Aiming at the problem of substation equipment point cloud recognition, this study proposed a new recognition method. Firstly, a local coordinate system of equipment point cloud was established. The symmetry and distribution density of equipment point cloud were used to determine the x and y axes of the local coordinate system. The coordinate system was invariant to translate and rotate, and was robust to noise. Then, a new feature descriptor was defined based on the difference between the shape and view of substation equipment point cloud, which was used to describe and recognize the point cloud of equipment. Also, a template library containing 54 kinds of electrical equipment such as lightning arrester, circuit breaker, and disconnecting switch was established, which contained the information of type, number and feature descriptor of each template device. The feature descriptor of the equipment point cloud to be identified was calculated and was used for matching the most similar model in the template library to realize the recognition of equipment point cloud. Finally, the method proposed in this study and another two substation point cloud recognition algorithms were tested on 90 equipment point clouds to be identified. The results showed that our method could achieve 90% recognition accuracy, and the average time to identify a device was 3.2 s, which could balance the recognition accuracy and recognition efficiency. And our method slightly outperformed the other two methods when equipment point cloud with noise and occlusion. Moreover, when the density of the point cloud to be identified was not uniform, our method could still maintain the recognition accuracy of more than 70%.
Addressing the issues of insufficient consideration for non-market ecosystem services and unclear definition of compensation control thresholds in forest ecosystem compensation, in this paper a regional ecological compensation mechanism oriented toward securing forest ecosystem service functions was proposed. Based on emergy analysis and integrated with multidisciplinary methods, a quantitative evaluation framework for forest ecosystem service functions was established. The M-K and Pettitt methods were jointly applied to the interannual sequence of functional benefits to identify abrupt changes and determine the compensation control threshold. A coupled humannature emergy system was constructed to trace ecological and economic flows within and beyond the system, compensation standards (i.e., the unit emergy value of forest resources) was defined, and the value compensation relationship between humans and forest systems was quantified with the constraint of functional benefit thresholds. This study provided a new perspective for forest ecological compensation research. The proposed theoretical framework and methodology were applied to Luoyang City in Henan Province. Results showed that the total emergy benefits of forest ecosystem functions exhibited a steady upward trend from 2010 to 2021, with a cumulative increase of 0.37%. The year 2016 was identified as the mutation point, with a compensation control threshold of 5.11E+22 sej. The years 2010—2015, during which the functional benefits were below the threshold, required compensation. During this period, the total compensation value showed a fluctuating decline, decreasing from 37.71 million yuan in 2010 to 12.64 million yuan in 2015. The compensation value was primarily driven by the magnitude of functional benefit reduction below the threshold, which was closely linked to the abundance of key ecological elements within the forest system. Therefore, continuously enhancing forest protection and sustainable management was essential to realize the goal of harmonious coexistence between humans and nature.
To address the challenges of capturing subtle features and the scarcity of data samples in learning emotion recognition, a facial emotion recognition method based on a generative diffusion model and multimodal multiscale visual encoding was proposed. Firstly, a learning emotion dataset integrating multi-scale global and local detailed features was constructed, and the generative diffusion model was used to augment scarce emotional samples, thereby alleviating data constraints in few-shot learning scenarios. Secondly, a multimodal multiscale visual encoding mechanism was designed, which achieved high-precision modeling and effective fusion of micro-expressions and fine-grained emotional features by combining global features of facial images with local details from salient regions. Finally, the experiments were conducted on various models, including CNNs, Vision Transformers, and hybrid architectures. The results showed that the proposed method achieved an overall recognition accuracy of 68.10%, with an average improvement of 2.98% and a maximum improvement of 5.30% compared with existing baseline methods. The ablation experiments further verified the effectiveness and synergistic contribution of the generative diffusion model and the multimodal multiscale fusion module in enhancing the model’s capability to capture micro-expression details and improving overall recognition robustness.
