Abstract:

Abstract: To address the issues of blind searching, redundant nodes, and non‑smooth paths inherent in the traditional rapidly‑exploring random tree connect algorithm for unmanned vehicles, a series of improvements were proposed in goal sampling, node expansion and trajectory optimization. Firstly, a goal‑guided dynamic probability sampling strategy was introduced to filter the randomly selected points, thereby improving sampling efficiency and accelerating convergence. Next, an improved artificial potential field component based on the escape force was incorporated into the node expansion process to help the unmanned vehicle avoid getting trapped in local minima while enhancing its target‑searching capability and node expansion efficiency. Finally, a trajectory quality evaluation function was constructed to assess the safety, deviation, and smoothness of the trajectories generated by the unmanned vehicle at different time steps. The trajectory with the minimum cost value was then selected to guide the vehicle’s motion. The enhanced algorithm was simulated and compared with the traditional RRT‑Connect algorithm in different testing environments. The simulation results showed that, compared to the traditional algorithm, the proposed algorithm could reduce the average path length by 9.83% and the average planning time by 85.40% in simple obstacle environments. In narrow passage environments, the average path length and planning time could be reduced by 10.56% and 64.63%, respectively. In U‑shaped obstacle environments, the average path length and planning time could be reduced by 22.82% and 66.92%, respectively. Furthermore, the proposed algorithm significantly improved the path planning success rate in complex environments, making it more suitable for autonomous vehicle path planning.

Abstract: In the task of 3D hand pose estimation from a single image in color, challenges such as occlusion and high self‑similarity of hand parts might lead to large prediction errors and unnatural hand structures. To address these issues, a graph convolution‑based 3D hand pose estimation method was firstly proposed. Visual features and 2D keypoint positions were extracted from the input image using Keypoint R‑CNN. These features were then fed into an improved adaptive kernel graph convolution module (AK_GraFormer). Subsequently, a residual‑connected AKGNN graph kernel was introduced to adaptively process graph‑structured data, thereby enhancing the model’s feature learning and representation. Finally, a dynamic training strategy was employed, which was monitored by a proposed evaluation metric, to optimize estimation performance. Experimental results on the HO‑3D_v3 and FreiHand datasets demonstrated that the proposed method outperformed existing approaches in monocular 3D hand pose estimation. Specifically, the procrustes‑aligned mean per joint position error (PA‑MPJPE) was reduced by up to 12.50 percentage points, and the area under the curve (AUC) of the percentage of correct keypoints metric was improved by up to 3.44 percentage points compared to state‑of‑the‑art methods.

Abstract: To address the issues of harsh environmental conditions, high labor intensity, and safety risks in manual hook removal operations in thermal power plant tipper unloading systems, in this study a tipping machine hook‑removal robot with a hybrid active‑passive control configuration was proposed. Firstly, the mechanical structure of the robot was designed, with three active moving joints, one active rotating joint, one passive moving joint, and one passive rotating joint. The movement adaptability of the passive structure was analyzed using a graphical method, and the forward and inverse kinematic models of the entire system were established using the D‑H method, providing a theoretical foundation for motion control. In terms of the control system, a hardware architecture was designed, and a hook handle recognition algorithm was proposed, based on measurement data from a linear laser displacement sensor. This algorithm was used to obtain the position information P_g(x_g,z_g) between the end effector and the hook handle, enabling precise positioning and gripping of the hook handle. A pilot test was conducted to verify the effectiveness of the proposed solution. The results showed that out of 50 hook removal operations, the robot achieved a 100% success rate, with all joint torques within the rated limits. The average time for high‑position hook removal was 25 seconds, and 30 seconds for low‑position hooks, both meeting the production requirements. The maximum instantaneous torque at each joint was 62.1 N·m, only 43.7% of the system’s limit, demonstrating sufficient safety margin. The experimental results reflected the robot’s adaptive ability to handle trajectory uncertainties of hooks, allowing it to accommodate deviations in the path of different models or variations of the same model. Furthermore, the precision of the proposed recognition algorithm was validated.

