Leo satellite network and federated learning model construction method therefor

Abstract

By combining a federated learning model and a low earth orbit (LEO) satellite network, the present invention provides a general federated learning framework over an LEO satellite network (FedSN) for achieving federated learning. The FedSN is composed of two main components: a sub-structure scheme and pseudo-synchronous model aggregation. The sub-structure scheme comprises sub-structure customization, distribution and aggregation methods, and respectively solves problems such as resource limitations, training imbalance, staleness of intra-group models. According to a pseudo-synchronous model aggregation strategy, the difference between weights of models is brought into a weight function, and a buffer-based aggregation method is developed, so that the staleness of inter-group models is reduced. Extensive experimental results show that a FedSN framework is superior to a state-of-the-art baseline. The present invention shows and improves the potential of deploying a FedSN on an LEO satellite network.

Description

A LEO satellite network and its federated learning model construction method Technical Field

[0001] The present invention belongs to the field of satellite data processing technology, and in particular relates to a LEO satellite network and a method for constructing a federated learning model thereof. Background Art

[0002] Low Earth Orbit (LEO) satellite networks (LEO) are a technology that provides communications services by deploying multiple satellites in low Earth orbit. Compared to traditional terrestrial infrastructure, LEO satellite networks offer lower latency and greater coverage, providing bandwidth, connectivity, and communications services to users worldwide. They also collect vast amounts of satellite imagery for analysis of geographic, ecological, and climate data across various fields.

[0003] Due to technical limitations, traditional satellite systems cannot process data in orbit and can only transmit large amounts of image data back to ground stations. The ground stations then transmit this data to cloud computing platforms for subsequent image data analysis and processing, or for training corresponding data processing machine learning models. Since LEO satellites cannot maintain constant communication with ground stations on Earth, large amounts of data downloads place significant pressure on downlink bandwidth. Overcrowding of satellites or ground stations also wastes resources, making downloading all raw data from satellites increasingly unfeasible. This is why federated learning has been proposed and deployed on LEO constellations.

[0004] Federated learning is a machine learning method that centrally trains distributed data. Its goal is to train models locally on devices without sending raw data to a central server. Federated learning performs local model updates on local devices and then aggregates the updated parameters to train a global model. This approach enables collective intelligence model training while protecting user privacy.

[0005] Summary of the Invention

[0006] One embodiment of the present invention, a general Federated Learning Framework over LEO Satellite Networks (FedSN), involves federated learning and low-Earth orbit satellite networks (LEOSNs). Federated learning is a distributed machine learning method designed to train machine learning models without sharing the original data. A low-Earth orbit satellite network (LEOSN) is a satellite network that provides communication services via a network of satellites in low Earth orbit.

[0007] FedSN consists of two main components: substructure scheme and pseudo-synchronous model aggregation.

[0008] The substructure solution includes substructure customization, distribution and aggregation methods, which respectively solve problems such as resource limitations, training imbalance and outdated models within the group.

[0009] A pseudo-synchronous model aggregation strategy incorporates the differences between model weights into the weight function and develops a buffer-based aggregation method to alleviate the problem of inter-group model staleness. Extensive experimental results show that the FedSN framework outperforms state-of-the-art baselines. This work demonstrates the potential of deploying FedSN on LEO satellite networks.

[0010] The FedSN federated learning framework on the LEO satellite network of the embodiment of the present disclosure includes:

[0011] 1. Federated learning (FL) modeling on a low-orbit satellite network, which includes two main components: LEO satellites and ground stations.

[0012] 2. Model the available budget for each satellite.

[0013] 3. Design the workflow for the substructure solution within the group, including substructure segmentation, allocation, downlink transmission, aggregation, and assembly methods. Split the complete data information into multiple smaller fragments, and then customize specific fragments through flexible assembly.

[0014] The FedSN disclosed in this paper combines the federated learning model with the low-Earth orbit satellite network, and proposes a general federated learning framework FedSN for implementing federated learning on the low-Earth orbit satellite network. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0016] FIG1 is a schematic diagram of satellite network composition according to one embodiment of the present invention.

[0017] FIG2 is a schematic diagram illustrating the impact of satellite budget resources on federated learning according to one embodiment of the present invention.

[0018] FIG3 is a schematic diagram of the impact of satellite uplink quality according to one embodiment of the present invention.

[0019] FIG4 is a schematic diagram of an inter-group and intra-group satellite network according to one embodiment of the present invention.

[0020] FIG5 is a schematic diagram showing the impact of satellite cluster out-of-dateness on federated learning according to one embodiment of the present invention.

[0021] FIG6 is a schematic diagram showing the composition of the FedSN framework system according to one embodiment of the present invention.

[0022] FIG7 is a schematic diagram of the FedSN framework training process according to one embodiment of the present invention.

[0023] FIG8 is a schematic diagram showing the convergence of a dataset in federated learning according to one embodiment of the present invention.

[0024] FIG. 9 is an example of substructure model assembly according to one embodiment of the present invention.

[0025] FIG. 10 shows an example of substructure model distribution according to one embodiment of the present invention.

[0026] FIG11 is an example of a pseudo-synchronous aggregation strategy training round workflow according to one embodiment of the present invention.

[0027] FIG12 is a schematic diagram of the test accuracy of FedSN and benchmarks on different data sets according to one embodiment of the present invention.

[0028] FIG. 13 illustrates examples of convergence test accuracy for different data sets according to one embodiment of the present invention.

