Satellite network federated learning method and system based on data similarity grouping

Abstract

The invention discloses a satellite network federated learning method and system based on data similarity grouping, and the method comprises the steps: initializing system key parameters and a data structure for a satellite network system of each orbit; selecting a satellite node as a coordinator, traversing the satellite node as a source node, and grouping according to the similarity between the nodes; taking the source node of each group as a representative node, executing local training, and selecting an adjacent node as a new representative node to continue alternate training; after the iteration training is completed, the coordinator is dominated to simultaneously superpose the parameters representing the nodes towards the coordinator in two directions to complete aggregation; and transmitting the model parameters after parameter aggregation to a satellite node, and performing the next round of cycle training. Through the technical scheme of the invention, the reasonability and adaptability of grouping are ensured, the dynamic balance of the intra-group calculation load is realized, the reliability of aggregation is ensured, the efficiency is improved, and the calculation and communication overhead is obviously reduced.

Description

Technical Field

[0001] The present invention relates to the field of satellite communication technology, and in particular to a satellite network federated learning method based on data similarity grouping and a satellite network federated learning system based on data similarity grouping. Background Art

[0002] With the rapid development of satellite Internet, satellite network security issues are becoming increasingly prominent. Cyber ​​attackers may disrupt the normal operation of satellite networks by launching malicious behaviors such as DDoS attacks and data theft, which seriously threatens the security of space infrastructure. Therefore, it is of great significance to develop an efficient satellite network malicious traffic detection system.

[0003] At present, machine learning, especially deep learning technology, has achieved remarkable results in the field of network traffic anomaly detection. However, directly applying traditional centralized machine learning methods to satellite networks faces the following challenges:

[0004] 1. Data privacy issues: Satellite network traffic data may involve sensitive information and is not suitable for direct transmission to the central server for training.

[0005] 2. High communication overhead: The communication bandwidth between satellite nodes is limited, and frequent transmission of large amounts of training data will result in huge communication overhead.

[0006] 3. Limited computing resources: The computing power and energy supply of satellite nodes are very limited, making it difficult to support local training of complex models.

[0007] 4. Uneven data distribution: The distribution of traffic data collected by satellite nodes in different orbits and locations may vary greatly.

[0008] As a new distributed machine learning paradigm, federated learning solves the problem of data privacy protection by training the model on the data side. However, the existing federated learning solutions still have the following shortcomings when applied to satellite networks:

[0009] 1. Ignoring the similarity of adjacent satellite data: Adjacent satellites in the same orbit cover similar areas, so the data collected often have strong correlation, a feature that has not been fully utilized.

[0010] 2. Low training efficiency: All nodes participate in each round of training, resulting in a waste of computing resources and increasing unnecessary communication overhead.

[0011] 3. Model aggregation has high overhead: Global model aggregation requires frequent inter-satellite communications and consumes a lot of energy.

[0012] 4. Unbalanced load distribution: The differences in computing power and energy levels of satellite nodes are not taken into account, which may cause some nodes to be overloaded. Summary of the invention

[0013] In view of the above problems, the present invention provides a satellite network federated learning method and system based on data similarity grouping. Through a dynamic grouping strategy based on distance and similarity, satellite nodes with similar data features are divided into a group and share the model, which not only reduces redundant training, but also ensures the rationality and adaptability of the grouping by limiting the grouping range and dynamically adjusting the similarity threshold. By adopting a representative node rotation strategy for model training, the dynamic balance of the computing load within the group is achieved, and the waste of resources caused by the need for all nodes to participate in training in traditional federated learning is avoided. Through the coordinator node, the parameter transmission path is reasonably planned, and the parallel aggregation of model parameters is achieved, which not only ensures the reliability of the aggregation, but also improves the efficiency through parallel transmission, effectively solving the problem that the chain transmission in the traditional method is susceptible to single point failure. Through the above-mentioned complete closed-loop adaptive cyclic training process from the initial dynamic grouping, representative node rotation training to parallel aggregation, model update and propagation, the entire system can be continuously optimized, while ensuring the model performance, significantly reducing the computing and communication overhead.

[0014] To achieve the above object, the present invention provides a satellite network federated learning method based on data similarity grouping, comprising:

[0015] Initialize key system parameters and data structures for each orbital satellite network system;

[0016] Select one of the satellite nodes in the track as a coordinator, traverse the satellite nodes as source nodes, and group them according to the similarity between the nodes;

[0017] The source node of each group is used as the representative node to perform the preset rounds of local training, and after the training is completed, the adjacent node is selected as the new representative node to continue the rotation training;

[0018] After completing the preset rounds of iterative training, the coordinator initiates model aggregation in two directions at the same time, and superimposes the parameters of the representative nodes in the direction of the coordinator to complete the aggregation;

[0019] The model parameters of the global model formed after parameter aggregation are transmitted to the satellite nodes in the orbit, and the next round of cyclic training is carried out.

