193 results about "Global topology" patented technology

The invention relates to the technical field of power system topology identification, in particular to a power distribution network frame topology identification method based on an improved graph neural network, and the method comprises the steps: collecting the real-time electric quantity data of nodes and edges of a power distribution network, and carrying out the modeling of a power distribution network graph structure; the method comprises the following steps: constructing a node, edge and hyperedge feature matrix by using real-time collected data, inputting an improved graph neural network topology identification model, dynamically weighting and adjusting an edge weight through a graph attention network, splicing and fusing local topological features and global topological features, and generating a prediction adjacency matrix; when the power distribution network is dynamically changed, a change area is positioned through adjacency matrix difference, and a sub-graph is extracted for incremental updating; and constructing a topological structure of the power distribution network based on the final adjacent matrix, and outputting a connection relationship between the physical positions of the nodes and the edges. Compared with the prior art, the real-time performance, generalization ability and applicability of the model are remarkably improved, and the problems of accuracy and efficiency of power distribution network topology identification in the dynamic environment are solved.

The invention relates to the technical field of industrial internet, in particular to an industrial product production management system based on the industrial internet, which comprises a multi-source sensing module, a dynamic topology module and an intelligent decision module, abnormal events are detected through an event pattern recognition engine, and structured event data are output; the dynamic topology module maps the production element entities into reconfigurable topology nodes and marks a real-time attribute set; activating a space-time constraint generator according to the structured event data, and creating a dynamic constraint edge; when the change of the dynamic constraint edge exceeds a preset threshold value, generating a global topology reconstruction signal; otherwise, generating a local constraint adjustment signal; the intelligent decision-making module responds to the signal and triggers multi-objective optimization calculation based on transfer learning or executes incremental rescheduling with a memory function in a process correlation subgraph, so that efficient production management is realized.

The invention provides a deep learning-based die process die surface automatic supplement generation method and system, and relates to the technical field of deep learning, and the method comprises the steps: extracting the contour line and feature point information of a die surface as a training sample; feature learning is carried out based on a multi-scale image convolutional network and a double-flow network, and optimized local geometric features and global topological features are extracted; generating a supplementary modular surface through the generative adversarial network; and carrying out stress field constraint verification and then fusing with an original model. According to the method, intelligent supplementary generation of the mold surface is realized, and the mold design efficiency and quality are improved.

The invention belongs to the related technical field of brain image processing, and provides an Alzheimer's disease classification method and system based on topology perception and a group hypergraph in order to solve the problem of inaccurate classification of the Alzheimer's disease in the prior art. Constructing a dynamic function connection network sequence through a sliding window strategy; a local topology perception encoder and a global topology perception encoder are respectively used for extracting local topology features and global topology features of each time window, deep interaction and fusion are carried out, and comprehensive feature representation of a tested level is generated; according to the method, each subject is used as a hypergraph node, hyperedges are constructed on the basis of comprehensive feature representation of a subject level and by combining feature similarity calculated by diffusion tensor imaging features and clinical embedded features of the subject, then a group hypergraph is constructed, a classification result is obtained by using a hypergraph neural network, and the early classification diagnosis accuracy of the Alzheimer's disease is effectively improved.

The invention relates to the technical field of unmanned aerial vehicle cooperative control, and discloses a multi-unmanned aerial vehicle cooperative control method based on data link networking. The method comprises the following steps of: constructing a cluster collaborative task data set comprising a flight path feature, a load state feature and a communication link feature; extracting dynamic cooperative nodes, including a cluster formation form switching node and a task target switching node, from the cluster formation form switching node and the task target switching node; generating a track cooperative trigger signal based on the dynamic cooperative node, and identifying a local topology adjustment node triggered by the formation form switching and a global topology reconstruction node triggered by the task target switching; generating a cooperative offset parameter containing a track offset and a communication delay offset according to the node classification; extracting cluster collaboration protocol parameters based on the collaboration offset parameters; and controlling the unmanned aerial vehicle cluster to execute a cooperative flight task according to the protocol parameters. The method can improve the cooperative efficiency and task adaptability of the unmanned aerial vehicle group in a dynamic environment, and is suitable for various multi-vehicle cooperative operation scenes.

