A dynamic aggregation federated learning method based on satellite and ground station connection density

Abstract

The present invention discloses a dynamic aggregation federated learning method based on the connection density between satellites and ground stations, comprising the following steps: in round i, when the satellite and the ground station establish a connection for the first time, the ground station sends a global model w i The satellite trains the model locally and returns the local model to the ground station upon its next connection. When the number of collected local models exceeds the buffer capacity V or reaches a predefined time period T0, the ground station updates the global model using a weighted average method. This invention fully utilizes satellite computing resources, improves model robustness, increases the number of global model aggregations or the number of aggregated gradients, and effectively improves model prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of satellite technology, and in particular to a dynamic aggregation federated learning method based on the connection density between satellites and ground stations. Background Art

[0002] Recent advances in satellite technology have reduced the cost of satellite deployment and promoted the large-scale development of low-Earth orbit (LEO) satellites. The large amount of image data collected by LEO satellites has enhanced machine learning capabilities and helped solve some global problems such as climate change and the spread of disease.

[0003] Traditional machine learning based on satellite imagery typically requires ground stations to download images generated by satellites, which then transmit the data to cloud computing platforms to train machine learning models. However, downloading all the raw image data generated from satellites is becoming increasingly infeasible.

[0004] This is mainly because: (1) unlike geosynchronous satellites, low-orbit satellites only communicate with ground stations a few times a day, and each communication time is limited; (2) due to the multi-million dollar cost of building ground stations and the difficulty of expanding them, the downlink bandwidth is limited; (3) satellite images are usually high-resolution and may contain private information, which may cause privacy leakage after being downloaded.

[0005] To address these challenges, some researchers have applied federated learning (FL) to satellite constellations. FL is a server-client architecture in which only model parameters need to be exchanged between the server and client, without requiring the client to transmit raw data. The concept of FL was proposed by McMahan et al., who combined FL with stochastic gradient descent (SGD) to propose FedAvg (the Federated Averaging Algorithm), a synchronous FL algorithm. Simply put, FedAvg selects a subset of clients to execute SGD in each round, then a central server aggregates the results of these clients to achieve global model optimization. To improve flexibility and scalability, Cong Xie et al. proposed a new asynchronous federated optimization algorithm, FedAsync, based on FL. Unlike synchronous federated learning, FedAsync allows server and client updates to occur at any time without synchronization, which is beneficial when devices are heterogeneous. To address potential privacy and security issues when aggregating individual clients in asynchronous FL, John Nguyen et al. proposed a novel buffered asynchronous aggregation method, FedBuff, which is independent of the optimizer choice and combines the best features of synchronous and asynchronous FL. In FedBuff, clients asynchronously enter and complete local training. However, the server model is not updated immediately upon receiving each client update. Instead, client updates are stored in a buffer. The server update occurs only when the number of clients whose updates are stored in the buffer meets the buffer size, which is a tunable parameter.

[0006] In FL for satellite constellations, the roles played by ground stations and satellites in FL can be mainly divided into the following two types:

