A heterogeneous federated learning method for low-orbit satellite internet based on dynamic aggregation

Abstract

The present invention provides a low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation, which relates to the field of federated learning technology and includes the following steps: S1, a central server assigns a local training task to each low-orbit remote sensing satellite through a shared data set method; S2, satellites participating in local training receive shared data and parameters of the current global model, and upload updated local model parameters to the central server; S3, the central server calculates a weighted value based on the frequency of each satellite's participation in aggregation and the time when the model last participated in aggregation; based on the weighted value, determines whether convergence has occurred; S4, outputs the current global model; S5, replaces the initial global model in S1 with the current global model and re-executes S1-S3 until convergence, outputting the current global model. The present invention can improve the efficiency, accuracy and stability of model training, enhance the generalization ability of the global model, thereby ensuring collaborative work between different satellites and efficient convergence of the model.

Description

Technical Field

[0001] The present invention relates to the field of federated learning technology, and more specifically, to a low-orbit satellite Internet heterogeneous federated learning method based on dynamic aggregation. Background Art

[0002] With the rapid development of satellite remote sensing technology, low-orbit satellite constellations are increasingly being used in fields such as urban planning, weather forecasting, and disaster management due to their wide coverage, lack of geographical restrictions, and ability to provide rich remote sensing information. However, traditional centralized model training methods rely on ground stations downloading satellite remote sensing data and then transmitting it to ground data centers for training. Federated learning allows data to remain local to the satellite and enables distributed model training by transmitting model parameters instead of raw data. Compared with traditional methods, federated learning significantly reduces communication overhead while protecting data privacy. Previous studies have introduced parameter averaging aggregation methods in federated learning, improving communication efficiency through local training and periodic global updates. A hierarchical aggregation architecture combines synchronous and asynchronous training to significantly improve model convergence speed and accuracy. Asynchronous update methods effectively address the problem of varying training progress between nodes in heterogeneous environments. Furthermore, clustering algorithms are used to alleviate data inequality, and asynchronous training efficiency and model quality are improved by allowing some nodes to participate in global updates.

[0003] While existing methods have achieved significant results in improving federated learning performance, applying federated learning algorithms to low-orbit remote sensing satellites presents significant challenges, including data heterogeneity, model bias, and uneven participation. When existing federated learning algorithms are applied to low-orbit remote sensing satellites, the data collected by these satellites can vary significantly. In particular, data collected by satellites in different regions and environments can exhibit significant distribution differences. This data disparity can affect the training of local models on each satellite, leading to client drift during training and limiting the generalization ability of the global model, resulting in reduced model performance. In particular, when data distribution is uneven, model training can face additional challenges, making it difficult to achieve the expected accuracy. During the federated learning process, each time a local model is updated, the satellite uploads its updated parameters and participates in the global aggregation. However, this global model update often overwrites the training results of local models, thereby losing the personalized characteristics unique to local training. To prevent this overlay from causing a loss of personalized features in local training, the activity and frequency of participation of low-orbit remote sensing satellites vary greatly. Some satellites participate frequently in training, generating large amounts of data, while others may not participate for extended periods. This can lead to the global model's over-reliance on certain satellites during aggregation, impacting the quality and efficiency of the aggregation. The lack of dynamic adjustment to the frequency and timeliness of participation can lead to insufficient accuracy and convergence of the global model. Summary of the Invention

[0004] This invention aims to address the challenges of uneven data distribution, the difficulty in balancing personalized and global coordination, and the frequency and timeliness of updates in existing federated learning systems for low-orbit remote sensing satellites. By providing a dynamic aggregation-based heterogeneous federated learning method for the low-orbit satellite internet, this approach improves the efficiency, accuracy, and stability of model training and enhances the generalization capabilities of the global model, thereby ensuring collaborative work between different satellites and efficient model convergence. By optimizing data distribution, personalized updates, and a dynamic weighted aggregation strategy, this invention improves the efficiency, accuracy, and stability of model training and enhances the generalization capabilities of the global model.

[0005] The technical means adopted in the present invention are as follows:

[0006] A low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation includes the following steps:

[0007] S1. Obtain global data from all low-orbit remote sensing satellites. The central server configures an initial global model. The central server assigns local training tasks to each low-orbit remote sensing satellite using a shared dataset method. The central server extracts a portion of the global data as shared data and sends the shared data to the low-orbit remote sensing satellite for local training.

[0008] S2. Satellites participating in local training first receive shared data and the parameters of the current global model, and use the exponential moving average mechanism to personalize and update their local models to obtain updated local model parameters, which are then uploaded to the central server.