The invention discloses an unmanned aerial vehicle automatic homeward voyage method based on a global topological map and a graph search algorithm, which comprises a global passable area topological connected graph module, a homeward voyage path search module, a homeward voyage path optimization module and a homeward voyage path error processing module. According to the method, global path searching is carried out in combination with a Dijkstra algorithm, path smoothness, flight time and obstacle avoidance are optimized by using a B spline curve, the precision and speed of path planning are effectively improved, and meanwhile, the system can cope with changes in the environment by updating a topological connected graph and optimizing the path in real time, so that the path planning efficiency is improved. And secondly, the path safety and the flight stability are both considered, and an obstacle distance penalty term is introduced, so that the flight path can avoid the obstacle, and the collision risk is reduced.

The invention discloses a three-dimensional physical field real-time prediction method and system based on geometric deep learning and physical constraint, and relates to the technical field of computer-aided engineering and artificial intelligence. The method comprises the following steps: directly extracting native boundary representation data (B-Rep) of a three-dimensional model from a computer aided design system; constructing a heterogeneous dual graph taking a parameterized curved surface as a graph node, skipping finite element grid division, and aggregating local and global topological features by using a graph neural network; in combination with a physical information driving mechanism, a partial differential equation (PDE) residual error is introduced as a loss function for constraint training, and generalization prediction of a novel geometric structure is realized; and finally, the physical field state quantity is predicted through direct regression and is rendered in real time. An incremental reasoning mechanism based on a local topology subgraph is adopted, millisecond-level physical field real-time feedback under design modification is achieved, and the method is suitable for scheme rapid screening and trend prediction in the initial stage of design.

The invention discloses a multi-parameter optimization method and system for a curved surface acoustic wave device. The method comprises the following steps: constructing theoretical models of the surface acoustic wave device under different strain conditions; solving the coupling wave equation to obtain the propagation velocity of the surface acoustic wave in the piezoelectric material under different strain conditions; acquiring at least one structural parameter of the curved surface acoustic wave device associated with the propagation speed; and performing global optimization on the at least one structure parameter according to a preset global topological optimization network to obtain a target structure parameter corresponding to the at least one structure parameter. A global topological optimization network is adopted, gradient back propagation is utilized, structural parameters of a device are intelligently designed, a quality factor value of the device is generated through finite element simulation calculation, and finally the quality factor distribution condition of the generated device is output. The problem that accurate real-time measurement of multiple parameters of the bent surface of the bearing cannot be achieved in the prior art can be solved.

The invention discloses an image processing method and system based on an artificial intelligence large model, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining an original image, segmenting the original image into overlapped blocks, and generating an image block feature set through a fusion strategy; constructing an undirected graph structure based on the image block feature set, calculating a node similarity weight and generating a global topology descriptor; generating an importance weight according to the dimension importance of the global topology descriptor, generating a noise vector in combination with noise intensity and structural constraint, and generating an enhancement vector and an offset matrix through a UNet encoder-decoder; and performing feature distillation on the enhanced vector by using a student network and a teacher network, and generating a final distillation feature through cosine similarity matching and an attention mechanism. According to the method, a dynamic noise intensity calibration mechanism is introduced in a knowledge distillation link, and accurate constraint of a noise direction is realized through variance parameter linkage control, so that generalization errors are reduced.

The invention discloses a manifold perception-based large model navigation method guided by a historical topological graph, and relates to a computer vision technology. The method aims at solving the challenges that in the navigation process, long-distance reasoning experience is insufficient, instruction fragments and dynamic visual observation are difficult to align, and large model reasoning is prone to illusion interference. Firstly, a large model based on an encoder-decoder structure is used for supplementing historical information coding for a visual observation sequence, and therefore global topological information guidance is provided for long-distance reasoning. And secondly, in order to effectively solve the problem that large model reasoning is subjected to illusion interference, significant space-time differences in a visual observation sequence are mined by using a multi-curvature manifold, so that the large model can accurately describe the current environment and make a decision according to a visual reference object in thinking. Besides, in order to strengthen the perception capability of the large model to the space structure and establish a graph self-attention mechanism, the node distance embedded in the constructed historical topological graph is combined with the visual similarity so as to model the space relationship between the nodes.