[0007] (1) Federated Learning Framework with Satellite as Client

[0008] Jinhyun So et al. proposed FedSpace, an efficient FL framework operating between satellites and ground stations. They formulated an optimization problem to maximize model convergence rate according to a global model aggregation schedule at the ground station, which captures the unique trade-off between satellite idleness and local model staleness in this environment. To solve this optimization problem, FedSpace first computes deterministic and time-varying satellite connectivity using satellite orbits and Earth rotation, and then uses this information to determine a global model aggregation schedule that better balances idleness and staleness. The model empirically demonstrates the effectiveness of their solution using real-world satellite networks and satellite imagery datasets, and experimental results show that FedSpace significantly reduces training time while achieving equivalent accuracy compared to synchronous FL (McMahan et al., 2017) and state-of-the-art buffered asynchronous FL (Nguyen et al., 2021). Nasrin Razmi, BhoMatthiesen et al. considered FL in LEO constellations, where satellites collaboratively train ML models without sharing local datasets. Building on FedAvg and FedAsync, a novel asynchronous variant of FedAvg is proposed to adapt to this setting, identifying and addressing unique challenges compared to terrestrial networks. Zubair Md Fadlullah and Nei Kato proposed asynchronous updates for federated learning models at edge nodes to build smart remote sensing local AI models with forest fire detection scenarios. Their proposal eliminates the need to explicitly exchange data with neighboring nodes and / or the cloud. As a result, their proposed approach jointly protects data privacy and alleviates network overhead by efficiently scheduling deep and shallow model parameters. Experiments were conducted using real-world datasets to demonstrate the performance of our proposed asynchronous federated learning model by simulating several edge nodes at different layers of the considered ensemble network. The results show that their proposal is feasible in terms of high remote sensing accuracy, low convergence time, and significantly low bandwidth overhead compared to contemporary methods.

[0009] (2) Federated Learning Framework with Satellite as Server

[0010] Quoc-Viet Pham, Ming Zeng, and others developed a novel concept, aerial FL, by integrating AAN (Aerial Access Networks) and FL (Federated Learning). They discussed the application of AAN to FL in mobile edge networks. Specifically, they outlined aerial FL and demonstrated its advantages over traditional FL networks with fixed AS deployments. They also introduced key applications in sky model aggregation, aerial relaying, aerial FL users, and FANETs. FL over wireless communication networks can significantly improve learning performance to meet the growing requirements for data privacy and communication overhead in future 6G networks. Hao Chen, Ming Xiao, and Zhibo Pang discussed the role of FL in addressing the challenges of LEO-based satellite communications and proposed a federated learning model that deploys parameter servers on satellites and uses user devices as clients. Experimental evaluation demonstrated that the proposed FL over LEO satellite constellations achieves practical accuracy on the ensemble MNIST dataset, and that the communication overhead of FL-based methods is significantly lower than that of CL-based methods. Furthermore, they noted several challenges for future research in terms of privacy and security, resource management, and communication overhead.

[0011] In FL for satellite constellations, the main problems are: 1. When the connection density between satellites and ground stations is too sparse, other satellites will spend a lot of time waiting; 2. When the connection density between satellites and ground stations is too dense, a large number of satellites will be idle, resulting in a waste of computing resources. Summary of the Invention

[0012] The technical problem to be solved by the embodiments of the present invention is to provide a dynamic aggregation federated learning method based on the connection density between satellites and ground stations, which can prevent a large number of satellites from being idle when the connection between satellites and ground stations is dense, thereby wasting computing resources.

[0013] To solve the above technical problems, an embodiment of the present invention provides a dynamic aggregation federated learning method based on the connection density between satellites and ground stations. The method is specifically implemented based on the Synchronous Periodic Aggregation Strategy (SPA), and the implementation of SPA specifically includes the following steps:

[0014] S1: i When the satellite establishes a connection with the ground station for the first time, the ground station sends the global model w i ;

[0015] S2: The satellite trains the model locally and returns the local model to the ground station when it connects next time;

[0016] S3: When the number of local models collected reaches a predefined time period T When 0, the ground station updates the global model using a weighted average method.

[0017] Wherein, the updating of the global model is expressed as:

[0018]

[0019] In the formula, B represents the satellite index set returned to the local model in this round, V is the buffer capacity of the buffer, and a(B) is the aggregation index. The subscript SPA is the abbreviation of Synchronous Periodic Aggregation, which means synchronous periodic aggregation; =1, indicating that the ground station can perform local update aggregation.

[0020] By combining the buffered asynchronous aggregation method FedBuff on the basis of SPA, the present invention is fully implemented, including the following steps:

[0021] S1: i When the satellite establishes a connection with the ground station for the first time, the ground station sends the global model w i ;

[0022] S2: The satellite trains the model locally and returns the local model to the ground station when it connects next time;

[0023] S3: When the number of local models collected exceeds the buffer capacity V or to a predefined time period T When 0, the ground station updates the global model using a weighted average method.

