An edge collaborative optimization method based on graph neural networks and digital twins
By constructing an edge collaborative optimization method using graph neural networks and digital twin technology, the problem of multi-node resource collaborative optimization in edge computing networks is solved, realizing dynamic collaborative scheduling and real-time optimization of resources, and improving computing efficiency and adaptability.
Patent Information
- Application Number
- CN202611556049.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-09
- Publication Date
- 2026-11-17
AI Technical Summary
There is a problem of multi-node resource collaborative optimization in edge computing networks. Existing technologies are insufficient in terms of computational complexity, dynamic adaptability and model evolution, making it difficult to achieve effective resource collaborative scheduling.
An edge collaborative optimization method is constructed using graph neural networks and digital twin technology. By building an edge computing network system model, the graph neural network is used to encode and decode the network state tensor. Combined with supervised learning and incremental update mechanism, real-time optimization of resource collaborative decision-making is achieved.
It enables multi-dimensional dynamic collaborative scheduling of resources among edge nodes, reduces online decision-making latency, adapts to network topology changes, and improves resource utilization and the real-time performance and reliability of decision-making.
Smart Images

Figure CN123123151A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of edge computing network resource management technology, and in particular relates to an edge collaborative optimization method based on graph neural networks and digital twins. Background Technology
[0002] In the development of edge computing network architecture, the early deployment model of a single edge server has gradually evolved into a multi-layered heterogeneous collaborative computing network that includes ground base station edge nodes, UAV-mounted edge computing platforms, satellite edge computing nodes, and vehicle edge nodes. This heterogeneous and distributed development trend has significantly increased the dimensionality of network resources, involving the computing resources, communication bandwidth resources, storage resources, and node energy resources of each node. There are strong coupling relationships between different types of resources: task offloading decisions directly affect the allocation of computing resources and communication bandwidth usage; computing resource allocation affects task processing latency and node energy consumption; and energy consumption status, in turn, restricts the available computing capacity of nodes, forming a complex multi-domain coupling relationship. This deep coupling of multi-dimensional resources makes the edge collaborative optimization problem extremely complex.
[0003] In the field of resource collaborative optimization methods, existing research mainly focuses on approaches such as resource optimization based on mathematical programming, resource collaborative scheduling based on heuristic algorithms, dynamic resource collaboration based on deep reinforcement learning, edge resource competition and collaboration based on game theory, and digital twin technology. Resource optimization based on mathematical programming offers solutions with guaranteed optimality, but its computational complexity increases exponentially or exponentially with the number of nodes and tasks, making it difficult to meet the timeliness requirements of online decision-making in large-scale dynamic networks. Resource collaborative scheduling based on heuristic algorithms exhibits good flexibility in resource collaborative scheduling and can handle the joint optimization of multi-dimensional resources, but it lacks theoretical guarantees of optimality, and algorithm parameter adjustments rely heavily on prior experience and repeated trials, resulting in unstable generalization performance under different network sizes and topologies. Dynamic resource collaboration based on deep reinforcement learning typically requires extensive online exploration and trial and error, leading to long training times, low sample efficiency, and problems such as policy oscillations and convergence difficulties in highly dynamic network environments. Most existing game theory methods are applicable to static or quasi-static scenarios. In edge computing networks where nodes dynamically join / leave and resource states change in real time, the timeliness of solving game equilibrium faces challenges.
[0004] Digital twins enable real-time mapping and interaction between the physical and information worlds by constructing virtual mirrors of physical systems. However, the application of existing digital twin technology in edge computing resource collaboration is still in its early stages, mainly focusing on passive tasks such as state awareness and prediction. It has not yet achieved the learning and approximation of the optimization decision-making process itself, and suffers from key problems such as a lack of effective mechanisms for resource collaboration between edge nodes, excessively high computational complexity in resource collaboration optimization, and a lack of effective solutions to the dynamic nature of network topology. In summary, resource collaboration optimization in edge computing networks urgently requires systemic technological breakthroughs in areas such as multi-node resource collaboration mechanisms, computational efficiency, dynamic adaptability, and model evolvability. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes an edge collaborative optimization method based on graph neural networks and digital twins, thereby resolving the issues present in the prior art.