The invention discloses a point cloud classification / segmentation method based on local geometric contour and global topological structure maintenance, which comprises the following steps of: firstly, constructing a neighborhood relation graph according to a spatial distance through a neighborhood relation construction module, and accurately capturing complex geometric and semantic information in point cloud data through spatial distance calculation and a k-nearest neighbor algorithm; meanwhile, a neighborhood relation graph self-diffusion mechanism further enhances the relevance between local and global features of the point cloud, so that the model can better capture the geometric contour (shape) and global topological structure features of the point cloud, the geometric distribution characteristics in the point cloud data are fully considered, and the accuracy and robustness of feature extraction are improved. The sensitivity of the model to the local geometric details of the point cloud is enhanced, the efficiency of local feature aggregation is effectively enhanced, the influence of redundant information is reduced, feature redundancy caused by a method which only depends on a nearest neighbor method to extract features is avoided, the precision of point cloud classification and segmentation tasks is improved, and the accuracy of point cloud classification and segmentation is improved. And meanwhile, sharing of global information is enhanced, and the accuracy and robustness of point cloud classification and segmentation tasks are remarkably improved.

The invention discloses a parameter determination method and device of an analog circuit, a medium and a product, and relates to the technical field of integrated circuit design. Local performance feedback and global topology perception of performance index parameters are used as input parameters, and pertinence and generalization ability of parameter adjustment are remarkably improved. Through a strategy function, a parameter candidate set of a target analog circuit is obtained, through multi-group sampling processing, variance of strategy gradient estimation is reduced, and training stability and convergence speed are improved. Depending on a value network is not needed, and training resource overhead is reduced. And applying the parameter candidate set output by the strategy function to the simulation processing scene to realize the accuracy of simulation processing evaluation. The target performance indexes are extracted from the simulation result, optimization of the performance indexes is achieved, the multiple target performance indexes are processed according to the multi-target reward function, balance selection and rejection are achieved in the multiple performance indexes, the parameter combination corresponding to the highest advantage value is obtained, and the reliability of the analog circuit parameter determination process is improved.

The invention relates to an optical network topology dynamic adjustment method and system, computer equipment, a medium and a product. The method comprises the following steps: acquiring a computing resource association degree parameter according to equipment resource allocation information, acquiring a data transmission delay parameter according to global topology information of a network topology structure, and transmitting the data transmission delay parameter to a computing node under the condition that communication activity data generated by computing equipment of each computing node in a target iteration process passes verification, acquiring a communication traffic matrix according to the communication activity data, acquiring actual communication traffic according to the communication traffic matrix and the global topological information, acquiring a network load balancing parameter according to the actual communication traffic, and acquiring an evaluation function value based on a computing resource association degree parameter, a data transmission delay parameter and the network load balancing parameter; and according to the evaluation function values corresponding to all the network topology structures, determining a target network topology structure, and adjusting the optical connection path to realize optical network topology adjustment. By adopting the method, the network performance can be improved.

The invention discloses an intelligent data quality repairing method based on dynamic rule evolution, which belongs to the technical field of data quality management, and comprises the following steps: constructing a knowledge graph based on physical storage structure information and service logic of structured data; performing deep learning on the knowledge graph by using a graph neural network, and dynamically generating a global topology view based on a learning result; in combination with a historical damage mode and a global topology view, an optimal structured data scanning path is generated by utilizing reinforcement learning, and association anomalies of abnormal partitions in an optimal path scanning result are identified based on a graph neural network; the method comprises the following steps: constructing a multi-modal association sub-graph based on association anomaly, repairing structural defects in the multi-modal association sub-graph through a correct physical structure reversely deduced by a graph neural network, and carrying out credibility scoring on a repairing result to form a structured data management closed loop. The method can adapt to continuously evolved data modes and novel anomalies, continuously improves the robustness and autonomy of the system to deal with complex data problems, and reduces the long-term operation and maintenance cost.

The invention discloses a fissure surface reconstruction method and device based on a dynamic graph neural network. The reconstruction method comprises the following steps: acquiring core parameters of a coal rock fracture source; the method comprises the following steps: preprocessing core parameters of a fracture source by adopting a standardized processing method, constructing a fracture space-time topology analysis model based on a DGCNN graph neural network, and taking fractures as graph nodes; an edge convolution mechanism is utilized, an adjacent graph is constructed through the space-time relation between nodes, neighborhood features are aggregated to update node representation, and a local structure and global topology are captured; crack feature similarity is calculated, and clustering is carried out according to a density threshold value by using a DBSCAN algorithm; and in combination with a crack magnitude optimization clustering result, identifying a main crack surface crack with high expansion potential. Reconstructing a macroscopic crack surface by adopting a B-spline surface fitting algorithm; a control point is selected through a principal component analysis method to determine a crack main direction, a B-spline curve is constructed based on crack point distribution to generate a boundary contour, and a crack surface is accurately fitted by using a B-spline algorithm.