[0020] In the above technical solution, preferably, the step of selecting one of the satellite nodes in the orbit as a coordinator, traversing the satellite nodes as source nodes, and grouping the nodes according to the similarity between the nodes includes:

[0021] Select one of the satellite nodes in the orbit as a coordinator, the coordinator uses an unvisited node as a source node, and traverses the satellite nodes that meet a preset distance condition in the same orbit in a clockwise direction;

[0022] If the satellite node has not been visited, determine whether the similarity between the satellite node and the source node reaches a preset similarity threshold, and if so, place the satellite node in the group where the source node is located;

[0023] If the satellite node has been accessed by other groups, determine whether the satellite node meets the group reallocation condition, and if so, adjust and allocate the original group to the group where the source node is located;

[0024] If the neighboring nodes of the source node are not assigned to the current group, stop investigating the more distant nodes in this direction.

[0025] In the above technical solution, preferably, the satellite network federated learning method based on data similarity grouping further includes:

[0026] The number of nodes in the current group is monitored. If the number of nodes in the current group reaches the maximum number of nodes in the preset group, the preset similarity threshold is modified to the minimum similarity in the current group.

[0027] In the above technical solution, preferably, the source node of each group is used as a representative node to perform a preset round of local training, and after the training is completed, an adjacent node is selected as a new representative node to continue the rotation training. The specific process includes:

[0028] Each group initially selects a source node as a representative node, which performs a preset round of local model training and sends the model parameters to other satellite nodes in the current group after the training is completed;

[0029] Select the next satellite node in the clockwise direction as the new representative node. If the next satellite node belongs to another group, select the first satellite node in the counterclockwise direction of the current group as the new representative node.

[0030] After other satellite nodes in the current group receive the model parameters of the representative node, they are directly used for prediction without the need for training.

[0031] In the above technical solution, preferably, after completing the preset rounds of iterative training, the coordinator initiates model aggregation in two directions at the same time, and superimposes the parameters of the representative nodes in the direction of the coordinator to complete the aggregation. The specific process includes:

[0032] After completing the preset rounds of iterative training, the coordinator initiates model aggregation in both clockwise and counterclockwise directions;

[0033] Send an aggregation request to the representative node of the group at a distance of n / 2 in the clockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the counterclockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0034] Send an aggregation request to the representative node of the group at a distance of n-1-(n / 2) in the counterclockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the clockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0035] The coordinator receives the aggregation parameters from two directions to complete the aggregation of parameters.

[0036] The present invention also proposes a satellite network federated learning system based on data similarity grouping, which applies the satellite network federated learning method based on data similarity grouping disclosed in any one of the above technical solutions, including:

[0037] The system initialization module is used to initialize key system parameters and data structures for the satellite network system in each orbit;

[0038] A similarity grouping module is used to select one of the satellite nodes in the orbit as a coordinator, traverse the satellite nodes as source nodes, and group them according to the similarity between the nodes;

[0039] A rotation training module is used to perform a preset round of local training using the source node of each group as a representative node, and select an adjacent node as a new representative node to continue the rotation training after the training is completed;

[0040] The model aggregation module is used to initiate model aggregation in two directions at the same time under the leadership of the coordinator after completing the preset rounds of iterative training, and superimpose the parameters of the representative nodes in the direction of the coordinator to complete the aggregation;

[0041] The global update module is used to transfer the model parameters of the global model formed by parameter aggregation to the satellite nodes in the orbit and perform the next round of cyclic training.

[0042] In the above technical solution, preferably, the similarity grouping is specifically used for:

[0043] Select one of the satellite nodes in the orbit as a coordinator, the coordinator uses an unvisited node as a source node, and traverses the satellite nodes that meet a preset distance condition in the same orbit in a clockwise direction;

[0044] If the satellite node has not been visited, determine whether the similarity between the satellite node and the source node reaches a preset similarity threshold, and if so, place the satellite node in the group where the source node is located;

[0045] If the satellite node has been accessed by other groups, determine whether the satellite node meets the group reallocation condition, and if so, adjust and allocate the original group to the group where the source node is located;

[0046] If the neighboring nodes of the source node are not assigned to the current group, stop investigating the more distant nodes in this direction.

[0047] In the above technical solution, preferably, the similarity grouping is also used for:

[0048] The number of nodes in the current group is monitored. If the number of nodes in the current group reaches the maximum number of nodes in the preset group, the preset similarity threshold is modified to the minimum similarity in the current group.

