Decentralization federated learning method and device for low earth orbit satellites, and storage medium

Abstract

The invention provides a decentralized federated learning method and device for a low earth orbit satellite and a storage medium, and relates to the technical field of satellite communication, and the method comprises the steps: employing a dual-stage adaptive model aggregation mechanism, employing an in-orbit aggregation algorithm in the first stage; in the second stage, cross-orbit plane propagation is adopted, and the propagation turns are dynamically adjusted through a multi-turn Gossip protocol diffusion model to balance the convergence speed and the communication cost; a model self-compensation mechanism is adopted, and for the problem of cross-orbit inter-satellite link packet loss, a receiver compensates lost data packets by using local parameters, so that retransmission requirements are reduced, and robustness and communication efficiency are improved. The low earth orbit satellite-oriented decentration federated learning method provided by the invention is superior to the existing decentration federated learning method in the aspects of convergence performance, communication efficiency and link robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite communication, and specifically, to a decentralized federated learning method, device, and storage medium for low-earth orbit satellites. Background Art

[0002] Traditional centralized federated learning (CFL) algorithms rely on ground stations (GSs) as central servers. However, the communication windows between low-earth orbit (LEO) satellites and GSs are short and irregular, resulting in slow convergence speed and low efficiency. Moreover, GS failures can lead to the paralysis of the entire system.

[0003] In addition, existing decentralized federated learning (DFL) algorithms do not consider the characteristics of satellite networks. Most studies assume ideal communication conditions and ignore the dynamic topology and link instability of actual LEO constellations. Summary of the Invention

[0004] Aiming at the defects in the prior art, the purpose of the present invention is to provide a decentralized federated learning method, device, and storage medium for low-earth orbit satellites, aiming to solve the problems of low efficiency, unreliable communication, and resource constraints in model training in satellite networks.

[0005] To solve the above problems, the technical solution of the present invention is as follows:

[0006] A decentralized federated learning method for low-earth orbit satellites, comprising the following steps:

[0007] Adopt a two-stage adaptive model aggregation mechanism. In the first stage, an in-orbit aggregation algorithm is adopted;

[0008] In the second stage, cross-orbit plane propagation is adopted. The model is diffused through multiple rounds of the Gossip protocol, and the propagation rounds are dynamically adjusted to balance the convergence speed and communication cost;

[0009] Adopt a model self-compensation mechanism. For the problem of packet loss in inter-satellite links across orbits, the receiver uses local parameters to compensate for lost data packets, reducing the need for retransmission and improving robustness and communication efficiency.

[0010] Preferably, in the step of adopting a two-stage adaptive model aggregation mechanism and adopting an in-orbit aggregation algorithm in the first stage, ring all-reduce is used to achieve efficient synchronization of satellites in the same orbit plane. The communication overhead is independent of the number of satellites in the plane and has strong scalability.

[0011] Preferably, the step of adopting a two-stage adaptive model aggregation mechanism and adopting an in-orbit aggregation algorithm in the first stage specifically includes:

[0012] The execution process of the ring all-reduce algorithm on three satellite nodes S0, S1, and S2 is divided into the Reduce-Scatter stage and the All-Gather stage.

[0013] Initial setup: Each node holds chunk data. S0 has [A0, A1, A2], S1 has [B0, B1, B2], and S2 has [C0, C1, C2].

[0014] Reduce-Scatter stage:

[0015] First round of communication: S0 sends A1 to S1 and receives C0 from S2; S1 sends B2 to S2 and receives A1 from S0; S2 sends C0 to S0 and receives B2 from S1. Each node accumulates the received data.

[0016] Second round of communication: S0 sends A2 to S2 and receives B0 + A1 from S1; S1 sends B0 + A1 to S0 and receives C1 + B2 from S2; S2 sends C1 + B2 to S1 and receives A2 from S0. Eventually, each node holds partial global sums.

[0017] All-Gather stage:

[0018] Third round of communication: S0 sends A0 + B0 + C0 to S1 and receives A2 + B2 + C2 from S2; S1 sends A1 + B1 + C1 to S2 and receives A0 + B0 + C0 from S0; S2 sends A2 + B2 + C2 to S0 and receives A1 + B1 + C1 from S1.