[0006] To achieve the above objectives, this invention provides an edge collaborative optimization method based on graph neural networks and digital twins, comprising: Based on heterogeneous edge nodes and user devices, an edge computing network system model is constructed, resulting in a resource collaborative optimization problem with the goal of minimizing the total system cost while satisfying resource capacity and task latency constraints. Based on the physical layer, digital twin layer, and decision layer, a digital twin architecture is constructed. A graph neural network model is used to encode and decode the input network state tensor to obtain the predicted resource coordination decision. An expert dataset is constructed based on historical best decision data, and the graph neural network model is trained using supervised learning. A heterogeneous graph is constructed based on the current network state data, and a trained graph neural network model is used to process the heterogeneous graph to obtain resource coordination decisions for the current time slot. When the utility deviation between the predicted decision and the true optimal decision meets the triggering condition, the model parameters are incrementally updated using mini-batch gradient descent combined with elastic weight consolidation, and resource coordination decisions for subsequent time slots are obtained based on the updated model.
[0007] Optionally, the process of constructing an edge computing network system model based on heterogeneous edge nodes and user devices includes: Based on the heterogeneous set of edge nodes and the set of user equipment, an edge node set containing computing capacity and communication bandwidth is obtained; Based on the set of edge nodes, a set of computing tasks containing data volume and computing density is obtained; Based on the set of edge nodes and the set of computing tasks, a multi-dimensional resource state vector is obtained, consisting of computing resource state, bandwidth resource state, energy resource state, and task requirement state. Based on the multidimensional resource state vector, the computational resource variables and bandwidth resource variables allocated to tasks by edge nodes are obtained; Based on the set of edge nodes, the set of computing tasks, the multidimensional resource state vector, and the computing resource variables and bandwidth resource variables, the edge computing network system model is obtained.
[0008] Optionally, the process of constructing a resource collaborative optimization problem with the objective of minimizing the total system cost and satisfying resource capacity and task latency constraints includes: The upload latency of the task is obtained based on the amount of data in the task and the allocated bandwidth. The computational latency of a task is obtained based on its data volume, computational density, and allocated computing resources. Based on the upload latency and the calculation latency, the task processing latency is obtained; The cost of edge node resource consumption is obtained by weighted summation based on the computing and bandwidth resource consumption of edge nodes. Based on the task processing latency and the resource consumption cost of the edge nodes, a weighted summation method is used to obtain the total system cost. Based on the total system cost, a resource collaborative optimization objective function is obtained with the goal of minimizing the total system cost. Based on the computing resource capacity, bandwidth resource capacity, task latency upper limit, and resource non-negativity of edge nodes, resource capacity constraints and task latency constraints are obtained. Based on the resource collaborative optimization objective function and the resource capacity constraint and task delay constraint, the resource collaborative optimization problem is obtained.
[0009] Optionally, the process of building a digital twin architecture based on the physical layer, digital twin layer, and decision layer includes: Based on the physical layer, a fixed time slot synchronization method is used to process the multi-dimensional resource status information of edge nodes to obtain synchronized network status data. Based on the synchronized network state data, a network state tensor is obtained, which includes the network topology graph structure, node computing resource state vectors, node bandwidth resource state vectors, and task requirement descriptions. Based on the network state tensor, a digital twin layer is constructed; Based on the network state tensor output by the digital twin layer, a decision layer is constructed, and the network state tensor is processed by solving the mapping method to obtain resource collaborative decision-making. The digital twin architecture is obtained based on the physical layer, the digital twin layer, and the decision layer.
[0010] Optionally, the process of encoding and decoding the input network state tensor using a graph neural network model to obtain the predicted resource collaborative decision includes: Based on the network state tensor, a heterogeneous graph containing computing task nodes and edge computing nodes is obtained; The heterogeneous graph is processed using a heterogeneous graph attention encoder to obtain the graph coding features of each node; A time-series encoder is used to process the graph coding features and global state features of each node to obtain the time-series coding features; The time-series coding features are processed using a resource cooperative decoder to obtain the predicted values for resource allocation and bandwidth resource allocation. Based on the predicted values of computing resource allocation and bandwidth resource allocation, the predicted resource coordination decision is obtained.