[0009] S3. The central server aggregates the global model based on the updated local model parameters. During the aggregation process, the central server merges the models of the satellites participating in the aggregation. The central server calculates a weighted value based on the frequency of each satellite participating in the aggregation and the time when the model last participated in the aggregation. Based on the weighted value, the central server determines whether the global model has converged. If so, S4 is executed; if not, S5 is executed.

[0010] S4, output the current global model;

[0011] S5: Replace the initial global model in S1 with the current global model and re-execute S1-S3 until convergence, and output the current global model.

[0012] Furthermore, S1 specifically includes the following steps:

[0013] S11. Initialize the global model and parameter configuration; the central server initializes the global model parameters , providing an initial global model for federated learning as the starting point for all low-orbit remote sensing satellites; initialization Low-orbit remote sensing satellites, initialize the number of times to participate in aggregation ; Set the initial model parameters for each low-orbit remote sensing satellite and record the frequency of each low-orbit remote sensing satellite participating in the global aggregation;

[0014] S12, create a shared data set; the central server from the global data Extract a part of it as a shared dataset , determine the proportion of shared data , The ratio of shared data extracted from global data is in the range of ; Randomly extract shared data: from global data The random sampling ratio is Data as a shared dataset ;make sure The data category distribution in is as close to the global data distribution as possible ; When extracting shared data, a stratified sampling method is used to ensure that each category There are enough samples in

[0015] S13. Distribute shared datasets to low-orbit remote sensing satellites.

[0016] Furthermore, S13 specifically includes the following steps:

[0017] S131. Determine the distribution ratio , The ratio of shared data received by each low-orbit remote sensing satellite, ranging from [0,1];

[0018] S132, allocating shared data; The ratio is Part of the data is sent to each participating low-orbit remote sensing satellite, so that the distribution of shared data categories received by each low-orbit remote sensing satellite is as similar as possible. Category distribution of

[0019] S133, combination of local data and shared data; each low-orbit remote sensing satellite uses its private data and shared data for joint training during local training; low-orbit remote sensing satellite Data distribution Expressed as:

[0020] ;

[0021] in: It is a low-orbit remote sensing satellite The original data distribution, is the global data distribution, is the proportion of shared data;

[0022] ;

[0023] in, Measured the The weight variance after the synchronization reflects the degree of customer drift since the previous synchronization round; is the attenuation factor, taking into account the regularization coefficient and learning rate Impact on customer drift; Measured by low-orbit remote sensing satellite New data distribution With global distribution the differences between; Measured each During step synchronization, client drift is caused by gradient updates; Indicates the Wheel time, low orbit satellite The maximum gradient norm of local model parameters.

[0024] Furthermore, S2 specifically includes the following steps:

[0025] The central server follows the set ratio A number of low-orbit remote sensing satellites are randomly selected to participate in this round of training, denoted as ;Collected from low-orbit remote sensing satellites Random selection low-orbit remote sensing satellites to ensure that each low-orbit remote sensing satellite has a certain probability of being selected; the central server will Sent to selected low-orbit remote sensing satellites ; Low-orbit remote sensing satellites receive the global model and perform exponential moving average personalized updates. Each low-orbit remote sensing satellite Receive global model parameters , and use the exponential moving average mechanism to update the local model parameters , the formula of the exponential moving average mechanism is as follows;

[0026] ;

[0027] in: is the exponential moving average decay rate; is a scaling factor used to adjust the impact of the difference between the global model and the local model;

[0028] The weight difference of the deep network is used to represent the weight difference of the entire model; each low-orbit remote sensing satellite Use its private data and shared data to Round of local training, update local model parameters ;

[0029] At each update, a regularization term is introduced , to prevent the local model from deviating too much from the global goal, the specific optimization problem is:

[0030] ;

[0031] in: Low-orbit remote sensing satellite The local loss function of is the regularization coefficient; each low-orbit remote sensing satellite Update the local model The data is uploaded to the central server, which collects local model updates from all participating low-orbit remote sensing satellites.

[0032] Furthermore, S3 specifically includes the following steps:

[0033] S31, for each low-orbit remote sensing satellite participating in this round of aggregation , increasing the number of times it participates in aggregation ;Record the frequency of each low-orbit remote sensing satellite participating in the global aggregation;

[0034] S32, adjusting the aggregation frequency and weight of the model;

[0035] S33, the central server will update the global model parameters It is sent to all participants to complete this round of federated learning process, and the new global model parameters are sent to all participating low-orbit remote sensing satellites through a secure communication protocol to ensure that each low-orbit remote sensing satellite can perform the next round of local training based on the latest global model; the central server selects a certain proportion of participants to participate in the next round of iterative training, and reselects low-orbit remote sensing satellites to participate in the next round of iteration according to the set strategy until the model converges.