Abstract: Given that traditional wind power cluster prediction methods failed to effectively account for the spatial meteorological correlations among stations and struggle to efficiently deduce the overall cluster power based on single‑station predictions, in this study a multi‑dimensional spatiotemporal information fusion framework was proposed for stations based on an attention‑based spatiotemporal embedding mechanism. This framework aimed to fully exploit the complex spatiotemporally coupled characteristics embedded within discrete numerical weather prediction (NWP) heterogeneous meteorological information. Firstly, a multi‑head self‑attention mechanism was employed to directly fuse spatial features, enhancing the model’s ability to capture spatial power correlations across multiple stations. Secondly, cluster location information was deconstructed using the maximal information coefficient to construct a non‑Euclidean graph data structure reflecting meteorological correlations. This was combined with a spatial‑temporal attention mechanism to achieve cross‑fusion of spatiotemporal features between stations and their neighborhoods, dynamically adjusting the influence weights among stations to capture spatiotemporal dynamic dependencies. Furthermore, an encoder‑decoder architecture was used to integrate spatial and spatiotemporal features into a unified semantic space to capture temporal continuity within sequences. Finally, the proposed model was verified based on the actual wind farm operation data of a certain region in Northwest China. Experimental results showed that the two error evaluation indexes of the proposed method RMSE and MAE were significantly lower than those of the other six prediction models, which effectively verified its advancement and adaptability.

Abstract: Aiming at the problem that the conventional PI excitation control was prone to bus voltage overshoot and oscillation in the fault recovery stage after the large‑capacity synchronous condenser was connected to the high voltage direct current (HVDC) receiving‑end system, a PI excitation control strategy of synchronous condenser based on grey prediction model GM(1,1) was proposed in this study. In this strategy the predicted value of voltage deviation was obtained in advance by grey modeling and trend prediction of the receiving bus voltage sequence, and was used as the input of the PI controller to realize the forward‑looking adjustment of the excitation voltage. Based on the CIGRE standard test system, a PSCAD/EMTDC and MATLAB/Simulink co‑simulation platform was constructed. Under various typical disturbance conditions, the dynamic response characteristics of grey prediction PI control and conventional PI control were compared and analyzed. The simulation results showed that the proposed control strategy could effectively compensate the large inertia lag of the excitation system, significantly suppress the bus voltage overshoot and shorten the system adjustment time. With the three‑phase short‑circuit fault, the minimum value of bus voltage was increased by about 4.5 kV, the minimum value of DC transmission power was increased by about 31 MW, and the system recovery time was significantly shortened. This method had the characteristics of low computational complexity and strong real‑time performance, which could provide an effective technical way for the optimization of excitation control of synchronous condenser in HVDC receiving‑end system.

Abstract: In order to provide more critical insights for the design and operational strategy of latent heat storage energy utilization systems based on molten salt, the heat release process of a phase change storage unit utilizing binary eutectic nitrates (\ce{NaNO_{3}} and \ce{KNO_{3}}, mass ratio of 23∶27) was investigated through numerical analysis in this study. The solidification and heat release processes in the phase change storage unit, with air serving as the heat transfer fluid, were simulated and calculated. Subsequently, the dynamic evolution characteristics of temperature and liquid phase fraction within the molten salt, the heat release rate of the phase change storage unit, and the outlet temperature of the air were analyzed and evaluated. The findings revealed that natural convection within the molten salt continued to play a significant role during the heat release process. As the phase transition progressed, the liquid phase fraction of the molten salt generally demonstrated a parabolic trend. Concurrently, both the heat release rate and the air outlet temperature exhibited linear decreases.

Abstract: To enhance the comprehensive utilization efficiency of cold energy at liquefied natural gas (LNG) receiving terminals, this study focuses on the Hainan Yangpu LNG receiving station and a cascade utilization system that integrates LNG cold energy with geothermal energy was proposed. The system aimed to achieve stepwise conversion and utilization of energy at different quality levels. The LNG gasification process was divided into four temperature stages, each matched with a Rankine cycle operating under corresponding conditions to recover waste heat for power generation. Meanwhile, geothermal steam was introduced as a high‑temperature heat source to drive the Rankine cycles, and after heat release, it was reinjected into the geothermal reservoir to ensure sustainable geothermal resource development. In addition to power generation, the system integrated a distillation‑based seawater desalination unit, utilizing waste heat from the Rankine cycles for seawater desalination, thereby enabling combined cooling, heating, and power generation. Process simulation of the system was conducted using HYSYS software, with an LNG inlet temperature of −162 ℃, a gasification pressure of 0.1 MPa, a geothermal steam temperature of 115 ℃, a pressure of 0.1 MPa, and a cooling water temperature of 25 ℃. Based on the simulation, exergy analysis was performed for each Rankine cycle subsystem and the overall system to evaluate energy utilization efficiency and exergy loss distribution. The results indicated that the system achieved a net power output of 25 372.1 kW, a freshwater production rate of 2 781.5 kg/h, and an overall exergy efficiency of 47%. The power generation contributions of the four temperature‑stage Rankine cycles accounted for 60.9 %, 23.2 %, 4.6 %, and 11.3 % of the total power output, respectively. Exergy loss analysis revealed that heat exchangers were the primary source of exergy loss in the system, with exergy losses significantly higher than those of pressure equipment, making them the most concentrated points of energy loss. By coupling geothermal energy with LNG cold energy through Rankine cycles for efficient power generation, the system not only improved overall energy utilization efficiency but also effectively mitigated the low‑temperature marine environmental pollution caused by the direct discharge of LNG cold energy.

