A network recovery strategy optimization method based on GNN-DQN

Abstract

The application discloses a network recovery strategy optimization method based on GNN-DQN, relates to the technical field of network recovery strategy optimization, and aims to solve the problems of the existing method, such as staticity, lack of physical and functional coupling consideration, and difficulty in adapting to dynamic constraints. The core steps are as follows: acquiring network topology, static and dynamic characteristic data of nodes and edges, and constructing a dynamic graph state; fusing features through a GNN model containing a GINEConv layer to output high-dimensional embedding representation; inputting a resource state vector into a DQN model to calculate component recovery Q values; generating an optimal recovery sequence based on a Top-K strategy and resource constraints, updating the network state after execution, training the model with the NPI performance increment as reward feedback, and repeating the process until the network recovery meets the standard. The method realizes dynamic intelligent decision-making, considers multi-objective optimization, improves the utilization rate of operation and maintenance resources and the network recovery speed, and effectively enhances the network resilience.

Description

Technical Field

[0001] This invention relates to the field of network recovery strategy optimization technology, specifically a network recovery strategy optimization method based on GNN-DQN. Background Technology

[0002] With the development of the digital economy, the scale of various information infrastructure networks, such as backbone communication networks, computing power scheduling networks, and cloud-edge collaborative business networks, continues to expand, with network nodes and transmission links being highly coupled and interconnected. These networks carry massive real-time business data streams, and once they encounter disturbances such as hardware failures, network attacks, bandwidth overload, or fiber optic cable interruptions, the functions of core business nodes and transmission links will be degraded or completely interrupted. Because information networks have strong spatiotemporal data correlation and strong business flow dependence, single-point failures can trigger cascading packet loss, cross-regional business interruptions, and global business congestion, directly causing business shutdowns and online service failures, resulting in significant economic losses.

[0003] Therefore, improving the network's self-healing capability after being disturbed has become a core concern for network operation and maintenance management departments. Current research and practice on network fault recovery mainly fall into two categories: one is static topology vulnerability assessment based on complex network theory, and the other is intelligent operation and maintenance decision optimization based on deep learning and reinforcement learning.

[0004] At the static topology analysis level, existing methods abstract the network into a node-link graph model, identifying weak links through indicators such as node degree, betweenness centrality, and global network efficiency. However, these methods can only perform offline static connectivity assessments and cannot adapt to dynamic service loads. For example, patent CN115734274B discloses a cellular network fault diagnosis method based on deep learning and knowledge graphs. This scheme uses graph convolutional networks to mine the topological relationships between 5G base stations and transmission nodes, but it is only used for fault type classification and identification, and only realizes fault alarm diagnosis. It does not design a dynamic step-by-step recovery decision process for multi-fault concurrent scenarios, and completely ignores the impact of dynamic service characteristics such as real-time service load and link throughput on recovery priority.

[0005] At the dynamic intelligent optimization level, existing research falls into two categories: The first category utilizes graph neural networks to process network monitoring data streams, completing fault location and status monitoring. For example, patent CN115860392B discloses an energy internet optimization and reconstruction method and system based on graph convolutional networks. This scheme achieves network topology reconstruction based on GCN, but the optimization objective is only single-point fault location and static topology reconstruction of power communication links. It does not combine the upper limit of operation and maintenance resources to generate multi-component collaborative repair sequences and does not support multi-cycle continuous dynamic recovery iteration. This type of technology essentially belongs to the category of fault perception and static topology adjustment, and cannot output a global repair scheduling scheme oriented towards business benefits. The other category introduces deep reinforcement learning into network resource scheduling. For example, patent CN118819776B discloses a mobile edge computing task scheduling method based on quantum heuristic DDQN. This scheme achieves computing power task allocation based on deep Q-networks, but the model is only oriented towards computing task scheduling and does not integrate communication layer characteristics such as transmission latency and multi-service requirements, and cannot establish a coupling quantitative relationship between hardware links and upper-layer transmission services. In addition, patent CN115333947A discloses a routing optimization method based on graph neural networks and deep reinforcement learning. Although it also combines GNN and DQN, its optimization goal is the real-time forwarding path of data packets. It is geared towards logical layer routing scheduling and does not involve the optimization of the batch repair sequence after physical node or transmission link hardware failure. The two have fundamental differences in state space, constraints, and optimization goals.