[0049] In the above technical solution, preferably, the rotation training module is specifically used for:

[0050] Each group initially selects a source node as a representative node, which performs a preset round of local model training and sends the model parameters to other satellite nodes in the current group after the training is completed;

[0051] Select the next satellite node in the clockwise direction as the new representative node. If the next satellite node belongs to another group, select the first satellite node in the counterclockwise direction of the current group as the new representative node.

[0052] After other satellite nodes in the current group receive the model parameters of the representative node, they are directly used for prediction without the need for training.

[0053] In the above technical solution, preferably, the model aggregation module is specifically used to:

[0054] After completing the preset rounds of iterative training, the coordinator initiates model aggregation in both clockwise and counterclockwise directions;

[0055] Send an aggregation request to the representative node of the group at a distance of n / 2 in the clockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the counterclockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0056] Send an aggregation request to the representative node of the group at a distance of n-1-(n / 2) in the counterclockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the clockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0057] The coordinator receives the aggregation parameters from two directions to complete the aggregation of parameters.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) The present invention makes full use of the similarity characteristics of adjacent satellite data in the same orbit. Through a dynamic grouping strategy based on distance and similarity, satellites with similar data characteristics are grouped together and share models. This grouping method not only reduces redundant training, but also ensures the rationality and adaptability of grouping by limiting the grouping range and dynamically adjusting the similarity threshold.

[0060] (2) The present invention innovatively designs a representative node rotation mechanism. Each group is only responsible for training by one representative node, and the rotation is performed according to preset rules. Other nodes directly use the trained model. This mechanism not only achieves dynamic balancing of the computing load within the group, but also avoids the waste of resources caused by the need for all nodes to participate in training in traditional federated learning.

[0061] (3) The present invention designs an efficient parallel aggregation mechanism. The coordinator node rationally plans the parameter transmission path to achieve parallel aggregation of model parameters. This design not only ensures the reliability of aggregation, but also improves efficiency through parallel transmission, effectively solving the problem that chain transmission in traditional methods is susceptible to single point failure.

[0062] (4) The present invention provides a complete closed-loop training process. From the initial dynamic grouping, representative node rotation training to parallel aggregation, model update and propagation, an adaptive cycle is formed. This design enables the entire system to be continuously optimized, significantly reducing computing and communication overhead while ensuring model performance.

[0063] Through the above innovative design, the present invention has effectively solved the key problems faced by federated learning in satellite networks, such as resource limitation, load balancing and communication efficiency, and provided a new solution for distributed machine learning in similar scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A process of a satellite network federated learning method based on data similarity grouping disclosed in an embodiment of the present invention;

[0065] Figure 2 A schematic diagram of the initial state of a satellite node disclosed in an embodiment of the present invention;

[0066] Figure 3 A schematic diagram of the selection and traversal sequence of a coordinator node disclosed in an embodiment of the present invention;

[0067] Figure 4 A schematic diagram of source node neighborhood traversal disclosed in an embodiment of the present invention;

[0068] Figure 5 A schematic diagram of grouping a satellite node for the first visit disclosed in an embodiment of the present invention;

[0069] Figure 6 A schematic diagram of reallocation of visited satellite nodes disclosed in an embodiment of the present invention;

[0070] Figure 7 A schematic diagram of a group termination condition disclosed in an embodiment of the present invention;

[0071] Figure 8 A schematic diagram of adjusting the threshold of dynamic similarity disclosed in an embodiment of the present invention;

[0072] Fig. 9 A schematic diagram of rotation training of representative nodes disclosed in an embodiment of the present invention;

[0073] Fig.10 A schematic diagram of parallel aggregation of model parameters disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0076] like Figure 1 As shown, a satellite network federated learning method based on data similarity grouping provided by the present invention includes:

[0077] Initialize key system parameters and data structures for each orbital satellite network system;

[0078] Select one of the satellite nodes in the orbit as the coordinator, traverse the satellite nodes as source nodes, and group them according to the similarity between the nodes;

[0079] The source node of each group is used as the representative node to perform the preset rounds of local training, and after the training is completed, the adjacent node is selected as the new representative node to continue the rotation training;

[0080] After completing the preset rounds of iterative training, the coordinator initiates model aggregation in two directions at the same time, and superimposes the parameters of the representative nodes in the direction of the coordinator to complete the aggregation;

[0081] The model parameters of the global model formed after parameter aggregation are transmitted to the satellite nodes in orbit, and the next round of cyclic training is carried out.