[0011] Optionally, the process of constructing an expert dataset based on historical best decision data and training the graph neural network model using supervised learning includes: The resource coordination optimization problem is processed using mathematical programming methods to obtain the optimal computational resource allocation value and the optimal bandwidth resource allocation value corresponding to each network state; Based on the network states and their corresponding optimal computing resource allocation values and optimal bandwidth resource allocation values, an expert sample is obtained; The expert dataset is obtained by repeatedly generating the expert samples under different network topologies, resource states, and task requirement modes. The graph neural network model is supervised learning using the expert dataset. A loss function is constructed based on the mean squared error between the predicted resource collaborative decision and the expert samples. The model parameters are then updated to obtain the trained graph neural network model.
[0012] Optionally, the process of constructing a heterogeneous graph based on current network state data, and processing the heterogeneous graph using a trained graph neural network model to obtain resource coordination decisions for the current time slot includes: Based on the resource usage of each node, pending task information, network topology connection status and node location update data collected in the current time slot, the current network status data is obtained. Based on the current network state data, a heterogeneous graph containing computing task nodes, edge computing nodes, and computing offloading relationship edges is obtained; The heterogeneous graph is processed by forward propagation using a trained graph neural network model to obtain the predicted values of computational resource allocation and bandwidth resource allocation for the current time slot. Based on the predicted values of computing resource allocation and bandwidth resource allocation for the current time slot, a resource coordination decision for the current time slot is obtained.
[0013] Optionally, the process of incrementally updating the model parameters using mini-batch gradient descent combined with elastic weight consolidation includes: The utility bias is obtained by comparing the resource coordination decision based on the current time slot prediction with the actual optimal decision. Based on the utility deviation, the cumulative deviation, the degree of decrease in prediction accuracy, or the state of network topology change are obtained. Triggering conditions are obtained based on the cumulative deviation exceeding a preset threshold, the prediction accuracy dropping below a set threshold, or a significant change in the network topology. When the triggering condition is met, an updated training dataset is obtained based on the newly generated expert samples and the original training dataset; The parameters of the graph neural network model are processed using mini-batch gradient descent to obtain the parameter update gradient; The gradient and elastic weight consolidation term are updated based on the parameters to obtain the incrementally updated model parameters.
[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides a unified modeling foundation for multi-dimensional joint scheduling of resources among heterogeneous edge nodes by constructing an edge computing network system model and forming a resource collaborative optimization problem. It achieves comprehensive perception and collaborative decision-making of the entire network's resource status by constructing a digital twin architecture comprising a physical layer, a digital twin layer, and a decision layer, and by employing a graph neural network model to encode and decode the network state tensor. Furthermore, it enables the model to have end-to-end mapping capability from network state to optimal collaborative decision by constructing an expert dataset based on historical optimal decision data and training the model using supervised learning, effectively reducing the computational complexity of online solutions. By constructing a heterogeneous graph based on current network state data and performing online inference, it improves the real-time performance of decisions. Finally, by combining elastic weights with incremental updates when utility deviation meets triggering conditions, it enables the model to adapt to dynamic changes in network topology and suppress the forgetting of old knowledge, thereby continuously outputting resource collaborative decisions for subsequent time slots. Attached Figure Description
[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating edge collaboration and resource optimization in an embodiment of the present invention. Figure 2 This is a graph-based digital twin architecture diagram according to an embodiment of the present invention. Detailed Implementation
[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0018] Example 1 like Figure 1 and Figure 2 As shown, this embodiment provides an edge collaborative optimization method based on graph neural networks and digital twins. This invention utilizes the mirroring characteristics of digital twins to achieve collaborative scheduling of edge nodes, online resource allocation, and improved resource utilization.