[0036] Furthermore, S32 specifically includes the following steps:

[0037] S321. The central server calculates each participant In the Frequency Factor in Round Aggregation , the calculation formula is:

[0038] ;

[0039] in: Indicates the The number of times a participant participated in the aggregation before this round, Indicates the total number of times all participants in this round participate in the aggregation; frequency factor Reflects the activity level of each low-orbit remote sensing satellite in the federated learning process;

[0040] S322, the central server according to the frequency factor and time factor , the models participating in global aggregation are divided into the following four cases:

[0041] When satellites participate in aggregation times When the number of satellites participating in aggregation is less than the number of the last aggregation round, the frequency of the model participating in aggregation is high. rounds, it is determined that the time of participating in global aggregation is recent; when the number of satellites participating in aggregation is the same as the number of rounds from the last aggregation round During the round, it is determined that the time of participating in the global aggregation is close;

[0042] When satellites participate in aggregation times And the number of times since the last aggregation round Wheel time, frequency factor Determined to be close, time factor For far, among them and To set the threshold; when the model participates in aggregation frequency and the model last participated in global aggregation time recently, close, Far, in this case give Weight; when the model participates in aggregation frequently and the model last participated in global aggregation a long time ago, close, Far, give Weight; when the model participates in aggregation less frequently and the model last participated in global aggregation recently, close, When the model participates in aggregation less frequently and the model last participated in global aggregation a long time ago, close, Far, reduce its weight;

[0043] The central server uses the frequency factor and time factor , use the following formula for global parameter aggregation:

[0044] ;

[0045] in, For the The amount of local data of each participant, is the total data volume of all participants, is a hyperparameter used to control the effect of time decay; For the Participants in Local model parameters after round aggregation.

[0046] The present invention also provides a storage medium, which includes a stored program, wherein when the program is run, any of the above-mentioned low-orbit satellite Internet heterogeneous federated learning methods based on dynamic aggregation is executed.

[0047] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor, wherein the processor executes any one of the above-mentioned low-orbit satellite Internet heterogeneous federated learning methods based on dynamic aggregation through the computer program.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] In existing federated learning systems, data distribution varies significantly between different low-orbit remote sensing satellites. This causes each satellite's local model to deviate from the global model during training, impacting the training effectiveness and generalization capability of the global model. To address this issue, this paper proposes a solution for creating a shared dataset. By extracting a certain percentage of global data and distributing it to each participating satellite, this approach reduces data distribution differences, mitigates client drift, and improves the generalization capability of the global model.

[0050] To effectively balance the personalized needs of local models with the consistency of the global model, this paper introduces an exponential moving average (EMA) mechanism. To prevent the loss of personalized features from local training due to this overwriting, the EMA mechanism smoothly adjusts the differences between the local and global models, ensuring that each satellite's local model retains its personalized features while maintaining consistency with the global model. By personalizing the global model, we ensure both the accuracy of local training and the stability of the global model, thereby improving the performance and adaptability of the overall model.

[0051] In existing systems, the frequency with which different satellites participate in model aggregation varies significantly, leading to infrequent model updates for some satellites, potentially compromising the quality of the global model. To address this issue, the present invention designs a dynamic weighted aggregation strategy that weights each satellite's model update based on both frequency and time. This ensures that data from active satellites has a greater impact on the global model, while satellites that have not participated in aggregation for a long time receive appropriate weight adjustments, thereby improving global model aggregation effectiveness and training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0053] Figure 1 It is the overall flow chart of the present invention.

[0054] Figure 2 This is the overall architecture diagram of the inter-satellite collaborative training method based on federated learning in the present invention.

[0055] Figure 3 Create and distribute shared data flows for the present invention.

[0056] Figure 4 Schematic diagram of shallow and deep networks of the present invention.

[0057] Figure 5 This is a diagram of the global parameter aggregation process of the present invention.

[0058] Figure 6 This is the global parameter aggregation flow chart of the present invention. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions 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 embodiments described are only 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 making creative efforts should fall within the scope of protection of the present invention.