[0082] In this implementation, satellite nodes with similar data features are divided into a group and share the model through a dynamic grouping strategy based on distance and similarity, which not only reduces redundant training, but also ensures the rationality and adaptability of the grouping by limiting the grouping range and dynamically adjusting the similarity threshold. By adopting a representative node rotation strategy for model training, dynamic balancing of the computing load within the group is achieved, and the waste of resources caused by the need for all nodes to participate in training in traditional federated learning is avoided. The coordinator node reasonably plans the parameter transfer path to achieve parallel aggregation of model parameters, which not only ensures the reliability of aggregation, but also improves efficiency through parallel transmission, effectively solving the problem that chain transmission in traditional methods is susceptible to single point failures. Through the above-mentioned complete closed-loop adaptive cyclic training process from the initial dynamic grouping, representative node rotation training to parallel aggregation, model update and propagation, the entire system can be continuously optimized, while ensuring model performance, significantly reducing computing and communication overheads.

[0083] Specifically, considering that adjacent satellites in the same orbit may collect similar traffic data, and the computing and energy resources of satellite nodes are limited, the present invention designs a data similarity grouping strategy and a representative node rotation mechanism to optimize the federated learning training process in satellite networks. This method is expected to reduce computing resource consumption and communication overhead while maintaining the detection performance of the model.

[0084] Specifically, the nodes are first grouped based on the data similarity between them. By setting reasonable distance and similarity thresholds, the coordinator node is responsible for organizing and managing the grouping process. Satellite nodes with similar data will be divided into the same group, and the dynamic threshold adjustment mechanism will be used to adapt to changes in the network environment. This similarity-based grouping strategy can effectively reduce redundant training. By introducing the maximum distance threshold and dynamic similarity threshold mechanism, satellite node grouping based on data similarity is realized. This strategy calculates the cosine similarity based on the model parameters obtained by node training, and makes full use of the similarity characteristics of adjacent satellite data in the same orbit, so that satellites with similar data are preferentially divided into the same group and share the same model, thereby reducing training overhead and improving system efficiency.

[0085] After grouping, the representative node rotation strategy is used for model training. Each group selects a representative node to be responsible for model training, and after training, the model is distributed to other nodes in the group for use. Based on the clockwise rotation representative node selection strategy, the representative node is responsible for completing the training of the model in the group and distributing the trained model to other satellite nodes in the group for use. The representative nodes rotate according to the preset rules to ensure balanced computing load within the group. This design fully considers the limited resources of satellite nodes and avoids the problem of excessive resource consumption by a single node. Other nodes in the group do not need to perform model training, and directly use the model trained by the representative node for prediction, which significantly reduces the computing overhead.

[0086] In order to further optimize the communication overhead, the present invention implements an efficient model aggregation mechanism and constructs a two-level aggregation architecture of global aggregation within the orbit and aggregation between ground stations. Based on the grouping information maintained by the coordinator node, the representative node of each group passes the model parameters to the coordinator node for global aggregation within the orbit. By rationally planning the parameter transfer path, the parallel aggregation of the model parameters is realized, and the aggregation results of each orbit are completed by the ground station to complete the final global model update. This design not only ensures the aggregation efficiency, but also provides better fault tolerance. Finally, the update of the global model is completed through the ground station, realizing the closed loop of the entire training process, effectively reducing the inter-satellite communication overhead and improving the overall efficiency of the system.

[0087] In the above implementation, initializing and setting key system parameters includes:

[0088] The maximum distance threshold (max_distance) is used to limit the neighborhood range considered by the source node;

[0089] Maximum group size (max_group_size), which controls the maximum number of nodes allowed in each group;

[0090] Similarity threshold array (similarity_threshold), used to control the similarity requirements for nodes to join a group;

[0091] Maximum group size threshold (max_group_size_threshold), used to trigger similarity threshold adjustment;

[0092] The number of training rounds (m) indicates the number of iterations required for one round of federated learning.

[0093] Initialize the data structure:

[0094] group array: group[i] = i means each node is initially a separate group;

[0095] Visited array: visited[i] = false means node i has not been visited;

[0096] similarity_threshold array: The initial value is set to 0.8.

[0097] In the above implementation, preferably, one of the satellite nodes in the orbit is selected as the coordinator, the satellite nodes are traversed as source nodes, and the nodes are grouped according to the similarity between them. The specific process includes:

[0098] Select one of the satellite nodes in the orbit as the coordinator. The coordinator takes the unvisited node i (visited[i]=false) as the source node, marks visited[i]=true, and traverses the satellite nodes j in the same orbit that meet the preset distance condition (|ij|≤max_distance) in a clockwise direction.

[0099] If satellite node j has not been visited (visited[j] = false), then determine whether the similarity similarity(i, j) between the satellite node and the source node reaches the preset similarity threshold similarity_threshold[i]. If so, node j is placed in the group where source node i is located.

