Satellite selection method based on multi-satellite resource equalization

Abstract

The invention belongs to the technical field of satellite communication, and relates to a satellite selection method based on multi-satellite resource equalization, which comprises the following steps that: each user i selects a satellite with the highest current signal strength for random access; each user i uploads a data set to the accessed satellite training global model; each user i calculates a utility value # imgabs0 # of the user i and sends the utility value # imgabs0 # to a ground controller, the ground controller judges whether a dynamic balance state is reached or not according to the utility value # imgabs1 #, and if the dynamic balance state is reached, the controller notifies all the users to reside on a currently accessed satellite; otherwise, the ground controller calculates an average utility value # imgabs3 # according to the utility value # imgabs2 # and broadcasts the utility values of # imgabs4 # and all accessed satellites j to all the users, and each user i re-selects the satellites for access and carries out the next round of global model training; according to the invention, appropriate satellites are dynamically selected for model updating and data transmission according to different requirements of satellite resources and federated learning, and balanced utilization of the satellite resources is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite communications, and in particular relates to a satellite selection method based on multi-satellite resource balancing. Background Art

[0002] In order to better respond to natural disasters, technicians usually use data collected by IoT sensors distributed in different isolated areas (such as forests, oceans and mountains) to train machine learning models. However, many IoT sensors used for natural disaster prediction are located in areas with poor communication infrastructure, with no ground base stations or network access points available. Therefore, satellite communications, as a seamless connection solution that can cover the world, have shown great potential in these areas.

[0003] Federated learning is a distributed machine learning paradigm that allows multiple devices to collaboratively train shared models while keeping data localized. This approach is particularly suitable for scenarios where data privacy is sensitive and communication is restricted, such as IoT devices in remote areas. With the assistance of satellite communications, federated learning FL can cross geographical boundaries, connect devices distributed around the world, jointly train models, and thus provide intelligent services.

[0004] However, under the coverage of low earth orbit satellites, each user is very likely to be covered by multiple satellites at the same time. When a user initiates an access request to a satellite, multiple satellites can be selected. However, since users do not know each other's information, it is very likely that a large number of users will choose the same satellite for FL tasks.

[0005] Therefore, how to dynamically select appropriate satellites for model updates and data transmission based on the satellite's resource status and the different requirements of the user's FL tasks, so as to efficiently and evenly utilize the computing resources of each satellite, is a complex decision-making problem. Summary of the invention

[0006] In order to solve the above-mentioned prior art problems, the present invention adopts a satellite selection method based on multi-satellite resource balancing, comprising:

[0007] S1: Each user i selects the satellite with the highest current signal strength for random access;

[0008] S2: Each user i uploads the local data set to the connected satellite to train the global model M t ;

[0009] S3: Each user i calculates the signal strength of the connected satellite, the size of the data set uploaded to the connected satellite, and the global model M t The training time is used to calculate the respective utility values And their respective utility values Sent to the ground controller; where t is the training round of the global model;

[0010] S4: Ground controller according to the utility value Determine whether a dynamic equilibrium state is reached, if so, execute step S5; otherwise, execute step S6;

[0011] S5: The ground controller notifies all users to stay on the currently connected satellite, completing the satellite selection;

[0012] S6: The ground controller calculates the utility value of each user Calculate the average utility value The utility value of each user i As the utility value of satellite j accessed by each user i The average utility value and the utility values ​​of all connected satellites j Broadcast to all users and execute step S7;

[0013] S7: Each user i calculates the current temperature T(t) according to the utility values ​​of all connected satellites j Average utility value And the current temperature T(t) selects a satellite for access, and executes step S2 to perform the next round of global model training.

[0014] The local data set uploaded by each user to the connected satellite is a non-sensitive data set; each user i uploads the local data set to the connected satellite to train the global model M t include:

[0015] S21: After the connection is established, each user i uploads a non-sensitive data set to the connected satellite;

[0016] S22: When each connected satellite successfully receives the non-sensitive data set, it sends a confirmation message to the corresponding user i;

[0017] S23: Each connected satellite trains a local model using the non-sensitive data sets received from multiple users;

[0018] S24: After each connected satellite has trained a local model, the trained local model is sent to one of the satellites for aggregation to obtain a trained global model, and the trained global model is sent back to each user i.