[0024] Wherein, the updating of the global model is expressed as:

[0025]

[0026] In the formula, B represents the satellite index set returned to the local model in this round, V is the buffer capacity of the buffer, and a(B) is the aggregation index. The subscript SDA is the abbreviation of Synchronous Dynamic Aggregation, which means synchronous dynamic aggregation, that is, the dynamic aggregation federated learning method proposed in this patent; =1, indicating that the ground station can perform local update aggregation.

[0027] The weighted average method includes weighting the loss function of the global model on the local data set on each satellite to obtain the final loss function :

[0028]

[0029]

[0030] Where |Ds| is the number of records in the private dataset Ds, s represents the satellite subscript, and S represents the set of satellites contained in the FL system; , Is the global model in the local dataset The loss function on , ℓ(w, x) is the training loss of model w for data sample x.

[0031] Wherein, the S2 specifically includes the steps of:

[0032] After receiving the global model, the satellite will use the global model to initialize the local model, and then use the local dataset to train the model:

[0033]

[0034] Where η is the learning rate, the subscript of w is the local training index, indicating the round of local training, and the superscript is the satellite index. represents the local model of the ith orbit of satellite s, represents the local model of the satellite s's i+1th round, Using a local dataset The loss function obtained from the last i-th round of training.

[0035] The satellite uses a local dataset to train a local model for multiple epochs to reduce the number of communication rounds required for convergence.

[0036] Wherein, the S3 further includes the steps of:

[0037] S31: The ground station aggregates local model parameters / gradients from the satellite and updates the global model when the update condition is met. The expression is:

[0038]

[0039] Where s represents the satellite subscript, represents the satellite index set of the local model returned in this round, |Ds| is the number of records in the private dataset Ds, It represents the sum of the number of records of the datasets of all satellites in the satellite index set returned to the local model in this round. a(B) is the aggregation index (a(B) = 1, indicating that the ground station can perform local update aggregation;

[0040] S32: After the ground station updates the global model, it coordinates the satellite to perform the next round of training and sends the latest global model.

[0041] The implementation of the embodiments of the present invention has the following beneficial effects: 1. The present invention uses synchronous dynamic aggregation (SDA) combined with synchronous periodic aggregation and synchronous buffer aggregation, which aggregates the local model once within a preset time period, thereby shortening the long update waiting time caused by sparse satellite connections; and in the case of dense satellite connections, the synchronous buffer aggregation strategy fully utilizes a large number of satellites that are idle when the satellite connections are dense, and fully utilizes the computing resources of the satellites; 2. The present invention can perform the first global model update faster, thereby improving the robustness of the model; 3. The present invention is superior to other methods in terms of the training time required for target accuracy; 4. The present invention adopts synchronous dynamic aggregation (SDA) with a dynamic aggregation strategy to increase the number of global model aggregations or the number of aggregation gradients, effectively improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of synchronous dynamic aggregation of the present invention;

[0043] Figure 2 This is a comparison of the shortest time required for four federated learning methods to reach the set accuracy under two dataset distributions;

[0044] Figure 3 This is a graph showing how the accuracy of the federated learning method on the test set changes with training time under the independent and identically distributed setting;

[0045] Figure 4 This is a graph showing how the accuracy of the federated learning method on the test set changes with training time under the non-IID setting;

[0046] Figure 5 It is a comparison chart of the number of global model updates and the number of local model gradients of the four FL algorithms. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to the accompanying drawings.

[0048] Combine Figure 1 As shown, a dynamic aggregation federated learning method based on the connection density of satellites and ground stations in an embodiment of the present invention is implemented by the following method.