[0006] In summary, existing network fault recovery technologies suffer from three core deficiencies: First, existing topology assessment, GNN fault diagnosis, and DRL resource scheduling schemes fail to establish a dynamic and quantitative coupling relationship between network hardware topology and upper-layer service transmission. Second, physical link connectivity does not equate to normal service transmission; models lack functional layer metrics such as data flow and link throughput, making it impossible to accurately quantify the service benefits of repair actions. Third, traditional static sorting and offline intelligent optimization algorithms cannot adapt to the dynamic network state during fault evolution; strategies cannot be adaptively adjusted in real time for scenarios such as post-repair service data flow reallocation and the addition of new concurrent faults. Fourth, existing DQN scheduling schemes lack a sequential decision-making framework for balancing short-term and long-term benefits under conditions of large-scale networks, multiple repair cycles, and resource constraints. Fifth, network fault recovery is a strongly constrained optimization problem, requiring simultaneous consideration of limited operational computing power, hardware spare parts, repair time, and pre-repair dependencies of components. Sixth, existing single models struggle to simultaneously accommodate multiple constraints and cannot batch output optimal recovery sets that satisfy resource limits through multi-component collaborative efforts.

[0007] The aforementioned shortcomings, when applied to operational and maintenance scenarios, can cause serious problems: emergency maintenance personnel relying solely on manual experience to determine the repair sequence are prone to getting trapped in local optima, resulting in delayed recovery of core transmission channels, prolonged low-level operation of network-wide service capabilities, and significantly extended service interruption duration and economic losses. Therefore, there is an urgent need for a closed-loop intelligent recovery scheduling method that integrates the network's full-time and spatial characteristics and adapts to dynamic operational and maintenance constraints. Summary of the Invention

[0008] The purpose of this invention is to provide a network recovery strategy optimization method based on GNN-DQN to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a network recovery strategy optimization method based on GNN-DQN, comprising the following steps:

[0010] S1. Obtain the topology data, static and dynamic feature data of nodes, and static and dynamic feature data of edges of the target network;

[0011] S2. Based on the data obtained in step S1, construct a dynamic graph state representation of the network;

[0012] S3. Input the dynamic graph state representation obtained in step S2 into the pre-trained graph neural network (GNN) feature fusion model for processing to obtain a high-dimensional graph embedding representation of the current network state. The graph neural network (GNN) feature fusion model uses the GINEConv layer, which can fuse edge features, as the core message passing operator. By performing multi-layer aggregation and updating of node features and edge features, it outputs node-level embedding vectors and graph-level embedding vectors.

[0013] S4. Input the node-level embedding vector, graph-level embedding vector, and currently available recovery resource state vector obtained in step S3 into the pre-trained deep Q-network DQN decision model. The DQN decision model calculates a Q value for each service node or transmission link component to be repaired based on the input embedding vector and resource state. The Q value represents the expected long-term cumulative performance benefit that can be brought about by choosing to restore the component at the current position.

[0014] S5. Based on the Q values ​​of each component calculated in step S4, and combined with the current recovery resource constraints, the optimal recovery action sequence for the current moment is generated using the Top-K selection strategy; the recovery action sequence refers to a set of one or more service nodes or transmission links that should be repaired first within the current decision cycle.

[0015] S6. Execute the recovery action sequence generated in step S5, update the damaged state of the network, and calculate the network performance recovery increment based on the network comprehensive performance index NPI. Use this as a reward signal to feed back to the training process of the DQN decision model. Then, return to step S2 and make a recovery decision for the next moment based on the updated network state until the network performance recovers to the preset threshold or reaches the maximum recovery period.

[0016] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] (1) It overcomes the limitations of existing methods that only assess static topology connectivity or only repair business data flow, and realizes an integrated and refined measurement of the status of network physical infrastructure and dynamic transmission service performance, providing a unified performance scale that accurately reflects the real transmission loss for recovery decisions.