[0019] Calculate the utility value of user i include:

[0020]

[0021] Among them, I j is the signal strength of satellite j, is the size of the data set uploaded by user i, is the training time of the global model in the tth training round, and α, β, κ, ξ are parameter weights.

[0022] Ground controls according to utility value Determining whether a dynamic equilibrium state has been reached includes:

[0023] Set the threshold ε, if the utility value of all users in the previous round and the utility value of the current round If the difference is less than the threshold ε, a dynamic equilibrium state is reached.

[0024] The calculation of the current temperature of each user i includes: setting the annealing coefficient, obtaining the temperature T(t-1) of the previous training round, multiplying the temperature T(t-1) of the previous training round by the annealing coefficient, and obtaining the current temperature T(t); wherein the acquisition process of the initial temperature T(0) includes: the ground controller calculates the utility value of each user in the first training round The variance σ u , according to the variance σ u Calculate the initial temperature T(0) and broadcast the initial temperature T(0) to all users.

[0025] User i’s utility value based on all satellite j accesses Average utility value And the current temperature T(t) to select the satellite for access includes:

[0026] User i calculates his own utility value and average utility value difference If bad If it is greater than 0, the currently connected satellite will be retained; otherwise, it is determined whether there is a utility value among all connected satellites. Higher than the utility value of the satellite currently accessed by user i If there is a satellite, then filter out the utility value among all connected satellites Higher than the utility value of the satellite currently accessed by user i satellites, and obtain the satellite set S better , using a weighted random sampling method from the satellite set S better A satellite is selected for access; otherwise, a weighted random sampling method is used to select a satellite from all the accessed satellites, and the acceptance probability mechanism of the simulated annealing algorithm is used to determine whether to access the selected satellite.

[0027] Use weighted random sampling method to select from the satellite set S better Selecting a satellite from includes: calculating the satellite set S betterThe utility difference between each satellite in and the satellite currently accessed by user i is normalized to obtain the selection probability p of each satellite. j , according to the selection probability p j From S better Select a satellite from the navigation pane.

[0028] The acceptance probability mechanism of the simulated annealing algorithm is used to determine whether to access the selected satellite, including: calculating the acceptance probability P according to the utility difference of the selected satellite and the current temperature T(t) accept , generate a random number r∈[0,1], if r <P accept , user i is connected to the selected satellite; otherwise, user i stays on the currently connected satellite.

[0029] Calculate the acceptance probability P accept include:

[0030]

[0031] Among them, j′ is the selected satellite, ΔU j′ is the utility difference of the selected satellite j′, and exp() is an exponential function with the Euler number e as the base.

[0032] Beneficial effects of the present invention:

[0033] The present invention dynamically adjusts the satellite resource status and the different requirements of the user's federated learning task, namely the satellite signal strength, the size of the data set and the global model M. t The training time is shortened and appropriate satellites are selected for model update and data transmission, which avoids the resource overload problem caused by a large number of users selecting the same satellite and achieves balanced utilization of satellite resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of a system model of a satellite selection method based on multi-satellite resource balancing provided by an embodiment of the present invention;

[0035] Figure 2 A flowchart of a satellite selection method based on multi-satellite resource balancing provided by an embodiment of the present invention;

[0036] Figure 3 A flowchart of four-step random access provided by an embodiment of the present invention;

[0037] Figure 4 A flowchart of a user's decision on whether to replace a satellite for the next round of global model training is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The framework of the satellite system model is as follows Figure 1 As shown in the figure, the whole system consists of three parts: low-orbit satellite S, user terminal UE and controller (Controller); among them, there is two-way communication between the user terminal and the low-orbit satellite, and there is two-way communication between the user terminal and the controller through the Low-Power Wide-Area Network (LPWAN), and the controller only communicates with each user terminal; LPWAN uses LoRa (Long Range) technology.

