A railway foreign object detection algorithm for complex weather conditions based on an improved RT-DETR, named FRP-DETR, was proposed. A feature complementary mapping module (FCM) and Pzconv units were introduced, which built complementary paths between shallow spatial details and deep semantic information to compensate for the limitations of single-scale feature representation, with the two components working collaboratively to enhance the perception capabilities of small and edge targets. A railway perception modulation fusion module (RMFM) was introduced, which was based on adaptive channel attention and spatial modulation mechanisms to enhance the model’s response to key semantic information in railway scenes. The original downsampling module was replaced with pinwheel-shaped convolution (PSConv), which enhanced edge texture extraction through multi-directional asymmetric padding and separable convolution. A railway foreign object detection dataset containing four weather conditions was constructed based on the SaMam style transfer method, by transferring sunny railway foreign object images to rainy, foggy, and snowy scenes. Experimental results showed that compared to the original RT-DETR-R18 model, this method achieved improvements of 1.65 percentage points and 3.4 percentage points in mAP@0.5 and mAP@0.5:0.95, respectively, with a 63.5% reduction in parameter count and an inference speed of 88 frames per second. The results verified that FRP-DETR achieved high accuracy, lightweight design, and real-time performance in railway foreign object detection under complex weather conditions.

To address the issue of inconsistent thermal stress between the upper and lower devices of the same bridge arm in converters under conditions of uneven aging or layout deviations, a novel thermal balancing method for upper and lower bridge arms based on dead-time shifting adjustment was proposed. By monitoring the on-state voltage of insulated gate bipolar transistor (IGBT) online, the real-time junction temperature of IGBTs was deduced based on the mapping relationship between the on-state voltage and junction temperature of IGBTs. Then, the temperature difference between the upper and lower devices of the same bridge arm was determined. When the thermal stress difference between the upper and lower bridge arms exceeded the threshold, the effective conduction time of the upper and lower bridge arms was adjusted by fine-tuning the position of the dead time, thereby actively regulating the power loss distribution of the upper and lower bridge arms. This achieved an increase in the loss of the bridge arm with lower temperature and a decrease in the loss of the bridge arm with higher temperature without affecting the total system loss, thus improving the thermal stress consistency of power devices in the converter and enhancing system reliability. Theoretical analysis and experimental results showed that the proposed method did not affect system efficiency and had no significant negative impact on output performance. Under typical operating conditions, the average temperature difference between the upper and lower devices was reduced by approximately 14% after adjustment, effectively improving the thermal stress consistency of the upper and lower bridge arms.

An optimized model KUNet based on Kolmogorov-Arnold network (KAN) and convolutional KAN (CKAN) was proposed to enhance the performance of the UNet model in order to address the limitation of the UNet model in complex feature extraction and generalization capability for medical image segmentation task. Traditional convolutional layer was replaced with CKAN, KAN feature enhancement module was introduced, and skip connection was optimized. Then the diversity and accuracy of feature extraction were improved while preserving structural information by incorporating an adaptive basis function learning mechanism. Comparative experiments were conducted against the UNet baseline model, nnUNet model and Swin-UNet model on four different multimodal datasets: LiTS, CORN, DRIVE and Lungs. Results showed that the average maximum absolute performance gap (MAPG) between the UNet baseline model and the KUNet model across the four datasets were 0.679 9 and 0.620 3 for Dice coefficient and IoU coefficient, respectively, and the KUNet model achieved average improvement metrics of 0.3213 and 0.2625 compared with the optimal or suboptimal model across the four datasets. The KUNet model was utilized to effectively extract more feature within short training cycle and improve the accuracy of image segmentation.

A spatiotemporal joint lane-change intention prediction model based on Transformer and graph attention neural network (GATv2) was proposed in order to accurately predict lane-changing intention of surrounding vehicles and improve the safety and reliability of autonomous vehicles. Symmetric exponential average filtering was applied to smooth the raw NGSIM data. Then velocity and acceleration were recalculated by using the difference method, and outliers were removed. The data of the target vehicle and its surrounding vehicles were unified by establishing an eight-neighborhood for each vehicle, providing a foundation for model training and validation. Transformer multi-head attention was used to mine temporal dependency, while GATv2 was employed to quantify the interaction relationship between vehicles at each time step. Then a Transformer-GATv2 lane-changing prediction model based on spatiotemporal feature and multi-vehicle interaction was constructed. The experimental results showed that the proposed model significantly outperformed other baseline models in accuracy, recall, F₁-score, and other metrics. A prediction accuracy of 98.04% was achieved, which demonstrated superior lane-changing intention prediction performance and strong early prediction capability. The model was deployed on a six-degree-of-freedom driving simulator integrated with the CARLA co-simulation platform for model-in-the-loop testing, verifying the usability of the model in real-world scenario.

