A satellite federated edge learning method based on satellite-ground collaborative transmission
By adopting the methods of coordinated transmission of satellites and ground and laser inter-star link aggregation in low-orbit satellite networks, the problems of short communication windows and sporadic satellite links in federated edge learning are solved, and efficient model training and accuracy improvement are achieved .
Patent Information
- Application Number
- CN202410015881.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-04
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-01-04
AI Technical Summary
The short communication window of federal edge learning in low-orbit satellite networks and sporadic satellite link problems lead to low model training efficiency. The existing solutions ignore the topological characteristics and collaborative transmission mechanism of satellite federated edge learning.
Adopt satellite federated edge learning method based on satellite-ground collaborative transmission, and perform multiple rounds of local model training and laser inter-star link aggregation in orbit, and efficient coordinated transmission of global models is combined with network flow algorithms to reduce the use of satellite-ground links. and communication delay.
It improves the efficiency and accuracy of model training, reduces convergence time, reduces communication costs, and effectively utilizes satellite computing capabilities.
Smart Images

Figure CN117914382B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wireless communication technology, and in particular to a satellite federated edge learning method based on satellite-ground collaborative transmission. Background Art
[0002] As 5G communication technology enters the commercial stage and is widely deployed worldwide, especially as we enter the 6G era, low-Earth orbit (LEO) satellite networks have become a key research direction and are being actively deployed around the world. Examples include Starlink, Kuiper, and China's recently announced LEO constellation project. These projects aim to leverage the advantages of low latency and high coverage, providing new solutions for remote sensing, geographic information systems, disaster monitoring, and other fields.
[0003] In addition to communication capabilities, modern low-orbit satellites are typically equipped with a variety of payloads, including computing and remote sensing. Simultaneously, the rapid development of various satellite-based computing hardware, such as onboard CPUs, GPUs, and other in-orbit computing chips, has led to a continuous increase in onboard computing capabilities. Furthermore, thanks to the low altitude of low-orbit satellites and the maturity of remote sensing technologies such as synthetic aperture radar and multispectral sensing, coupled with the explosive growth in the number of low-orbit satellites, a large amount of high-resolution remote sensing data is being collected on board, undoubtedly containing enormous value for information extraction and utilization. As the most important data mining method currently in use, machine learning technology has been widely used in remote sensing data processing, including land cover classification, cloud detection and precipitation estimation, oil spill detection, and many other applications.
[0004] However, the traditional approach to training machine learning models on large amounts of remote sensing data is to transmit the raw data from satellites to ground stations and perform further processing in cloud computing centers. However, with the dramatic increase in onboard data volumes and the limited and expensive spectrum resources between satellites and the ground, this approach incurs significant communication costs. Furthermore, the use of raw onboard data raises serious privacy concerns, particularly regarding the use of high-resolution satellite imagery, which is subject to strict regulatory restrictions.
[0005] Given the challenges of communication bottlenecks and data privacy, leveraging satellite computing power for on-orbit model training is becoming a major trend. Federated edge learning is a new type of collaborative training of global models by multiple edge devices without requiring any raw data to leave their devices. The edge device first trains the local model using various optimization methods, and then transmits the local model parameters to the parameter server, which aggregates the received model parameters and broadcasts them back to all edge devices, creating an iterative model training process. Although the deployment of federated edge learning in terrestrial networks has been widely studied and the solutions are relatively mature, the deployment of federated edge learning in large low-orbit constellation networks faces unique challenges, mainly manifested in the extremely short communication windows caused by the high mobility of satellites and sporadic satellite-to-ground connections (satellite-to-ground links).
[0006] To address this issue, most existing work focuses on designing asynchronous mechanisms to reduce the latency of satellite-to-ground links, which results in slower global model convergence. Another research direction is to leverage mature laser intersatellite link technology for satellite relay to reduce latency during satellite-to-ground link outages. However, these studies ignore the topological characteristics of satellite federated edge learning and lack a coordinated transmission mechanism between satellite and ground. Summary of the Invention
[0007] The purpose of this invention is to further improve the efficiency of satellite federated edge learning and effectively solve problems such as short communication windows and sporadic satellite-to-ground links.