[0040] A communication scenario consisting of N low-orbit satellites, K ground terminal users and 1 ground controller is constructed. A wildfire prediction model is trained through collaborative federated learning between satellites and ground terminal users. Ground terminal users will be distributed in different isolated areas (such as forests and mountains) to collect natural environment data, including air temperature, wind speed, wind direction and soil moisture.

[0041] Based on the above communication scenario, the embodiment of the present invention proposes a satellite selection method based on multi-satellite resource balancing, such as Figure 2 As shown, including:

[0042] S1: At the beginning of the first round of global model training, each user i selects the satellite with the highest current signal strength for random access;

[0043] In the communication scenario, each ground terminal user will be covered by n low-orbit satellites at a certain moment (n∈N). When conducting the first round of global model training, the user terminal will select the satellite with the highest current signal strength and use Figure 3 The four-step random access method shown establishes a connection with the satellite;

[0044] The process of establishing a connection with a satellite using four-step random access includes:

[0045] S11: The ground terminal user randomly selects a preamble sequence (Mg1) and sends it to the satellite through the physical random access channel (PRACH) to request to establish a communication connection;

[0046] S12: After receiving the request, the satellite sends a response message (Msg2) to the ground terminal user through the physical downlink shared channel (PDSCH), including uplink resource allocation and timing adjustment information, so that the ground terminal user can perform uplink synchronization;

[0047] S13: The ground terminal user sends a connection establishment request (Msg3) to the satellite via the physical uplink shared channel (PUSCH) according to the resource allocation in the response message, which contains the necessary connection information and user data;

[0048] S14: After receiving the connection establishment request, the satellite verifies and processes it, and then sends a connection establishment confirmation message (Msg4) to the ground terminal user through the physical downlink shared channel (PDSCH), thereby successfully establishing a communication connection.

[0049] S2: Each user i uploads the local data set to the connected satellite to train the global model M t ;

[0050] Upload the local dataset to the satellite to train the global model M t include:

[0051] S21: After the connection is established, the user uploads the non-sensitive data set to the connected satellite;

[0052] The total amount of data samples that user i needs to process is D i , where the non-sensitive data sets that need to be unloaded to the satellite are Right now The remaining data is updated by user i for local model update. Each user needs to determine the non-sensitive data set that needs to be unloaded to the satellite before training begins. After the connection is established, the user terminal unloads the non-sensitive data set to satellite j for model training.

[0053] S22: When each connected satellite successfully receives the non-sensitive data set, it sends a confirmation message to the corresponding user i;

[0054] S23: Each connected satellite trains a local model using the non-sensitive data sets received from multiple users;

[0055] S24: After each connected satellite has trained a local model, the trained local model is sent to one of the satellites for aggregation to obtain a trained global model, and the trained global model is sent back to each user i.

[0056] S3: Each user i calculates the signal strength of the connected satellite, the size of the data set uploaded to the connected satellite, and the global model M t The training time is used to calculate the respective utility values And their respective utility values Sent to the ground controller; where t is the training round of the global model;

[0057] The process of calculating the utility value of the user terminal includes:

[0058] After receiving the updated global model downloaded from the satellite, user i calculates its own utility value using the following formula:

[0059]

[0060] Among them, I j is the signal strength of satellite j, is the size of the non-sensitive data set uploaded by user i, is the training time of the global model in the tth training round, α and β are positive constants of parameter weights, which are used to adjust the contribution of each part, and κ and ξ are small positive numbers to prevent the denominator from being zero.

[0061] S4: Ground controller according to the utility value Determine whether a dynamic equilibrium state is reached, if so, execute step S5; otherwise, execute step S6;

[0062] The ground controller determines whether the dynamic balance state is reached based on the utility value, including:

[0063] Set a threshold ε, preferably ε is 0.05; if the utility value of all users in the previous round and the utility value of the current round The differences are all less than the threshold ε:

[0064]

[0065] This indicates that a dynamic equilibrium state has been reached.