[0019] (2) It solves the problem that existing graph neural network models are mostly used for data prediction or repair, but cannot effectively represent the complex relationship between infrastructure damage and repair decisions, and realizes a deep understanding and representation of high-dimensional, dynamic network states and their repair value.

[0020] (3) It solves the problem that existing static sorting or offline optimization algorithms cannot make real-time adaptive decisions in dynamic environments, and that general deep reinforcement learning models are difficult to handle complex engineering constraints. It realizes the generation of near-optimal recovery sequences that take into account both immediate benefits and long-term resilience under strict resource and engineering constraints.

[0021] (4) It provides data-driven, closed-loop intelligent decision support for the entire recovery cycle of large-scale networks after extreme disturbances. It can dynamically output the optimal recovery instructions that can be executed based on the real-time network status and resource conditions, thereby significantly improving the utilization rate of maintenance spare parts / computing resources and accelerating the recovery speed of network services. It provides a feasible systematic technical solution for solving the core decision-making problems in improving network resilience. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0024] Figure 2 This is a structural diagram of the GNN feature fusion model of the present invention;

[0025] Figure 3 This is a schematic diagram of the DQN decision model and action generation of the present invention;

[0026] Figure 4 This is a schematic diagram of the model training and simulation interaction framework of the present invention;

[0027] Figure 5 This is a bar chart comparing network performance recovery speeds under different recovery strategies of the present invention;

[0028] Figure 6 A pie chart showing the optimization effect of resource allocation recovery according to the present invention;

[0029] Figure 7 This is a line graph comparing the network performance recovery process curves of the present invention;

[0030] Figure 8 This is a radar chart showing the multi-target optimization effect of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figures 1-4 As shown, this invention provides a network recovery strategy optimization method based on GNN-DQN, comprising the following steps:

[0033] S1. Obtain the topology data, static and dynamic feature data of nodes, and static and dynamic feature data of edges of the target network; wherein, the topology data includes a set of nodes and a set of edges, the static features of nodes include at least node service level and service scale attributes, the dynamic features of nodes include at least node fault level and real-time service traffic load, the static features of edges include at least edge length, transmission rate and link baseline throughput, and the dynamic features of edges include at least edge fault level and real-time service traffic load;

[0034] S2. Based on the data obtained in S1, construct a dynamic graph state representation of the network; the dynamic graph state representation is as follows: Where V is the set of nodes and E is the set of edges. Let be the characteristic matrix of all nodes at time t. Let be the feature set of all edges at time t;

[0035] S3. Input the dynamic graph state representation obtained in S2 into the pre-trained graph neural network (GNN) feature fusion model for processing to obtain a high-dimensional graph embedding representation of the current network state. The graph neural network (GNN) feature fusion model uses the GINEConv layer, which can fuse edge features, as the core message passing operator. By performing multi-layer aggregation and updating of node features and edge features, it outputs node-level embedding vectors and graph-level embedding vectors.

[0036] S4. Input the node-level embedding vector, graph-level embedding vector, and currently available recovery resource state vector obtained in S3 into the pre-trained deep Q-network DQN decision model. The DQN decision model calculates a Q value for each service node or transmission link component to be repaired based on the input embedding vector and resource state. The Q value represents the expected long-term cumulative performance benefit that can be brought by choosing to restore the component at the current position.

[0037] S5. Based on the Q values ​​of each component calculated in S4, and combined with the current recovery resource constraints, the optimal recovery action sequence for the current moment is generated using a Top-K selection strategy; the recovery action sequence refers to a set of one or more service nodes or transmission links that should be repaired first within the current decision cycle.

[0038] S6. Execute the recovery action sequence generated in S5, update the damaged state of the network, and calculate the network performance recovery increment based on the network comprehensive performance index NPI. Use this as a reward signal to feed back into the training process of the DQN decision model. Then, return to S2 and make a recovery decision for the next time step based on the updated network state until the network performance recovers to the preset threshold or reaches the maximum recovery period.