[0019] Fourth round of communication: S0 sends A2 + B2 + C2 to S1 and receives A1 + B1 + C1 from S1; S1 sends A1 + B1 + C1 to S0 and receives A2 + B2 + C2 from S2; S2 sends A1 + B1 + C1 to S1. All nodes obtain the complete global sum.

[0020] Final calculation: If average gradients / parameters are required, each node divides the global sum by 3.

[0021] Preferably, the step of adopting the model self-compensation mechanism to address the packet loss problem in the inter-satellite link across orbits, where the receiver uses local parameters to compensate for lost data packets, reduce the retransmission requirement, and improve robustness and communication efficiency specifically includes: The receiver first counts the missing data packets, then performs packet splitting on the local model parameters, and finally uses the corresponding local data packets to compensate for the missing data.

[0022] Furthermore, the present invention also provides a decentralized federated learning device for low-earth orbit satellites, including a processor and a memory for storing executable instructions of the processor. The processor is configured to execute the decentralized federated learning method for low-earth orbit satellites as described above by executing the executable instructions.

[0023] Furthermore, the present invention also provides a computer-readable storage medium for storing program code for executing the decentralized federated learning method for low-earth orbit satellites as described above.

[0024] Compared with the prior art, the decentralized federated learning method for low-earth orbit satellites of the present invention achieves efficient model synchronization through in-orbit aggregation (Orbit Reduce) and inter-orbit dissemination (Gossip Dissemination), and improves the robustness under unstable inter-satellite links by combining a self-compensation mechanism. It is superior to existing decentralized federated learning algorithms in terms of convergence performance, communication efficiency, and link robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Other features, objects, and advantages of the present invention will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0026] Figure 1 is a flowchart of the decentralized federated learning method for low-earth orbit satellites of the present invention;

[0027] Figure 2 is another flowchart of the decentralized federated learning method for low-earth orbit satellites of the present invention;

[0028] Figure 3 is a schematic diagram of the execution process of the Orbit Reduce algorithm of the present invention;

[0029] Figure 4 is a specific example diagram of multiple rounds of Gossip between M = 5 orbital planes;

[0030] Figure 5 is a schematic diagram of the satellite model Gossip self-compensation mechanism between M = 3 orbital planes. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0032] Specifically, the present invention provides a decentralized federated learning method DFedSat for low-earth orbit satellites, as Figure 1 and Figure 2 shown. The method includes the following steps:

[0033] S1: Adopt a two-stage adaptive model aggregation mechanism. In the first stage, use the in-orbit aggregation algorithm;

[0034] Specifically, the in-orbit aggregation (Orbit Reduce) algorithm: Based on ring all-reduce, it realizes the efficient synchronization of satellites in the same orbital plane. The communication overhead is independent of the number of satellites in the plane and has strong scalability. The execution process of the OrbitReduce algorithm is as Figure 3 shown, and the specific detailed process is as follows:

[0035] The execution process of the ring all-reduce algorithm on three satellite nodes (S0, S1, S2) can be divided into two main stages: Reduce-Scatter and All-Gather.

[0036] Initial setup: For clarity, we define the symbols as follows. Each node holds chunk data. For example, S0 has [A0, A1, A2], S1 has [B0, B1, B2], and S2 has [C0, C1, C2]. The goal is to calculate the global sum (such as gradient / parameter aggregation).

[0037] Reduce-Scatter stage:

[0038] First-round communication: S0 sends A1 to S1 and receives C0 sent by S2; S1 sends B2 to S2 and receives A1 sent by S0; S2 sends C0 to S0 and receives B2 sent by S1. Each node accumulates the received data. For example, S0 gets A0 + C0.

[0039] Second-round communication: S0 sends A2 to S2 and receives B0 + A1 sent by S1; S1 sends B0 + A1 to S0 and receives C1 + B2 sent by S2; S2 sends C1 + B2 to S1 and receives A2 sent by S0. Finally, each node holds a partial global sum. For example, S0 gets A0 + B0 + C0, S1 gets A1 + B1 + C1, and S2 gets A2 + B2 + C2.