The invention relates to the technical field of 3D scanning, and discloses a real-time alignment highlight curved surface metallographic 3D scanning method and system, and the method comprises the steps: extracting normal vector features from a multi-view point cloud, calculating an approximate geodesic distance, and constructing a normal vector geodesic distance correlation graph; constructing a constraint propagation intensity matrix through a geodesic distance inverse ratio and a normal vector included angle exponential attenuation function, and enabling the constraint propagation intensity matrix to act on a multi-view point cloud registration residual error to generate a closed-loop error energy field; constructing a graph cut network node at the gradient extreme value of the energy field, and executing a minimum cut algorithm under geodesic constraint by taking the integral value of the energy field as the side capacity to complete global point cloud fine registration; according to the method, the local geometry and the global topology are combined through the normal vector geodesic association diagram, the problem that the registration accumulative error breaks out at the closed loop position due to the fact that the highlight curved surface metallographic surface lacks texture features is solved, and real-time accurate alignment of the highlight curved surface metallographic workpiece multi-view point cloud on an automatic detection production line is achieved.

The invention discloses an underground mine pipeline type point cloud completion method based on axial slicing, which relates to the technical field of point cloud data processing, is used for performing data completion on pipeline type point clouds, and comprises the steps of axially slicing preprocessed pipeline point cloud data, performing dimension reduction on three-dimensional point clouds to two-dimensional plane processing, and calculating related geometric parameters of pipelines. The method comprises the following steps: extracting a pipeline center line based on improved DBSCAN clustering and least square circle fitting, constructing a skeleton model under geometric topology constraints, combining local geometric features and global topology constraints, realizing geometric reconstruction of a missing region through rotating surface generation and a local interpolation algorithm, and realizing robust repair of the missing region. According to the method, a convex hull and an RANSAC method are combined to extract a pipeline center line and geometric parameters, pipeline point cloud data complementation is realized through a continuous rotation correction method, and the pipeline point cloud complementation quality is remarkably improved.

The invention provides a power distribution network topological structure automatic identification and optimization reconstruction method and system, and relates to the technical field of power systems, and the method comprises the steps: obtaining power distribution network information, and generating an initial topological graph; setting heterogeneous sensing nodes to form a sensing domain, and identifying a topological association relationship by an edge computing unit; generating a global topological feature matrix; dynamic response characteristics and load characteristics are extracted, and association strength is calculated; and establishing a Bayesian network analysis causal relationship, calculating an intervention effect value and generating a candidate reconstruction scheme. Accurate identification and intelligent reconstruction of the topological structure of the power distribution network are realized, and the reliability and the operation efficiency of the system are improved.

The invention discloses a novel power system-oriented power distribution network topology optimization configuration method and system. The method comprises the steps of obtaining current power grid topology data of a target power distribution network and supply and demand state data of a period of time in the future; the power grid topological data comprises a global topological graph and a current topological structure; the supply and demand state data comprises power supply data and load data; performing iterative optimization on the topological structure of the target power distribution network by adopting an optimization algorithm based on a particle swarm to obtain a new topological structure; wherein in the iterative optimization process, a fitness function is constructed by taking minimization of power grid network loss and node voltage deviation as targets, a population is initialized according to a current topological structure, the fitness of particles is calculated according to the supply and demand state and the fitness function, and particle position updating is performed according to the global topological graph under radial constraints.

The invention relates to the technical field of display, and discloses an LED screen correction coefficient adjusting method and system, and the method comprises the steps: collecting the original brightness and chromaticity data of each LED unit; constructing a global topological connection relation graph; calculating an initial correction coefficient; the corrected residual error of the boundary pixels of the adjacent units is analyzed; establishing a consistency constraint equation set aiming at minimizing the difference between the boundary brightness and the chromaticity; and performing joint solution to obtain a global optimization correction matrix and writing the matrix into a driving module. The system comprises a data acquisition module, a topology construction module, an initial correction module, a residual analysis module, a constraint construction module, a global optimization module and a coefficient writing module, and integrates a temperature drift compensation and visual verification mechanism. Through cross-unit collaborative optimization, seamless fusion of splicing areas is realized, and brightness uniformity, chromaticity consistency and long-term operation stability are improved.