[0029] FIG. 14 illustrates the total computational and communication overheads for different data sets according to one embodiment of the present invention.

[0030] FIG. 15 illustrates an example of the relationship between the width of customized models for different data sets and the convergence test accuracy according to one embodiment of the present invention.

[0031] FIG16 is a schematic diagram showing the effect of the number of satellites on the convergence test accuracy of different data sets according to one embodiment of the present invention.

[0032] FIG17 illustrates an example of the impact of the substructure customized distribution method on dataset training performance according to one embodiment of the present invention. DETAILED DESCRIPTION

[0033] In practice, deploying federated learning on LEO satellite networks encounters many problems, including the following aspects.

[0034] 1. Issues with heterogeneous computing and storage budgets.

[0035] When deploying federated learning in satellite networks, existing federated learning frameworks assume that each satellite can provide sufficient device resources (computing power and storage size) to train a local model. Therefore, the ground station can always aggregate all trained local models to achieve good performance.

[0036] However, in practice, different satellites have different device resources, and their allocated resources for federated learning training may change dramatically during runtime because these allocated resources depend on the priority allocation of device resources by the programs running on the satellites. Therefore, due to the heterogeneous device resources of different satellites, federated learning aggregation in the ground station may lead to insufficient local model training on some satellites.

[0037] 2. Uplink communication bottleneck problem.

[0038] An uplink bottleneck is a limitation or bottleneck encountered when sending data from a user device to a network server in a communication system. In federated learning and low-Earth orbit satellite networks, uplink bottlenecks can impact system performance and efficiency. In low-Earth orbit satellite networks, uplink bottlenecks can arise from several factors: bandwidth limitations, signal attenuation, and bidirectional link asymmetry.

[0039] 3. Obsolescence in LEO satellite networks.

[0040] Due to the unstable connection between satellites and ground stations, satellites orbiting the Earth can be divided into two sets: inter-sets and intra-sets. The inter-sets represent all satellites visible to a ground station at different contact times, while the intra-sets represent satellites in different orbits within each contact time. For example, if the contact time between a low-Earth orbit satellite and a ground station is 9 minutes, and the maximum orbital period is 90 minutes, in this case, the satellites are divided into 10 inter-sets, each of which contains multiple intra-sets. Federated learning models can become outdated in both inter-sets and intra-sets. Inter-set model outdation is caused by inconsistent model versions due to unstable connections, while intra-set model outdation is caused by the imbalanced participation of satellites in different orbits in the training of the federated learning model.

[0041] To address these challenges, this paper proposes a general framework for federated learning over low-Earth orbit satellite networks. This framework aims to overcome the challenges of federated learning in terms of communication latency, bandwidth limitations, and network instability, enabling efficient distributed model training and updates across satellite networks. In this paper, FedSN stands for "A General Federated Learning Framework over LEO Satellite Networks."

[0042] There are multiple drawings involved in the embodiments of the present disclosure, wherein:

[0043] Figure 1 shows a typical FedSN scenario on a satellite network, which consists of two components: a LEO satellite and a ground station. In the figure, "Satellite" is translated into "satellite" in Chinese, and "Ground station" is translated into "ground station" in Chinese.

[0044] Figure 2 illustrates the impact of heterogeneous compute and storage resources on federated learning (FL). Figures 2(a) and (b) show the performance of test accuracy and untrained satellite rate under compute and storage constraints. Here, GBSense is experimented with VGG-16 in an IID setting. In the figure, Computation refers to computation, and Storage refers to storage.

[0045] Figure 3 shows how uplink communication between LEO satellites and the ground station (GS) becomes the primary bottleneck for FL. Figure 3(d) shows the relationship between test accuracy and contact time, obtained from an experiment on the GBSense dataset using VGG-16 in an IID setting. (a) GS of Starlink, (b) Experimental setup, (c) Uplink and downlink rate, and (d) Local models updating failure.

[0046] Figure 4 shows that satellites orbiting Earth can be divided into two sets: the inter-group set and the intra-group set. The inter-group set represents all satellites visible from a ground station at different contact times, while the intra-group set represents the satellites in different orbits at each contact time. In the figure, the inter-group set is the inter-group set, and the intra-group set is the intra-group set.

[0047] Figure 5 shows the impact of intra-group set staleness and inter-group set staleness on FL. Figure 5(a) and Figure 5(b) respectively show the impact of different intra-group satellite orbit period ratios and inter-group model aggregation on test accuracy.

[0048] Figure 6 is a schematic diagram of the system composition of the FedSN framework in an embodiment of the present disclosure, in which a sub-structure scheme is disclosed.

[0049] Figure 7 is an example of the intra-group training process of the FedSN framework for the k-th inter-group set, where three satellites in the group participate in model training and the available budget is B k ={0.25,1,0.75}.

[0050] Figure 8 is a schematic diagram of the convergence of models with different widths 8(a) and different distribution schemes 8(b) on the GBsense dataset under the IID setting on FL.

[0051] Figure 9 is an example of model assembly for the k-th intergroup set: the first basic substructure model of the aggregation The layer weight matrix is ​​expressed as Customize the global model The layer weight matrix is ​​denoted as W k,l .

[0052] Figure 10 shows a left-side horizontal representation, which includes examples of substructure combinations extracted by random substructure distribution (left part), static substructure distribution (middle part), and the proposed adaptive rolling substructure distribution (right part): represents the substructure combination selected by the i-th satellite in the k-th intergroup set. The global model is divided into four × 0.25 basic substructure models (i.e. L = 4). The available budget is B k ={0.25,0.5,0.5,0.75}.