[0039] In step S1, the process of obtaining and constructing the static features of nodes and edges is as follows: extracting nodes V and edges E from the network topology graph; reading the service priority level of each node from the network operation and maintenance management platform and performing numerical encoding; statistically analyzing the total number of nodes in the entire network, static global connectivity, and other basic topology indicators, and performing Min-Max normalization processing to obtain the basic network topology features; and obtaining the network transmission rate of each edge from technical standard data. Combined with the geometric length of the side and the minimum forwarding interval of the network scheduling system Calculate the baseline throughput of the link according to the formula:

[0040]

[0041] In the formula Represents the physical length of the link. Represents the rated transmission rate of the link. This represents the minimum forwarding interval for data packets;

[0042] The node's service priority level, service scale attribute, link physical length, transmission rate, and link baseline throughput together constitute the static feature vector of the node and link.

[0043] In step S2, the specific process of constructing the dynamic graph state representation of the network is as follows: defining the node set. ,in Define the total number of nodes; define the edge set. A node pair belongs to the edge set if and only if there is a topological link between two nodes; at time step Integrate the feature vectors of all nodes to form a node feature matrix. :

[0044]

[0045] in, The total dimension of the node feature vectors. Static characteristics of nodes, such as node level and business scale attributes. For node dynamic characteristics, such as node fault level and real-time service traffic load; integrate the feature vectors of all edges. Forming an edge feature set ,in These are static edge characteristics, such as edge length, transmission rate, and link baseline throughput. For edge dynamic characteristics, such as edge fault level and real-time service traffic load, Let be the feature vector of the edge connecting node i and node j at time t; and let be the dynamic graph state at time t. By quadruple Complete characterization.

[0046] In S3, the graph neural network (GNN) feature fusion model adopts a three-layer GINEConv structure; for the first... The layer's message passing and node update process strictly follow the formula:

[0047]

[0048] in, It is a node In the Hidden representation of layers, Representation: The hidden representation of neighbor node j in layer l, initial layer , It is the original feature vector of node i at time t; It is a node The set of neighboring nodes; It is the edge At any moment eigenvectors; It is a function that performs a linear transformation on the edge features; It is the ReLU nonlinear activation function; It is a learnable scalar parameter; It is a multilayer perceptron; after three layers of computation, the final embedded representation of all nodes is obtained. and for all nodes Perform global average pooling to obtain the graph-level embedding representation. :

[0049]

[0050] In step S4, the training process of the pre-trained Deep Q-Network (DQN) decision model is based on a Markov decision process framework; the state of this framework... The dynamic graph state representation defined for step S2; actions To extract from the set of damaged components The selected subset that satisfies the resource constraints; the reward function is defined as:

[0051]

[0052] That is, the difference in the network performance index (NPI) between adjacent states; the optimization objective of the model is to maximize the expected cumulative discounted reward. ,in Let t be the discount factor γ raised to the power of t, where the discount factor determines the present value of the future reward. To minimize the impact of future rewards on current decisions, a value of 0.9 to 0.99 is used. For immediate rewards, T is the termination time step. The decision rules for the agent (DQN); the model minimizes the temporal difference loss. Training is performed, where C is a candidate recovery component currently being evaluated in state St, and the target value is... , and These are the Q values ​​output by the online network and the target network, respectively. For the next state Next candidate recovery components, For the next state Below is a collection of all feasible repair components.

[0053] The Deep Q-Network (DQN) decision model supports multi-objective decision-making; a preset objective weight vector is introduced during the training and decision-making process. ,in and These correspond to the relative importance of the resilience loss objective, the recovery cost objective, and the recovery period objective, respectively; the reward function is extended to a weighted multi-objective form:

[0054]

[0055] in, For the incremental toughness loss, NPI t Let NPI0 be the network performance index at time step t, where NPI0 is the network performance index at the initial time t=0, i.e., before recovery begins after the disturbance occurs. For cost increment, The resources / costs required to repair component c, The incremental value for the schedule penalty is 1 if any component is still in a faulty state, and 0 otherwise; the target weight vector As an additional input along with the state representation, the DQN decision model is fed into it, enabling the model to learn and generate recovery strategies that adapt to different decision preferences.