[0008] The technical solution of the present invention is to provide a satellite federated edge learning method based on satellite-ground coordinated transmission, comprising the following steps:
[0009] Step 1: At the beginning of training, the k-th satellite in the m-th orbit uses the local dataset corresponding to the k-th satellite and performs multiple rounds of local model training starting from the same starting point, namely:
[0010]
[0011] Among them, the local model of round r can be expressed as η is the learning rate, For the rth round of global iteration and the tth round of orbital iteration, the eth local training is based on The stochastic gradient value of ;
[0012] Step 2: After completing local training, Using the orbital model aggregation algorithm based on ring full reduction of inter-satellite laser links, efficient aggregation is performed between satellites in the same orbit to obtain the average model within the orbit. Right now:
[0013]
[0014] Among them, K m is the number of satellites in each orbit;
[0015] Step 3: The local model of each satellite in orbit m is the average model A satellite-ground collaborative transmission mechanism based on network flow is used to aggregate the global model. After the global model is updated, the ground signal station broadcasts it back to all satellites to start a new round of global iteration.
[0016] Preferably, in step 1, the stochastic gradient descent method is used to perform multiple rounds of local model training starting from the same initial point.
[0017] Preferably, the intra-track model aggregation algorithm comprises the following steps:
[0018] Step 201: Determine the number of satellites in orbit K m and the dimension d of the model parameters of federated edge learning;
[0019] Step 202: The satellite in orbit m divides the local model into 2K equal parts m model blocks of equal size, i.e. The model nugget is represented as:
[0020]
[0021] Where u=0,1,...,K m -1, Indicates rounding up;
[0022] Step 203: In the i-th iteration, the k-th satellite in orbit m will Send to (k+1)%K m Satellites will Send to (k-1)%K m Satellites, where % means remainder;
[0023] When 0≤i <K m When -1, the model block update criterion for each satellite is:
[0024]
[0025]
[0026] When K m -1≤i<2K m When -2, the model block update criteria for each satellite is:
[0027]
[0028]
[0029] After i iterations, the average model within the track is obtained
[0030] Preferably, the satellite-ground coordinated transmission mechanism includes the following steps:
[0031] Step 301: construct a network flow problem, and construct a graph including a node part and an edge part in each time slot in the network flow problem;
[0032] Step 302: The ground parameter server solves the network flow problem in step 1 using the Ford-Fulkerson algorithm to obtain a quantitative model of satellite-ground coordinated transmission in the current time slot and informs the satellites establishing satellite-ground links with the ground signal station of the model ratio required for transmission.
[0033] Step 303: In different time slots, the ground parameter server performs dynamic coordination according to steps 1 and 2 based on the current satellite-to-ground link status, thereby achieving aggregation of the global model.
[0034] The technical solution of the present invention proposes a satellite federated edge learning method based on satellite-ground collaborative transmission, which achieves the purpose of keeping data local through multiple rounds of local model training on the satellite, and adds a method of averaging multiple intra-orbit model rounds in the global round to reduce the use of satellite-ground links with high overhead, short windows, and long waiting times, while accelerating model convergence. Taking advantage of the higher data transmission rate and stability of intra-orbit laser inter-satellite links compared to satellite-ground links, a method of adding multiple intra-orbit training rounds in the global round is proposed to reduce the use of satellite-ground links. Utilizing the stable ring topology structure formed by laser inter-satellite links in the same low-orbit satellite orbit, the inventors proposed an efficient intra-orbit aggregation communication scheme based on the ring full reduction algorithm. In order to quickly perform global model aggregation, a satellite-ground efficient collaborative model transmission mechanism based on network flow is also proposed to further reduce the total communication delay. The architecture and communication mechanism design for satellite federated edge learning effectively reduces the convergence time of satellite federated edge learning model training and improves the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0036] Figure 1 A schematic diagram of a satellite federated edge learning method based on satellite-ground coordinated transmission provided by one embodiment of the present invention;
[0037] Figure 2A system architecture diagram of a satellite federated edge learning method based on satellite-ground coordinated transmission provided by one embodiment of the present invention;
[0038] Figure 3 A flow chart of an intra-orbit model aggregation algorithm for ring full reduction based on inter-satellite laser links provided in one embodiment of the present invention;
[0039] Figure 4 A schematic diagram of an algorithm for a satellite-ground efficient collaborative model transmission mechanism based on network flow provided in one embodiment of the present invention;
[0040] Figure 5 This is a comparison chart of the learning performance of the method proposed in the present invention, the satellite federated edge learning method based on hybrid local gradient descent, and the satellite federated edge learning method based on federated averaging on the EuroSAT satellite remote sensing dataset. DETAILED DESCRIPTION
[0041] The present invention will be further described below in conjunction with specific examples. It should be understood that these examples are only intended to illustrate the present invention and are not intended to limit the scope of the present invention. In the following embodiments, many detailed descriptions are provided to enable the present application to be better understood. However, those skilled in the art will readily appreciate that some of the features may be omitted under different circumstances, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the present application are not shown or described in the specification. This is to prevent the core of the present application from being overwhelmed by excessive descriptions, and for those skilled in the art, it is not necessary to describe these related operations in detail. They can fully understand the related operations based on the description in the specification and the general technical knowledge in the art. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art may make various changes or modifications to the present invention, and these equivalents also fall within the scope defined by the appended claims of the present application.