Involute gear transmissions faced challenges such as uneven distribution of contact stress on tooth surfaces that leading to pitting wear, and insufficient bending resistance when transmitting high torque and power. To address these issues, in this study a novel herringbone planetary gear transmission system, featuring a point-contact meshing form consisting of concave tooth profiles (central gear)-convex tooth profiles (planetary gear)-concave tooth profiles (internal ring gear) was proposed. Based on spatial geometric relationships, the equations for the convex and concave tooth profiles of each component were derived. The mathematical description of the conjugate tooth surfaces was completed using the helical motion method, and the geometric equations for the formed tooth surfaces were established. MATLAB and UG software were utilized to perform operations such as surface stitching, extrusion, merging, and Boolean operations, constructing a concave parabolic central gear, a convex circular-arc planetary gear, and a concave parabolic internal ring gear. The assembly of three transmission models—the novel herringbone planetary gear transmission, the involute gear transmission, and the involute-circular-arc gear transmission—was completed. Adaptive meshing technology was employed to gradually refine the mesh, and a detailed sensitivity analysis determined the relationship between mesh density and computational accuracy. A convergence curve of mesh density versus stress was plotted, showing that the equivalent stress stabilized when the mesh density reached 2.0. To balance computational accuracy and efficiency, a mesh density of 2.1 was selected for simulation analysis. With operating conditions of 360 kW input power and a sun gear speed of 2 000 r·min-1, finite element analysis was conducted to compare the meshing contact characteristics of the three configurations. Results demonstrated that compared with the involute gear system, the novel herringbone planetary gear system exhibited a 24.5% reduction in static equivalent stress, more uniform tooth surface stress distribution, and a 4.63% improvement in bending resistance, significantly outperforming standard involute gear transmissions.
Network node embedding is mapping nodes in a network to a low-dimensional vector representation, so that vector space-based learning methods can be directly applied to handle downstream tasks such as link prediction. Most of the existing network node embedding models were for unsigned networks and could not be directly used to deal with signed networks (usually need to be converted into unsigned networks for processing, thus discarding a lot of valuable information embedded in the positive and negative signs on the edges).A node embedding model (SNEDA)based on graphical neural networks was proposed that could directly deal with symbolic networks. Based on structural balance theory and status theory, the paths between nodes were divided into 20 different motif structures according to the direction and the positive and negative information on the edges. A network propagation model was designed with two levels of attention mechanism, which could capture the contribution and influence of different neighboring nodes to the vector representation of node i by node-level attention mechanism when aggregating the direct neighboring information of node i, and captured the vector representation of different motif to node i by path-level attention when aggregating the second-order and higher-order neighboring information of node i. A two-level attention mechanism was introduced to integrate different contributions and influences at the node level and path level, which could it not only improve the time efficiency of the algorithm but also make the final vector representation of node i more beneficial to improve the prediction accuracy of the downstream link prediction task. Through experiments conducted on four real social network datasets, the SNEDA model improved the AUC and F1 metrics by about 3.1% and 1.1%, respectively, compared with the benchmark model, and the results showed that the network representation obtained by the model could improve the accuracy of link prediction.
To improve the effectiveness of the active collision avoidance strategy, a risk assessment method for collision time margin was proposed. The three-degree-of-freedom vehicle model and the Dugoff tire model were established, and the state parameters were calculated to obtain the normalized tire force. The pavement adhesion coefficient estimator was designed based on the traceless Kalman filtering algorithm, and the effectiveness of the adhesion coefficient estimator was verified through simulation. The road adhesion coefficient was added to the safety distance model to address the limitations of the traditional collision avoidance model, which only considered the position and the vehicle movement conditions. The active collision avoidance path was generated by using a fivetic polynomial, and the required safe steering distance was calculated. Based on the risk assessment method, a collision avoidance mode selection strategy was designed, enabling intelligent vehicles to select the appropriate collision avoidance mode according to the kinematic relationship with obstacles. The longitudinal control based on the vehicle inverse dynamic model and the lateral control using MPC were adopted to decouple the control of intelligent vehicles. The effectiveness of the collision avoidance strategy was verified through the joint simulation experiment of Carsim and Simulink and the real vehicle test.