Abstract: Neural Architecture Search (NAS) was an interdisciplinary study in the field of deep learning, which aimed to automate the design of neural network structures. NAS required repeated training and evaluation of a large number of candidate networks, which was computationally expensive. Differentiable Neural Network Architecture Search (DNAS) transformed the discrete architecture search problem into a differentiable continuous optimization problem, which reduced the computational cost. Firstly, a search algorithm framework for differentiable network architecture was constructed from three aspects: search space, search strategy and performance evaluation strategy. Secondly, the performance estimation bias, architecture overfitting and search stability problems of parameterization operation, as well as the improvement strategies of optimizing search space and improving efficiency were analyzed, compared and summarized. Then, the error rate, parameter quantity, search time and experimental hardware conditions of typical DNAS algorithms on image classification datasets were compared and analyzed. Finally, it pointed out the application potential of DNAS in complex scenarios such as edge device deployment, medical signal analysis, and cross‑modal matching, and proposed future research directions toward multi‑objective optimization, task‑driven search space design, and cross‑task transfer and reuse.

Abstract: To address the limitations of existing traffic flow prediction methods in fully utilizing node attributes to guide graph structure learning and capturing complex spatio‑temporal dependencies, in this study an Adaptive Spatio‑Temporal Graph Convolutional Network (AdpSTGCN) integrating adaptive graph structure learning with spatio‑temporal convolutional architecture was proposed. Firstly, an adaptive graph structure learning method based on node attributes was designed to dynamically capture spatial relationships in road networks from both global and local perspectives. Secondly, a dedicated spatio‑temporal convolutional architecture was developed to effectively model spatio‑temporal correlations in traffic flow patterns, further enhancing the model’s capability to handle complex spatio‑temporal relationships. A progressive training strategy was introduced to address challenges of excessive learnable parameters and data sparsity during model training. Finally, experimental evaluations on highway traffic datasets (METR‑La and PEMS‑Bay) demonstrated the model’s performance in 15, 30, and 60 minutes traffic flow prediction tasks. Experimental results showed that the AdpSTGCN model achieved the best performance among multiple baseline models in terms of three prediction error metrics: MAE, RMSE, and MAPE. These findings indicate the model’s superior modeling capabilities for both short‑term and long‑term traffic flow prediction tasks, providing a theoretical foundation for urban traffic management strategies.

Abstract: To address the issues of insufficient pretrained knowledge representation, ambiguous boundaries, and category confusion in Chinese medical few‑shot named entity recognition, a multi‑stage recognition approach guided by triple prompts was proposed in this study. Firstly, a triple‑prompt learning mechanism was constructed, in which a stepwise reasoning paradigm, comprising boundary localization, type discrimination, and contextual verification was adopted. By integrating the interpretability of discrete templates with the optimizability of continuous vectors, the capability of the model to parse entity structures was enhanced. Secondly, a bidirectional gating network was designed to dynamically regulate feature interactions between prompt information and textual representations, thereby strengthening semantic capture for complex terminology. Finally, contrastive learning was introduced to optimize the distribution of the feature space, improving intra‑class compactness and inter‑class separability. Experimental results indicated that F1 scores of 91.86%, 84.06%, and 88.93% were achieved on the IMCS‑V2‑NER, cMedQANER, and CCKS2019 datasets, respectively. These findings demonstrated that the proposed method consistently improved entity recognition performance and exhibited strong generalization capability under low‑resource settings.

