GCN-based multi-layer satellite network control node optimization deployment method

Abstract

The invention discloses a multi-layer satellite network control node optimization deployment method based on a GCN, relates to the technical field of satellite network management and graph neural network application, and solves the problems that in an existing large-scale network, the deployment calculation complexity of a controller is high; in order to solve the problems that an MCD model lacks research on inter-layer interactive modeling in a multi-layer control plane and the like, the cooperative work of a super controller and a controller is realized by constructing a multi-layer satellite network architecture, so that the network response capability of a cross-domain service is improved. And a mathematical model which aims at reducing the network delay and balancing the load of the controller is established, and solving is carried out through an SA algorithm. Besides, a multi-controller deployment algorithm of the GCN network is provided, a solution process of a neural network learning SA algorithm is utilized, a rapid approximate optimal deployment strategy is realized, and technical support is provided for application of a large-scale satellite network. Experimental results show that the method is superior to other related schemes in the aspects of network response time delay, load balancing and expandability.

Description

Technical Field

[0001] This invention relates to the fields of satellite network management and graph neural network application technology, specifically to a multi-layer satellite network controller selection method based on graph convolutional networks (GCN). It aims to address the challenge of efficiently selecting multiple controllers in multi-layer satellite networks by introducing inter-controller layer interactions and GCN solution methods, thereby reducing network response latency, balancing controller load, and improving the applicability and computational efficiency of the solution in large-scale satellite constellations. Background Technology

[0002] In Software Defined Network (SDN) architectures, controller deployment is a critical issue. Improper deployment can severely impact the performance and reliability of the entire network, leading to increased network response latency, reduced data transmission efficiency, and increased network management complexity. Unlike terrestrial networks, the connections between controllers and switches in satellite networks are constantly changing, requiring frequent updates to node deployment plans. This necessitates highly efficient deployment calculations. Furthermore, to achieve global control, whether controllers are placed on ground stations or satellites, a large number of control nodes are needed. A large number of control nodes imposes significant computational overhead on deployment plans, making satellite network controller deployment a challenging problem.

[0003] Existing research on multi-controller deployment (MCD) in satellite networks mainly includes single-layer and multi-layer network architectures. In a single-layer architecture, controllers are typically deployed on low Earth orbit (LEO) satellites. However, due to limited coverage, a large number of controllers are required, and the high dynamism of LEO satellites leads to frequent switching between controllers and switches. These characteristics increase the complexity of network control, especially in large-scale satellite constellations, where the limitations of this single-layer architecture are more pronounced. In a two-layer architecture, high Earth orbit (GEO) satellites typically act as controllers, while LEO satellites act as switches. Although GEO satellites have wide coverage, the greater distance between them and LEO satellites increases control latency. Therefore, a multi-layer control approach is considered more suitable for improving the control efficiency of satellite networks. In this architecture, satellites above LEO act as the first-layer controllers, and satellites in higher orbits act as the second-layer controllers. The first-layer controllers are closer to the LEO satellites, ensuring lower control latency, while the second-layer controllers expand the control range and further optimize overall latency through cooperation with the first-layer controllers. This multi-layer control plane architecture offers greater flexibility, and the hierarchical collaborative control approach is more suitable for large-scale satellite constellation scenarios. However, there is currently little research on multi-layered control surfaces for satellite networks, and no MCD studies have been found that model the interactions between control surface layers. Moreover, the applicability of most MCD schemes in large-scale satellite networks is still unclear. Summary of the Invention

[0004] To address the problems of high computational complexity in controller deployment in existing large-scale networks; lack of research on inter-layer interaction modeling in multi-layer control planes in existing MCD models; and unclear applicability of most MCD models in large-scale satellite networks, this invention provides an optimized deployment method for control nodes in multi-layer satellite networks based on GCN.

[0005] A method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN is proposed. This method involves the following steps: First, given the current time t, the method calculates the deployment scheme of the controller at time t+1.

[0006] Step 1: Calculate the average network response delay dl at time t+1 based on the network response delay at time t. t+1 The average network response delay dl t+1 Set the sum of the average flow delay and the average switch migration delay;

[0007] Step 2: Calculate the controller load after the switch migration caused by controller overload or failure at time t+1. The load includes controller c j Calculate the load on the existing switch and the load on the newly migrated switch, and calculate the load variance.