In the past, the judgment of the timing of the opening and changing of the disc cutter mainly relied on sensor data and human experience, which led to the serious wear of the disc cutter, or affected the tunneling speed. In order to accurately judge the timing of disc cutter opening and tool replacement, in this study the excavation efficiency calculation method was summarized to characterize the use value of disc cutter, and machine learning method was used to predict it. Based on the Shenzhen CFL tunnel project, the influencing factors of the shield tunneling efficiency were analyzed in the previous literature, 15 characteristics were selected as the input parameters, and the hob tunneling efficiency was used as the output parameters, and a total of 37 849 data series were obtained as the total sample set after data processing. The machine learning method was used to train on datasets, and the algorithmic models used include Random Forest, Extra Tress, GBDT and XGBOOST. The results showed that the machine learning model could predict the tunneling efficiency of the hob cutter well, and the XGBOOST model had the best prediction effect, with a determination coefficient of 0.955, an average absolute error of 7.053, and a root mean square error of 13.249.
Aiming at the problem of high complexity of path planning and difficulty in generating high-quality paths in effective time for UAVs in complex threat environments, a multi-strategy fusion particle swarm-butterfly optimization improvement algorithm (IPSOBOA) was proposed. The initial population was optimized through tent chaotic mapping combined with inverse learning strategy to enhance the diversity of the population; nonlinear parameter adjustment and dynamic conversion probability mechanism were introduced to balance the global search and local exploitation; combined with the particle swarm algorithm, the velocity term was introduced in the local search phase, and the position update equation with dynamic change of velocity was proposed to improve the search efficiency. Based on four benchmark test functions and three different threat scenarios respectively, IPSOBOA was compared with the butterfly optimization algorithm and various other optimization algorithms. The experimental results showed that, compared with the butterfly optimization algorithm in three scenarios of the static environment, IPSOBOA optimized the optimal fitness value by 1.8%, 17%, and 44% respectively, and optimized the path length by 1.8%, 42.4%, and 61.3% respectively; in the dynamic environment, it combined global path tracking and real-time obstacle avoidance to generate smoother and safer paths.
Aiming at the problem that existing convolutional neural network (CNN)-based modulation recognition methods are highly dependent on single modal data (e.g., IQ sequences) and difficult to adequately extract multidimensional features of signals, in this study a multimodal parallel structural modulation recognition method was proposed based on bidirectional long short-term memory network (BiLSTM) and residual network (ResNet), termed the BiLSTM-ResNet (BLR network). Firstly, the temporal features of IQ data were extracted by BiLSTM in the upper branch, and the spatial features of constellation maps were extracted by ResNet-18 in the lower branch. Secondly, serial feature fusion was used in the decision fusion module to better exploit the complementary nature of the multimodal data. Lastly, the signal modulation styles were recognised with the help of the model’s feature extraction capability. In this study, experimental validation was carried out on the publicly available dataset RML2018.01a. The experimental results showed that the overall recognition accuracy of BLR network in the 6-30 dB SNR interval was stable at 96.48%, 2.61% and 3.91% higher than that of the single-modal ResNet and BiLSTM models, respectively, and 1.25% higher than that of the CNN-LSTM model with concatenated structure, which verified that the model proposed in this paper had the modulation recognition problem Effectiveness.
To understand the evolution of vegetation community characteristics and driving forces in different landscape types, the evolvement rule of the growth and diversity characteristics of vegetation communities in different landscape types of the Jili Yellow River Wetland downstream of the Xixia Yuan Dam in the middle reaches of the Yellow River were studied. Through on-site investigation, remote sensing image interpretation, and constructing aboveground biomass inversion model of herbaceous vegetation based on random forest, the growth characteristics and diversity characteristics of vegetation communities in different landscape types were analyzed, and the driving effects of environmental factors on the evolution of vegetation community characteristics in different landscape types were identified. The results showed that the vegetation coverage grade in the study area was mainly high vegetation coverage, followed by relatively high vegetation coverage. The average aboveground biomass of herbaceous vegetation was generally in the low and medium plant zones, and showed a trend of increase-decrease-increase from 2018 to 2023, reaching the maximum and minimum values in 2019 and 2022, respectively, with 331 g·m-2 and 244 g·m-2. The vegetation in the tidal flat wetland showed the characteristics of coexistence of hygrophytes, mesophytes and halophytes. Soil nutrient factors had a significant correlation with the intra-annual characteristics changes of the vegetation communities in tidal flat wetland. While the land use types changes had a significant impact on vegetation coverage, especially in the areas with medium and low vegetation coverage that were more significantly.