[0060] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0061] This invention aims to address the challenges of uneven data distribution, the difficulty in balancing personalized and global coordination, and the frequency and timeliness of updates in existing federated learning systems for low-orbit remote sensing satellites. By providing a dynamic aggregation-based heterogeneous federated learning method for the low-orbit satellite internet, this approach improves the efficiency, accuracy, and stability of model training and enhances the generalization capabilities of the global model, thereby ensuring collaborative work between different satellites and efficient model convergence. By optimizing data distribution, personalized updates, and a dynamic weighted aggregation strategy, this invention improves the efficiency, accuracy, and stability of model training and enhances the generalization capabilities of the global model.

[0062] The present invention proposes a low-orbit satellite Internet heterogeneous federated learning method based on dynamic aggregation, which is mainly composed of k low-orbit remote sensing satellite constellations and a central server located in medium orbit. Under the entire federated learning architecture, the distributed nodes composed of low-orbit remote sensing satellites perceive ground information through the payloads they carry and collect data for training local models. The gradient information obtained during the training process is first uploaded to the central server in medium orbit through the inter-satellite transmission link composed of the access network, transmission network and core network, and then the central server sums all the received model parameters through the parameter aggregation algorithm to obtain global gradient information. Afterwards, the central server transmits the updated gradient information to each distributed node through the same transmission path for the remote sensing satellite to complete a model update. After multiple rounds of gradient updates, the global model converges and the entire training process is completed. The overall architecture diagram is as follows Figure 2 As shown in the figure, the entire training process usually includes two main steps: parameter aggregation and model update. Parameter aggregation refers to the weighted summation of the model parameters of all low-orbit remote sensing satellites to generate a new global model. Model update refers to sending the updated global model parameters to each low-orbit remote sensing satellite and using it to update the satellite model. In addition, since all data is only used for local model training, federated learning can also protect user privacy and data security. The overall process is as follows Figure 1 shown.

[0063] A low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation includes the following steps:

[0064] S1. Obtain global data from all low-orbit remote sensing satellites. The central server configures an initial global model. The central server assigns local training tasks to each low-orbit remote sensing satellite using a shared dataset method. The central server extracts a portion of the global data as shared data and sends the shared data to the low-orbit remote sensing satellite for local training.

[0065] S1 Specifically, to address the problem of data distribution differences among low-orbit remote sensing satellites, a method for creating and distributing shared data sets is proposed. In the initial stage of federated learning, the central server needs to configure an initial global model for the entire system and assign local training tasks to each low-orbit remote sensing satellite.

[0066] First, initialize the global model and parameter configuration. The central server initializes the global model parameters. , providing an initial global model for federated learning as the starting point for all low-orbit remote sensing satellites. Initialize the central server for each low-orbit remote sensing satellite , initialize the number of times to participate in aggregation , set the initial model parameters for each low-orbit remote sensing satellite, and record the frequency of each low-orbit remote sensing satellite participating in the global aggregation so as to perform weighted processing in the subsequent aggregation process.

[0067] Next, create a shared dataset, and the central server will Extract a part of it as a shared dataset , determine the proportion of shared data , The ratio of shared data extracted from global data is in the range of . The choice of needs to find a balance between data distribution differences, communication overhead and model complexity. Usually, It can be set to a smaller value to reduce communication overhead while ensuring that each low-orbit remote sensing satellite can receive enough diverse data. The random sampling ratio is Data as a shared dataset .make sure The data category distribution in is as close to the global data distribution as possible , in order to ensure the quality of shared data. When extracting shared data, a stratified sampling method can be used to ensure that each category This ensures that there are sufficient samples in the shared data. This prevents certain categories from being overrepresented or underrepresented in the shared data, which could affect the model's generalization ability. By introducing shared datasets, the data distribution gap between low-orbit remote sensing satellites is narrowed, reducing client drift.

[0068] Finally allocate the shared dataset:

[0069] (1) Determine the allocation ratio , It is the ratio of each low-orbit remote sensing satellite receiving shared data, and its value range is [0,1]. The choice of also needs to find a balance between data distribution differences, communication overhead and storage costs. Usually, It can be set to a smaller value to ensure that each low-orbit remote sensing satellite can receive sufficient shared data without increasing excessive communication and storage burdens.

[0070] (2) Allocate shared data: Allocate shared data sets The ratio is Part of the data is sent to each participating low-orbit remote sensing satellite. Ensure that the shared data category distribution received by each low-orbit remote sensing satellite is as close as possible to ensure the validity of shared data.

[0071] (3) Combination of local data and shared data: Each low-orbit remote sensing satellite uses its private data and shared data for joint training during local training. Data distribution It can be expressed as:

[0072] ;

[0073] in: It is a low-orbit remote sensing satellite The original data distribution. is the global data distribution. is the proportion of shared data.