[0053] Figure 11 shows an example of a training-wheel workflow for the proposed pseudo-synchronous aggregation strategy, where the satellites orbiting the Earth are divided into four groups.

[0054] Figure 12 shows the test accuracy of FedSN and other five baselines on the GBSense and EuroSAT datasets.

[0055] Figure 13 shows the convergence test accuracy on the GBSense and EuroSAT datasets.

[0056] Figure 14 illustrates the total computational and communication overheads for the GBSense and EuroSAT datasets.

[0057] Figure 15 shows the relationship between customized model width and convergence test accuracy on the GBSense and EuroSAT datasets.

[0058] Figure 16 shows the effect of the number of participating satellites on the convergence test accuracy for the GBSense and EuroSAT datasets.

[0059] Figure 17(a) shows the impact of substructure customization methods on the training performance of the GBsense dataset, and Figure 17(b) shows the impact of different substructure distribution methods on the training performance of the GBsense dataset.

[0060] According to one or more embodiments, a FL modeling method on a low-orbit satellite network appropriately models federated learning on a LEO satellite network, providing a theoretical basis for further substructure schemes and pseudo-synchronous model aggregation. As shown in Figure 1, consider a typical FedSN scenario on a satellite network, which includes two main components:

[0061] (1) LEO satellite

[0062] Considering that all satellites have a specific budget for model training (i.e., computing and storage resources on the device), the set of satellites participating in model training is represented by N = {1, 2, ..., N}, where N is the number of satellites. The set of computing and storage resources on the device are represented by C = {c1, c2, ..., c N} and S={s1,s2,…,s N}, where c i ∈[0,1] and s i ∈[0,1] represents the resources available to the ith satellite (i.e., the portion of the computational workload and storage space of the global model that the ith satellite can afford). Similarly, the effect of uplink rate on model training is given by U={u1,u2,…,u N} means that u i ∈[0,1] represents the part of the global model that the ground station can upload to the i-th satellite within the contact time with the i-th satellite. The local dataset residing on the i-th satellite is denoted as D i ={x i,k ,y i,k}, where x i,k and y i,k Represents the local dataset D i The kth input data and its corresponding label in . Therefore, the total data set is

[0063] (2) Ground Station

[0064] The ground station, a central server with powerful computing capabilities, is responsible for executing the FedSN substructure scheme and pseudo-synchronous model aggregation. It is also responsible for collecting parameters of each visible satellite, including computing and storage resources, as well as channel state information required for the operation of the FedSN framework.

[0065] The global model is represented by W∈R b Represented by, where b is the dimension of the model parameters. i,k The derived predicted values ​​are expressed as where f(x;ω) maps the relationship between the input data x (we abuse x here) and the predicted value given the model parameters ω. Therefore, the local loss function of the i-th satellite is expressed as Among them L i,k (x i,k ,y i,k ; W) represents the local data set D i The global loss function is a weighted average of the local loss functions, where the weights are proportional to the size of the local dataset.

[0066] The goal of federated learning is to find the optimal model parameter W in the following optimization problem: * :

[0067] To solve Equation (1), traditional federated learning (e.g., FedAvg) employs synchronous model aggregation to find the optimal model parameters. However, as observed, the intermittent connectivity between low-Earth orbit satellites and ground stations makes it infeasible to directly implement FedAvg. In addition, in traditional federated learning, the withdrawal of satellites with limited available equipment resources (ci < 1 or si < 1 or ui < 1), as well as the imbalanced participation of satellites due to different orbital periods, can lead to severe model bias, thereby degrading training performance.

[0068] According to one or more embodiments, a federated learning (FL) modeling method on a low-orbit satellite network includes modeling the available budget for each satellite. The modeling process is described as follows.

[0069] The low earth orbit satellites in various orbits around the earth are divided into K groups. For the k-th satellite group, the satellites participating in the model training are represented as where N k is the number of satellites. The local data set on the i-th satellite in the k-th inter-group set is expressed as Therefore, the total dataset is represented as The computing and storage resource set of the satellite in the kth intergroup set is expressed as and in and They represent the available computing and storage resources of the i-th satellite in the k-th inter-group set, respectively. The calculation formulas of c and s are as follows:

[0070] Where j = {m|i∈P m}, P m represents the set of satellites in the mth orbit (i.e., the mth intra-group set), represents the computational resources allocated to the i-th satellite in the k-th inter-group set for federated learning model training (i.e., floating-point operations per second), η and θ represent the computational workload (in floating-point operations) of the local update period and global model training, respectively. and denote the available storage resources of the i-th satellite in the k-th inter-group set and the data size of the global model (in bits), respectively.

[0071] To describe the impact of uplink quality on the satellite's available link budget, define To represent the proportion of hidden layer channels that the ground station can upload to the substructure model within a period of contact time with the i-th satellite in the k-th inter-group set. Expressed as:

[0072] in is the uplink rate from the ground station to the i-th satellite in the k-th inter-group set, Δt and φ represent the contact time and the data size of the global model (in bits), respectively. In the embodiment of the present disclosure, considering the heterogeneous computing and storage resources of the satellites and the uplink rate, the available budget of the i-th satellite in the k-th inter-group set is expressed as:

[0073] Before model training begins, the ground station collects the available budget of each satellite, and based on this information, FedSN can perform the following intra-group substructure scheme: including substructure customization, allocation, and aggregation methods. The basic idea of ​​the disclosed embodiment is to divide the complete knowledge into multiple smaller fragments, and then customize specific fragments through flexible assembly. For the sake of simplicity, the analysis here focuses on the k-th inter-group set. As shown in Figure 7, in any training round 1 For t∈T={1,2,…,T}, the intra-group training process includes the following steps. For simplicity, the training round number t is omitted here.