Abstract: It was difficult to capture high‑order interactions among multiple nodes by graph neural networks based on binary‑edge structures, and static graph topologies were not adaptable to dynamic data distributions. To address these limitations, a few‑shot regression model based on adaptive multi‑head hypergraph convolutional networks (AM‑HGCN) was proposed. In this model, feature similarity and topological structure were integrated through a dynamic hypergraph construction method, multi‑scale hyperedges were generated using k\_hop neighbors and k-nearest neighbors (k\_NN) strategies to enable adaptive capture of feature interactions. Then, a multi‑head hypergraph convolutional network was designed to extract heterogeneous features via parallel attention heads and fuse multi‑granularity information through a dynamic gating mechanism to enhance expressive capability. Finally, a model‑agnostic meta‑learning framework was introduced to achieve rapid task adaptation through inner‑ and outer‑loop optimization. Experiments were conducted on the Boston Housing, Energy Efficiency, IMDB, and MinilmageNet datasets. For structured datasets, AM‑HGCN outperform mainstream baseline models significantly in evaluation metrics, The coefficient of determination was 0.775, which verifies its effectiveness in modeling complex relationships. Ablation studies further demonstrated that the collaborative effect of dynamic hypergraphs and multi‑head attention mechanisms was crucial for the enhancement of few‑shot regression performance, overall validating the effectiveness of the proposed method.

Abstract: To overcome the limitations of existing infrared object detection algorithms, which mainly arise from inadequate exploitation of temporal information and inter‑frame dependencies in dynamic target detection, thereby resulting in suboptimal detection accuracy, a real‑time infrared dynamic object detection framework based on YOLO‑IDOD, incorporating a dynamic attention module（DAM）and a channel attention convolution（CACONV）module, were proposed. The YOLOv12s architecture was employed as the baseline network, in which a dynamic attention mechanism was integrated at the input stage to extract short‑term optical flow features via an optical flow network, effectively suppressing background motion interference and enhancing the network’s sensitivity to target motion characteristics. Furthermore, a channel attention convolution module was embedded within the network architecture, where channel‑wise attention mechanisms was introduced at both the input and output stages to facilitate more discriminative feature representation and selection for the DAM‑enhanced features. The proposed modules was designed as plug‑and‑play components, enabling spatiotemporal feature aggregation and adaptive feature selection, thereby improving the generalization capability of the network for infrared dynamic target detection. Experimental evaluations demonstrated that the improved YOLO‑IDOD model achieved a precision of 79.9 %, a recall of 62.5 %, an mAP@50 of 77.7 %, and an mAP@95 of 57.3 % on a mixed dataset composed of a self‑constructed dataset（IRDA）and the public FLIR_ADAS_v2 dataset. Compared with the baseline YOLOv12s model, precision, mAP@50, and mAP@95 were improved by 5.2 percentage points, 4.6 percentage points, and 2.4 percentage points, respectively, while maintaining a comparable recall rate, thereby effectively enhancing detection accuracy and generalization performance for infrared dynamic targets.

Abstract: To investigate the influence mechanism of polymer density on the contact morphology and interfacial shear properties of the polymer‑silt interface, a combined method of interface nondestructive separation and three‑dimensional laser scanning was employed, and scanning tests on the three‑dimensional topographic features of the polymer‑silt interface were carried out, by which the interfacial roughness parameters under different densities were obtained. In conjunction with interfacial direct shear tests, the variations of interfacial roughness parameters, interfacial shear stress‑displacement relationships, interfacial shear mechanical strength parameters, and the correlation between roughness parameters and interfacial shear strength under the influence of different polymer densities were studied. Based on this, a stochastic three‑dimensional rough interface finite element model was established using COMSOL software, and the shear damage mechanism of the polymer‑silt interface under different interfacial roughness levels was further explored. The results showed that the finite element numerical model based on three‑dimensional interface reconstruction could be used to accurately predict the shear stress‑displacement relationships of the polymer‑silt interface at different densities, and the primary damage characteristic of the polymer‑silt interface under direct shear loading was manifested as stress concentration formed at interfacial asperities, with higher polymer density leading to more asperities and more pronounced stress concentration. Both the interfacial roughness and shear strength of the polymer‑silt interface were increased with increasing polymer density, and a linear relationship was satisfied between the interfacial roughness and shear strength of the polymer‑silt interface.

Abstract: The problem of environmental vibration and secondary structural noise caused by subway train operation along the line is becoming increasingly prominent, and so it is necessary to carry out the optimization research on vibration reduction performance of damping pad floating slab track bed. Firstly, One operating metro line of Zheng‑zhou was selected as an engineering example, and the vibration signals of ordinary track bed and damping pad floating slab track bed were in‑situ tested, the Z‑weighted vibration acceleration levels were obtained with 1/3 octave spectrum analysis, and the damping pad floating slab had better vibration reduction performance compared to ordinary integral track bed, which could reach 12.80 dB. Secondly, The simulation model of damping pad floating slab track bed‑tunnel‑soil was established with ABAQUS software. The accuracy of the model and rationality of assumption were verified compared with measured data, the single factor analysis method was adopted. The results showed that: by appropriately reducing the vertical stiffness of the fasteners, increasing the density of the floating slab and reducing the static modulus of the damping pad vibration reduction performance of the damping pad floating slab track bed can be improved. Finally, an optimized combination scheme for damping pad floating slab track bed was proposed for parameters such as fastener stiffness, static modulus of vibration damping pad and floating slab density. The optimized damping pad floating slab track bed could achieve better vibration reduction performance with comparative analysis through finite element simulation.