[0008] Step 3: Construct an objective function based on the average network response delay obtained in Step 1 and the load variance between controllers obtained in Step 2, and set constraints.

[0009] Step 4: Input the obtained time-slot topology data into the GCN model for training to obtain the trained GCN model;

[0010] Step 5: Use the trained GCN model to derive the controller deployment scheme, output and save the optimal controller deployment strategy.

[0011] The beneficial effects of this invention are:

[0012] The multi-layer satellite network controller selection method described in this invention constructs a two-layer spatial control plane architecture and introduces an inter-layer cooperation mechanism to improve overall control efficiency. To address the high computational complexity of controller deployment in large-scale networks, this invention introduces a Graph Convolutional Network (GCN) to solve for controller deployment schemes. This method uses Simulated Annealing (SA) as a reference, learning its optimization process to effectively improve the solution speed and significantly reduce the computational overhead of deployment strategies. Furthermore, considering that the applicability of most existing MCD schemes in large-scale satellite network environments is not yet clear, this invention systematically evaluates the proposed method in low-Earth orbit satellite constellation scenarios of different scales. Experimental results verify that this invention still possesses good performance and scalability under large-scale networks.

[0013] The multi-layer satellite network controller selection method described in this invention improves the network responsiveness of cross-domain services. A mathematical model aimed at reducing network latency and balancing controller load is established and solved using the SA algorithm. Furthermore, a multi-controller deployment algorithm based on the GCN model is proposed. This algorithm utilizes a neural network to learn the SA algorithm's solution process, achieving a fast and approximately optimal deployment strategy, thus providing technical support for the application of large-scale satellite networks. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the SDN-based hierarchical satellite network architecture in the satellite network controller deployment method described in this invention.

[0015] Figure 2 This is a diagram illustrating the process of a new stream request.

[0016] Figure 3 The diagrams show physical models of satellite constellations of different sizes; (a) is a physical model of a low-Earth orbit (LEO) satellite constellation with a size of 100, (b) is a physical model of a LEO satellite constellation with a size of 200, and (c) is a physical model of a LEO satellite constellation with a size of 400.

[0017] Figure 4 This is a comparison chart of average network response latency;

[0018] Figure 5 A comparison chart of load variance;

[0019] Figure 6 A comparison chart of the average load of the controllers;

[0020] Figure 7 A comparison chart of average network response latency under different satellite constellation scenarios;

[0021] Figure 8 A comparison chart of controller load variance under different satellite constellation scenarios;

[0022] Figure 9 This is a comparison chart of runtime for satellite constellations of different sizes. Detailed Implementation

[0023] Specific Implementation Method 1: Combination Figures 1 to 6 This embodiment describes an optimized deployment method for control nodes in a multi-layer satellite network based on Gaussian Convergence Nodes (GCN). This method constructs a two-layer control plane architecture in space and introduces an inter-layer cooperation mechanism to improve overall control efficiency. To address the high computational complexity of controller deployment in large-scale networks, this embodiment introduces a GCN model to solve for controller deployment schemes. This method uses the Optimal Array (SA) algorithm as a reference, learning its optimization process to effectively improve the solution speed and significantly reduce the computational overhead of deployment strategies. Furthermore, considering that the applicability of most existing Mid-Card Deployment (MCD) schemes in large-scale satellite network environments is not yet clear, this embodiment systematically evaluates the proposed method in low-Earth orbit (LEO) satellite constellation scenarios of different scales. Experimental results verify that the present invention still possesses good performance and scalability in large-scale networks.

[0024] This method, by setting the current time as t, solves for the controller deployment scheme at time t+1. The specific steps are as follows:

[0025] Step 1: Calculate the average network response delay at time t+1 based on the network response delay at time t; the network response delay includes flow setup delay and switch migration delay.

[0026] In this embodiment, the average flow setting delay calculation process is as follows:

[0027] When a new flow from the source switch does not match any rules, it sends a new flow processing request to the controller that matches the rule. The controller responds to the switch's request and installs the forwarding rules for the new flow. Subsequently, the switch updates its flow table and forwards the new flow to the destination switch. During this process, the following three scenarios may occur: Figure 2 As shown:

[0028] (1) Intra-domain flow request (flow f1 in the figure): When the source switch and the destination switch are located in the same control domain, the flow request process is an intra-domain flow request. At this time, the flow setup cost only includes the basic flow processing latency and does not involve information synchronization latency.