[0074] After the introduction of shared data, the data distribution of low-orbit remote sensing satellites is closer to the global data distribution, reducing customer drift. The customer drift (i.e. weight variance) of each category after synchronization can be expressed as:

[0075] ;

[0076] in, Measured the The weight variance after the synchronization reflects the degree of customer drift after the previous synchronization. is the attenuation factor, taking into account the regularization coefficient and learning rate Impact on customer drift. Measured by low-orbit remote sensing satellite New data distribution With global distribution The difference between Measured each During the step synchronization process, client drift is caused by gradient updates.

[0077] By distributing shared datasets, each low-orbit remote sensing satellite can not only use its own private data during local training, but also have access to more diverse shared data, thereby improving the generalization ability of the model. Figure 3 shown.

[0078] S2. Satellites participating in local training first receive shared data and the parameters of the current global model, and use the exponential moving average mechanism to personalize and update their local models to obtain updated local model parameters, which are then uploaded to the central server.

[0079] Specifically, S2 proposes an EMA mechanism for personalized updates to address the coordination issues between personalized training of low-orbit remote sensing satellites and the global model. In each round of federated learning training, the central server randomly selects a certain percentage of low-orbit remote sensing satellites to participate in local training. Participating satellites first receive the parameters of the current global model and use the exponential moving average (EMA) mechanism to perform personalized updates on their local models. This mechanism enables each low-orbit remote sensing satellite to conduct local training, combining its private and shared data without relying entirely on the global model, thereby gradually optimizing the local model. The EMA mechanism helps balance the differences between the global and local models, ensuring the personalization of the local model is coordinated with the global model, thereby improving the accuracy and effectiveness of training.

[0080] The central server follows the set ratio A number of low-orbit remote sensing satellites are randomly selected to participate in this round of training, denoted as . From low-orbit remote sensing satellite collection Random selection Low-orbit remote sensing satellites are randomly selected to ensure that each low-orbit remote sensing satellite has a certain probability of being selected. By randomly selecting low-orbit remote sensing satellites, it is ensured that the participants of each iteration are different, and some low-orbit remote sensing satellites are prevented from excessively affecting the global model. The central server will send the current global model parameters Sent to selected low-orbit remote sensing satellites Ensure that each participating low-orbit remote sensing satellite can perform local training based on the latest global model. The low-orbit remote sensing satellite receives the global model and performs EMA personalized updates. Each low-orbit remote sensing satellite Receive global model parameters , and use the EMA mechanism to update the local model parameters .

[0081] ;

[0082] in: is the EMA decay rate, is a scaling factor used to adjust the impact of the difference between the global model and the local model. By constraining the global model parameters, the impact of large differences between the global aggregation parameters and the local model parameters on the local model performance is reduced, so as to maintain the personalization of the local model.

[0083] Typically, models involved in local training consist of multiple layers, each containing a significant number of weights. Consequently, the number of parameters per layer is substantial. Therefore, when updating parameters, it is unrealistic to calculate the weight difference between all local parameters and the global aggregated parameters. In experiments, we observed the following:

[0084] (1) The shallow layers of a deep neural network learn general features that are applicable to different datasets, which means that there will not be much difference between the global model and the local model for this part.

[0085] (2) In contrast, deep networks in deep neural networks are more likely to learn some specific features related to the data.

[0086] Based on this, the present invention uses the weight difference of the deep network to represent the weight difference of the entire model, thereby reducing the amount of calculation. Figure 4 shown.

[0087] Each low-orbit remote sensing satellite Use its private data and shared data to Round of local training, update local model parameters .

[0088] At each update, a regularization term is introduced , to prevent the local model from deviating too much from the global goal. The specific optimization problem is:

[0089] ;

[0090] in: It is a low-orbit remote sensing satellite The local loss function is usually cross entropy loss or mean square error. is the regularization coefficient, which is used to control the strength of regularization. Update the local model The data is uploaded to a central server, which collects local model updates from all participating low-orbit remote sensing satellites for global aggregation. A regularization term is introduced to limit the amplitude of local model updates and prevent them from deviating too much from the global target.

[0091] S3. The central server aggregates the global model based on the updated local model parameters. During the aggregation process, the central server merges the models of the satellites participating in the aggregation. The central server calculates a weighted value based on the frequency of each satellite participating in the aggregation and the time when the model last participated in the aggregation. Based on the weighted value, the central server determines whether the model has converged. If so, S4 is executed; if not, S5 is executed.