[0074] a) Substructure segmentation: In this step, the global model is first divided into several basic substructure models by segmenting by channel based on the minimum constrained budget in the k-th inter-group set. in Denotes the lth basic substructure model, and L denotes the number of basic substructure models. Then, the ground station determines the number of basic substructure models that each satellite can afford based on the available budget of the heterogeneous satellites.

[0075] b) Substructure allocation: The ground station then extracts the basic substructure model from each satellite The selected combination is then assigned to the corresponding satellite via uplink. k The problem of extracting subsets.

[0076] c) After receiving the basic substructure model combination provided by the ground station, each satellite uses the local resident data to train the corresponding basic substructure model in parallel.

[0077] d) Substructure downlink transmission: After local model training, each satellite sends its basic substructure model combination to the ground station within the contact time.

[0078] e) Substructure Aggregation: In this step, the ground station aggregates the same basic substructure models collected from different satellites to form a single model for model assembly. To address the staleness of models within a group caused by differences in orbital periods, the substructure aggregation method described in this disclosure is applied.

[0079] f) Substructure assembly: After aggregating the same basic substructure models, the ground station will assemble a customized global model based on these aggregated basic substructure models.

[0080] The above summarizes the workflow of FedSN intra-group training. The customized substructure solution combines the substructure customization, allocation, and aggregation methods in the previous stages a), d), and e). In the following sections, the disclosed embodiments continue to provide the corresponding intra-group substructure solution (SS) implementation details.

[0081] 1) Substructure customization method (SCM).

[0082] In response to the challenges posed by heterogeneous available budgets, it is necessary to design flexible and scalable model customization methods. Before customizing a model, it is important to assess its complexity. The main metric used to quantify the complexity of a model is the width or depth of the model (i.e., the number of channels or layers in the model). Although there is a learning framework called separation learning that can separate the global model into a client model and a server-side model and train both sides separately, during the training phase, the client must share the features of its client model with the server. Considering the low-Earth orbit satellite network, the limited contact time is too short to train a good and robust model through separation learning. Therefore, in the embodiment of the present disclosure, the design of the substructure model focuses on the width of the model.

[0083] As shown in Figure 7, different basic substructure models are constructed using different hidden layer channels as the model width, while maintaining the same input and output layers. By flexibly controlling the model width, diverse substructure complexity can be achieved. This is because different substructure models have different numbers of parameters, which can meet the heterogeneous available budget of satellites. To this end, this paper studies the impact of different model widths on training performance to understand this issue.

[0084] We first studied budget-free FL on LEO satellite networks. Using FedAvg, we trained local models with varying widths, but limited their maximum widths to ×1, ×0.5, and ×0.25. Their training performance is shown in Figure 8(a). It is clear that wider models converge faster and generalize better than narrower ones. Therefore, these experiments demonstrate the desirability of customizing wider models rather than narrower ones.

[0085] Furthermore, the performance of budget-constrained federated learning training is studied by using HeteroFL. HeteroFL is a current federated learning framework for heterogeneous clients with different computing capabilities, but it only assumes that the local model of each satellite has a predefined fixed width. For example, if the i-th satellite has a specific computing resource HeteroFL can only allocate the corresponding fixed-width ×0.75 model, but not other fixed-width models.

[0086] Therefore, the ×0.75 model cannot be trained using the datasets on other satellites. Therefore, during the training phase, HeteroFL fails to fully utilize the data scattered across the satellites. We improve HeteroFL to form a variant called VHeteroFL and evaluate its training performance, in which each satellite can train as many affordable models as possible. This means that when the satellite has When , it can train {×0.25, ×0.5, ×0.75} models.

[0087] To simplify the analysis, the available budget of the satellite is configured to follow a discrete uniform distribution, namely P(b i =0.25,0.5,0.75,1)=0.25. The results shown by the dashed line are shown in Figure 8(a). Clearly, the wider model not only exhibits lower test accuracy than its FedAvg counterpart, but even the narrower model is inferior. This is because only 1 / 4 of the satellite data can be used to train the ×1 model, while 3 / 4 of the satellite data information is incorporated into the ×0.25 model. The increased data participation enables the ×0.25 model to generalize better than the ×1 model, suggesting that customizing the widest model may not be a wise choice.

[0088] Based on the above two observations, it can be found that substructure customization needs to be well designed, otherwise it may lead to deterioration rather than improvement of training performance. Based on this, the embodiment of the present disclosure proposes a flexible substructure customization method. First, the basic substructure model is determined according to the minimum budget constraint within the satellite group to ensure that each satellite can train these models. Then, the global model is divided into multiple basic substructure models through channel partitioning, and the substructure models are divided according to the available budget of the satellite. Customize different combinations of these models. This can eliminate the under-training phenomenon of the basic substructure model, thereby effectively improving the convergence speed of model training and test accuracy. An example in Figure 7 can be used to illustrate how FedSN determines the number of basic substructure models. The global model is divided into four × 0.25 basic substructure models, and there are three satellites in the k-th intergroup set, whose available budget B k ={0.25,0.75,0.75}. In this case, the system decides to set 1, 3, and 3 basic substructure models for each satellite respectively.