[0029] (2) Cross-control domains but belonging to the same super control domain (flow f2 in the figure): When the source switch and the destination switch do not belong to the same control domain, but their corresponding controllers are located in the same super control domain, the flow setup cost is increased by information synchronization delay on the basis of case (1). At this time, if there is a link between the corresponding controllers, information synchronization can be performed directly; otherwise, a cross-domain request needs to be sent to its super controller to complete the information synchronization.

[0030] (3) Crossing super control domains (flow f3 in the figure): When the controllers of the source switch and the destination switch are located in different super control domains, the information synchronization delay not only involves the synchronization between controllers, but also the synchronization delay between super controllers in order to obtain routing information from other control domains.

[0031] In this embodiment, a corresponding average flow setting delay formula is constructed based on the new flow request process.

[0032] In the first scenario: When a switch receives a new flow request, it needs to send a packet-in message to the controller. Subsequently, the controller sends installation rule F to the switches on the flow path within the domain. rule The average flow setting delay generated by this process is shown in equation (1).

[0033]

[0034] In the formula, Let i be the distance between satellite i and satellite j at time t+1; Indicates controller c j Whether it is in normal working condition at t+1, 0 indicates fault, 1 indicates normal; r is the link transmission rate; c is the speed of light; γ is the time it takes for the controller to process a packet-in message; F packet For packet-in messages;

[0035] Indicates the controller c at time t+1 j With switch s i The connections between nodes. There are only two connection relationships between nodes: 0 indicates no connection, and 1 indicates a connection. The connection relationships between nodes are shown below:

[0036]

[0037] The second scenario: If there are links between controllers, information synchronization is performed directly; otherwise, each controller will collect intra-domain network information F sync The information is sent to the super controller. Then, the super controller sends the synchronized global network information to the controller. The average flow setting delay generated in this process is shown in equation (2).

[0038]

[0039] In the formula, It is matrix Z t+1 A binary (0 or 1) variable. When At that time, controller c j and controller cj' There is no connection between them; otherwise, controller c j and controller c j' There are connections between them; For controller c j and controller c j' Distance at t+1; Indicates controller c j At time t+1, whether the system is in normal working condition is indicated by 0 for fault and 1 for normal. Indicates controller c j and super controller c k Distance at t+1. For time t+1, the super controller k and controller c j' The connection relationship between them; For controller c j' Is it in normal working condition at t+1? For the super controller c k With controller c j' Distance at time t+1; F sync For information within the domain network; For time t+1, the super controller c k With controller c j The connections between nodes; there are only two types of connections between nodes: 0 indicates no connection, and 1 indicates a connection. The connection relationships between nodes are shown below:

[0040]

[0041] The third scenario: The controllers of the source and destination switches need to send intra-domain network information F to the corresponding super controller. sync Since the super controllers are located on high-orbit satellites and are visible to each other, they can directly synchronize information. The average synchronization delay generated by this process is shown in equation (3).

[0042]

[0043] In the formula, For the super controller c k and c k' Distance at t+1; For the super controller c k' With controller c j' The connection relationship between them;

[0044] This invention selects the information synchronization method by setting control variables β1, β2 ∈ {0, 1}. Finally, the average flow is set with a delay, as shown in equation (4).

[0045]

[0046] In this embodiment, the average switch migration delay is calculated as follows:

[0047] Switch migration typically occurs in two situations: the controller controlling the switch is overloaded or fails. If a switch needs to be migrated at the start of time slot t+1, the original controller c... j Send migration information F to the switch to be migrated migra The average time delay generated by this process is shown in equation (5).

[0048]

[0049] Then, the controllers associated with the migrated switch need to synchronize information. If the original controller and the new controller belong to the same super control domain, there are two situations: when there is a link between the controllers, the original controller and the new controller can directly synchronize information without the participation of the super controller; when there is no link between the controllers, the original controller needs to synchronize information with the new controller through the super controller. If the two do not belong to the same super control domain, then information synchronization needs to be completed through their respective super controllers. The information synchronization delay calculation method in these three cases is the same as described above, and an additional control variable η∈{0,1} is set to select the information synchronization method. Finally, the average switch migration delay is shown in equation (6).

[0050]

[0051] Finally, the average network response delay is the sum of the average flow setup delay and the average switch migration delay, as shown in Equation (7).