To address the issues of existing deep learning-based meter reading algorithms on edge devices, such as high resource consumption, the lack of robustness, error accumulation, and difficulty in end-to-end pointer extraction in traditional image processing methods in complex scenarios, a lightweight improved YOLOv8 model was proposed for meter detection. Meanwhile, the YOLOv8-pose keypoint model was employed to extract keypoints from the meter dial, and the reading was calculated using an angle-based method by fitting the pointer and scale lines. Firstly, a lightweight RGELAN module was designed to replace the C2f module, reducing the complexity of the backbone and neck networks. Then, the CASC-Head detection head, which considered multi-scale feature contributions, replaced the decoupled head, reducing detection parameters. Finally, the Shape-IoU optimized regression loss was introduced to improve detection accuracy. Experimental results showed that the improved YOLOv8-RSS model achieved 98.5% precision and 90.6% mAP@50:95, with only 0.3% and 0.4% losses compared with the original YOLOv8, while reducing parameters, computation, and model size by 48.3%, 44.4%, and 46%, respectively. In complex scenarios, it achieved an average relative error of 1.425%, average absolute error of 0.557%, 3.08 MB parameters, and 78 frame per second. Compared with existing methods, the proposed algorithm reduced space consumption and reading errors, and improved detection speed.
Aiming at the problem of inertia estimation of regional interconnected power systems with a high proportion of renewable energy connected to the grid, a new estimation algorithm named kernel-Attention-Transformer based on the combination of Transformer and kernel attention network was proposed to accurately and efficiently estimate the regional inertia of the system. Then, the self-attention mechanism of Transformer was used to extract the dynamic features of the system, and KAN’s kernel attention mechanism was combined to replace the traditional fully connected layer and softmax layer, which enhanced the adaptability and robustness of the model for complex nonlinear and strong random data. In addition, through the visualization of inertia distribution, real-time monitoring of inertia changes was realized, which provided intuitive decision-making basis for system operators. Verified in the improved Australian 14-machine 59-node system, compared with the traditional RNN, LSTM, GRU and Transformer, the proposed kernel attention-Transformer algorithm significantly improved the noise filtering ability and estimation accuracy under different simulation backgrounds. At the same time, the visualization results of inertia distribution clearly showed the spatio-temporal variation characteristics of the system inertia, which provided strong support for the safe and stable operation of the power system.
In learning with instance-dependent label noise (IDN), semi-supervised methods could mitigate noise interference and leverage feature information, but their effectiveness depended on accurate noise identification and was susceptible to the choice of recognition technique. To address this limitation, a robust feature-centroid mechanism was designed to weaken the influence of unreliable samples and a distribution-adaptive dynamic mixture model (DMM) was proposed based on feature similarity. Pairwise feature similarities was extracted, both Gaussian Mixture Models (GMM) and Beta Mixture Models (BMM) were used to fit these similarity distributions, and dynamically to fuse their outputs to achieve more accurate noise identification. A semi-supervised learning strategy was then integrated to complete the training process. On artificially corrupted CIFAR-10 and CIFAR-100 datasets, our method achieved state-of-the-art performance. On real-world noisy benchmarks Animal-10N and Clothing1M, it attained classification accuracies of 84.21% and 75.80%, respectively, outperforming representative existing approaches and demonstrating the effectiveness and applicability of our approach for IDN learning tasks.