[0089] After segmenting the basic substructure models, they need to be assembled into a customized global model. Since each satellite trains its own combination of basic substructure models based on its available budget, the budget-constrained federated learning problem becomes budget-free. Recalling Figure 8(a), when there is no budget constraint, the wide model shows faster convergence and better generalization than the narrow model. Therefore, it is necessary to assemble the global model as broadly as possible.

[0090] As shown in Figure 9, the global model is divided into four × 0.25 substructures based on the hidden layer channels. To reconstruct the global model, the weight matrices of the aggregated basic substructures are concatenated into a matrix of the same size as the global model weight matrix. During the matrix concatenation process, non-trainable parameters (gray areas) are set to 0 to avoid interference between different basic substructures. Therefore, for any given Layer, the weight matrix of the assembled global model is expressed as:

[0091] 2) Substructure distribution method (SDM).

[0092] When determining the number of basic substructure models for different satellites and their combinations, it is necessary to select and distribute the basic substructure models. In the embodiment of the present disclosure, a substructure customization method is set up, where L=4, and two different substructure distribution methods are adopted: random distribution (RD) and static distribution (SD). For RD and SD, the combination of substructures is extracted from the global model in a random and fixed manner, respectively. As shown in Figure 10, RD randomly selects basic substructure models to form the combination of each satellite, while SD maintains a fixed combination. Figure 8(b) shows the significant performance gap between RD, SD and the ideal case, where the ideal case is the ideal performance achieved by deploying a global model on each satellite for local model training. There are two insights into these results:

[0093] For RD, it randomly extracts different combinations in consecutive training rounds, resulting in unbalanced training of the basic substructure model. This unbalanced training leads to an imbalance in the update frequency of different parts of the global model, inevitably leading to slower convergence speed and lower test accuracy.

[0094] For SD, the performance degradation stems from two main inherent flaws:

[0095] First, as shown in Figure 10, in most cases, satellites cannot train all basic substructure models unless some of them have sufficient available budget to train a complete global model. Second, due to the fixed combination method, any basic substructure model cannot be fully trained on scattered data on all four satellites.

[0096] It can be seen that the substructure distribution method is crucial for improving training performance. To this end, the disclosed embodiments propose an adaptive rolling substructure distribution method that adheres to the following two design principles: First, different basic substructure models should be trained as equally as possible. Second, each basic substructure model should be trained on datasets distributed across different satellites, regardless of the satellite's available budget. In each round, an adaptive rolling window is used to select consecutive basic substructure models based on the available budget of each satellite. However, for different satellites, it is necessary to ensure that the starting position of the window in the basic substructure model sequence is different.

[0097] Taking Figure 10 as an example, assuming that the available budget of the k-th satellite group is B k ={0.25,0.5,0.5,0.75}, the corresponding combination cardinalities are 1, 2, 2 and 3. For the first satellite, a window of length 1 is used to select the basic substructure model Then, the rolling window is moved forward one step and the length is adjusted to 2 according to the available budget of the second satellite, resulting in the combination Similarly, select the combination For the third satellite, select the combination To the fourth satellite.

[0098] In summary, the rolling window mechanism ensures that all base substructure models are trained with approximately equal weight. Furthermore, the adaptive window size adjusts in real time based on the available budget for each satellite. Furthermore, this approach allows for full training on data dispersed across different satellites. The substructure aggregation method is described next.

[0099] 3) Substructure Aggregation Method (SAM).

[0100] The same basic substructure models from different satellites are aggregated for model assembly. However, as mentioned earlier, the orbital period changes of satellites in different orbits in the satellite constellation lead to an imbalance of satellites participating in model training, thereby reducing the effectiveness of model training. In order to compensate for the model bias in FL model training, it is crucial to develop an appropriate substructure aggregation method. The aggregated basic substructure model is given by:

[0101] Therefore, the embodiment of the present disclosure proposes a pseudo-synchronous model aggregation strategy (PMAS) between groups. The core idea of ​​this strategy is to first asynchronously aggregate local models with lower staleness to speed up the convergence of model training, and then synchronously aggregate models with higher staleness at the end of each round of training to enhance generalization ability. To this end, unlike existing methods, this strategy considers the similarity between model weights to characterize the staleness of the model. At the same time, the strategy in the embodiment of the present disclosure does not always perform asynchronous aggregation, but adopts a buffer-based approach. Without loss of generality, we focus on any training round t∈T={1,2,...,T} consisting of K inter-group sets. For simplicity, the training round number t is omitted. The pseudo-synchronous model aggregation strategy includes the following steps:

[0102] a) Weight function design. During the communication time between the kth inter-group set and the ground station, the customized model W of the group is evaluated. k The weight similarity with the latest model W* stored in the ground station. The cosine distance is used to evaluate the difference between the model weights and can be calculated as:

[0103] According to formula (8), the expression of the model staleness function is: S(d k , τ k )=d k s(τ k ), (9)

[0104] b) Asynchronous model aggregation with low staleness.

[0105] For the k-th intergroup set, the latest model stored in the ground station is updated by:

[0106] Where α is a hybrid hyperparameter for asynchronous aggregation. The threshold γth for model staleness is determined based on training task experience. According to formula (10), the ground station switches the appropriate strategy based on the staleness function of the customized model of any inter-group set. Specifically, when the value of the staleness function of the customized model is lower than the threshold γth, it is classified as a low-staleness model and asynchronously aggregated into the latest model; otherwise, it is classified as a high-staleness model and stored in the cache.