Abstract: It was found that buried segmented pipelines crossing active faults are prone to leak at joints with surface fault rupture displacements. However, no quantitative assessment study was conducted on joint failure risk that considered uncertainties such as surface rupture location and the magnitude of surface rupture displacement. In this study, an assessment framework for joint failure risk combining Probabilistic Fault Displacement Hazard Analysis and Monte Carlo simulation was proposed to quantify the surface rupture hazard in the pipeline crossing area and the annual failure probability of joints. A ductile iron pipeline with a nominal diameter of 1 400 mm and a segment length of 6 m crossing the Qujiang Fault in Yunnan was analyzed as a case study. The surface rupture displacements in the pipeline crossing area for recurrence periods of 475 years, 975 years, and 2 475 years were determined to be 0.70 m, 1.49 m, and 3.08 m, respectively. When the fault‑pipeline intersection angle was set to 70°, the annual failure probability of the joint was found to reach its minimum value of approximately 3.43×10⁻³ with the failure mode identified as rotational failure. It was revealed that the joint failure mode gradually transitioned from tensile failure to rotational failure as the fault‑pipeline intersection angle and pipe diameter increased. It was recommended that the most conservative surface displacement distribution model should be selected based on specific parameters such as the fault‑pipe intersection angle and pipe diameter to assess joint failure risk in practical engineering.

Abstract: Magnesium and its alloys have shown great application potential in medical implants such as vascular stents, biliary stents, bone tissue engineering stents, bone nails, bone plates, and porous dental implants due to their excellent biocompatibility and mechanical matching properties, and have attracted much attention in the field of biomedical materials. However, in the face of the complex and ever‑changing physiological environment of the human body, magnesium alloys have poor corrosion resistance and magnesium alloy devices are prone to degradation, leading to premature performance decline and insufficient reliability. Therefore, in the design of magnesium alloys, it is necessary not only to take into account different internal environments but also to consider the performance changes and reliability of the devices during long‑term service. In this study the high‑reliability design strategies of magnesium alloys were reviewed, including alloy composition design, process control, surface modification, computer simulation, etc. The current applications of magnesium alloys in orthopedics, cardiovascular surgery, general surgery, stomatology and other fields, as well as the design work of related materials, were summarized. It was proposed that the future development of biomedical magnesium alloys could focus on controllable degradation, material functionalization and intelligent design, etc., providing reference and inspiration for the clinical use of magnesium alloys.

Abstract: To address the limitation in mechanical property enhancement caused by the poor compatibility between the polypropylene (PP) matrix and polyethylene terephthalate (PET) microfibers in‑situ microfibrillar PP/PET composites, an "in‑situ fibrillation via melt blending‑high‑speed hot stretching" technique was employed in this study. Ternary PP/PET/PP‑g‑MAH microfibrillar composites were successfully fabricated by introducing maleic anhydride‑grafted polypropylene (PP‑g‑MAH) as a compatibilizer. The effects of compatibilizer content on the microstructure, crystallization behavior, rheological properties, and mechanical performance of the composites were systematically investigated. Results showed that the addition of PP‑g‑MAH significantly reduced the phase domain size of PET spherical particles before fibrillation, and the interfacial compatibility was improved. After in‑situ fibrillation, a high draw ratio of 14.2 was achieved by the PET microfibers, and well‑dispersed microfibrils with a minimum diameter of 202 nm were obtained. The synergistic effect of PET microfibrils and a small amount of compatibilizer significantly accelerated the crystallization rate of the PP matrix and the melt viscoelasticity was enhanced. Compared to neat PP, the tensile strength of the composites was improved by 11.5% and 24.5% through compatibilizer addition and in‑situ fibrillation, respectively, and by up to 30% through their combined effect. Additionally, the tensile fracture energy was increased by 217% compared to conventional PP/PET blends. These findings showed that the PP matrix could be effectively enhanced and toughened by the synergistic approach of using a compatibilizer in conjunction with in‑situ fibrillation.