The sensor's timely and accurately data transmission is a guarantee for the decision-making unit to obtain effective data for decision-making (e.g., estimation, inference, or control) in IoT. To reduce estimation distortion, the decision unit uses multiple packets concurrently for joint estimation by using the best linear unbiased estimator (BLUE). Age upon decisions (AuD) and mean-squared-error (MSE) were introduced as metrics to measure the timeliness and the distortion of the information at the decision moments of the system, respectively. Two decision-making strategies were proposed, and the information timeliness and the distortion performance of the proposed strategies were investigated. In the strategy of using a fixed number of packets for decision making, the monitoring center performed an estimation after per fixed number of packets were received. In the strategy of using fixed time intervals for decision making, the monitoring center made an estimation at fixed intervals. The relationship between the system timeliness and the distortion was balanced by scheduling the decision process of the system to minimize the weighted sum of average AuD and average distortion. Simulation results show that the proposed strategies can improve the system timeliness and reduce the distortion performance by scheduling the decision-making process of the system.

A detection model named dynamic kernel perception YOLO (DKP-YOLO) was constructed to address the problems of contextual information loss, multi-scale distress coexistence, and insufficient detail restoration in pavement distress detection from a UAV perspective. A lightweight crack context module was proposed to mitigate the variability of crack morphology and the loss of contextual information in complex scenes. This module enhanced feature discriminability through heterogeneous parallel depthwise separable convolutions and a dynamic feature fusion mechanism. In order to tackle the challenges of complex backgrounds and multi-scale distress, a pavement distress perception module was designed, which improved the multi-scale distress perception capability using parallel multi-scale convolutions and a dual-attention mechanism. A large-kernel perception feature fusion module was introduced to overcome the limitations of traditional convolutions in modeling long-range dependencies and responding to small targets. This module captured global contextual relationships by integrating extra-large receptive field convolution with channel-spatial attention. Finally, a lightweight dynamic upsampling operator was incorporated into the neck network to reduce detail loss and semantic ambiguity during upsampling. Experiments on the UAV-PDD2023 dataset showed that the proposed model significantly improved detection accuracy while effectively reducing model complexity. Compared with the baseline model, DKP-YOLO achieved a better balance between being lightweight and maintaining high detection performance, making it more suitable for UAV-based pavement distress detection.

A novel parameter optimization method for model predictive control was proposed to address the problems of insufficient tracking accuracy and poor real-time performance of model predictive controllers in trajectory tracking of autonomous vehicles. An improved non-dominated sorting whale optimization algorithm (NSWOA) based on the improved Sinusoidal mapping and a Lévy flight strategy was proposed to solve the problem of relatively concentrated and out-of-bound optimal solutions obtained by the NSWOA. The parameter optimization problem of model predictive controllers was formulated as a multi-objective optimization problem. The predictive horizon, control horizon, and sampling time were used as optimization variables. The sum of squared lateral trajectory errors and total computation time were used as optimization objectives. The improved NSWOA was employed to solve the multi-objective optimization problem and obtain the Pareto optimal solution set. The optimal controller parameter combination was determined by using the expert scoring method, the continuous ordered weighted averaging operator method, the game theory-based combined weighting method, and the technique for order preference by similarity to an ideal solution. The tracking accuracy of the proposed method was improved by an average of 56.27%, and the computation time was reduced by an average of 21.54%. This method provides a new idea that balances high tracking accuracy and high real-time performance for the tuning strategy of model predictive control parameters.

An improved non-dominated sorting genetic algorithm II (NSGA-II) was proposed in order to address the challenges of time-consuming manual airspace sectorization and the difficulty in comparing the quality of different sectorization schemes. A three-dimensional multi-objective optimization model for sectorization was established by using a grid-region-sector hierarchy in order to balance controllers’ workload within sectors and reduce workload differences between sectors. A fitness evaluation operator, a probability-adaptive combination crossover operator and a dynamic mutation operator were incorporated in the NSGA-II algorithm in order to enhance the number of feasible solutions, solution diversity and computational efficiency. A simulation was conducted for the automatic 3D sectorization of Xi'an high-altitude airspace. Results showed that the optimized scheme improved workload balance within sectors by 37% and reduced inter-sector workload by 24% compared with the current sectorization configuration. The proposed improved NSGA-II provided a broader range of options for decision-makers with varying preferences compared with traditional weighted multi-objective optimization algorithms.