[0107] c) Synchronous model aggregation with high staleness.

[0108] At the end of the training round, the highly stale custom models stored in the cache are aggregated into a single global model represented as:

[0109] Then, model W' is aggregated with the latest model with a smaller weight, expressed as:* =βW * +(1-β)W′ (12)

[0110] Here, β is a hybrid hyperparameter for synchronous aggregation. Below, we use an example in Figure 11 to illustrate the entire process of PMAS.

[0111] Four inter-group sets are considered in this training round, and the initial latest model version is L0. When the satellite in the first inter-group set establishes a connection with the ground station, all the basic substructure models from the satellite are assembled into a customized global model A1 through the substructure customization in Section 3. The customized model A1 is classified as a high-staleness model and is therefore stored in the cache instead of being aggregated into the latest model. When the satellites in the second and third inter-group sets establish a connection with the ground station, their customized models A2 and A3 are classified as low-staleness models. Therefore, both customized models are asynchronously aggregated into the latest model. Subsequently, similar to A1, the customized model A4 from the fourth inter-group set is stored in the cache due to its high model staleness. Finally, all customized models stored in the cache (i.e., A1, A4) are synchronously aggregated into the latest model with a smaller weight.

[0112] In short, unlike existing strategies that only rely on time-related weighting functions for asynchronous aggregation, the strategy of the present disclosure incorporates the similarity of model weights into the weighting function and asynchronously aggregates low-staleness models in a timely manner to speed up the convergence of the model, while storing high-staleness models in a cache and aggregate them at the end of the training round to enhance generalization ability.

[0113] To further validate the solution of this disclosure, this disclosure experimentally tested the above technical solution. FedSN was implemented using Python 3.7 and PyTorch 1.9.1, and training was performed on a ThinkPad P17 Gen1 laptop equipped with an NVIDIA Quadro RTX 3000 GPU, an Intel i9-10885H CPU, and a 4TB SSD. FedSN uses the widely adopted and recognized VGG-16 network. VGG-16 is a classic deep convolutional neural network consisting of 13 convolutional layers and 3 fully connected layers. VGG-16 utilizes a stack of multiple convolutional layers to effectively extract features from images, while the fully connected layers are responsible for classifying and predicting the extracted features. The learning rate for each satellite was uniformly set to 0.005, and the batch size was 128. The loss function was cross-entropy loss, and the weighting function s(τ) was a polynomial function.

[0114] In the experiments, it is assumed that N satellites are deployed orbiting the Earth, distributed in J different altitude orbits. Unless otherwise specified, N and J are set to 24 and 3 by default. The default orbital period ratio is p1:p2,...,:pJ=1:1.5:2, and the orbit with the longest period is set to 120 minutes. The computation and storage resource budgets between satellites follow a discrete uniform distribution (i.e., P(bi=0.25,0.5,0.75,1)=0.25). The uplink rate is set according to the actual traces collected from Starlink (e.g., RTT).

[0115] The spatial modulation recognition dataset GB-Sense and the remote sensing image dataset EuroSAT are used to evaluate the training performance of FedSN. GB-Sense contains sampled signals of 13 modulation types, including 16,000 training samples and 4,000 test samples. EuroSAT contains remote sensing images of 10 different categories, such as industry, highway, and forest. EuroSAT has 21,600 training samples and 5,400 test samples. In addition, experiments are conducted under IID and non-IID data settings. In the IID setting, the training samples are randomly divided into 24 equal slices, and each slice is assigned to one satellite. In the non-IID setting, the data is sorted by label, divided into 240 slices, and 10 slices are assigned to one of each of the 24 satellites.

[0116] Furthermore, ablation experiments are conducted on micro-benchmarks to demonstrate the effectiveness of each carefully designed component in FedSN (i.e., substructure customization, substructure distribution, substructure aggregation, and pseudo-synchronous model aggregation).

[0117] 1) Substructure Customization. Figure 17(a) shows the impact of the substructure customization method on the training performance of the GBsense dataset. For FedAvg, resource constraints on the satellite hinder the effective integration of scattered information into the training model, resulting in the worst training performance. However, by deploying the proposed substructure customization method on FedAvg, the satellite can utilize more local data to train its basic substructure combination, thereby improving the test accuracy from 75.3% to 94.4%. In addition, the smaller performance gap between the ideal case and SCM+FedAvg further confirms the effectiveness of the proposed substructure customization method.

[0118] 2) Substructure Distribution. Figure 17(b) shows the impact of different substructure distribution methods on the training performance of the GBsense dataset. SCM+RD and SCM+SD have lower test accuracy due to the random and fixed selection of substructure combinations, which leads to unbalanced training of different parts of the global model. The comparison between SCM+SDM, SCM+RD, and SCM+SD highlights the superiority of the adaptive rolling substructure distribution method proposed in this disclosure, which better ensures more balanced model training, thereby improving training accuracy.

[0119] 3) Substructure Aggregation. Figure 17(c) shows the impact of the substructure aggregation method on the training performance of the GBsense dataset. The comparison between SAM and FedAvg demonstrates the effectiveness of the substructure aggregation method proposed in this disclosure in alleviating the lag in the model within the group caused by differences in orbital period, thereby improving the generalization ability of model training. By modifying the aggregation weights, satellites with longer orbital periods are given greater weights, while satellites with shorter orbital periods are given smaller weights, thus ensuring balanced participation of satellites.