The invention relates to the technical field of artificial intelligence, in particular to a knowledge distillation acceleration method and system for a multi-modal large model, and the method comprises the steps: obtaining multi-modal input data; feature mapping of topology maintenance is executed, topological representation of a multi-modal feature space is constructed, and a feature mapping module is designed based on a residual module; a topological hierarchical attention mechanism is applied, and multi-head topological attention fusion is realized through sensing of a local topological attention unit and a global topological structure; topology-driven multi-task knowledge distillation is executed, five cross-modal distillation tasks are designed, and distillation loss is calculated based on topological similarity measurement; a topology-aware model acceleration technology is adopted, adaptive quantization is performed based on feature topology importance, neural network structure search for topology maintenance is executed, a knowledge distillation process is guided through a topology principle, a topological structure relationship of a feature space is maintained, and multi-modal knowledge is effectively transmitted. Experiments show that compared with an original large model, the method has the advantage that the size of the model is reduced by 80-90%.

The invention discloses a wiring diagram recognition method and device based on deep learning and electronic equipment, and the method comprises the steps: carrying out the correction and mask processing of an initial wiring diagram, obtaining a first wiring diagram, and fusing mask parameters and mask skeleton parameters; extracting circuit elements in the first wiring diagram based on fusion mask parameters and mask skeleton parameters, and classifying the circuit elements to obtain a first element set of the first wiring diagram; performing port connection line pairing processing and topological optimization processing on the first element set, and determining a second element set and a corresponding second wiring diagram; according to the second element set, performing multi-dimensional verification and global topological structuring processing on the second wiring diagram to obtain a topological graph data model; and determining a JSON file, a visual topological graph, a VF2 check report and a Neo4j graph database of the wiring diagram according to the structured data in the topological graph data model. According to the invention, a more accurate wiring diagram can be identified.

The invention discloses a fault diagnosis method for a rotating part in a phase modifier based on a self-adaptive multi-channel graph neural network, and belongs to the technical field of industrial equipment fault diagnosis. According to the method, the problems of low fault diagnosis precision, long training time and high training cost when the real distribution of the actual signals and the distribution of the training samples are deviated in the traditional method are solved. According to the method, vibration signals of a rotating part of a phase modifier are collected and preprocessed, then preprocessed vibration signal fragments are input into a constructed feature space and topological space two-channel graph convolutional network structure, and the two-channel graph convolutional network structure models local features and global topological relations of the signals respectively; local and global information of signals can be comprehensively extracted, an edge-enhanced attention mechanism and an adaptive module fusion strategy are introduced to carry out weighted distribution on feature contributions of different modules, and then features after weighted fusion are input into a classifier to realize fault diagnosis. The method can be applied to phase modifier fault diagnosis.

The invention discloses an action sports scoring method and device based on big data analysis, and relates to the technical field of computer vision, and the method comprises the following steps: S1, constructing a human body posture tensor manifold; s2, generating aligned action track characteristics; s3, generating symmetrical positive definite manifold features; s4, generating deep manifold distribution parameters; s5, quantifying global distribution deviation characteristics; s6, calculating a residual vector based on the deep manifold distribution parameters and the standard action distribution parameters, inputting the residual vector into the improved AGCN model for processing, constructing an adaptive topology based on the residual vector, and performing aggregation analysis to obtain a joint-level physical angle error after multi-scale space-time convolution and manifold enhancement; and S7, outputting a comprehensive score vector. According to the method, the limitation that a traditional method only depends on local geometric features and ignores global topological structure constraints and statistical distribution priori is overcome, and an efficient solution is provided for intelligent scoring of sports actions.

The invention discloses a transformer substation anti-error entity alignment method and system based on multi-modal structure fusion, and belongs to the technical field of intelligent power grids, and the method comprises the steps: carrying out the cross-system multi-modal data collection of a power grid, carrying out the preprocessing and feature extraction, and generating multi-modal features; multi-modal features are fused to generate joint embedding, a multi-modal similarity matrix is calculated, an initial coupling matrix is solved in combination with a Sinkhorn algorithm, and an initial anchor point set is screened; iteratively updating the initial anchor point set in combination with multi-modal consistency verification to obtain a verification anchor point set; and taking the iterated verification anchor point set as priori knowledge, injecting the priori knowledge into a Gromov-Wasserstein structure optimization model, and carrying out block calculation on the structure similarity of an adjacent matrix to realize global topology alignment. According to the method, cross-system entity alignment is realized through three-stage progressive optimization, the alignment result is fused with equipment entities, attributes, relationships and dynamic states, and the method plays an important role in scenes such as real-time intelligent detection of substation anti-misoperation and the like in combination with a logical reasoning engine.