[0052]

[0053] Step 2: Calculate the load on the controllers after the switch is migrated due to partial controller overload or failure at time t+1. It includes controller c j The load of the original switch and the load of the newly migrated switch are shown in Equation (8).

[0054]

[0055] in, Indicates switch s i The number of stream requests at time t.

[0056] The load variance is calculated similarly to the variance calculation formula. As shown in equation (9).

[0057]

[0058] Where σ is the number of fault controllers.

[0059] Step 3: Construct the objective function based on average network latency and load balancing between controllers, as shown in the following formula:

[0060]

[0061] Its constraints are:

[0062]

[0063] Among them, U j For controller c j The maximum processing capacity. The weight coefficients ω1 and ω2 of the objective function (10) satisfy ω1+ω2=1. Constraint (11) states that a switch can only be controlled by one controller, and a controller can only be controlled by a higher-level super controller. Constraint (12) ensures that the controller matched to the switch is in normal working condition. Constraint (13) states that the number of faulty controllers is less than the total number of controllers. Constraint (14) ensures that controller c j It has the capability to handle the total traffic requests sent by the switches it is connected to.

[0064] Step 4: Export the topology of 4,000 time slots from STK (Satellite Toolkit), and randomly generate 10 different loads in each time slot to obtain 40,000 sets of data. Use 80% of this data as input to train the GCN model; use 20% (8,000 sets) as the test set to test the trained GCN model. The specific implementation process is as follows:

[0065] Step 41: Before training the GCN model, the following input data needs to be prepared: visibility matrix, edge feature matrix (i.e., propagation delay matrix), and node feature vectors.

[0066] By simulating the satellite network using the STK tool, the visibility matrix and range matrix for each time slot can be derived. The visibility matrix A at time t is... t Sum distance matrix D t This represents the visibility and distance between satellite nodes at time t+1, and their specific representations are as follows:

[0067] Visibility matrix A t :

[0068]

[0069] Where p is the total number of controllers and switches in the satellite network, i.e., p = m1 + m2 + n. ijThis represents the communication relationship between node i and node j (satellite i and satellite j). If bidirectional communication exists between the two nodes, then a ij =1. Otherwise, a ij =0.

[0070] Distance matrix D t =[d ij ] p×p With visibility matrix A t Same type. Among them, d ij Represents the distance between node i and node j, if a ij =0, then d ij =0.

[0071] Propagation delay matrix PL t It is based on the distance matrix D t Dividing by the speed of light c, we get the following result:

[0072]

[0073] Node load is selected as a node feature, and a p-dimensional node feature vector LN is constructed as shown below. t+1 :

[0074]

[0075] Among them, ln0, ln i ln p-1 These are randomly generated node load values;

[0076] To fully utilize the two types of edge weight information, the propagation delay matrix and the visibility matrix, a GCN model with four convolutional layers is adopted. The first two layers are used to process the edge weights of the propagation delay matrix, and the last two layers are used to process the edge weights of the visibility matrix.

[0077] Step 42: Based on the multi-objective function (10), the SA algorithm is used to generate the optimal deployment scheme of the controller for the input data as the real label, forming a training dataset of 32,000 groups;

[0078] The actual label is a p×p relation matrix. To more intuitively present the controller deployment, this relation matrix has been simplified, retaining only the correspondence between switches and controllers, and reorganized into an n×m relation matrix R as shown below, where m = m1 + m2. In relation matrix R, each row labels a controller corresponding to a switch.

[0079]

[0080] Where, r ij Indicates switch si The relationship with the controller c j If r ij = 0, it means that the switch s i is not controlled by the controller c j If r ij = 1, it means that the switch s i is controlled by the controller c j .

[0081] Step 43: Initialize epoch = 200 and count = 0;

[0082] Step 44: If count < epoch, execute Step 45; otherwise, save the trained GCN model and execute Step 48;

[0083] Step 45: Execute four convolution operations in sequence to gradually achieve message passing and feature fusion between layers. The convolution operation of each layer is as follows:

[0084]

[0085] In the first two convolution operations, I is the identity matrix, is the adjacency matrix; in the last two convolution operations, is 's degree matrix, PL t is the propagation delay matrix; H l is the feature matrix of the l-th layer (the 0-th layer is the input p-dimensional feature vector LN t+1 ); W l is the weight of the l-th layer, which is adjusted by the neural network itself during the training process; σ() is the non-linear activation function, and the ReLU function is adopted in this invention. Step 46: Use the cross-entropy loss function (CrossEntropy Loss) to calculate the log-likelihood difference between the predicted value and the true value, then calculate the gradient of the model parameters based on this error, and use these gradients to update the weights of the model.