[0019] This invention addresses the challenges of heterogeneous node collaboration, high computational complexity in multi-dimensional resource joint optimization, dynamic changes in network topology, and the lack of continuous evolution capabilities in edge computing networks. It proposes an edge collaborative optimization architecture that deeply integrates graph neural networks and digital twins. First, it establishes an edge computing network system model encompassing heterogeneous edge nodes and user devices, and solves the multi-dimensional resource collaborative optimization problem. Second, it constructs a three-layer digital twin architecture comprising a physical layer, a digital twin layer, and a decision layer, abstracting the edge computing network as a heterogeneous graph model. Then, it designs a digital twin model based on a heterogeneous graph attention network and a time-gated recurrent unit, using expert datasets to supervise the learning of the mapping relationship between network states and optimal collaborative decisions. Finally, it introduces online inference, feasibility verification, and an incremental learning mechanism based on elastic weight consolidation. This invention achieves dynamic collaborative scheduling of computing, bandwidth, and energy resources among edge nodes, reduces online decision latency, and is applicable to various heterogeneous edge computing scenarios such as space-ground integrated networks, vehicle-to-everything (V2X) networks, and industrial IoT.
[0020] An edge collaborative optimization method based on graph neural networks and digital twins includes the following steps: First, based on heterogeneous edge nodes and user devices, an edge computing network system model is constructed to obtain a resource collaborative optimization problem with the objective of minimizing the total system cost while satisfying resource capacity and task latency constraints. Second, based on the physical layer, digital twin layer, and decision layer, a digital twin architecture is constructed, and a graph neural network model is used to encode and decode the input network state tensor to obtain the predicted resource collaborative decision. Third, based on historical optimal decision data, an expert dataset is constructed, and supervised learning is used to train the graph neural network model. Fourth, based on the current network state data, a heterogeneous graph is constructed, and the trained graph neural network model is used to process the heterogeneous graph to obtain the resource collaborative decision for the current time slot. Fifth, when the utility deviation between the predicted decision and the true optimal decision meets the triggering condition, mini-batch gradient descent combined with elastic weight consolidation is used to incrementally update the model parameters, and the updated model is used to obtain the resource collaborative decision for subsequent time slots.
[0021] S1. Construct an edge computing network system model and resource collaboration architecture, and establish an edge computing network system model that includes a heterogeneous set of edge nodes and a set of user devices.
[0022] Furthermore, the process of constructing an edge computing network system model based on heterogeneous edge nodes and user devices includes: obtaining an edge node set containing computing capacity and communication bandwidth based on the heterogeneous edge node set and user device set; obtaining a set of computing tasks containing data volume and computing density based on the edge node set; obtaining a multi-dimensional resource state vector composed of computing resource state, bandwidth resource state, energy resource state, and task requirement state based on the edge node set and computing task set; obtaining computing resource variables and bandwidth resource variables allocated to tasks by edge nodes based on the multi-dimensional resource state vector; and obtaining the edge computing network system model based on the edge node set, computing task set, multi-dimensional resource state vector, and computing resource variables and bandwidth resource variables.
[0023] As a specific implementation of this embodiment, S1 specifically includes the following steps: S11. Let the set of edge nodes be defined. This includes computing capacity and communication bandwidth; S12, Design the set of computational tasks The task includes a large amount of data. and calculate density ; S13. Define a multidimensional resource state vector. Each time slot The system resource status consists of computing resource status, bandwidth resource status, energy resource status, and task requirement status. S14. Define resource allocation variables, edge nodes. Assigned to task The computing resources and bandwidth are respectively used , express.
[0024] S2. Construct a resource coordination optimization problem for multi-dimensional resource allocation to maximize the overall system utility.
[0025] Furthermore, the process of constructing a resource collaborative optimization problem with the objective of minimizing the total system cost while satisfying resource capacity and task latency constraints includes: obtaining the upload latency of the task based on the data volume and allocated bandwidth; obtaining the computation latency of the task based on the data volume, computation density, and allocated computation resources; obtaining the task processing latency based on the upload latency and computation latency; obtaining the edge node resource consumption cost by weighted summation based on the computation resource consumption and bandwidth resource consumption of edge nodes; obtaining the total system cost by weighted summation based on the task processing latency and edge node resource consumption cost; obtaining the resource collaborative optimization objective function with the objective of minimizing the total system cost based on the total system cost; obtaining resource capacity constraints and task latency constraints based on the computation resource capacity, bandwidth resource capacity, task latency upper limit, and resource non-negativity of edge nodes; and obtaining the resource collaborative optimization problem based on the resource collaborative optimization objective function, resource capacity constraints, and task latency constraints.