[0056] In step S5, the specific process of adopting the Top-K selection strategy in combination with the current recovery resource constraints is as follows: First, based on the fault repair prerequisite dependencies and the total amount of currently available recovery resources... From all damaged component sets Filtering out at time step A set of recoverable feasible components For sets Each component in The component-level embedding representation is extracted using a GNN feature fusion model. and with graph-level embedding Resource status By concatenating the data and inputting it into the DQN decision model, we can obtain its Q-value estimate. Sort all feasible components in descending order of their Q-values; finally, starting from the top of the sorted list, select components one by one to add to the recovery action set. Until the cumulative resource consumption of the selected component reaches the current period's resource limit. The selected set of components at this time is the optimal recovery action sequence for the current moment.

[0057] In step S6, the calculation method for the network comprehensive performance index (NPI) is as follows: calculate the network performance at each time step... Structural layer performance indicators and functional layer performance indicators The structural layer performance indicators Weighted network efficiency Weighted natural connectivity And level coupling redundancy The weights of the linear combination of the three indicators are determined using the entropy weight method, specifically as follows:

[0058]

[0059] The functional layer performance indicators It is a weighted average of the network-wide data flow-latency coupled service delivery efficiency calculated based on business data flow; NPI is calculated by the formula:

[0060]

[0061] in

[0062]

[0063]

[0064] and This represents the baseline performance value under normal network conditions. and These are the preset weighting coefficients.

[0065] The functional layer performance indicators The specific calculation process is as follows: First, based on the edge equivalent transmission delay matrix under the current network state... Business request frequency matrix For any source-destination node pair (S, D) in the network, representing the transmission requirement from node S to node D, find the forwarding path with the lowest arrival time and delay. ,in, Let S be the set of all possible paths from S to D. Given the equivalent transmission delay of edge (i,j) under the current network state, extract the bottleneck service request frequency on the forwarding path with the lowest delay as the effective supply. ;

[0066] Then, calculate the data flow and latency-coupled service delivery efficiency of the SD service source and destination nodes. Next, a weight matrix reflecting importance is constructed. Finally, the performance indicators of the entire network's functional layers are calculated using the formula:

[0067]

[0068] in The reachability efficiency matrix, This represents the weighted inner product.

[0069] In step S6, the specific process of executing the recovery action sequence and updating the network state is as follows: based on the recovery action sequence generated in step S5... For each component contained in the sequence Determine its fault level Updated to a healthy state, i.e. For those not included The components within maintain their fault level unchanged, that is The fault level update of a component directly determines its capacity. For transmission link components, their capacity is reduced according to the formula:

[0070]

[0071] in, The link baseline throughput reduction factor corresponding to the fault level; the updated fault level vector. Together with capacity information, they constitute the network state at the next time step. The foundation; based on Recalculate overall network performance metrics and before the action was performed Perform tasks and receive immediate rewards. The reward is used for subsequent training of the DQN decision model.

[0072] Figure 5 The bar chart comparing network performance recovery speed under different recovery strategies of the present invention shows that the network performance recovery speed of the method of the present invention is better than that of existing methods. Figure 6 The pie chart showing the optimization effect of resource allocation in this invention demonstrates that the allocation ratio of the method is reasonable. Figure 7 This is a line graph comparing the network performance recovery process curves of the present invention. The three curves in the graph, from top to bottom, represent the GNN-DQN strategy curve, the static ranking method curve, and the random recovery strategy curve, respectively. Figure 8 The radar chart showing the multi-objective optimization effect of this invention demonstrates that the GNN-DQN method proposed in this invention is the optimal solution that simultaneously achieves high recovery performance, low cost, and short construction period.