[0100] If satellite node j has been visited by other groups (visited[j] = true), then determine whether the satellite node meets the group reallocation conditions similarity(i,j)≥similarity_threshold[i] and similarity(i,j)>similarity(group[j],j). If so, adjust node j from the original group to the group where source node i is located.

[0101] If the neighboring nodes of the source node are not assigned to the current group, stop investigating the more distant nodes in this direction.

[0102] In the above implementation, preferably, the satellite network federated learning method based on data similarity grouping further includes:

[0103] The number of nodes in the current group is monitored. If the number of nodes in the current group reaches the maximum number of nodes in the preset group, the preset similarity threshold is modified to the minimum similarity in the current group.

[0104] In the above implementation, preferably, the source node of each group is used as a representative node to perform a preset round of local training, and after the training is completed, an adjacent node is selected as a new representative node to continue the rotation training. The specific process includes:

[0105] Each group initially selects a source node as a representative node, which performs local model training for a preset round m / max_group_size and sends the model parameters to other satellite nodes in the current group after the training is completed. Ordinary nodes in the group use the received model directly for prediction without training.

[0106] Select the next adjacent satellite node in the clockwise direction as the new representative node. If the next satellite node belongs to another group, select the first satellite node in the counterclockwise direction of the current group as the new representative node;

[0107] After other satellite nodes in the current group receive the model parameters of the representative node, they are directly used for prediction without training.

[0108] In the above implementation, preferably, after completing the preset rounds of iterative training, the coordinator initiates model aggregation in two directions at the same time, and superimposes the parameters of the representative nodes in the direction of the coordinator to complete the aggregation. The specific process includes:

[0109] After completing the preset rounds (m rounds) of iterative training, the coordinator initiates model aggregation in both clockwise and counterclockwise directions. Based on the maintained grouping information, the total number of groups n in the track can be known.

[0110] Send an aggregation request to the representative node of the group at a distance of n / 2 in the clockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the counterclockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0111] Send an aggregation request to the representative node of the group at a distance of n-1-(n / 2) in the counterclockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the clockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0112] The coordinator receives aggregation parameters from both directions to complete parameter aggregation.

[0113] In the above implementation, each ground station receives the aggregation parameters of the orbit it is responsible for and aggregates them to form a new version of the global model. When the ground station is connected to the corresponding orbital satellite, it passes the new model parameters to a satellite node in the orbit. The node acts as a new "coordinator" and is responsible for propagating the new parameters in the orbit and subsequent grouping and aggregation tasks. When all satellites in the orbit receive the new parameters, they enter the next round of training cycle and repeat the entire process of grouping, training, and aggregation to continuously optimize the model performance.

[0114] The present invention further proposes a satellite network federated learning system based on data similarity grouping, which applies the satellite network federated learning method based on data similarity grouping disclosed in any one of the above embodiments, including:

[0115] The system initialization module is used to initialize key system parameters and data structures for the satellite network system in each orbit;

[0116] The similarity grouping module is used to select one of the satellite nodes in the orbit as the coordinator, traverse the satellite nodes as source nodes, and group them according to the similarity between the nodes;

[0117] A rotation training module is used to perform a preset round of local training using the source node of each group as a representative node, and select an adjacent node as a new representative node to continue the rotation training after the training is completed;

[0118] The model aggregation module is used to initiate model aggregation in two directions at the same time after completing the preset rounds of iterative training, and superimpose the parameters of the representative nodes in the direction of the coordinator to complete the aggregation;

[0119] The global update module is used to transfer the model parameters of the global model formed by parameter aggregation to the satellite nodes in orbit and carry out the next round of cyclic training.

[0120] In the above implementation, preferably, similarity grouping is specifically used for:

[0121] Select one of the satellite nodes in the orbit as the coordinator. The coordinator uses the unvisited node as the source node and traverses the satellite nodes that meet the preset distance conditions in the same orbit in a clockwise direction.

[0122] If the satellite node has not been visited, it is determined whether the similarity between the satellite node and the source node reaches a preset similarity threshold. If so, the satellite node is included in the group where the source node is located.

[0123] If the satellite node has been accessed by other groups, determine whether the satellite node meets the group reallocation conditions. If so, adjust the original group to the group where the source node is located.

[0124] If the neighboring nodes of the source node are not assigned to the current group, stop investigating the more distant nodes in this direction.

[0125] In the above implementation, preferably, similarity grouping is also used for:

[0126] The number of nodes in the current group is monitored. If the number of nodes in the current group reaches the maximum number of nodes in the preset group, the preset similarity threshold is modified to the minimum similarity in the current group.

[0127] In the above implementation, preferably, the rotation training module is specifically used for:

[0128] Each group initially selects a source node as a representative node, which performs a preset round of local model training and sends the model parameters to other satellite nodes in the current group after the training is completed;

[0129] Select the next satellite node in the clockwise direction as the new representative node. If the next satellite node belongs to another group, select the first satellite node in the counterclockwise direction of the current group as the new representative node.