[0026] As a specific implementation method of this embodiment, the specific implementation process of S2 includes: S21. Task processing latency includes upload latency and computation latency. express, ,in, For data transmission latency, To calculate processing delay; S22. Edge node resource consumption costs comprehensively consider computational resource consumption, bandwidth resource consumption, and energy consumption. express, ,in, , For the weighting coefficients, satisfying ; S23. The total system cost is the weighted sum of the processing latency of all tasks and the resource consumption cost of all edge nodes. express, , in, and For the weighting coefficients, satisfying These respectively reflect the system's emphasis on latency performance and resource consumption; S24. The constraints of the optimization problem include computational resource capacity constraints, bandwidth resource capacity constraints, task latency constraints, and resource non-negativity constraints.
[0027] S3. Establish a three-layer digital twin architecture that includes a physical layer, a digital twin layer, and a decision layer.
[0028] Furthermore, the process of constructing a digital twin architecture based on the physical layer, digital twin layer, and decision layer includes: First, based on the physical layer, processing the multi-dimensional resource state information of edge nodes using a fixed-time-slot synchronization method to obtain synchronized network state data; second, based on the synchronized network state data, obtaining a network state tensor containing the network topology graph structure, node computational resource state vectors, node bandwidth resource state vectors, and task requirement descriptions; third, constructing the digital twin layer based on the network state tensor; fourth, constructing the decision layer based on the network state tensor output by the digital twin layer, processing the network state tensor using a solution mapping method to obtain resource collaborative decisions; and finally, obtaining the digital twin architecture based on the physical layer, digital twin layer, and decision layer.
[0029] As a specific implementation method of this embodiment, the specific implementation process of S3 includes: S31. Construct a physical layer. The physical layer synchronizes multi-dimensional resource status information to the digital twin layer in fixed time slots, including the computing resource usage, bandwidth utilization, remaining energy, task queue length, and location changes of each node. S32. Construct a digital twin layer and a network state tensor. The network state tensor includes the network topology graph structure, the computational and bandwidth resource state vectors of each node, and a description of task requirements. express, , Represents the network topology structure. This represents the resource state vector calculated by each node. This represents the bandwidth resource status vector of each node. This indicates a description of the task requirements; S33. Construct a decision-making layer to receive the output of the digital twin layer, generate and distribute resource coordination decisions, and use resource coordination decisions for... This indicates that the core mapping relationship between the digital twin layer and the decision-making layer is... .
[0030] S4. Design a digital twin model based on graph neural networks to generate an expert dataset covering diverse scenarios.
[0031] As a specific implementation method of this embodiment, the specific implementation process of S4 includes: S41. Construct an edge computing network graph, using... It means that, among them, This represents a set of nodes, including computing task nodes and edge computing nodes. Represents a set of edges, containing possible computational unloading relation edges. This represents a set of node types. Different types of nodes have different feature dimensions and semantic meanings. In subsequent graph neural network processing, a type-specific parameterization method is adopted.
[0032] S42. Design a graph neural network digital twin model, using... The model comprises a heterogeneous graph attention encoder, a temporal encoder, and a resource collaborative decoder. Specifically, 1) the heterogeneous graph attention encoder uses a heterogeneous graph attention network to differentiate the processing of different types of nodes, in order to capture the complex interaction relationships and resource propagation paths between heterogeneous nodes. It means that, among them, Represents a node In the Features of the layer Represents a node The neighborhood group, The attention coefficient is calculated using a multi-head attention mechanism and reflects the attention of neighboring nodes. For the target node The importance of Represents the pair of node types The relevant learnable weight matrix enables type-specific feature transformation. This represents a non-linear activation function. The attention coefficient is calculated as follows: ,in, 1) Represents vector concatenation. 2) The time encoder uses gated recurrent units (GRUs) to capture the temporal evolution characteristics of the network state, using... It means that, among them, This represents the set of all node features output by the graph encoder. Representing global state characteristics. The temporal encoder enables the digital twin model to understand and predict the dynamic evolution of the network, and can infer the current optimal decision based on historical state trends. 3) The resource collaborative decoder decodes the encoded graph representation and temporal features into multi-dimensional resource collaborative decisions, using... It means that, among them, , . and These are two independent multilayer perceptrons that share underlying coding features. To ensure the feasibility of resource allocation output, in and Non-negative activation and normalization are applied to the output of the node so that the sum of the resources allocated to each node does not exceed its capacity. S43. Generate a training dataset using the historical optimal solutions to the resource collaborative optimization problem. It means that, among them, Supervised learning is used to train a graphical neural network, with the loss function being the sum of the mean squared error losses in resource allocation. It means that, among them, and This represents the weighting coefficients, and the last term is the L2 regularization term.