[0042] like Figure 1 As shown, in this embodiment, a satellite federated edge learning method based on satellite-ground coordinated transmission includes the following steps:
[0043] like Figure 2 As shown in the figure, step 1, on-board local training: At the beginning of training, each satellite (the kth satellite in the mth orbit) uses its local dataset and performs multiple rounds of local model training starting from the same initial point through an optimization algorithm (such as stochastic gradient descent), namely:
[0044]
[0045] Among them, the local model of round r can be expressed as η is the learning rate, For the rth round of global iteration and the tth round of orbital iteration, the eth local training is based on The random gradient value of .
[0046] Step 2: In-orbit model aggregation: After completing local training, in order to obtain the average model of all satellites in each low-orbit satellite orbit as the starting point for the next round of local training to improve the performance and generalization ability of the model, it is planned to use high-speed laser intersatellite links to efficiently aggregate between satellites in the same orbit to obtain the average model in the orbit.
[0047] At this time, the local model of the kth satellite in the mth orbit is To obtain the average model within the orbit, we can make full use of the ring topology formed by the stable high-speed laser intersatellite link in each orbit, and use the ring full reduction to perform efficient aggregation between satellites in the same orbit to obtain the average model within the orbit. Right now
[0048]
[0049] Among them, K m is the number of satellites in each orbit.
[0050] like Figure 3 As shown, the intra-orbit model aggregation algorithm based on the Ring All-Reduce of the inter-satellite laser link includes the following steps:
[0051] Step 201, parameter initialization: First, determine the number of satellites in orbit K m and parameters such as the dimension d of the model parameters of the trained federated edge learning.
[0052] Step 202: Model block division: All satellites in track m divide their local models into 2K blocks. m model blocks of equal size, i.e. And these blocks can be represented as:
[0053]
[0054] where u=0,1,...,K m -1, Indicates rounding up.
[0055] Step 203, iterative update of model blocks: In each iteration, each satellite uses a full-duplex laser intersatellite link to simultaneously send a specific model block to the two satellites in front and behind in the orbital direction, and receives the model blocks sent to it by these two satellites, thereby updating its own local model. That is, in the i-th iteration, the k-th satellite in orbit m will Send to (k+1)%Km Satellites will Send to (k-1)%K m Satellites, where % means remainder. When 0≤i <K m When -1, the model block update criterion for each satellite is:
[0056]
[0057] When K m -1≤i<2K m When -2, the model block update criteria for each satellite is:
[0058]
[0059] After i=2K m -2 iterations, the local model of each satellite in the orbit is updated to the average of the original local models of all satellites in the orbit, that is, the average model in the orbit is obtained
[0060] Step 3: Global model aggregation and broadcasting: After multiple rounds (T rounds) of model aggregation within the orbit, the local model of each satellite in orbit m is (i.e., average model). At this time, in order to aggregate the in-orbit average models of all low-orbit satellite orbits through the link between satellites and ground signal stations, a satellite-ground collaborative transmission mechanism based on network flow is adopted to complete the aggregation of the global model.
[0061] The satellite-ground collaborative transmission mechanism based on network flow includes the following steps:
[0062] Step 301: Construct a network flow problem: To maximize the number of model parameters received by the ground parameter server in each window, a network flow problem is constructed. In this problem, the following graph needs to be constructed for each time slot:
[0063] In the node part, there is a source node s, a sink node D, and a track node O for each track m. m , set up a satellite node S for each satellite m,k , set up a ground station node GS for each ground signal station g g; In terms of edges, there is an edge pointing to the latter between the source node and each orbital node. In terms of edges, there is an edge pointing to the latter between the source node s and each orbital node, with the weight being the average proportion of the orbital model of the orbit that has not been transmitted to the ground before the start of the current time slot. There is an edge pointing to the latter between each orbital node and each satellite node in the orbit, with a weight of 1. There is an edge pointing to the latter between each satellite node and each ground station node, with a weight being the proportion of the total amount of data that can be transmitted by the satellite-to-ground link in the current time slot divided by the model size. There is an edge pointing to the former between the sink node D and each ground signal station node, with a weight being the proportion of the total amount of data that can be transmitted by the link between the ground station and the ground parameter server in the current time slot divided by the model size, as shown in the figure. Figure 4 shown.