[0120] 4) Pseudo-synchronous model aggregation. Figure 17(d) shows the impact of different inter-group aggregation strategies on the training performance of the GBsense dataset. By comparing SS+PMAS and SS+FedAsync, it can be clearly seen that the proposed customized pseudo-asynchronous model aggregation strategy significantly accelerates the model convergence speed and improves the test accuracy. This is mainly due to two reasons: first, the proposed strategy aggregates the customized models with low lag asynchronously in a timely manner, accelerating the model convergence speed; second, at the end of the training round, the disclosed strategy aggregates the customized models with high lag with smaller weights to enhance the generalization ability.

[0121] This paper evaluates the overall performance of the FedSN framework based on test accuracy, computational overhead, and communication overhead. It also evaluates the performance of the FedSN framework when customizing models of varying widths and the number of satellites N on training performance. To validate the advantages of the FedSN framework, it is compared with other baselines:

[0122] Ideal: The ideal case is a resource- and lag-unconstrained FedAvg. In this ideal case, it is assumed that all low-Earth orbit satellites have sufficient resources to support the intra-satellite training of the global model, and are not constrained by the intra-group and inter-group model lags.

[0123] SS+FedAsync: The SS+FedAsync benchmark uses the substructure solution of the disclosed embodiment for intra-group model training. For the aggregation of inter-group models, FedAsync is used with a polynomial weighting function.

[0124] SS+FedAvg: The SS+FedAvg benchmark uses a substructure scheme for within-group model training and FedAvg for between-group model aggregation.

[0125] FedAsync: The trained models of all satellites are consistently aggregated into the latest model in an asynchronous manner.

[0126] FedAvg: The trained models of all satellites are stored in the ground station and synchronized into the latest updated model at the end of each training round.

[0127] The substructure scheme consists of three components, each of which mitigates the adverse effects of resource limitations, uneven training, and intra-group model staleness on the training process. Comparing FedSN with SS+FedAsync reveals that the proposed pseudo-asynchronous model aggregation strategy effectively compensates for inter-group staleness caused by model version differences, thereby improving the test accuracy of model training. Figure 12 shows the test accuracy of FedSN and five other baselines on the GBSense and EuroSAT datasets. As can be seen, as the model converges, FedSN's test accuracy remains comparable to the ideal case. Clearly, FedSN outperforms SS+FedAsync, SS+FedAsync, FedAsync, and FedAvg. Notably, due to the substructure scheme, FedSN, SS+FedAsync, and SS+FedAvg achieve significantly better test accuracy than FedAsync and FedAvg. In addition, the comparison of Figure 12(a) with Figure 12(b) and the comparison of Figure 12(c) with Figure 12(d) show that the convergence speed of FedSN and the other five benchmarks is slower in the non-IID setting than in the IID setting.

[0128] Figure 13 shows the converged test accuracy on the GBSense and EuroSAT datasets, where the accuracy is recorded at 250 epochs (GBSense) and 70 epochs (EuroSAT), when these learning frameworks have converged. Figures 13(a) and 13(b) show that FedSN achieves test accuracy of 98.19% and 92.88% (94.11% and 83.26%) on the GBSense (EuroSAT) dataset under the independent and identically distributed (IID) and non-IID settings, respectively, which is only a negligible decrease of 0.54% and 0.78% (0.79% and 0.8%) compared to the ideal case. This shows that although FedSN cannot provide global model training for each satellite due to resource limitations, the carefully designed substructure scheme alleviates this problem and integrates more useful information scattered among satellites into model training, resulting in training performance close to the ideal case. Furthermore, by comparing FedSN with SS+FedAvg and SS+FedAsync, we observe that the proposed pseudo-synchronous model aggregation strategy outperforms traditional strategies. For FedAvg and FedAsync, without specifically addressing resource heterogeneity, some satellites are consistently excluded from model training, resulting in poor generalization performance. Furthermore, Figures 13(a) and 13(b) show that in the non-IID setting, FedSN and the other five benchmarks achieve higher convergence test accuracy than in the IID setting.

[0129] Figure 14 illustrates the total computational and communication overhead for the GBSense and EuroSAT datasets. It can be seen that the FedSN framework reduces computational and communication overhead by approximately 38% while maintaining test accuracy similar to the ideal case. This is due to two reasons: First, the substructure scheme significantly reduces the number of network parameters by partitioning the network by channels, thereby reducing overall computational and communication overhead. Second, the substructure scheme effectively mitigates information loss caused by resource limitations, uneven training, and model staleness within a group. The pseudo-asynchronous model aggregation strategy eliminates the adverse effects of inter-group model staleness by designing a weighting function and a buffer-based model aggregation scheme. Overall, FedSN is a more efficient communication and computation framework with only a slight loss in accuracy.

[0130] Figure 15 shows the relationship between customized model width and converged test accuracy on the GBSense and EuroSAT datasets. For VHeteroFL, we only show performance evaluated over the trained width. Clearly, the proposed FedSN framework consistently achieves test accuracy similar to the ideal case on customized models of varying widths, demonstrating the excellent scalability of the FedSN framework. For VHeteroFL, satellite resource limitations prevent wider models from fully utilizing the dispersed data, resulting in poor training performance. In contrast, FedSN overcomes the undertraining issue caused by resource limitations through the split-and-combine principle. Specifically, the global model is first split into basic substructure models according to a minimum budget constraint, and then these are recombined into a customized global model. As a result, FedSN achieves superior test accuracy to VHeteroFL at every model width.