[0033] S5. Use a hierarchical optimization method to generate labels for the training dataset.
[0034] Furthermore, the specific steps for generating expert datasets in S5 are as follows: S51. Use mathematical programming to generate labels for the training dataset, and for each network state... Solve the resource coordination optimization problem defined in step S2, and obtain the optimal allocation of computational resources using the Lagrange duality method, interior point method, or alternating direction multiplier method. and bandwidth allocation ; S52. Repeat sub-step S51 under different network topologies, resource states, and task requirements to generate an expert dataset covering diverse scenarios.
[0035] S6. Collect current network status data, including resource usage of each node, information on tasks to be processed, network topology connection status, and node position updates. Construct a graph of the current time slot, input the constructed heterogeneous graph into the trained graph neural network model, perform forward propagation calculation, and output the predicted resource coordination decision.
[0036] Furthermore, the process of constructing a heterogeneous graph based on the current network state data and processing the heterogeneous graph using a trained graph neural network model to obtain the resource coordination decision for the current time slot includes: obtaining the current network state data based on the resource usage of each node, the information of tasks to be processed, the network topology connection status, and the node location update data collected in the current time slot; obtaining a heterogeneous graph containing computing task nodes, edge computing nodes, and computing offloading relationship edges based on the current network state data; performing forward propagation processing on the heterogeneous graph using a trained graph neural network model to obtain the predicted values of computing resource allocation and bandwidth resource allocation for the current time slot; and obtaining the resource coordination decision for the current time slot based on the predicted values of computing resource allocation and bandwidth resource allocation for the current time slot.
[0037] As a specific implementation method of this embodiment, the specific implementation process of S6 includes: S61, in each decision slot Collect current network status data, including resource usage of each node, pending task information, network topology connection status, and node location updates. Construct the graph of the current time slot according to the heterogeneous graph construction method in sub-step S41. ; S62, The heterogeneous graph to be constructed Input the trained graph neural network model Perform a forward propagation calculation and directly output the predicted resource coordination decision, i.e. .
[0038] S7. Monitor the utility deviation between the predicted decision and the true optimal decision, add the newly generated expert samples to the training dataset when incremental learning is triggered, and use mini-batch gradient descent to incrementally update the model parameters.
[0039] Furthermore, the process of incrementally updating the model parameters using mini-batch gradient descent combined with elastic weight consolidation includes: obtaining the utility bias based on the resource collaborative decision and the true optimal decision predicted in the current time slot; obtaining the cumulative bias, the degree of decline in prediction accuracy, or the network topology change state based on the utility bias; obtaining the trigger condition based on any one of the indicators of cumulative bias, prediction accuracy, and network topology change state; obtaining the updated training dataset based on the newly generated expert samples and the original training dataset when the trigger condition is met; processing the parameters of the graph neural network model using mini-batch gradient descent to obtain the parameter update gradient; and obtaining the incrementally updated model parameters based on the parameter update gradient and the elastic weight consolidation term.
[0040] As a specific implementation method of this embodiment, the specific implementation process of S7 includes: S71. In actual implementation, there is a cost deviation between the monitoring and predictive decision and the actual optimal decision. express; S72. When the cumulative deviation exceeds the preset threshold, or the prediction accuracy of the model drops below the set threshold in multiple consecutive time slots, or the network topology changes significantly, incremental learning is triggered. S73, The newly generated expert sample Add the training dataset and use mini-batch gradient descent to incrementally update the model parameters. The update function is: During incremental training, elastic weight consolidation techniques are employed to prevent catastrophic forgetting of old knowledge. It means that, among them, Represents the diagonal elements of the Fisher information matrix, measuring parameters. The importance of old tasks This represents the reinforcement strength coefficient.