[0086] Step 47: count = count + 1, execute Step 44;

[0087] Step 48: Use the test set and use the SA algorithm to generate the corresponding switch-controller relationship matrix as the true label. Load the trained GCN model, predict the relationship matrix, calculate the accuracy rate and save the result. If the controller allocation accuracy rate reaches more than 85%, execute Step 5; otherwise, after adjusting the parameters (such as the hidden layer dimension and learning rate, etc.), return to Step 43 to retrain.

[0088] Step 5: Export 4,000 time-slot topologies of low-Earth orbit satellite constellations with sizes of 100, 200, and 400 from STK. Use the trained GCN model to infer the controller deployment scheme, output and save the optimal controller deployment strategy.

[0089] Specific Implementation Method Two: Combination Figures 1 to 9 This embodiment describes a verification example of the optimized deployment method for multi-layer satellite network control nodes based on GCN described in Specific Embodiment 1:

[0090] This method is based on a multi-layered satellite network architecture using SDN. For example... Figure 1 As shown, in this architecture, the data plane consists of satellites at the lowest orbital altitude (usually LEO satellites), responsible for data forwarding and processing. Due to the limited coverage of satellites, a multi-layered control plane composed of satellites at different altitudes can be constructed to achieve global control. Taking a two-layer control plane as an example, the first layer of satellite nodes closest to the switching node forms the first control plane, where each satellite acts as a regular controller. Several controllers can form a super control domain, managed by satellite nodes at a higher level. The highest layer of satellite nodes forms the second control plane, where each satellite acts as a super controller.

[0091] To comprehensively evaluate the performance of this embodiment, performance analysis was first conducted in a three-layer satellite network with a LEO satellite constellation size of 100, focusing on average flow setting latency and controller load balancing. To further verify the scalability of the invention, i.e., its applicability in large-scale satellite networks, the LEO satellite constellation size was gradually expanded to 200 and 400 satellites. The performance of each scheme was compared and evaluated in satellite networks with LEO numbers of 100, 200, and 400, respectively. Schematic diagrams of the physical models for satellite constellations of different sizes are shown below. Figure 3 As shown in (a)-(c).

[0092] When setting up the comparison scheme, the scheme of the present invention is compared with Dynamic and static controller placement in software-defined satellite networking (ASPO) and On-Demand Dynamic Controller Placement in Software Defined Satellite-Terrestrial Networking (ODAA).

[0093] ASPO addresses the MCD problem within a three-layer network architecture based on SDSN. The controller can be deployed in LEO, MEO, or GEO, while switches are primarily distributed in LEO. Optimization targets include latency, load, and fault tolerance, and an accelerated particle swarm optimization algorithm is used to solve the MCD model.

[0094] ODAA proposes a redundancy-based LEO satellite subnet partitioning method under a LEO / GEO two-layer network architecture to meet terminal coverage requirements. The number and location of controllers in the LEO satellite subnet are determined by optimizing network response latency, and an on-demand dynamic approximation algorithm is proposed to obtain an approximate solution.

[0095] like Figure 4 As shown, this invention outperforms the other two schemes in terms of average flow setting latency. Specifically, the SA scheme achieves 7.94% lower average flow setting latency than ASPO and 5.95% lower latency than ODAA. This result is attributed to incorporating the interaction between controllers at different layers when handling new flow requests across control domains, providing two path options: communication between controllers and communication between the super controller and the controller. In contrast, ASPO and ODAA do not consider the communication path selection problem under cross-domain requests when constructing their latency optimization models. This invention is close to the SA scheme, with an average latency 0.82% higher. This demonstrates that the GCN model possesses the ability to obtain the optimal approximate solution.

[0096] like Figure 5 As shown, over five consecutive time slots, this invention demonstrates better load balancing compared to the other two schemes. The SA scheme is on average 49.82% lower than the ASPO scheme and 73.21% lower than the ODAA scheme. ODAA performs the worst because its optimization model lacks load constraints, potentially leading to controller overload. While ASPO constrains the load, it doesn't consider load balancing. This invention overcomes the shortcomings of ASPO in controller placement by coordinating the super controller and controllers and optimizing load balancing based on the number of new flow requests to the switch. This invention is close to the SA scheme, but on average 17.7% higher.