An approximate Nash equilibrium policy optimization algorithm that simultaneously updated the policy of both players was proposed, in order to resolve the problem of low learning efficiency of the policy-based reinforcement learning method in the two-player zero-sum Markov game. The two-player zero-sum Markov game problem was described as a maximum-minimum optimization problem. The policy gradient theorem of the Markov game was given for the parameterized policy, and it provided a feasibility basis for algorithm implementation through the derivation of the approximate stochastic policy gradient. Different gradient update methods for the maximum-minimum problem were compared and analyzed, and it was found that the extragradient had better convergence performance than other methods. An approximate Nash equilibrium policy optimization algorithm based on the extragradient was proposed based on this finding, and the convergence proof of the algorithm was given. The tabular softmax parameterized policy and the neural network were used as parameterized policy on the Oshi-Zumo game, to verify the effectiveness of the algorithm in different game scale scenarios. The convergence and superiority of the algorithm compared to other methods were verified through comparative experiments.

A many-core parallel optimization scheme for large-point FFT was proposed according to the structural characteristics and programming specifications of the domestic Sunway 26010 processor, which was used in the Sunway Taihu Light supercomputer. The scheme was derived from the classic Cooley-Tukey FFT algorithm, and was accelerated in parallel by iteratively decomposing the one-dimensional large-point data into two-dimensional small-scale matrices. The "column-sharing, row-continuity" strategy was specially proposed in order to solve the problem of reading, writing, transposing and calculating of the "column FFT" of the matrix. The computing resources and transmission bandwidth of the many-core processor were fully utilized by reasonable data allocation, rearrangement and exchange combined with other optimization methods such as SIMD vectorization, twiddle factor optimization, double-buffering, register communication and stride transmission. The experimental results prove that the single core-group of 64 slave cores running parallel program can achieve a maximum speed-up of 65x and an average speed-up of more than 48x compared with the main core running the FFTW library.

The development of underwater manipulators was reviewed from the perspective of actuation methods. The key roles of dynamic modeling, motion control, and autonomous intelligence in the operations of underwater manipulators were discussed, and the trend of end-effectors evolving from rigid to flexible structures was analyzed. Problems of existing underwater manipulators in aspects such as structural design, dynamic modeling, and autonomous intelligent control were summarized. The aim of the dynamic modeling, motion control, and autonomous intelligence, which are the key technologies for realizing the operations of underwater manipulators, is to deal with the complexity and uncertainty of underwater operational environments. The intelligent underwater manipulators with the capabilities of autonomous operation and precise motion control have broad application prospects in marine engineering, deep-sea exploration, and ocean resource development.

Significant progress in multimodal large language models (MLLMs) has driven advances in visual question answering, visual understanding, and reasoning tasks, and their potential for deployment on resource-constrained edge devices is increasingly recognized. However, large model sizes and the substantial costs of deployment and inference remain major barriers to practical adoption. Optimizing MLLMs for edge devices has become a critical research direction in this field. A comprehensive survey of recent advances in optimizing MLLMs for edge deployment was presented, along with the associated challenges and development trends. The research evolution of MLLMs on edge devices was reviewed, with particular emphasis on model architecture optimization and inference scheduling strategies. In model architecture optimization, techniques including visual information compression, sparse attention, and mixture-of-experts models were specifically analyzed. System-level optimizations involving computation scheduling, hardware adaptation, compilation optimization, and cloud-edge collaboration were investigated to enhance inference efficiency and energy efficiency. Furthermore, the key challenges of these models in practical applications were discussed, and a variety of task scenarios ranging from assistive to collaborative and autonomous types were covered, categorized by the perspective of autonomy levels. Finally, current limitations were summarized and future research directions regarding standardized deployment, efficient computing and storage, and multi-modal fusion optimization were outlined.