[0131] Figure 16 shows the impact of the number of participating satellites on the convergence test accuracy for the GBSense and EuroSAT datasets. Observe that the curves for all satellite numbers in each figure follow similar trends. Experimental results for satellite numbers ranging from 5 to 20 show negligible performance impact on FedSN, while the test accuracy of VHeteroFL decreases with increasing number of participating satellites, highlighting FedSN's superior robustness relative to VHeteroFL. As the number of satellites increases, the distribution of resource constraints becomes more heterogeneous. For VHeteroFL, the fixed substructure model customization hinders its ability to effectively handle the high heterogeneity of resource constraints. In contrast, FedSN's substructure model customization, through flexible substructure partitioning and assembly, enables it to overcome resource heterogeneity, thereby enhancing robustness. Furthermore, based on practical observations, FedSN consistently outperforms VHeteroFL in training performance across a wide range of satellite numbers.

[0132] Ablation experiments of the disclosed embodiments are conducted to demonstrate the effectiveness of each carefully designed component in FedSN (i.e., substructure customization, substructure distribution, substructure aggregation, and pseudo-synchronous model aggregation).

[0133] The foregoing description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art will readily conceive of various equivalent modifications or substitutions within the technical scope disclosed herein, and such modifications or substitutions are intended to be encompassed within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A model construction method based on LEO satellite network, characterized in that: Assume that the satellite set is represented by N = {1, 2, ..., n}, correspondingly, The computing resources of the satellite set are C = {c1, c2, ..., c N }, the storage resources are S = {s1,s2,…,s N }, where c i ∈[0,1] and s i ∈[0,1] represents the computing resources and storage resources available to the i-th satellite, The influence of the uplink rate from the ground station to the satellite on the model training is given by U = {u1,u2,…,u N } means that u i ∈[0,1] represents the part of the global model that the ground station can upload to the i-th satellite within the contact time with the i-th satellite, The local dataset residing on the i-th satellite is denoted as D i ={x i,k ,y i,k },in, x i,k and y i,k Represents the local dataset D i The kth input data and its corresponding label in , so that, The global model is denoted as W∈R b , where b is the dimension of the model parameters, The total dataset of the global model is From the input data x i,k The derived predicted values ​​are expressed as where f(x;ω) maps the relationship between the input data x and the predicted value given the model parameters ω, The local loss function of the i-th satellite is expressed as Where L i,k (x i,k ,y i,k ; W) represents the local data set D i The sampling loss function for the kth data sample in is, The global loss function is a weighted average of local loss functions, where the weights are proportional to the size of the local dataset. Through the server configured by the ground station, a FedSN based on the federated learning model is built.

2. The model building method according to claim 1, characterized in that: The satellite set is divided into K inter-group sets, For the kth satellite set, the satellite set participating in model training is represented as Where N k is the number of satellites, The local data set on the i-th satellite in the k-th inter-group set is expressed as Therefore, the total data set is represented as The computing and storage resource set of the satellite in the kth inter-group set is and in, and denote the available computing resources and storage resources of the ith satellite in the kth inter-group set, respectively. Where j = {m|i∈P m }, P m represents the set of satellites in the mth orbit, represents the computing resources allocated to the i-th satellite in the k-th inter-group set for federated learning model training, η and θ represent the computational workload of the local update cycle and global model training, respectively. and They represent the available storage resources of the i-th satellite in the k-th inter-group set and the data size of the global model respectively.

3. The model building method according to claim 2, characterized in that: set up is used to represent the ratio of the hidden layer channel of the ground station uploading to the intra-group substructure model during the contact period with the i-th satellite in the k-th inter-group set. This parameter is used to describe the impact of the uplink quality on the satellite's available link resources. Will It is expressed as: in, is the uplink rate from the ground station to the i-th satellite in the k-th inter-group set, Δt and φ represent a contact time and the data size of the global model, respectively.

4. The model building method according to claim 3, characterized in that: The available resources of the i-th satellite in the k-th inter-group set are expressed as:

5. The model building method according to claim 3, characterized in that: The construction process of the substructure model includes substructure segmentation, substructure allocation, satellite training of substructure model, satellite sending of trained substructure model, ground station aggregation of substructure model, and substructure assembly.

6. The model building method according to claim 5, characterized in that: The substructure segmentation step specifically includes: The global model first divides the channels based on the minimum restricted resource in the k-th inter-group set. After segmentation, it is divided into several basic substructure models in represents the lth basic substructure model, L represents the number of basic substructure models, The ground station then determines the number of basic substructure models that each satellite can afford based on the satellite's available resources.

7. The model building method according to claim 5, characterized in that: The step substructure allocation specifically includes: The ground station extracts the basic substructure model from each satellite A combination of The selected combination is then distributed to the corresponding satellites via uplink.

8. The model building method according to claim 6, characterized in that: While keeping the input and output layers unchanged, different basic substructure models are constructed using different channels of the hidden layer as the width of the model.

9. The model building method according to claim 5, characterized in that: The substructure model is aggregated using pseudo-synchronous pairing.

10. A LEO satellite network, characterized in that: Adopt any model building method as claimed in claims 1 to 9 to build a FedSN based on the federated learning model.

Images