[0130] After other satellite nodes in the current group receive the model parameters of the representative node, they are directly used for prediction without training.

[0131] In the above implementation, preferably, the model aggregation module is specifically used to:

[0132] After completing the preset rounds of iterative training, the coordinator initiates model aggregation in both clockwise and counterclockwise directions;

[0133] Send an aggregation request to the representative node of the group at a distance of n / 2 in the clockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the counterclockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0134] Send an aggregation request to the representative node of the group at a distance of n-1-(n / 2) in the counterclockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the clockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path;

[0135] The coordinator receives aggregation parameters from both directions to complete parameter aggregation.

[0136] According to the satellite network federated learning method and system based on data similarity grouping disclosed in the above embodiment, the specific implementation process includes:

[0137] Step 1: System Initialization

[0138] Step 1-1: Set key system parameters within the track. For each track, set the following parameters:

[0139] max_distance is the maximum distance threshold. If it is set to 2, it means that each source node considers at most 2 adjacent nodes on the left and right.

[0140] max_group_size: Maximum group size. If it is set to 5, each group will contain at most 5 satellite nodes.

[0141] max_group_size_threshold is the maximum threshold of the group. If it is set to 4, the similarity threshold adjustment will be triggered when the number of satellite nodes in the group reaches 4;

[0142] m is the number of training rounds, which indicates the number of iterations required for one round of federated learning. For example, it is set to 20.

[0143] Step 1-2: Initialize the data structure. Create arrays group, visited, and similarity_threshold of length n (the total number of satellites in the orbit). Initially, group[i] = i means each satellite node is a separate group, visited[i] = false means satellite node i has not been visited, and similarity_threshold[i] = 0.8 means the initial similarity threshold is 0.8. Figure 2 As shown, the initial states of the satellite nodes in the orbit are independent of each other.

[0144] Step 1-3: Initial model training. Each satellite node uses local data for initial training to obtain initial model parameters for similarity calculation.

[0145] Step 2: Grouping data by similarity

[0146] Step 2-1: Select a coordinator node. Select a satellite node in orbit as a coordinator to manage the grouping process and maintain global information. Figure 3 As shown, the coordinator is responsible for traversing the ungrouped satellite nodes in clockwise order.

[0147] Step 2-2: Model similarity calculation. For any two satellite nodes i and j, their similarity is calculated using the model parameters obtained through their respective training:

[0148] (1) Flatten the model parameters into vector form, denoted as vectors vi and vj

[0149] (2) Using the cosine similarity calculation formula:

[0150] similarity(i,j)=(vi·vj) / (||vi||*||vj||)

[0151] Where vi·vj represents the vector inner product, and ||vi|| represents the L2 norm of vector vi. In this way, a similarity value in the range of [-1,1] can be obtained. The larger the value, the more similar the models of the two satellite nodes are.

[0152] Step 2-3: Source node traversal and grouping. For each unvisited satellite node i (visited[i] = false), treat it as a source node:

[0153] (1) Flag visited[i] = true

[0154] (2) Traverse its neighboring node j (|ij|≤max_distance). Figure 4 As shown, satellite node 1, as a source node, will traverse satellite nodes 10 and 9 on its left, and satellite nodes 2 and 3 on its right, to determine whether they will join group 1.

[0155] (3) Determine whether satellite node j meets the joining conditions:

[0156] If visited[j] = false and similarity(i, j) ≥ similarity_threshold[i], add j to i's group. Figure 5 As shown, satellite node 10 is accessed for the first time and the similarity between satellite node 10 and satellite node 1 is greater than the similarity threshold of satellite node 1, so satellite node 10 joins the group where satellite node 1 is located.

[0157] If visited[j]=true but similarity(i,j)≥similarity_threshold[i] and is greater than the similarity of the group j currently belongs to, then j will be reclassified into i's group. Figure 6 As shown, satellite node 4, as the source node, will traverse satellite nodes 3, 2 and satellite nodes 5, 6 respectively. When traversing satellite node 3, because visited[3] = true, but the similarity threshold of similarity group 4 between satellite node 3 and satellite node 4 is greater than the similarity between satellite node 3 and the representative node (satellite node 1) of the group to which satellite node 3 currently belongs, then satellite node 3 will be divided into group 4.

[0158] (4) Update group[j]=i, visited[j]=true.

[0159] (5) If a neighbor satellite node is not assigned to a group, stop considering satellite nodes farther away. Figure 7 As shown, satellite node 6, as a source node, will traverse satellite nodes 5, 4 and satellite nodes 7, 8 respectively. When traversing satellite node 5, satellite node 5 does not meet the conditions for joining group 6, so satellite node 4 is no longer considered.