An intelligent generation method of electromechanical equipment operation and maintenance schemes with enhanced Chain-of-Thought was proposed, aiming at the problems of low efficiency and poor traceability of operation and maintenance schemes based on manual experience caused by the complex structure of electromechanical equipment and the high coupling degree of faults. Utilizing the capabilities of multi-source knowledge fusion and knowledge reasoning of large language models, the preprocessing process of multi-source heterogeneous operation and maintenance domain knowledge of electromechanical equipment was designed, and the knowledge ontology model of the operation and maintenance domain of electromechanical equipment with enhanced Chain-of-Thought was established. Through the injection of fault knowledge with enhanced Chain-of-Thought and the fine-tuning of large models, the Chain-of-Thought enhanced domain model with causal chain reasoning ability was constructed. The fault traceability reasoning of “fault phenomenon - cause ranking - scheme generation” for electromechanical equipment has been realized. The graph retrieval-augmented generation technology was introduced to construct a components knowledge graph with community division. The multi-component maintenance knowledge was deeply integrated and reasoned, which improved the generation quality of operation and maintenance schemes and achieved an intelligent operation and maintenance closed loop from fault tracking to operation and maintenance scheme generation. Finally, the performance evaluation and application verification of the Chain-of-Thought enhanced domain model were carried out. The results show that the proposed method demonstrates excellent performance in tasks such as fault tracking and operation and maintenance scheme generation, significantly improving the accuracy of fault tracking and the rationality of operation and maintenance schemes.

A rebar inspection method based on improved mask region with convolutional neural network (Mask R-CNN) model and stereo vision technology was proposed in order to promote the transformation of reinforcement inspection to intelligence. The improved model Mask R-CNN with channel attention and spatial attention (Mask R-CNN+CA-SA) was formed by adding a bottom-up path with attention mechanism in Mask R-CNN. The diameter and spacing of rebar can be obtained by combining stereo vision technology for coordinate transformation, thereby achieving intelligent rebar inspection. The training was conducted on a self-built dataset containing 3450 rebar pictures. Results showed that the Mask R-CNN+CA-SA model increased the F1 score and mean average precision (mAP) by 2.54% and 2.47% compared with the basic network of Mask R-CNN, respectively. The rebar mesh verification test and complex background test showed that the absolute error and relative error of rebar diameter were basically controlled within 1.7 mm and 10%, and the absolute error and relative error of rebar spacing were controlled within 4 mm and 3.2% respectively. The proposed method is highly operable in practical applications. The intelligent rebar inspection technology can greatly improve work efficiency and reduce labor costs while ensuring sufficient inspection accuracy.

A heterogeneous smoke and flame visual detection method integrating visible and infrared images was proposed, in order to address the problems of low detection efficiency, poor accuracy, and consequent delays in fire disposal and misallocation of resources in the early stage of tunnel fires caused by lighting interference and structural environment. For infrared images, an OTSU threshold segmentation optimized by a grey wolf optimizer fused with differential evolution was applied, and the maximum entropy algorithm was combined to improve flame region segmentation accuracy. Considering the tunnel scene characteristics, YOLO11 was adapted, and the backbone network was optimized using a Swin-Transformer modified with a neural attention mechanism to enhance weak feature extraction. During feature fusion, a lightweight BiFPN was employed to adapt to multi-scale smoke and flame features, and an adaptive WIoU loss function was introduced to improve localization accuracy under fuzzy boundaries. At the decision layer, weighted fusion and conflict resolution were used to achieve dual-modal collaborative decision-making. Experimental results demonstrated that the proposed method achieved at least a 23.55% reduction in average localization error compared with single-modal detection. Compared with nine fusion algorithms such as evidence theory, the recall, F1, and mAP values of the proposed method reached 92.52%, 90.83%, and 93.47%, respectively, all exceeding those of the compared algorithms, while maintaining the smallest localization error fluctuation. The inference speed of the model increased by an average of 14.61%. The research results not only contribute to tunnel fire prevention and control but also effectively improve overall tunnel safety management and emergency response capability.

A lightweight traffic police gesture recognition method for autonomous driving was proposed to address the challenge of maintaining high recognition accuracy while achieving lightweight deployment, particularly given the subtle variations in gestures and complex application scenarios. An efficient adaptive weight downsampling module was designed on the basis of the YOLOv8n algorithm. This module replaced the standard convolutions in the backbone network to capture spatial differences, reducing both parameter count and computational complexity. Additionally, a C2f-G module was introduced to better leverage local and contextual features, improving the recognition accuracy in complex backgrounds. A triplet attention mechanism was incorporated into the neck network to comprehensively capture fine-grained feature details. Experimental validation on the Chinese traffic police gesture dataset showed that compared to the baseline model, the proposed method reduced the parameter count by 44.0%, increased the mean average precision by 3 percentage points, and achieved a detection speed of 294 frames/s. The proposed method effectively addresses the challenge of dynamic traffic police gesture recognition in complex environments, achieves fast and high-precision recognition, realizes a lightweight architecture, and significantly reduces the deployment complexity.