[0073] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A network recovery strategy optimization method based on GNN-DQN, characterized in that, Includes the following steps: S1. Obtain the topology data, static and dynamic feature data of nodes, and static and dynamic feature data of edges of the target network; S2. Based on the data obtained in step S1, construct a dynamic graph state representation of the network; S3. Input the dynamic graph state representation obtained in step S2 into the pre-trained graph neural network (GNN) feature fusion model for processing to obtain a high-dimensional graph embedding representation of the current network state. The graph neural network (GNN) feature fusion model uses the GINEConv layer, which can fuse edge features, as the core message passing operator. By performing multi-layer aggregation and updating of node features and edge features, it outputs node-level embedding vectors and graph-level embedding vectors. S4. Input the node-level embedding vector, graph-level embedding vector, and currently available recovery resource state vector obtained in step S3 into the pre-trained deep Q-network DQN decision model. The DQN decision model calculates a Q value for each service node or transmission link component to be repaired based on the input embedding vector and resource state. The Q value represents the expected long-term cumulative performance gains that can be achieved by choosing to restore the component at the current location. S5. Based on the Q values ​​of each component calculated in step S4, and combined with the current recovery resource constraints, the optimal recovery action sequence for the current moment is generated using the Top-K selection strategy; the recovery action sequence refers to a set of one or more service nodes or transmission links that should be repaired first within the current decision cycle. S6. Execute the recovery action sequence generated in step S5, update the damaged state of the network, and calculate the network performance recovery increment based on the network comprehensive performance index NPI, which is then used as a reward signal to feed back into the training process of the DQN decision model. Then, return to step S2 and make a recovery decision for the next moment based on the updated network status until the network performance recovers to the preset threshold or reaches the maximum recovery period.

2. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, The topology data includes a set of nodes and a set of edges. The static characteristics of nodes include node level and service scale attributes, and the dynamic characteristics of nodes include node fault level and real-time service traffic load. The static characteristics of edges include edge length, transmission rate and link baseline throughput, and the dynamic characteristics of edges include edge fault level and real-time service traffic load.

3. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, In step S1, the process of obtaining and constructing the static features of nodes and edges is as follows: extracting nodes from the network topology graph. ,side The system reads the service priority level of each node from the network operation and maintenance management platform and encodes it numerically; it calculates the total number of nodes and static global connectivity of the entire network, and performs Min-Max normalization to use it as the basic network topology feature; and it obtains the network transmission rate of each edge from the technical standard data. Combined with the geometric length of the side and the minimum forwarding interval of the network scheduling system Calculate the baseline throughput of the link according to the formula: ； In the formula This represents the rated transmission rate of the link.

4. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, The dynamic graph state is represented as follows: Where V is the set of nodes and E is the set of edges. Let be the characteristic matrix of all nodes at time t. Let be the feature set of all edges at time t; The specific process of constructing the dynamic graph state representation of the network is as follows: Define the set of nodes. Where N is the total number of nodes; define the edge set. A node pair belongs to the edge set if and only if there is a topological connection between two nodes; at time step Integrate the feature vectors of all nodes to form a node feature matrix. : ； in, The total dimension of the node feature vectors. For node static features, Dynamic features of nodes; integration of feature vectors from all edges Forming an edge feature set ,in As a static feature of the edge, The dynamic characteristics of the edge; the state of the dynamic graph at time t. By quadruple Complete characterization.

5. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, The graph neural network (GNN) feature fusion model described in step S3 adopts a three-layer GINEConv structure; for the first... The layer's message passing and node update process strictly follow the formula: ； in, It is a node In the Hidden representation of layers, This represents the hidden representation of neighbor node j in layer l, in the initial layer. , It is the original feature vector of node i at time t; It is a node The set of neighboring nodes; It is the edge At any moment eigenvectors; It is a function that performs a linear transformation on the edge features; It is the ReLU nonlinear activation function; It is a learnable scalar parameter; It is a multilayer perceptron; after three layers of computation, the final embedded representation of all nodes is obtained. and for all nodes Perform global average pooling to obtain the graph-level embedding representation. : 。 6. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, The training process of the pre-trained Deep Q-Network (DQN) decision model described in step S4 is based on a Markov decision process framework; the state of this framework The dynamic graph state representation defined for step S2; actions To extract from the set of damaged components The selected subset that satisfies the resource constraints; the reward function is defined as: ； That is, the difference in the network performance index (NPI) between adjacent states; the optimization objective of the model is to maximize the expected cumulative discounted reward. ,in The discount factor γ raised to the power of t. For immediate rewards, T is the termination time step. The decision rules for the agent; the model minimizes the temporal difference loss. Training was conducted, among which In the state Below, for a candidate recovery component currently being evaluated, the target value is... , and These are the Q values ​​output by the online network and the target network, respectively. For the next state Next candidate recovery components, For the next state The following is a collection of all feasible repair components; The Deep Q-Network (DQN) decision model supports multi-objective decision-making; during training and decision-making, a preset objective weight vector is introduced, where... and These correspond to the resilience loss target and the recovery cost target, respectively. The relative importance of the recovery schedule objective; the reward function is extended to a weighted multi-objective form: ； in, For the incremental toughness loss, NPI t Let NPI0 be the network performance index at time step t, where NPI0 is the network performance index at the initial time t=0, i.e., before recovery begins after the disturbance occurs. For cost increment, To repair the resources required for component c, The incremental penalty for project timeline is set to 1 if any component is still in a faulty state, and 0 otherwise; target weight vector. As an additional input along with the state representation, the DQN decision model is fed into it, enabling the model to learn and generate recovery strategies that adapt to different decision preferences.

7. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, In step S5, the specific process of adopting the Top-K selection strategy in combination with the current recovery resource constraints is as follows: First, based on the fault repair prerequisite dependencies and the total amount of currently available recovery resources... From all damaged component sets Filtering out at time step A set of recoverable feasible components For sets Each component in The component-level embedding representation is extracted using a GNN feature fusion model. and with graph-level embedding Resource status By concatenating the data and inputting it into the DQN decision model, we can obtain its Q-value estimate. Sort all feasible components in descending order of their Q-values; finally, starting from the top of the sorted list, select components one by one to add to the recovery action set. Until the cumulative resource consumption of the selected component reaches the current period's resource limit. The selected set of components at this time is the optimal recovery action sequence for the current moment.

8. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, The calculation method for the network comprehensive performance index (NPI) in step S6 is as follows: calculate the network performance at each time step. Structural layer performance indicators and functional layer performance indicators The structural layer performance indicators Weighted network efficiency Weighted natural connectivity And level coupling redundancy The weights of the linear combination of the three indicators are determined using the entropy weight method, specifically as follows: ； The functional layer performance indicators It is a weighted average of the network-wide data flow-latency coupled service delivery efficiency calculated based on business data flow; NPI is calculated by the formula: ； in: ， ， and This represents the baseline performance value under normal network conditions. and These are preset weighting coefficients; The functional layer performance indicators The specific calculation process is as follows: First, based on the edge equivalent transmission delay matrix under the current network state... Business request frequency matrix For any source-destination node pair (S, D) in the network, representing the transmission requirement from node S to node D, find the forwarding path with the lowest arrival time and delay. ,in, Let S be the set of all possible paths from S to D. Given the equivalent transmission delay of edge (i,j) under the current network state, extract the bottleneck service request frequency on the forwarding path with the lowest delay as the effective supply. ; Then, calculate the traffic-latency coupled service delivery efficiency of the SD service source and destination nodes. Next, a weight matrix reflecting importance is constructed. Finally, the performance indicators of the entire network's functional layers are calculated using the formula: ； in The reachability efficiency matrix, This represents the weighted inner product.

9. The network recovery strategy optimization method based on GNN-DQN according to claim 1, characterized in that, The specific process of executing the recovery action sequence and updating the network state in step S6 is as follows: based on the recovery action sequence generated in step S5... For each component contained in the sequence Determine its fault level Updated to a healthy state, i.e. For those not included The components within maintain their fault level unchanged, that is The fault level update of a component directly determines its capacity. For transmission link components, their capacity is reduced according to the formula: ； in, The link baseline throughput reduction factor corresponding to the fault level; the updated fault level vector. Together with capacity information, they constitute the network state at the next time step. The foundation; based on Recalculate overall network performance metrics and before the action was performed Perform tasks and receive immediate rewards. The reward is used for subsequent training of the DQN decision model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 9.

Images