[0040] All-Gather stage:

[0041] Third round of communication: S0 sends A0 + B0 + C0 to S1 and receives A2 + B2 + C2 sent by S2; S1 sends A1 + B1 + C1 to S2 and receives A0 + B0 + C0 sent by S0; S2 sends A2 + B2 + C2 to S0 and receives A1 + B1 + C1 sent by S1. At this time, each node has partially completed the global sum.

[0042] Fourth round of communication: S0 sends A2 + B2 + C2 to S1 and receives A1 + B1 + C1 sent by S1; S1 sends A1 + B1 + C1 to S0 and receives A2 + B2 + C2 sent by S2; S2 sends A1 + B1 + C1 to S1 (no need to receive). Finally, all nodes obtain the complete global sum [A0 + B0 + C0, A1 + B1 + C1, A2 + B2 + C2].

[0043] Final calculation: If average gradients / parameters are required, each node divides the global sum by 3. The entire process optimizes the bandwidth through ring communication to ensure efficient synchronization.

[0044] S2: In the second stage, cross-orbit plane propagation is adopted, and the model is diffused through multiple rounds of the Gossip protocol, dynamically adjusting the number of communication rounds to balance the convergence speed and communication cost;

[0045] Specifically, Figure 4 shows an example of model propagation between M = 5 orbit planes. In the figure represents the model parameters of the satellites on the m-th orbit plane in the C-th round of Gossip communication. Taking the red satellite on the orbit plane m = 3 as an example, in the first round of Gossip communication c = 1, its model parameters are:

[0046]

[0047] This parameter is composed of the model parameters of the other two orbit planes. And in the second round of Gossip communication c = 2, its model parameters are:

[0048]

[0049] This parameter is composed of the model parameters of all other orbit planes, which indicates that more rounds of Gossip communication can better approximate the average model, thereby enhancing model consistency.

[0050] S3: Adopt a model self-compensation mechanism. For the problem of packet loss in the cross-orbit inter-satellite link, the receiver uses local parameters to compensate for the lost data packets, reducing the need for retransmission and improving robustness and communication efficiency.

[0051] The decentralized federated learning method DFedSat for low-earth orbit satellites adopts a two-stage model synchronization mechanism. After orbit computing (Orbit Computing), that is, after the satellite performs stochastic gradient descent (SGD) updates locally, two-stage model synchronization is carried out, including in-orbit plane synchronization and inter-orbit plane propagation. To ensure efficient and reliable model aggregation in these processes, a self-compensation mechanism is introduced to cope with unstable inter-orbit communication. The model self-compensation mechanism is adopted to address the problem of packet loss in the inter-orbit satellite link. The receiver uses local parameters to compensate for lost data packets, reducing the need for retransmission and enhancing robustness and communication efficiency. Specifically: the receiver first counts the missing data packets, then slices the local model parameters into packets, and finally uses the corresponding local data packets to compensate for the missing data.

[0052] Figure 5 Figure 4 shows a schematic diagram of the satellite model Gossip self-compensation mechanism between M = 3 orbital planes. Taking SAT1 receiving parameter packets from SAT2 and SAT3 as an example, focus on the transmission process of the first four data packets. SAT1 detects anomalies when receiving the first packet from SAT2 and the fourth packet from SAT3, where m 2→1 = [0, 1, 1, 1], m 3→1 = [1, 1, 1, 0]. To solve these anomalies, SAT1 uses the first packet and the fourth packet of its own model for filling and compensation.

[0053] In summary, the DFedSat method of the present invention achieves efficient model synchronization through in-orbit aggregation (Orbit Reduce) and inter-orbit propagation (Gossip Dissemination), and combines the self-compensation mechanism to enhance the robustness under unstable inter-satellite links. The present invention conducts a comparative analysis of DFedSat with existing decentralized federated learning (DFL) algorithms (including decentralized stochastic gradient descent DSGD, decentralized DFedAvg, and DFedSAM based on sharpness-aware optimization) through numerical experiments. The performance of DFedSat in terms of convergence performance, communication efficiency, and link robustness is mainly evaluated. The experiments prove that in terms of convergence speed: DFedSat is superior to DSGD, DFedAvg, and DFedSAM, and the two-stage aggregation (in-orbit synchronization and inter-orbit propagation) of DFedSat can share model information more effectively. In terms of communication efficiency, the overhead is only 50% or even lower than that of the baseline algorithm. In terms of link robustness, DFedSat still maintains stable performance under low power (high packet loss rate).