Direct numerical simulations were performed to investigate the interaction between a horseshoe vortex and a free surface at various Weber numbers. The incompressible Navier-Stokes equations were solved using the volume-of-fluid method coupled with the piecewise linear interface construction scheme. The continuum surface force model was incorporated to account for surface tension effects during interface deformation. The simulations covered a range of Weber numbers from 0.24 to 0.84. By quantitatively analyzing the evolution of total kinetic energy, dissipation rate, total helicity, and surface energy, the influence of Weber number on the vortex-interface interaction was elucidated. At low Weber numbers, surface tension dominated, maintaining the interface as a coherent ring-like structure and suppressing deformation and topological changes, while the vorticity field remained stable with negligible variations in helicity. Conversely, at high Weber numbers, inertial forces dominated over capillary effects, leading to significant interface deformation, breakup, and reconnection. This process resulted in vortex breakdown, accelerated kinetic energy decay, distinct peaks in energy dissipation, and a continuous decrease in helicity. These findings underscored the critical role of the Weber number in energy transfer and dissipation mechanisms, providing theoretical support for multiphase flow dynamics and guidelines for engineering applications such as the design of surface and trans-media vehicles.

A small target detection algorithm (DLSRF-Net) for multi-scale complex scenarios from UAV viewpoint was proposed by reconstructing the YOLOv11, to address the insufficient feature extraction and poor detection performance of existing algorithms in small target detection under UAV viewpoint due to small target sizes, complex backgrounds, and multi-scale information in the scenarios. The adaptive depthwise separative receptive field attention convolution module (DWRFAConv) was proposed to improve the model’s ability to extract the receptive field features of small targets and reduce the model load. The multi-branch lightweight multi-scale linear attention mechanism was designed to enhance the model’s attention to small targets. The RSCDI module was designed as the upsampling layer and fully connected layer of the model to solve the problem of feature information loss, suppress the useless information, and improve the model’s detection accuracy. The model sizes were classified into two categories based on parameter count and computational complexity, and experimental validation was carried out on the VisDrone2021 dataset. The results showed that the proposed algorithm achieved the optimal performance under both model size categories, and the generalization ability of the proposed algorithm was verified on the DOTA and the SSDD datasets.

Discrete element method (DEM) and Tavares breakage model were employed to investigate the rock fragmentation mechanism and the statistical characteristics of fragments under impact loading. Drop hammer impact crushing process of limestone specimens with different aspect ratios (1.0—2.0) and sizes (diameter 30—50 mm) was systematically simulated. Image J image processing technology was used to quantitatively analyze the fragment count, particle size distribution, and shape indices. The results showed that increasing the aspect ratio and the size significantly raised the crushing energy threshold, which led to a delay in the peak normal force, a reduction in its amplitude, and a slowdown of the overall fragmentation process. The equivalent particle size of fragments showed poor alignment with the sieve analysis curve, which indicated that the fragments commonly exhibited irregular geometric configurations such as flaky and rod-like shapes. Fragments from specimens with lower aspect ratios (1.0) and smaller sizes (diameter 30 mm) had lower roundness, more pronounced angularity, and greater surface roughness. A quantitative relationship model among geometric parameters–crushing response–fragment morphology can provide theoretical basis and regulatory pathways for the design of impact-resistant structures, the optimization of blasting parameters, and the control of particle morphology in rockfill grading.

A fault diagnosis method for electro-hydraulic servo valve based on multi-source signal fusion and hybrid attention mechanism was analyzed in order to improve the diagnostic accuracy and robustness for concealed leakage and blockage fault in electro-hydraulic servo valve under heavy-load, time-varying-speed and strong-noise condition, aiming to enhance the capability of the model for weak fault feature representation and recognition. A dual-channel multiscale Transformer fusion network for pressure and control signal, named DC-MTF-Net, was constructed oriented toward engineering application scenario. A convolutional attention module was employed to enhance multiscale temporal feature extraction, while cross-attention gated fusion was introduced to realize adaptive weighting of multi-source feature. A two-stage learning framework combining autoencoder pretraining and classification fine-tuning was adopted to improve feature representation and generalization capability in complex production environment. Multi-fault experiments conducted on an electro-hydraulic servo valve test platform under heavy-load and time-varying-speed condition showed that the proposed method improved diagnostic accuracy by at least 6%, compared with baseline models. Higher stability and robustness were demonstrated in comparative experiments involving single-channel diagnosis and various noise disturbance, verifying the effectiveness of the method in engineering application.