[0160] Step 2-4: Dynamic threshold adjustment. When the number of satellite nodes in the group changes, especially when the number of satellite nodes reaches max_group_size_threshold, the threshold adjustment is triggered:

[0161] (1) Monitor the current group size current_group_size.

[0162] (2) If current_group_size ≥ max_group_size_threshold, set similarity_threshold[i] to the minimum similarity within the group. Figure 8 As shown, satellite node 1, as the source node, will traverse satellite nodes 10, 9 and satellite nodes 2, 3 respectively. After traversing satellite node 3, the satellite nodes in group 1 are [10, 1, 2, 3], reaching max_group_size_threshold = 4, triggering the threshold adjustment strategy, similarity_threshold[1] = 0.81 (minimum similarity within the group).

[0163] (3) Ensure that the threshold value always remains between 0.6 and 1.

[0164] Step 3: Training and rotating representative nodes

[0165] Step 3-1: Initial representative node selection. Each group uses the source node as the initial representative node, which is responsible for performing m / max_group_size rounds of local training. After the training is completed, the model parameters are sent to other satellite nodes in the group for use.

[0166] Step 3-2: Representative node rotation. Select the next satellite node as the new representative node in a clockwise direction. If the next satellite node belongs to another group, select the leftmost node of this group (i.e. the first node in the counterclockwise direction) as the new representative node. After receiving the model of the representative node, other satellite nodes in the group are directly used for prediction without training.

[0167] like Fig. 9 As shown, each group uses the source node as the initial representative node, which is responsible for executing m / max_group_size=20 / 5=4 rounds of local training, sending the model parameters to other representative nodes in the group, and then selecting the next satellite node as the new representative node in a clockwise direction.

[0168] Step 4: Model aggregation and update

[0169] Step 4-1: In-track aggregation. After completing the specified rounds of training, the coordinator node (located on the far left) starts the aggregation process:

[0170] (1) The coordinator calculates the total number of groups n in the track based on the maintained group array;

[0171] (2) Initiate aggregation requests in two directions at the same time:

[0172] Send a request to the representative node of the (n / 2)th group in the clockwise direction;

[0173] Send a request to the representative node of the n-1-(n / 2)th group in the counterclockwise direction;

[0174] (3) The representative node that receives the request passes its own parameters to the coordinator, and the representative nodes passing through add their own parameters to it;

[0175] (4) The coordinator receives aggregation parameters from both directions and completes the final aggregation.

[0176] like Fig.10 As shown, coordinator 1 will send a model parameter request to representative node 6 of group 6 and representative node 9 of group 9 respectively, and w6 and w9 will be propagated to coordinator 1 in the counterclockwise direction and clockwise direction respectively for aggregation.

[0177] Step 4-2: Ground station aggregation and model update. Each ground station receives the aggregation results when connected to the coordinator of the corresponding track, and performs the final aggregation between ground stations. The new global model is distributed to the track through the ground station and propagated within the track until all nodes are updated.

[0178] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A satellite network federated learning method based on data similarity grouping, characterized in that: include: Initialize key system parameters and data structures for each orbital satellite network system; Select one of the satellite nodes in the track as a coordinator, traverse the satellite nodes as source nodes, and group them according to the similarity between the nodes; The source node of each group is used as the representative node to perform the preset rounds of local training, and after the training is completed, the adjacent node is selected as the new representative node to continue the rotation training; After completing the preset rounds of iterative training, the coordinator initiates model aggregation in two directions at the same time, and superimposes the parameters of the representative nodes in the direction of the coordinator to complete the aggregation; The model parameters of the global model formed after parameter aggregation are transmitted to the satellite nodes in the orbit, and the next round of cyclic training is carried out.

2. The satellite network federated learning method based on data similarity grouping according to claim 1, characterized in that: The process of selecting one of the satellite nodes in the orbit as a coordinator, traversing the satellite nodes as source nodes, and grouping the nodes according to the similarity between the nodes includes: Select one of the satellite nodes in the orbit as a coordinator, the coordinator uses an unvisited node as a source node, and traverses the satellite nodes that meet a preset distance condition in the same orbit in a clockwise direction; If the satellite node has not been visited, determine whether the similarity between the satellite node and the source node reaches a preset similarity threshold, and if so, place the satellite node in the group where the source node is located; If the satellite node has been accessed by other groups, determine whether the satellite node meets the group reallocation condition, and if so, adjust and allocate the original group to the group where the source node is located; If the neighboring nodes of the source node are not assigned to the current group, stop investigating the more distant nodes in this direction.

