAV tracker based on knowledge distillation. Specifically, we design a fully transformer-based anti‑UAV teacher network (TransADT) and develop a lightweight Siamese student network (SiamGLT) constructed with on a gated linear transformer, enabling accurate real-time tracking. In addition, we introduce a heterogeneous model knowledge distillation framework (HKD) to facilitate knowledge transfer from the large transformer teacher to the convolutional student network. This framework comprises three types of knowledge transfer: foreground and background feature decoupled distillation, instance-wise relation distillation, and response knowledge distillation. Comprehensive experiments on Anti-UAV benchmarks and real-world scenarios demonstrate that the proposed tracker surpasses other state-of-the-art methods. Our SiamGLT model achieves an optimal balance between tracking accuracy and computational efficiency.