[0092] S4, output the current global model;

[0093] S5: Replace the initial global model in S1 with the current global model and re-execute S1-S3 until convergence, and output the current global model.

[0094] Specifically, S3-S5 address the frequency and timeliness of model updates for low-orbit remote sensing satellites by proposing a weighted aggregation method. After a low-orbit remote sensing satellite completes local training and uploads its updated model, the central server aggregates the global model. During this aggregation process, the central server not only merges the models of participating satellites but also calculates a weighted value based on the frequency of each satellite's participation and the last time the model participated in the aggregation.

[0095] First, for each low-orbit remote sensing satellite participating in this round of aggregation , increasing the number of times it participates in aggregation , record the frequency of each low-orbit remote sensing satellite participating in the global aggregation, so as to perform weighted processing in the subsequent aggregation process.

[0096] Next, adjust the aggregation frequency and weight of the model.

[0097] (1) The central server calculates each participant In the Frequency Factor in Round Aggregation , the calculation formula is:

[0098] ;

[0099] in: Indicates the The number of times a participant participated in the aggregation before this round, Indicates the total number of times all participants in this round participate in the aggregation. Frequency factor It reflects the activity level of each low-orbit remote sensing satellite in the federated learning process. Low-orbit remote sensing satellites that frequently participate in aggregation will obtain higher weights, while low-orbit remote sensing satellites that participate less frequently will obtain lower weights.

[0100] (2) The central server calculates the frequency factor and time factor , the models participating in global aggregation are divided into the following four cases:

[0101] When the model participates in aggregation frequently and the model last participated in global aggregation recently, Larger, In this case, the model parameters are relatively fresh and should be given a larger weight. When the model participates in aggregation frequently and the model last participated in global aggregation a long time ago, Larger, Large. Although the model participates frequently, its parameters may be outdated due to the long time since its last participation, so its weight should be appropriately reduced. When the model participates in aggregation less frequently and the model last participated in global aggregation more recently, Smaller, In this case, although the model participates less frequently, its parameters are relatively fresh and its weight can be appropriately increased. When the model participates in aggregation less frequently and the model last participated in global aggregation a long time ago, Smaller, Since the model is inactive and has not participated in aggregation for a long time, its parameters may be seriously outdated, so its weight should be significantly reduced.

[0102] By combining the frequency factor and the time factor, the central server can more finely control the aggregation of global model parameters, ensuring that those active models with large data volumes receive higher weights, while appropriately increasing the weights of those less involved but recently updated models, to avoid negative impacts on the global model caused by models that have not participated in aggregation for a long time. and time factor , use the following formula for global parameter aggregation:

[0103] ;

[0104] in, It is The amount of local data of each participant, is the total data volume of all participants, is a hyperparameter that controls the effect of time decay. It is Participants in Local model parameters after round aggregation.

[0105] from Figure 5 It can be seen in During the round aggregation process, the time when the local model aggregated by the global model last participated in the global aggregation is 、 , have not yet participated in global model training. The uploaded local model parameters can be regarded as a newer model, while the 、 It can be considered that the last update time is relatively long, The model is further away from the last update time. At the same time, the frequency of each local model participating in the aggregation is also different. 、 、 Participate in aggregation more frequently than other participants.

[0106] By introducing frequency factors and time factors, the central server can more flexibly adjust the weight of each low-orbit remote sensing satellite, ensuring that the aggregation of the global model is more reasonable and effective.

[0107] Finally, the central server sends the updated global model parameters The new global model parameters are sent to all participants, completing this round of federated learning. The new global model parameters are then transmitted to all participating low-orbit remote sensing satellites via a secure communication protocol, ensuring that each satellite can perform the next round of local training based on the latest global model. The central server selects a certain percentage of participants for the next round of iterative training and, based on the set strategy, reselects the next round of low-orbit remote sensing satellites until the model converges.

[0108] This invention addresses the issues of uneven data distribution and client drift by dynamically extracting global data and creating a shared dataset. The shared dataset is created not only based on the global data distribution but also takes into account the varying requirements of low-orbit remote sensing satellites. A stratified sampling approach is used to ensure data diversity and representativeness, thereby reducing the impact of data distribution differences on training and improving the generalization capability of the global model.

[0109] The present invention introduces the EMA mechanism for personalized update of local models of low-orbit remote sensing satellites, avoiding over-reliance on the global model while ensuring that each low-orbit remote sensing satellite can be effectively updated based on local data, balancing the collaborative optimization of the global model and the personalized needs of the local model, ensuring training accuracy and global coordination.