To tackle the problems of heightened search blindness, elevated memory usage, and redundant nodes in the output path resulting from the access of a large number of irrelevant extension nodes in traditional JPS algorithms, an improved JPS algorithm was proposed. This algorithm based on the priority of the target-point search direction and a dynamic-weight evaluation function. Firstly, according to the positions of the target point and the mobile robot, the priority of the direction of target point in the algorithm’s path-finding process was enhanced. A distance-based dynamic-weight evaluation function was introduced to mitigate the resource waste and efficiency loss caused by searching for irrelevant nodes. Meanwhile, the global path output by the improved JPS algorithm underwent secondary planning. Redundant nodes in the original path were eliminated, making the global path smoother. Secondly, the DWA algorithm was introduced and improved as a local path-planning algorithm. The improved DWA algorithm adopted a dynamic priority strategy based on collision distance to automatically avoid mobile robots on cross-paths. Finally, the improved JPS algorithm and DWA algorithm were separately simulated and verified. The results indicated that, compared with traditional jump-point algorithms, the improved algorithm in this study reduced the average number of searched extended nodes by 60.0%, the average number of trajectory nodes by 43.6%, and the average number of path turning points by 23.9%. The improved DWA algorithm could effectively address the drawbacks and limitations of traditional DWA algorithms in dynamic environmental problems such as path conflicts, and improve the collaboration and adaptability of DWA algorithms in multi-robot path planning.
To address the issue of collaborative enhancement between global smoothness and local textures in traditional image enhancement techniques, in this study an MDFD image enhancement model based on multi-scale dynamic filtering decomposition was proposed. Initially, learnable low-pass and high-pass filters were utilized to extract the low-frequency and high-frequency image components, respectively. Subsequently, by combining these two frequency-domain image components, the cross low-frequency channel attention fusion module (LFCA) and cross high-frequency spatial attention fusion module (HFSA) were introduced to achieve collaborative enhancement of image global and local features. Finally, a multi-scale fusion strategy was introduced to comprehensively utilize high-frequency and low-frequency information at different scales for feature fusion. The advantage of multi-scale fusion lay in its ability to effectively integrate details and global features at different scales, significantly enhancing the image at multiple levels. Experimental results showed that the MDFD model performed excellently in the validation on the FiveK and PPR10K datasets, with peak signal-to-noise ratio (PSNR) reaching 25.90 and 27.35, structural similarity index (SSIM) being 0.964 and 0.945, and ΔEab being 7.38 and 6.50, respectively. These results indicated that the MDFD model could offer superior image enhancement performance in complex environments and color-rich scenes.
In the process of 3D point cloud data acquisition, due to the factors such as the limitations of the accuracy of 3D laser scanning equipment and the interference from external environmental, the collected point cloud data is often contaminated with noise. To effectively remove noise while accurately preserving the geometric features of the 3D point cloud, in this study a denoising method based on 3D point cloud curvature and normal information segmentation was proposed. Firstly, the curvature and normal information of the point cloud were estimated using Singular Value Decomposition (SVD) and distance-weighted method, respectively, to divide the point cloud into flat and non-flat regions. Subsequently, improved statistical filtering approach was applied to flat regions, utilizing dynamic neighborhood size adjustment and curvature-weighted distance to optimize outlier detection. For non-flat regions, an improved bilateral filtering method was combined to enhance the weight function of spatial distance and normal difference, effectively preserving local geometric features. Experimental results on Stanford Bunny point cloud demonstrated that the denoising rate of the proposed algorithm reached by 97.83%, outperforming traditional methods such as statistical filtering, bilateral filtering, and DBSCAN. Experiments on spur gear point clouds showed that the deviation analysis between the denoised point cloud and the standard model indicated that more than 90% of the points in the overall spur gear point cloud had a distance deviation of less than 0.2 mm from the standard model, and more than 90% of the points on the gear teeth had a distance deviation within 0.25 mm. This confirmed that the algorithm could effectively preserve the geometric details and edge features of the point cloud model while removing noise, demonstrating its effectiveness.