3. The satellite network federated learning method based on data similarity grouping according to claim 2 is characterized in that: Also includes: The number of nodes in the current group is monitored. If the number of nodes in the current group reaches the maximum number of nodes in the preset group, the preset similarity threshold is modified to the minimum similarity in the current group.

4. The satellite network federated learning method based on data similarity grouping according to claim 2 is characterized in that: The source node of each group is used as a representative node to perform a preset round of local training, and after the training is completed, an adjacent node is selected as a new representative node to continue the rotation training. The specific process includes: Each group initially selects a source node as a representative node, which performs a preset round of local model training and sends the model parameters to other satellite nodes in the current group after the training is completed; Select the next satellite node in the clockwise direction as the new representative node. If the next satellite node belongs to another group, select the first satellite node in the counterclockwise direction of the current group as the new representative node. After other satellite nodes in the current group receive the model parameters of the representative node, they are directly used for prediction without the need for training.

5. The satellite network federated learning method based on data similarity grouping according to claim 4 is characterized in that: After completing the preset rounds of iterative training, the coordinator initiates model aggregation in two directions at the same time, and superimposes the parameters of the representative nodes in the direction of the coordinator to complete the aggregation. The specific process includes: After completing the preset rounds of iterative training, the coordinator initiates model aggregation in both clockwise and counterclockwise directions; Send an aggregation request to the representative node of the group at a distance of n / 2 in the clockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the counterclockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path; Send an aggregation request to the representative node of the group at a distance of n-1-(n / 2) in the counterclockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the clockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path; The coordinator receives the aggregation parameters from two directions to complete the aggregation of parameters.

6. A satellite network federated learning system based on data similarity grouping, characterized in that: The method for satellite network federated learning based on data similarity grouping as claimed in any one of claims 1 to 5 is applied, comprising: The system initialization module is used to initialize key system parameters and data structures for the satellite network system in each orbit; A similarity grouping module is used to select one of the satellite nodes in the orbit as a coordinator, traverse the satellite nodes as source nodes, and group them according to the similarity between the nodes; A rotation training module is used to perform a preset round of local training using the source node of each group as a representative node, and select an adjacent node as a new representative node to continue the rotation training after the training is completed; The model aggregation module is used to initiate model aggregation in two directions at the same time under the leadership of the coordinator after completing the preset rounds of iterative training, and superimpose the parameters of the representative nodes in the direction of the coordinator to complete the aggregation; The global update module is used to transfer the model parameters of the global model formed by parameter aggregation to the satellite nodes in the orbit and perform the next round of cyclic training.

7. The satellite network federated learning system based on data similarity grouping according to claim 6, characterized in that: The similarity grouping is specifically used for: Select one of the satellite nodes in the orbit as a coordinator, the coordinator uses an unvisited node as a source node, and traverses the satellite nodes that meet a preset distance condition in the same orbit in a clockwise direction; If the satellite node has not been visited, determine whether the similarity between the satellite node and the source node reaches a preset similarity threshold, and if so, place the satellite node in the group where the source node is located; If the satellite node has been accessed by other groups, determine whether the satellite node meets the group reallocation condition, and if so, adjust and allocate the original group to the group where the source node is located; If the neighboring nodes of the source node are not assigned to the current group, stop investigating the more distant nodes in this direction.

8. The satellite network federated learning system based on data similarity grouping according to claim 7, characterized in that: The similarity grouping is also used for: The number of nodes in the current group is monitored. If the number of nodes in the current group reaches the maximum number of nodes in the preset group, the preset similarity threshold is modified to the minimum similarity in the current group.

9. The satellite network federated learning system based on data similarity grouping according to claim 7, characterized in that: The rotation training module is specifically used for: Each group initially selects a source node as a representative node, which performs a preset round of local model training and sends the model parameters to other satellite nodes in the current group after the training is completed; Select the next satellite node in the clockwise direction as the new representative node. If the next satellite node belongs to another group, select the first satellite node in the counterclockwise direction of the current group as the new representative node. After other satellite nodes in the current group receive the model parameters of the representative node, they are directly used for prediction without the need for training.

10. The satellite network federated learning system based on data similarity grouping according to claim 9, characterized in that: The model aggregation module is specifically used for: After completing the preset rounds of iterative training, the coordinator initiates model aggregation in both clockwise and counterclockwise directions; Send an aggregation request to the representative node of the group at a distance of n / 2 in the clockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the counterclockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path; Send an aggregation request to the representative node of the group at a distance of n-1-(n / 2) in the counterclockwise direction, and after receiving the aggregation request, the representative node transmits the model parameters to the coordinator in the clockwise direction, and at the same time adds the model parameters of the representative nodes of each group in the path; The coordinator receives the aggregation parameters from two directions to complete the aggregation of parameters.

Images