[0054] Furthermore, the present invention also provides a decentralized federated learning device for low-earth orbit satellites, including a processor and a memory for storing executable instructions of the processor, and the processor is configured to execute the decentralized federated learning method for low-earth orbit satellites as described in the above embodiments by executing the executable instructions.

[0055] Furthermore, the present invention also provides a computer-readable storage medium, and the computer-readable storage medium is used to store program codes, and the program codes are used to execute the decentralized federated learning method for low-earth orbit satellites as described in the above embodiments.

[0056] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other.

Claims

1. A decentralized federated learning method for low-earth orbit satellites, characterized in that, The method includes the following steps: Adopt a two-stage adaptive model aggregation mechanism. In the first stage, an in-orbit aggregation algorithm is used; In the second stage, cross-orbit plane propagation is adopted. The model is diffused through multiple rounds of Gossip protocol, and the propagation rounds are dynamically adjusted to balance the convergence speed and communication cost; Adopt a model self-compensation mechanism. Aiming at the packet loss problem in the cross-orbit inter-satellite link, the receiver uses local parameters to compensate for the lost data packets, reduces the retransmission requirement, and improves the robustness and communication efficiency.

2. The decentralized federated learning method for low-earth orbit satellites according to claim 1, characterized in that In the step of adopting the two-stage adaptive model aggregation mechanism and using the in-orbit aggregation algorithm in the first stage, ring all-reduce is adopted to achieve efficient synchronization of satellites on the same orbit plane. The communication overhead is independent of the number of satellites in the plane and has strong scalability.

3. The decentralized federated learning method for low-Earth orbit satellites according to claim 2, wherein The step of adopting the two-stage adaptive model aggregation mechanism and using the in-orbit aggregation algorithm in the first stage specifically includes: The execution process of the ring all-reduce algorithm on three satellite nodes S0, S1, and S2 is divided into the Reduce-Scatter stage and the All-Gather stage. Initial setting: Each node holds chunk data. S0 has [A0, A1, A2], S1 has [B0, B1, B2], and S2 has [C0, C1, C2]; Reduce-Scatter stage: First round of communication: S0 sends A1 to S1 and receives C0 sent by S2; S1 sends B2 to S2 and receives A1 sent by S0; S2 sends C0 to S0 and receives B2 sent by S1. Each node accumulates the received data; Second round of communication: S0 sends A2 to S2 and receives B0 + A1 sent by S1; S1 sends B0 + A1 to S0 and receives C1 + B2 sent by S2; S2 sends C1 + B2 to S1 and receives A2 sent by S0. Finally, each node holds a partial global sum; All-Gather stage: Third round of communication: S0 sends A0 + B0 + C0 to S1 and receives A2 + B2 + C2 sent by S2; S1 sends A1 + B1 + C1 to S2 and receives A0 + B0 + C0 sent by S0; S2 sends A2 + B2 + C2 to S0 and receives A1 + B1 + C1 sent by S1; Fourth round of communication: S0 sends A2 + B2 + C2 to S1 and receives A1 + B1 + C1 sent by S1; S1 sends A1 + B1 + C1 to S0 and receives A2 + B2 + C2 sent by S2; S2 sends A1 + B1 + C1 to S1. All nodes obtain the complete global sum; Final calculation: If the average gradient / parameters are required, each node divides the global sum by 3.

4. The decentralized federated learning method for low-Earth orbit satellites according to claim 1, wherein The step of adopting the model self-compensation mechanism, aiming at the packet loss problem in the cross-orbit inter-satellite link, where the receiver uses local parameters to compensate for the lost data packets, reduces the retransmission requirement, and improves the robustness and communication efficiency specifically includes: The receiver first counts the missing data packets, then cuts the local model parameters into packets, and finally uses the corresponding local data packets to compensate for the missing data.

5. A decentralized federated learning device for low-earth orbit satellites, characterized in that, The device includes a processor and a memory for storing executable instructions of the processor, and the processor is configured to execute the decentralized federated learning method for low-earth orbit satellites according to any one of claims 1 to 4 by executing the executable instructions.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for executing the decentralized federated learning method for low-earth orbit satellites according to any one of claims 1 to 4.

Images