[0110] The proposed aggregation mechanism, based on both timeliness and frequency, dynamically weights participating low-orbit remote sensing satellites, precisely controlling the contribution of each satellite model to the global model aggregation. The frequency factor reflects the satellite's activity level, while the time factor considers the timeliness of its last aggregation participation. By combining these two factors, the central server can more rationally adjust the weights of satellite models, optimizing global model quality and training efficiency.

[0111] Through the above technological innovations, the present invention not only improves the performance and effect of federated learning of low-orbit remote sensing satellites, but also effectively overcomes the shortcomings of traditional methods in data distribution, model personalized update and global aggregation, and provides an efficient and stable solution for the collaborative training of future low-orbit remote sensing satellites in complex dynamic environments.

[0112] This invention significantly enhances the generalization capability of the global model by optimizing data allocation strategies and reducing data distribution differences between different satellites. By using a shared dataset to balance data differences between satellites, it effectively reduces client drift caused by data imbalance, thereby improving the training effectiveness and accuracy of the global model. In this way, the invention enables the coordinated operation of low-orbit remote sensing satellites in a variety of geographical environments, ensuring the stability and accuracy of the global model in complex environments.

[0113] This paper addresses the coordination issues between low-orbit remote sensing satellites and the global model during personalized training by introducing an exponential moving average (EMA) mechanism. This mechanism effectively adjusts the differences between the local and global models during the aggregation process, avoiding issues such as slow training speed and low accuracy caused by personalized differences in local models. Through the EMA mechanism, this paper ensures that the personalized characteristics of each satellite are preserved while simultaneously accelerating the convergence of the global model, ensuring high efficiency and stability across various tasks.

[0114] The dynamic weighted aggregation strategy proposed in this paper significantly improves the model's aggregation quality and training efficiency by dynamically adjusting the weight of each satellite in the global model based on the satellite's participation frequency and timeliness. Active satellites contribute more updates during training, allowing the global model to more accurately reflect the data characteristics of active satellites during aggregation while reducing interference from satellites that have not participated in training for a long time. Through this dynamic weighting mechanism, the present invention improves the timeliness and adaptability of the federated learning system, enabling low-orbit remote sensing satellite systems to more efficiently respond to environmental changes and fluctuations in the frequency of satellite participation, thereby enhancing the stability and efficiency of the overall training process.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation, characterized by: The steps include: S1. Obtain global data from all low-orbit remote sensing satellites. The central server configures an initial global model. The central server assigns local training tasks to each low-orbit remote sensing satellite using a shared dataset method. The central server extracts a portion of the global data as shared data and sends the shared data to the low-orbit remote sensing satellite for local training. S2. Satellites participating in local training first receive shared data and the parameters of the current global model, and use the exponential moving average mechanism to personalize and update their local models to obtain updated local model parameters, which are then uploaded to the central server. S3, the central server aggregates the global model based on the updated local model parameters; During the aggregation process, the central server merges the models of the satellites participating in the aggregation. The central server calculates the weighted value based on the frequency of each satellite participating in the aggregation and the time when the model last participated in the aggregation. Based on the loss function, it is determined whether the global model has converged. If it has converged, S4 is executed; if not, S5 is executed. S4, output the current global model; S5: Replace the initial global model in S1 with the current global model and re-execute S1-S3 until convergence, and output the current global model.

2. The low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation according to claim 1 is characterized in that: S1 specifically includes the following steps: S11. Initialize the global model and parameter configuration; the central server initializes the global model parameters , providing an initial global model for federated learning as the starting point for all low-orbit remote sensing satellites; initialization Low-orbit remote sensing satellites, initialize the number of times to participate in aggregation ; Set the initial model parameters for each low-orbit remote sensing satellite and record the frequency of each low-orbit remote sensing satellite participating in the global aggregation; S12, create a shared data set; the central server from the global data Extract a part of it as a shared dataset , determine the proportion of shared data , The ratio of shared data extracted from global data is in the range of ; Randomly extract shared data: from global data The random sampling ratio is Data as a shared dataset ;When extracting shared data, a stratified sampling method was used; S13. Distribute shared datasets to low-orbit remote sensing satellites.