[0097] While ensuring load balancing, it is also necessary to closely monitor the load status of each controller to prevent overload or high load issues. For example... Figure 6 As shown, this embodiment compares and analyzes the average load of each controller under different deployment schemes during a monitoring period of 15 time slots. Controllers 4, 7, and 10 in the ODAA scheme are under high load. Compared with the ASPO and ODAA schemes, the load distribution among controllers in the SA scheme is more uniform. This invention is similar to the SA scheme, showing relatively uniform load distribution.

[0098] To further verify that the present invention still has good performance in large-scale satellite constellation scenarios, the performance of each scheme was evaluated in low-Earth orbit satellite constellation scenarios of different scales.

[0099] like Figure 7 As shown, the average network response latency of this invention is consistently superior to the other two schemes in satellite constellation scenarios of different scales. The SA scheme is on average 5.39% lower than the ASPO scheme and on average 3.79% lower than the ODAA scheme. Meanwhile, this invention is close to the SA scheme, and on average 1.41% higher than the SA algorithm.

[0100] like Figure 8 As shown, compared to the other two schemes, this invention exhibits better load balancing across satellite constellation scenarios of varying scales. The SA scheme is on average 99.08% lower than the ASPO scheme and 99.45% lower than the ODAA scheme. Simultaneously, the GCN model achieves load balancing performance close to that of the SA scheme, approximately 1.5 times that of the SA scheme, further validating its good stability while maintaining performance.

[0101] Finally, to verify that it is possible to quickly output a near-optimal multi-controller deployment scheme, the running time of each scheme was compared in satellite networks of different sizes. Figure 9 As shown, the runtime of this invention is the shortest, averaging 96.28% lower than the SA scheme, 94.72% lower than the ASPO scheme, and 92.78% lower than the ODAA scheme. This is because the trained GCN model no longer requires multiple iterations to converge to the optimal solution, and can achieve parallel computation using the computing power of GPUs. In contrast, the SA, ASPO, and ODAA schemes all involve multiple iterations in the solution process; therefore, the larger the satellite network scale, the more the advantages of the GCN scheme become apparent.

[0102] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN, characterized by: This method solves the deployment plan of the controller at time t+1 through the current time t. This method is implemented by the following steps: Step 1: Calculate the average network response delay dl at time t+1 based on the network response delay at time t. t+1 The average network response delay dl t+1 Set the sum of the average flow delay and the average switch migration delay; Step 2: Calculate the controller load after the switch migration caused by controller overload or failure at time t+1. The load includes controller c j Calculate the load on the existing switch and the load on the newly migrated switch, and calculate the load variance. Step 3: Construct an objective function based on the average network response delay obtained in Step 1 and the load variance between the controllers obtained in Step 2, and set the constraint conditions; Step 4: Input the obtained time-slot topology data into the GCN model for training to obtain the trained GCN model; Step 5: Use the trained GCN model to derive the controller deployment plan, and output and save the optimal controller deployment strategy.

2. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 1, characterized in that: In Step 1, the calculation of the average flow setup delay is divided into three cases: intra-domain flow requests, flow requests across control domains but belonging to the same super control domain, and flow requests across super control domains; In the first scenario, when a switch receives a new flow request, it needs to send a packet-in message to the controller. Subsequently, the controller sends installation rule F to the switches on the flow path within the domain. rule The average stream delay for message generation is expressed by the following formula: In the formula, Let i be the distance between satellite i and satellite j at time t+1; Indicates controller c j At time t+1, whether the system is in normal working condition is indicated by 0 for fault and 1 for normal. r is the link transmission rate; c is the speed of light; γ is the time it takes for the controller to process one packet-in message; F packet For packet-in messages; For controller c at time t+1 j With switch s i The connection relationship between them; In the second scenario, if there are links between controllers, information synchronization is performed directly; otherwise, each controller will collect intra-domain network information F. sync The information is sent to the super controller; then, the super controller sends the synchronized global network information to the controller to set the average flow delay, expressed as follows: In the formula, It is matrix Z t+1 A binary (0 or 1) variable; when At that time, controller c j and controller c j' There is no connection between them; otherwise, controller c j and controller c j' There are connections between them; For controller c j and controller c j' Distance at t+1; Indicates controller c j At time t+1, whether the system is in normal working condition is indicated by 0 for fault and 1 for normal. Indicates controller c j and super controller c k Distance at t+1. For time t+1, the super controller k and controller c j' The connection relationship between them; For controller c j' Is it in normal working condition at t+1? For the super controller c k With controller c j' Distance at time t+1; F sync For information within the domain network; For time t+1, the super controller c k With controller c j The connection relationship between them; In the third scenario, the controllers of both the source and destination switches need to send intra-domain network information F to the corresponding super controller. sync Information synchronization is performed directly; the average stream setting delay is expressed by the following formula: In the formula, For the super controller c k and c k' Distance at t+1; For the super controller c k' With controller c j' The connection relationship between them; Set control variables β1, β2 ∈ {0, 1} to select the information synchronization method. The final average flow setup delay is expressed by the following formula:

3. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 2, characterized in that: In Step 1, the calculation process of the average switch migration delay is as follows: When the switch controller is overloaded or malfunctions, at the beginning of time slot t+1, if the switch needs to be migrated, the original controller sends migration information F to the switch to be migrated. migra The resulting average switch migration delay Expressed as follows: Then, the controllers related to the migrated switches need to synchronize information.

4. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 3, characterized in that: Judge whether the original controller and the new controller belong to the same super control domain. If so, there are two cases: when there is a link connection between the controllers, the original controller and the new controller directly synchronize information; when there is no link connection between the controllers, the original controller needs to synchronize information with the new controller through the super controller; Otherwise, the synchronization information needs to be completed by the super controllers to which the original controller and the new controller belong respectively; Set control variable η ∈ {0, 1} to select the information synchronization method; finally, the average switch migration delay is expressed by the following formula:

5. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 4, characterized in that: The final average network response delay is the sum of the average flow setup delay and the average switch migration delay, and is expressed by the following formula:

6. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 5, characterized in that: In step two, calculate the controller load after the switch migration caused by controller overload or failure at time t+1. Including controller c j The load on the existing switch and the load on the newly migrated switch can be expressed by the following formula: In the formula, Indicates switch s i The number of stream requests at time t; Calculate load variance This can be expressed as follows: In the formula, σ is the number of faulty controllers.

7. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 6, characterized in that: In Step 3, the multi-objective function and constraint conditions are as follows: Multi-objective function: Among them, the weight coefficients ω1 and ω2 satisfy ω1+ω2=1; Constraint 1: Constraint 2: Constraint condition 3: σ < m1 Constraint 4: In the formula, U j For controller c j The maximum processing capacity, constraint 1 means that a switch can only be controlled by one controller, and a controller can only be controlled by one higher-level controller; constraint 2 ensures that the controller matched to the switch is in normal working condition; constraint 3 means that the number of faulty controllers is less than the total number of controllers; constraint 4 ensures that controller c j It has the capability to handle the total traffic requests sent by the switches it is connected to.

8. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 7, characterized in that: In Step 4, the process of training the GCN model is as follows: Step 4-1: Divide the load data into a training set and a test set. The training set is used as input data. Based on the multi-objective function, use the SA algorithm to generate the optimal deployment plan of the controller for the input data as the true label; Step 4-2: Initialize the GCN model, the number of training epochs epoch = 200, and the number of iterations count = 0; Step 4-3: If count < epoch, execute Step 4-4; otherwise, save the trained GCN model and execute Step 4-6; Step 4-4: Execute four convolution operations to gradually achieve message passing and feature fusion between layers; Use the cross-entropy loss function to calculate the logarithmic likelihood difference between the predicted value and the true value, and calculate the gradient of the GCN model parameters based on the likelihood difference to update the weights of the model; Step 4-5: count = count + 1, and return to Step 4-3; Step 4-6: Combine the test set and use the SA algorithm to generate the corresponding switch-controller relationship matrix as the true label; test the trained GCN model and output the controller deployment plan.

9. The method for optimizing the deployment of control nodes in a multi-layer satellite network based on GCN according to claim 1, characterized in that: In Step 4-1, before training the GCN model, the input data is processed as follows: The satellite network was simulated using the STK tool, and the visibility matrix and distance matrix for each time slot were exported; the propagation delay matrix was then calculated based on the distance matrix. Node load is selected as node feature, and node feature vectors are constructed. The first two layers of the GCN model are used to process the edge weights of the propagation delay matrix, and the last two layers are used to process the edge weights of the visibility matrix.

Images