[0049] S1: i When the satellite establishes a connection with the ground station for the first time, the ground station sends the global model w i ;

[0050] S2: The satellite trains the model locally and returns the local model to the ground station when it connects next time;

[0051] After receiving the global model, the satellite will use the global model to initialize the local model, and then use the local dataset to train the model:

[0052]

[0053] Where η is the learning rate, the subscript of w is the local training index, indicating the round of local training, and the superscript is the satellite index. represents the local model of the ith orbit of satellite s, represents the local model of the satellite s's i+1th round, Using a local dataset The loss function obtained from the last i-th round of training.

[0054] The satellite can use the local dataset to train the local model for multiple epochs to reduce the number of communication rounds required for convergence. After the local training is completed, the satellite uploads the local model parameters / gradients to the ground station the next time it establishes a connection with the ground station.

[0055] S3: When the number of local models collected exceeds the buffer capacity V or to a predefined time period T At 0, the ground station updates the global model using a weighted average method. This process is expressed as the following equation:

[0056]

[0057] In the formula represents the satellite index set returned to the local model in this round, V is the buffer capacity of the buffer, a(B) is the aggregation index, where The subscript SDA is the abbreviation of Synchronous Dynamic Aggregation, which means synchronous dynamic aggregation, that is, the dynamic aggregation federated learning method proposed in this patent; =1, indicating that the ground station can perform local update aggregation.

[0058] The weighted average method includes weighting the loss function of the global model on the local data set on each satellite to obtain the final loss function :

[0059]

[0060]

[0061] Where |Ds| is the number of records in the private dataset Ds, s represents the satellite subscript, and S represents the set of satellites contained in the FL system; , Is the global model in the local dataset Where ℓ(w, x) is the training loss of model w for data sample x. It should be noted that the communication delay between ground stations is lower than the satellite-to-ground communication delay.

[0062] When a satellite is visible to a ground station, it can establish a connection with the ground station. In addition, since the satellite is orbiting according to a given orbit, the connection and disconnection moments between the satellite and the ground station can be accurately predicted. This is achieved through the following communication model:

[0063] Assume the satellite trajectory is , the ground station trajectory is When the ground station g is at the minimum elevation angle When satellite s can be seen within .

[0064] in and It refers to the satellite trajectory and ground station trajectory in the geocentric inertial coordinate system (the geocentric inertial coordinate system refers to the coordinate system that remains stationary or in uniform linear motion (no acceleration) in space. Its origin is at the center of the earth, the Z axis coincides with the earth's rotation axis, and the X and Y axes are in the equatorial plane, pointing to the prime meridian and the meridian of 90° east longitude respectively), where t is a continuous wall clock time. Represents the trajectory difference between the ground station trajectory and the satellite trajectory and the angle between the ground station trajectory and the satellite trajectory. and The equal sign between them also represents the angle between the satellite and the ground station. Represents the minimum elevation angle of the ground station, so the above formula means that the angle between the satellite and the ground station is less than or equal to , that is, the ground station can see satellite s within the minimum elevation angle.

[0065] The FL task of the ground station is divided into two steps: (1) aggregate local model parameters / gradients from the satellite and update the global model when the update conditions are met; (2) send the global model to the satellite. When the ground station receives the local model from the satellite, it needs to determine whether the update conditions are met. If the conditions are met, the ground station will use all the local models collected in this round to update the global model. Otherwise, the ground station will continue to wait for the satellite to upload local updates. The above process can be expressed as follows:

[0066]

[0067] Where s represents the satellite subscript, represents the satellite index set of the local model returned in this round, |Ds| is the number of records in the private dataset Ds, It represents the sum of the number of records in the datasets of all satellites in the satellite index set that returns the local model in this round. a(B) is the aggregation index (a(B)=1 indicates that the ground station can perform local update aggregation). After the ground station updates the global model, it coordinates the satellites for the next round of training and sends the latest global model.

[0068] The embodiments of the present invention have the following advantages:

[0069] 1. Synchronous Dynamic Aggregation (SDA) combines synchronous periodic aggregation and synchronous buffered aggregation. It aggregates the local model once within a pre-set time period, thereby shortening the long update wait time caused by sparse satellite connections. In the case of dense satellite connections, the synchronous buffered aggregation strategy fully utilizes the large number of satellites that are idle when satellite connections are dense, fully utilizing the satellite computing resources.