[0064] Step 302: Solve the network flow problem: The ground parameter server uses the Ford–Fulkerson algorithm to solve the network flow constructed in Step 1. This solves the network flow model for satellite-ground coordinated transmission in the current time slot and informs the satellites establishing the satellite-ground link with the ground signal station of the required transmission ratio. The key idea of the Ford–Fulkerson algorithm is as follows: As long as there is an augmenting path from source node s to sink node D, a flow is sent along that path, with the flow rate limited by the minimum capacity on the path. Another path is then found until no such path exists in the entire network.
[0065] Step 303, dynamic adjustment: In different time slots, the ground parameter server needs to dynamically coordinate the satellite-ground collaborative transmission scheme according to steps 1 and 2 based on the current satellite-ground link status, thereby achieving global model aggregation.
[0066] After the global model is updated, the ground signal station broadcasts it back to all satellites to start a new global iteration round.
[0067] like Figure 5 As shown, Figure 5 The learning performance of an embodiment of the present invention is compared with two other satellite federated edge learning methods (a satellite federated edge learning method based on hybrid local gradient descent, and a satellite federated edge learning method based on federated averaging) on the EuroSAT satellite remote sensing dataset. Figure 5 (a) is a schematic diagram of accuracy. Figure 5 (b) in the figure is a schematic diagram of the loss function value. As can be seen from the figure, compared with the two comparison methods, the proposed method can achieve lower loss function values and higher test accuracy at the same system latency. In other words, it has a faster convergence speed. Therefore, the number of global iterations required to achieve the same loss function value or test accuracy is smaller, achieving high-communication-efficiency satellite federated edge learning.
[0068] An embodiment of the present invention proposes an efficient federated edge learning framework for low-Earth orbit mega-constellation networks. Taking advantage of the higher data transmission rate and stability of intra-orbital laser inter-satellite links compared to satellite-to-ground links, a method is proposed to add multiple intra-orbital training rounds to the global round to reduce the use of satellite-to-ground links. Leveraging the stable ring topology formed by laser inter-satellite links within the same low-Earth orbit, an efficient intra-orbital aggregation communication scheme is proposed based on a ring full reduction algorithm. To rapidly aggregate global models, a network flow-based efficient satellite-to-ground collaborative model transmission mechanism is also proposed, further reducing the overall communication delay.
Claims
1. A satellite federated edge learning method based on satellite-ground collaborative transmission, characterized in that: The following steps are involved: Step 1: At the beginning of training, the kth satellite in the mth orbit uses the local data set corresponding to the kth satellite, starting from the same starting point, to perform multiple rounds of local model training, namely: Among them, the local model of the rth round can be expressed as η is the learning rate, The e-th local training is based on the r-th global iteration and the t-th orbital iteration. The random gradient value of ; Step 2: After completing local training, use the local model Using the orbital model aggregation algorithm based on the ring reduction of inter-satellite laser links, satellites in the same orbit are efficiently aggregated to obtain the average model in the orbit. Right now: Among them, K m is the number of satellites in each orbit; The intra-track model aggregation algorithm includes the following steps: Step 201: Determine the number of satellites K in orbit m and the dimension d of the model parameters of federated edge learning; Step 202: The satellite in orbit m divides the local model into 2K equal parts m equal-sized model blocks, i.e. The model nugget is represented as: Where u=0,1,...,K m -1, Indicates rounding up; Step 203: In the i-th iteration, the k-th satellite in orbit m will Send to (k+1)%K m satellites, will Send to (k-1)%K m satellites, where % means remainder; When 0≤i <K m When -1, the model block update criterion for each satellite is: When K m -1≤i<2K m When -2, the model block update criterion for each satellite is: After i iterations, the average model within the track is obtained Step 3: The local model of each satellite in orbit m is the average model A satellite-ground collaborative transmission mechanism based on network flow is used to aggregate the global model; The satellite-ground coordinated transmission mechanism comprises the following steps: Step 301: construct a network flow problem, and construct a graph including a node part and an edge part in each time slot in the network flow problem; Step 302: The ground parameter server solves the network flow problem in step 1 by using the Ford-Fulkerson algorithm to obtain a quantitative model of satellite-ground coordinated transmission in the current time slot and informs the satellite that establishes a satellite-ground link with the ground signal station of the model ratio required for transmission; Step 303: In different time slots, the ground parameter server performs dynamic coordination according to steps 1 and 2 based on the current satellite-to-ground link status, thereby achieving aggregation of the global model; After the global model is updated, the ground signal station broadcasts it back to all satellites to start a new round of global iterations.
2. A satellite federated edge learning method based on satellite-ground collaborative transmission as claimed in claim 1, characterized in that: In step 1, the stochastic gradient descent method is used to perform multiple rounds of local model training starting from the same initial point.