Due to the heterogeneity in entity representation, relationship definition, and semantic structure between different knowledge graphs, it is difficult to effectively improve alignment quality in the presence of structural differences and information loss in the graphs by relying solely on graph structure. Therefore, in this study a graph convolutional entity alignment method that integrates multiple information was proposed. Firstly, the improved PageRank algorithm was used to filter triplets and alleviate the impact of differences in knowledge graph structure. Next, we learn the embedding representations of entities and attributes were learnt through graph convolutional networks, and the relationships between entities were iteratively updated by using these representations. Finally, based on the PBAB method, text description information was integrated and weighted with graph structure information to enhance the effectiveness of entity alignment. The experimental results showed that the proposed method improved the Hits@1 metric by approximately 3% compared to the optimal baseline, with corresponding improvements observed in other evaluation metrics as well.
Battery overheating is one of the key issues to the performance and safety of special vehicles operating with complex conditions. To address this, inspired by the nutrient delivery capabilities of arterial and venous capillaries, in this study a novel vascular biomimetic flow channel structure was designed. Using computational fluid dynamics (CFD) numerical simulation methods, transient simulations of the battery module were conducted, and a preliminary validation of the simulation model was performed using an experimental platform. A comparative analysis of different structural schemes and an in-depth exploration of the relationship between the heat dissipation effect of the liquid-cooled plate, pressure loss, structural parameters, cooling media, discharge rates, and environmental temperatures were all conducted. And based on orthogonal experimental design, an optimized design of the liquidcooled plate structure was achieved.The results indicated that the novel channel structure significantly enhanced cooling performance compared to parallel channel structures, achieving a 40.7% reduction in pressure loss. The physical properties of the cooling medium directly affected its cooling performance and pressure loss, and increasing the inlet flow rate of the mass flow improved the cooling effect of the cold plate, though the improvement become limited after a flow rate of 30 L/min. With working environment temperatures ranging from 38℃ to 70℃, the maximum temperature of the battery module remained around 45 ℃, with a surface temperature difference of less than 2℃, all within a reasonable range. This study contributed to advancement of battery thermal management technology in special vehicles with various environmental conditions, providing data support for studies aimed at improving battery pack temperature uniformity and cooling rates while reducing energy consumption.
Existing deep learning methods could not fully consider the impact of holidays when predicting network traffic. To address this issue, based on an analysis of the influence of historical data on prediction performance, a spatio-temporal network traffic prediction method that was proposed to account for the impact of holidays. Firstly, a method for generating historical data that incorporates holidays as an external factor was introduced, enabling the semantic feature of holidays within the historical data to play a role in network traffic prediction. Secondly, spatiotemporal learning blocks were employed to capture the temporal correlations, complex spatial correlations, and spatiotemporal heterogeneity of network traffic, thereby obtaining the comprehensive spatio-temporal features embedded in the historical data. Finally, the results from multiple blocks were fused through skip connections to output the final prediction results. Experimental results on the Milan (Italy) and Taiwan (China) datasets demonstrated that, compared with the recent similar model AHSTGNN, the proposed model reduced the MAE by 0.9 and 24.35 on the two datasets, respectively, and decreased the RMSE by 1.81 and 58.25, respectively. This illustrated the effectiveness of the proposed approach.
In order to improve the multi-step prediction of wind speed, a hybrid prediction model based on data signal decomposition and grey wolf optimization algorithm was proposed to optimize extreme learning machine. Firstly, the original wind speed time series was decomposed into several intrinsic mode functions and a residual sequence using the complete ensemble empirical mode decomposition with adaptive noise, and the partial autocorrelation function model input. Then, the model was built and the prediction was made on the decomposition subsequence. An extreme learning machine neural network with multi-input-multi-output strategy was constructed, and grey wolf algorithm was used to solve the weight and bias of the optimal hidden layer. Finally, the subsequence was reconstructed and the final prediction result was obtained. Simulation experiments were conducted using multiple sets of measured data with a time resolution of 15 minutes. The root mean square errors of the proposed model in the three wind farms were 0.859, 0.925, and 0.927, respectively, which were lower than other comparative models, verifying the effectiveness of the model in predicting wind speed in the next four hours,i.e. 16 steps prediction.