[0041] Compared to other edge collaboration and resource allocation methods, this invention uses a heterogeneous graph attention network to uniformly model multiple types of nodes in the edge computing network. By capturing the differentiated interaction relationships between different types of nodes through type-specific attention parameters, it achieves comprehensive perception and collaborative scheduling of the multi-dimensional resource status of the entire network. Simultaneously, this invention captures the temporal evolution characteristics of the network state through a time-series encoder, enabling the digital twin model to perceive and predict topology change trends. Combined with an incremental learning mechanism, the model can quickly and adaptively adjust when the network topology undergoes significant changes, coping with dynamic environments without retraining.
[0042] At the modeling level, this invention constructs a multi-dimensional resource state tensor and a heterogeneous edge network graph model incorporating computation, bandwidth, and energy, achieving a comprehensive characterization of resource relationships between edge nodes. At the optimization level, it establishes a system utility maximization problem under multiple resource constraints, generating a high-quality expert dataset. At the learning level, it designs a digital twin model based on a heterogeneous graph attention network and a time-gated recurrent unit, capturing differentiated interactions between heterogeneous nodes through type-specific attention parameters, achieving an end-to-end mapping from network state to optimal resource collaborative decision-making. At the deployment level, it introduces online inference and an incremental learning mechanism based on elastic weight consolidation, ensuring the real-time performance, reliability, and long-term adaptability of decisions. This invention provides a systematic technical solution for resource collaborative optimization of edge computing networks, exhibiting significant advantages in multi-node collaboration, dynamic adaptation, computational efficiency, and model maintainability.
[0043] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. An edge collaborative optimization method based on graph neural networks and digital twins, characterized in that, Includes the following steps: Based on heterogeneous edge nodes and user devices, an edge computing network system model is constructed, resulting in a resource collaborative optimization problem with the goal of minimizing the total system cost while satisfying resource capacity and task latency constraints. Based on the physical layer, digital twin layer, and decision layer, a digital twin architecture is constructed. A graph neural network model is used to encode and decode the input network state tensor to obtain the predicted resource coordination decision. An expert dataset is constructed based on historical best decision data, and the graph neural network model is trained using supervised learning. A heterogeneous graph is constructed based on the current network state data, and a trained graph neural network model is used to process the heterogeneous graph to obtain resource coordination decisions for the current time slot. When the utility deviation between the predicted decision and the true optimal decision meets the triggering condition, the model parameters are incrementally updated using mini-batch gradient descent combined with elastic weight consolidation, and resource coordination decisions for subsequent time slots are obtained based on the updated model.
2. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of constructing an edge computing network system model based on heterogeneous edge nodes and user devices includes: Based on the heterogeneous set of edge nodes and the set of user equipment, an edge node set containing computing capacity and communication bandwidth is obtained; Based on the set of edge nodes, a set of computing tasks containing data volume and computing density is obtained; Based on the set of edge nodes and the set of computing tasks, a multi-dimensional resource state vector is obtained, consisting of computing resource state, bandwidth resource state, energy resource state, and task requirement state. Based on the multidimensional resource state vector, the computational resource variables and bandwidth resource variables allocated to tasks by edge nodes are obtained; Based on the set of edge nodes, the set of computing tasks, the multidimensional resource state vector, and the computing resource variables and bandwidth resource variables, the edge computing network system model is obtained.
3. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of constructing a resource coordination optimization problem with the objective of minimizing the total system cost while satisfying resource capacity and task latency constraints includes: The upload latency of the task is obtained based on the amount of data in the task and the allocated bandwidth. The computational latency of a task is obtained based on its data volume, computational density, and allocated computing resources. Based on the upload latency and the calculation latency, the task processing latency is obtained; The cost of edge node resource consumption is obtained by weighted summation based on the computing and bandwidth resource consumption of edge nodes. Based on the task processing latency and the resource consumption cost of the edge nodes, a weighted summation method is used to obtain the total system cost. Based on the total system cost, a resource collaborative optimization objective function is obtained with the goal of minimizing the total system cost. Based on the computing resource capacity, bandwidth resource capacity, task latency upper limit, and resource non-negativity of edge nodes, resource capacity constraints and task latency constraints are obtained. Based on the resource collaborative optimization objective function and the resource capacity constraint and task delay constraint, the resource collaborative optimization problem is obtained.
4. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of building a digital twin architecture, based on the physical layer, digital twin layer, and decision layer, includes: Based on the physical layer, a fixed time slot synchronization method is used to process the multi-dimensional resource status information of edge nodes to obtain synchronized network status data. Based on the synchronized network state data, a network state tensor is obtained, which includes the network topology graph structure, node computing resource state vectors, node bandwidth resource state vectors, and task requirement descriptions. Based on the network state tensor, a digital twin layer is constructed; Based on the network state tensor output by the digital twin layer, a decision layer is constructed, and the network state tensor is processed by solving the mapping method to obtain resource collaborative decision-making. The digital twin architecture is obtained based on the physical layer, the digital twin layer, and the decision layer.
5. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of encoding and decoding the input network state tensor using a graph neural network model to obtain the predicted resource collaborative decision-making includes: Based on the network state tensor, a heterogeneous graph containing computing task nodes and edge computing nodes is obtained; The heterogeneous graph is processed using a heterogeneous graph attention encoder to obtain the graph coding features of each node; A time-series encoder is used to process the graph coding features and global state features of each node to obtain the time-series coding features; The time-series coding features are processed using a resource cooperative decoder to obtain the predicted values for resource allocation and bandwidth resource allocation. Based on the predicted values of computing resource allocation and bandwidth resource allocation, the predicted resource coordination decision is obtained.
6. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of constructing an expert dataset based on historical best decision data and training the graph neural network model using supervised learning includes: The resource coordination optimization problem is processed using mathematical programming methods to obtain the optimal computational resource allocation value and the optimal bandwidth resource allocation value corresponding to each network state; Based on the network states and their corresponding optimal computing resource allocation values and optimal bandwidth resource allocation values, expert samples are obtained; The expert dataset is obtained by repeatedly generating the expert samples under different network topologies, resource states, and task requirement modes. The graph neural network model is supervised learning using the expert dataset. A loss function is constructed based on the mean squared error between the predicted resource collaborative decision and the expert samples. The model parameters are then updated to obtain the trained graph neural network model.
7. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of constructing a heterogeneous graph based on current network state data, processing the heterogeneous graph using a trained graph neural network model, and obtaining resource coordination decisions for the current time slot includes: Based on the resource usage of each node, pending task information, network topology connection status and node location update data collected in the current time slot, the current network status data is obtained. Based on the current network state data, a heterogeneous graph containing computing task nodes, edge computing nodes, and computing offloading relationship edges is obtained; The heterogeneous graph is processed by forward propagation using a trained graph neural network model to obtain the predicted values of computational resource allocation and bandwidth resource allocation for the current time slot. Based on the predicted values of computing resource allocation and bandwidth resource allocation for the current time slot, a resource coordination decision for the current time slot is obtained.
8. The edge collaborative optimization method based on graph neural networks and digital twins according to claim 1, characterized in that, The process of incrementally updating model parameters using mini-batch gradient descent combined with elastic weight consolidation includes: The utility bias is obtained by comparing the resource coordination decision based on the current time slot prediction with the actual optimal decision. Based on the utility deviation, the cumulative deviation, the degree of decrease in prediction accuracy, or the state of network topology change are obtained. The triggering condition is obtained based on any one of the following indicators: the cumulative deviation, the prediction accuracy, and the network topology change status. When the triggering condition is met, an updated training dataset is obtained based on the newly generated expert samples and the original training dataset; The parameters of the graph neural network model are processed using mini-batch gradient descent to obtain the parameter update gradient; The gradient and elastic weight consolidation term are updated based on the parameters to obtain the incrementally updated model parameters.