[0066] S5: The ground controller notifies all users to stay on the currently connected satellite, completing the satellite selection;

[0067] S6: The ground controller calculates the utility value of each user Calculate the average utility value The utility value of each user i As the utility value of satellite j accessed by each user i The average utility value and the utility values ​​of all connected satellites j Broadcast to all users and execute step S7;

[0068] The ground controller calculates the average utility value include:

[0069]

[0070] Where N is the total number of users, and t is the current training round;

[0071] S7: Each user i calculates the current temperature T(t) and calculates the utility value of all connected satellites j Average utility value And the current temperature T(t) selects a satellite for access, and executes step S2 to perform the next round of global model training.

[0072] like Figure 4 As shown in the figure, user i has the utility value of all satellite j accessed Average utility value And the current temperature T(t) selects the satellite to access including:

[0073] User i calculates his own utility value and average utility value difference If bad If it is greater than 0, the currently connected satellite will be retained; otherwise, it is determined whether there is a utility value among all connected satellites. Higher than the utility value of the satellite currently accessed by user i If there is a satellite, then filter out the utility value among all connected satellites Higher than the utility value of the satellite currently accessed by user i satellites, and obtain the satellite set S better , using a weighted random sampling method from the satellite set S better A satellite is selected for access; otherwise, a weighted random sampling method is used to select a satellite from all the accessed satellites, and the acceptance probability mechanism of the simulated annealing algorithm is used to determine whether to access the selected satellite.

[0074] Use weighted random sampling method to select better Selecting a satellite includes:

[0075] Calculate S better The utility difference ΔU of each candidate satellite j in the set j :

[0076]

[0077] Normalize the utility difference of each candidate satellite to obtain the selection probability p of each candidate satellite j :

[0078]

[0079] Where s is S better The index of the satellite in the collection.

[0080] According to the selection probability p j From S better Select a satellite to conduct the next round of global model training.

[0081] The acceptance probability mechanism of the simulated annealing algorithm is used to determine whether to access the selected satellite for global model training, including:

[0082] According to the utility difference ΔU of the selected satellite j′ j′ and the current temperature T(t) to calculate the acceptance probability P accept :

[0083]

[0084] The user generates a random number r∈[0,1], if r <P accept , the selected satellite will be used in the next round of training, otherwise it will continue to stay on the currently connected satellite.

[0085] Each user calculates the temperature of the current training round:

[0086] T(t)=δT(t-1)

[0087] Initial temperature T(0):

[0088] T(0)=c·σ u

[0089] Where c is a proportional constant. The larger the c is, the higher the initial temperature is and the more possibilities that can be explored are. The smaller the c is, the lower the initial temperature is, which is conducive to faster convergence to the local optimal solution. In this embodiment, c is set to 1; σ u is the utility value of each user in the first training round The variance of is:

[0090]

[0091] Wherein, δ is the annealing coefficient, which affects the speed of temperature drop. A slower temperature drop is conducive to fully searching the solution space. In this embodiment, δ is set to 0.85.

[0092] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A satellite selection method based on multi-satellite resource balancing, characterized in that: include: S1: Each user i selects the satellite with the highest current signal strength for random access; S2: Each user i uploads the local data set to the connected satellite to train the global model M t ; S3: Each user i calculates the signal strength of the connected satellite, the size of the data set uploaded to the connected satellite, and the global model M t The training time is used to calculate the respective utility values And their respective utility values Sent to the ground controller; where t is the training round of the global model; S4: Ground controller according to the utility value Determine whether a dynamic equilibrium state is reached, if so, execute step S5; otherwise, execute step S6; S5: The ground controller notifies all users to stay on the currently connected satellite, completing the satellite selection; S6: The ground controller calculates the utility value of each user Calculate the average utility value The utility value of each user i As the utility value of satellite j accessed by each user i The average utility value and the utility values ​​of all connected satellites j Broadcast to all users and execute step S7; S7: Each user i calculates the current temperature T(t) according to the utility values ​​of all connected satellites j Average utility value And the current temperature T(t) selects a satellite for access, and executes step S2 to perform the next round of global model training.