3. The low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation according to claim 2 is characterized in that: S13 specifically includes the following steps: S131. Determine the distribution ratio , The ratio of shared data received by each low-orbit remote sensing satellite, ranging from [0,1]; S132, allocating shared data; The ratio is Part of the data is sent to each participating low-orbit remote sensing satellite, so that the distribution of shared data categories received by each low-orbit remote sensing satellite converges. Category distribution of S133, combination of local data and shared data; each low-orbit remote sensing satellite uses its private data and shared data for joint training during local training; low-orbit remote sensing satellite Data distribution Expressed as: in: It is a low-orbit remote sensing satellite The original data distribution, is the global data distribution, is the proportion of shared data; in, Measured the The weight variance after the synchronization reflects the degree of customer drift since the previous synchronization round; is the attenuation factor, taking into account the regularization coefficient and learning rate Impact on customer drift; Measured by low-orbit remote sensing satellite New data distribution With global distribution the differences between; Measured each During step synchronization, client drift is caused by gradient updates; Indicates the Wheel time, low orbit satellite The maximum gradient norm of local model parameters.

4. The low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation according to claim 1 is characterized in that: S2 specifically includes the following steps: The central server follows the set ratio A number of low-orbit remote sensing satellites are randomly selected to participate in this round of training, denoted as ;Collected from low-orbit remote sensing satellites Random selection low-orbit remote sensing satellites to ensure that each low-orbit remote sensing satellite has a certain probability of being selected; the central server will Sent to selected low-orbit remote sensing satellites ; Low-orbit remote sensing satellites receive the global model and perform exponential moving average personalized updates. Each low-orbit remote sensing satellite Receive global model parameters , and use the exponential moving average mechanism to update the local model parameters , the formula of the exponential moving average mechanism is as follows; in: is the exponential moving average decay rate; is a scaling factor used to adjust the impact of the difference between the global model and the local model; The weight difference of the deep network is used to represent the weight difference of the entire model; each low-orbit remote sensing satellite Use its private data and shared data to Round of local training, update local model parameters ; At each update, a regularization term is introduced , to prevent the local model from deviating too much from the global goal, the specific optimization problem is: in: Low-orbit remote sensing satellite The local loss function of is the regularization coefficient; each low-orbit remote sensing satellite Update the local model The data is uploaded to the central server, which collects local model updates from all participating low-orbit remote sensing satellites.

5. The low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation according to claim 1 is characterized in that: S3 specifically includes the following steps: S31, for each low-orbit remote sensing satellite participating in this round of aggregation , increasing the number of times it participates in aggregation ;Record the frequency of each low-orbit remote sensing satellite participating in the global aggregation; S32, adjusting the aggregation frequency and weight of the model; S33, the central server will update the global model parameters Send it to all participants to complete this round of federated learning. Send the new global model parameters to all participating low-orbit remote sensing satellites through a secure communication protocol to ensure that each low-orbit remote sensing satellite can perform the next round of local training based on the latest global model. The central server selects a certain proportion of participants to participate in the next round of iterative training, and reselects low-orbit remote sensing satellites to participate in the next round of iteration according to the set strategy until the model converges.

6. The low-orbit satellite internet heterogeneous federated learning method based on dynamic aggregation according to claim 5 is characterized in that: S32 specifically includes the following steps: S321. The central server calculates each participant In the Frequency Factor in Round Aggregation , the calculation formula is: in: Indicates the The number of times a participant participated in the aggregation before this round, Indicates the total number of times all participants in this round participate in the aggregation; frequency factor Reflects the activity level of each low-orbit remote sensing satellite in the federated learning process; S322, the central server according to the frequency factor and time factor , the models participating in global aggregation are divided into the following four cases: When satellites participate in aggregation times When the number of satellites participating in aggregation is less than the number of the last aggregation round, the frequency of the model participating in aggregation is high. During the round, it is determined that the time of participating in the global aggregation is close; When satellites participate in aggregation times And the number of times since the last aggregation round Wheel time, frequency factor Determined to be close, time factor For far, among them and To set the threshold; when the model participates in aggregation frequency and the model last participated in global aggregation time recently, close, Far, in this case give Weight; when the model participates in aggregation frequently and the model last participated in global aggregation a long time ago, close, Far, give Weight; when the model participates in aggregation less frequently and the model last participated in global aggregation recently, close, When the model participates in aggregation less frequently and the model last participated in global aggregation a long time ago, close, Far, reduce its weight; The central server uses the frequency factor and time factor , use the following formula for global parameter aggregation: in, For the The amount of local data of each participant, is the total data volume of all participants, is a hyperparameter used to control the effect of time decay; For the Participants in Local model parameters after round aggregation.

7. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is run, the low-orbit satellite Internet heterogeneous federated learning method based on dynamic aggregation as described in any one of claims 1 to 6 is executed.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The processor executes the low-orbit satellite Internet heterogeneous federated learning method based on dynamic aggregation as described in any one of claims 1 to 6 through the computer program.

Images