[0070] 2. Such as Figure 2 As shown, the synchronous dynamic aggregation (SDA) proposed in this patent can update the first global model faster, thereby improving the robustness of the model (where Non-IID represents the non-independent and identically distributed setting, and IID represents the independent and identically distributed setting).

[0071] 3. If Figure 3 、 4 The data shows that Synchronous Dynamic Aggregation (SDA) outperforms other methods in terms of the training time required to reach the target accuracy under different dataset distribution settings. Specifically, in the IID (independent and identically distributed) setting, SDA achieves 92% accuracy in just 6.10 hours (4 times less than FedBuff and 10 times less than FedAvg). In the non-IID (non-independent and identically distributed) setting, SDA achieves 78% accuracy in 12.40 hours (2 times less than FedBuff and 5 times less than FedAvg).

[0072] 4. Such as Figure 5As shown, Synchronous Dynamic Aggregation (SDA) achieves more global model updates and more aggregated gradients than Synchronous Periodic Aggregation (SPA). Furthermore, SDA achieves a significantly higher number of global model updates (approximately 12%) than FedBuff, while slightly lower number of aggregated local gradients (approximately 1%). SDA, employing a dynamic aggregation strategy, increases the number of global model updates or aggregated gradients, effectively improving the model's prediction accuracy.

[0073] The above disclosure is only a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A dynamic aggregation federated learning method based on the connection density between satellites and ground stations, characterized in that: The following steps are involved: S1: i When the satellite establishes a connection with the ground station for the first time, the ground station sends the global model w i ; S2: The satellite trains the model locally and returns the local model to the ground station when it connects next time; S3: When the number of local models collected exceeds the buffer capacity V or to a predefined time period T At 0, the ground station updates the global model using a weighted average method; The updating of the global model is expressed as: In the formula, B represents the satellite index set returned to the local model in this round, V is the buffer capacity of the buffer, is the aggregation index, where The subscript SDA is Synchronous Dynamic Aggregation The abbreviation of , which means synchronous dynamic aggregation; =1, indicating that the ground station can perform local update aggregation; Said S3 further comprises the steps of: S31: The ground station aggregates local model parameters / gradients from the satellite and updates the global model when the update condition is met. The expression is: Where s represents the satellite subscript, represents the satellite index set of the local model returned in this round, |Ds| is the number of records in the private dataset Ds, It represents the sum of the number of records in the datasets of all satellites in the satellite index set returned to the local model in this round, and a(B) is the aggregation index; S32: After the ground station updates the global model, it coordinates the satellite to perform the next round of training and sends the latest global model.

2. The method for dynamic aggregation federated learning based on satellite and ground station connection density according to claim 1, characterized in that: The weighted average method includes weighted summing the loss function of the global model on the local data set on each satellite to obtain the final loss function : Where |Ds| is the number of records in the private dataset Ds, s represents the satellite subscript, and S represents the set of satellites contained in the FL system; , Is the global model in the local dataset The loss function on , ℓ(w, x) is the training loss of model w for data sample x.

3. The method for dynamic aggregation federated learning based on satellite and ground station connection density according to claim 2, characterized in that: The S2 specifically includes the steps of: After receiving the global model, the satellite will use the global model to initialize the local model, and then use the local dataset to train the model: Where η is the learning rate, the subscript of w is the local training index, indicating the round of local training, and the superscript is the satellite index. represents the local model of the ith orbit of satellite s, represents the local model of the satellite s’s i+1th round, Using a local dataset The loss function obtained from the last i-th round of training.

4. The method for dynamic aggregation federated learning based on satellite and ground station connection density according to claim 3, characterized in that: The satellite trains a local model for multiple epochs using a local dataset to reduce the number of communication rounds required for convergence.

Images