2. The satellite selection method based on multi-satellite resource balancing according to claim 1, characterized in that: The local data sets uploaded by each user to the connected satellite are non-sensitive data sets; Each user i uploads the local data set to the connected satellite to train the global model M t include: S21: After the connection is established, each user i uploads a non-sensitive data set to the connected satellite; S22: When each connected satellite successfully receives the non-sensitive data set, it sends a confirmation message to the corresponding user i; S23: Each connected satellite trains a local model using the non-sensitive data sets received from multiple users; S24: After each connected satellite has trained a local model, the trained local model is sent to one of the satellites for aggregation to obtain a trained global model, and the trained global model is sent back to each user i.

3. The satellite selection method based on multi-satellite resource balancing according to claim 1, characterized in that: Calculate the utility value of user i include: Among them, I j is the signal strength of satellite j, is the size of the data set uploaded by user i, is the training time of the global model in the tth training round, and α, β, κ, ξ are parameter weights.

4. The satellite selection method based on multi-satellite resource balancing according to claim 1, characterized in that: Ground controls according to utility value Determining whether a dynamic equilibrium state has been reached includes: Set the threshold ε, if the utility value of all users in the previous round and the utility value of the current round If the difference is less than the threshold ε, a dynamic equilibrium state is reached.

5. The satellite selection method based on multi-satellite resource balancing according to claim 1, characterized in that: The calculation of the current temperature of each user i includes: setting the annealing coefficient, obtaining the temperature T(t-1) of the previous training round, multiplying the temperature T(t-1) of the previous training round by the annealing coefficient, and obtaining the current temperature T(t); wherein the acquisition process of the initial temperature T(0) includes: the ground controller calculates the utility value of each user in the first training round The variance σ u , according to the variance σ u Calculate the initial temperature T(0) and broadcast the initial temperature T(0) to all users.

6. The satellite selection method based on multi-satellite resource balancing according to claim 1, characterized in that: User i’s utility value based on all satellite j accesses Average utility value And the current temperature T(t) to select the satellite for access includes: User i calculates his own utility value and average utility value difference If bad If it is greater than 0, the currently connected satellite will be retained; otherwise, it is determined whether there is a utility value among all connected satellites. Higher than the utility value of the satellite currently accessed by user i If there is a satellite, then filter out the utility value among all connected satellites Higher than the utility value of the satellite currently accessed by user i satellites, and obtain the satellite set S better , using the weighted random sampling method from the satellite set S better A satellite is selected for access; otherwise, a weighted random sampling method is used to select a satellite from all the accessed satellites, and the acceptance probability mechanism of the simulated annealing algorithm is used to determine whether to access the selected satellite.

7. The satellite selection method based on multi-satellite resource balancing according to claim 6, characterized in that: Use weighted random sampling method to select from the satellite set S better Selecting a satellite from includes: calculating the satellite set S better The utility difference between each satellite in and the satellite currently accessed by user i is normalized to obtain the selection probability p of each satellite. j , according to the selection probability p j From S better Select a satellite from the navigation pane.

8. The satellite selection method based on multi-satellite resource balancing according to claim 7, characterized in that: The acceptance probability mechanism of the simulated annealing algorithm is used to determine whether to access the selected satellite, including: calculating the acceptance probability P according to the utility difference of the selected satellite and the current temperature T(t) accept , generate a random number r∈[0,1], if r <P accept , user i is connected to the selected satellite; otherwise, user i stays on the currently connected satellite.

9. The satellite selection method based on multi-satellite resource balancing according to claim 8, characterized in that: Calculate the acceptance probability P accept include: Among them, j′ is the selected satellite, ΔU j′ is the utility difference of the selected satellite j′, and exp() is an exponential function with the Euler number e as the base.

Images