The invention provides a power distribution network topology construction splicing method based on recursive grouping and a graph attention network. The method comprises the following steps: acquiring electrical attribute information of node equipment in a power grid by using a monitoring sensor; recursive grouping is carried out, nodes and branches in the feeder line of the power distribution network are identified by using the constructed impedance matrix and distance matrix, an initial topological graph is constructed, and a node feature matrix and an initial adjacent matrix are obtained; a GATv2 graph attention network is introduced, dynamic attention calculation is performed by using the node feature matrix and the initial adjacent matrix, and a final topological graph of the feeder is generated; a landmark matching method and Pearson's correlation coefficient calculation are adopted, and boundary points of each independent topology are identified and matched; performing accurate alignment on the topological graph generated by each feeder line by using a graph embedding synchronization technology, and finally performing global topology processing; and carrying out topology verification on the constructed global topological graph. Construction and efficient splicing of line topologies in different areas of the power distribution network are achieved, and efficient operation and intelligent development of a modern power distribution network are supported.

The application relates to a dynamic particle swarm optimization-based X structure Steiner minimum tree construction method. The method mainly comprises the following three effective strategies: (1) a dynamic subgroup and information exchange strategy enables subgroups to exchange information with other subgroups while maintaining independence, and increases subgroup diversity; (2) an improved particle learning strategy can combine the advantages of local topological structure in particle diversity and optimization precision with the advantages of global topological structure in algorithm convergence speed; and (3) a transition from multi-group local learning to single-group global learning strategy enables particles to obtain a better line length optimization rate. The application takes optimization of line length as the target, and finally optimizes the important target of line length.

The application provides a distributed remote sensing interpretation vector post-processing method based on a grid DCEL topology. The method comprises: performing grid slicing on global grid data to obtain grid data of a plurality of grid units; constructing a bidirectional edge connection table topology based on the grid data of each grid unit respectively; performing elimination processing, thinning processing and smoothing processing on the topology of each grid unit respectively; merging the topology of each grid unit to obtain a global topology; and exporting the global topology as vector data. In this way, a high-performance grid topology engine based on a distributed architecture can be provided, a DCEL topology structure can be constructed from grid data, and distributed computing can be supported through a topology level merging algorithm; the grid is used to construct the topology, thereby reducing the steps of grid vectorization, and the grid-based topology construction method can naturally support distributed computing.

The invention relates to the technical field of data processing, and provides a data task scheduling method and system and a computer readable storage medium. The method comprises the following steps: constructing an initial task graph model according to a task definition; dynamically receiving a topology change instruction to perform online adjustment on the current task graph model; performing data consistency verification on the adjusted task graph model; after verification is passed, effective switching of the adjusted new task graph model is achieved, and the task being executed in the system is not interrupted in the effective switching process; task scheduling and resource allocation are carried out based on the new task graph model and the heterogeneous resource state after the new task graph model takes effect, and a target execution task adaptive to the heterogeneous environment is determined; and executing the target execution task in the heterogeneous environment. A dynamic topology maintenance mechanism is introduced, a topology change instruction is received and responded in real time, and finally seamless effective switching of a new model is realized, so that the problem of service interruption caused by static topology stiffness of a traditional system is solved, and the flexible adjustment capability of coping with global topology is improved.

A distributed inference communication method of a super-large-scale graph neural network comprises the following steps: dividing a global topology to be processed into a plurality of local sub-domains matched with the number of GPUs, and determining a local master node and a boundary master node allocated to each GPU; calculating coloring priorities of nodes, and performing parallel graph coloring based on the coloring priorities to construct a global independent set sequence; mapping the global independent set sequence to the local sub-domains of the GPUs to generate a local task slice sequence, which comprises boundary node slices and internal node slices, evaluating communication load and calculation load of the slices, and taking the internal node slices as calculation fillers to be executed in the communication waiting period of the boundary node slices; and each GPU executes an inference task by using a multi-stream asynchronous pipeline constructed locally according to the local task slice sequence. The method can improve the throughput of distributed inference.