Hyperparameter optimization and resource allocation method for federal element learning of low earth orbit satellite network
By constructing a decentralized federated meta-learning system model and optimizing hyperparameters and resource allocation, the problem of limited learning performance caused by heterogeneous communication bandwidth and hardware resources in low-Earth orbit satellite networks is solved, and efficient satellite network learning is achieved.
Patent Information
- Application Number
- CN202510831345.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-11-07
AI Technical Summary
In low-Earth orbit satellite networks, limited communication bandwidth, large propagation delay, and heterogeneous hardware resources limit the performance of federated learning, making it difficult to achieve efficient Earth observation missions.
A decentralized federated meta-learning system model is constructed. By jointly optimizing hyperparameters and resource allocation, it is decomposed into sub-problems of optimizing the number of global rounds, the amount of sampled data, the transmission power, and the computational power. This optimizes the satellite's computation and transmission strategies to reduce the overall learning time.
Despite energy constraints and heterogeneous hardware resources, this approach reduces overall learning time, improves global model accuracy, and enables efficient federated learning for low-Earth orbit satellite networks.
Smart Images
Figure CN120915352A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of wireless communication, in particular to a low-orbit satellite network federated meta-learning hyperparameter optimization and resource allocation method. BACKGROUND
[0002] Through a large number of earth images collected by satellites equipped with high-resolution cameras and advanced sensors, combined with advanced machine learning models, complex earth observation tasks can be achieved. However, due to the limited bandwidth of satellite communication, it is difficult to download all images to the ground station, while federated learning can avoid transmitting raw data, reduce transmission bandwidth and protect privacy.
[0003] Due to the scene characteristics of low-orbit satellites, the earth observation task requires real-time of the model. However, there is a communication window and a large propagation delay between the ground station and the satellite, and the satellites in the low-orbit satellite network are always moving at high speed, making it difficult to find a fixed central server. In addition, the satellite orbits are relatively fixed, resulting in highly non-independent and identically distributed data for different orbits, affecting the accuracy of the global model. Although the parallel training structure of federated learning can improve learning efficiency, the different hardware resources of different satellites result in large differences in training time, causing the straggler effect to limit the performance of federated learning. Therefore, it is worth paying attention to how hyperparameters and resource allocation affect federated learning and to what extent they can reduce the overall learning time of federated learning. SUMMARY
[0004] To solve the above technical problems, the present application provides a low-orbit satellite network federated meta-learning hyperparameter optimization and resource allocation method.
[0005] To achieve the above purpose, the present application adopts the following technical solutions:
[0006] A low-orbit satellite network federated meta-learning hyperparameter optimization and resource allocation method, comprising the following steps:
[0007] Step A, constructing a federated meta-learning process containing M orbits and N satellites running on the M orbits, and constructing a federated meta-learning system model based on the federated meta-learning process;
[0008] Step B, for the federated meta-learning process, establishing a hyperparameter and resource allocation joint optimization problem with the goal of minimizing the overall delay, and based on the relationship between the global model preset accuracy and the hyperparameters, decomposing the hyperparameter and resource allocation joint optimization problem into a global round optimization subproblem, a sampling data amount optimization subproblem, a transmission power optimization subproblem, and a computing power optimization subproblem, and obtaining the optimal hyperparameters and the optimal resource allocation scheme through solving calculation.
[0009] Further, the step A establishes a federated meta-learning process based on the decentralized federated meta-learning system model, and specifically includes the following steps:
[0010] Step A1, using a preset ground station to broadcast an initial preset global model and its model parameters related to satellite communication transmission to N satellites running on M orbits;
[0011] Step A2, each satellite respectively uses its local data set to iteratively train the received initial preset global model to obtain the local model and its model parameters corresponding to each satellite in the kth iteration training;
[0012] Step A3, based on the local model and its model parameters corresponding to each satellite, the inter-satellite link between satellites in a single orbit is used to aggregate the local models in the same orbit to obtain the single-orbit model and its model parameters corresponding to M orbits in the kth iteration training;
[0013] Step A4, based on the single-orbit model corresponding to each orbit, the inter-satellite link between satellites in different orbits is used to aggregate the M single-orbit models to obtain the global model and its model parameters in the kth iteration training, and the global model and its model parameters in the kth iteration training are used as the initial preset global model for the k+1th iteration training;
[0014] Step A5, iteratively execute steps A1 to A4 until the global model reaches a preset iteration number K or meets a preset model accuracy.
[0015] Further, the decentralized federated meta-learning system model further includes a federated meta-learning computing model constructed based on the federated meta-learning process, a federated meta-learning transmission model, and an energy consumption model, and specifically includes:
[0016] The federated meta-learning computing model is constructed based on a preset target low-orbit satellite network satellite set Then the satellite After receiving the initial preset global model, a personalized model is obtained by one-step gradient descent on the initial preset global model according to the loss function, and then a local model and its model parameters of satellite a in the kth iteration training are obtained by using a stochastic gradient descent, and further the calculation delay of satellite a is obtained:
[0017] F a (w)=f a (φ a (w));
[0018]
[0019] where F a (w) is the personalized model loss function of satellite a, and φa (w) is a personalized model of satellite a, w is a model parameter, a is a local learning rate of satellite a, represents the gradient of the loss function of satellite a;
[0020]
[0021] where w k+1,a is the local model parameter of satellite a in the k+1th round of iterative training, w k is the global model parameter in the kth round, and β is a model aggregation learning rate, is the estimated gradient;
[0022]
[0023]
[0024] where, is the calculation delay of satellite a in the kth round of iterative training, τ is the number of local iterations of the satellite, C is the number of CPU rounds required for a unit of bit sample data, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, D a is the local data set of satellite a, i.e., the learning sample of satellite a; is the sample subset sampled by satellite a from its local data set in the kth round of iterative training, is the sample subset sampled by satellite a from its local data set in the kth round of iterative training, is the sample subset sampled by satellite a from its local data set in the kth round of iterative training; q k,a is the CPU frequency of satellite a in the kth round of iterative training, is the satellite computing power of satellite a in the kth round of iterative training, ζ a is the effective capacitance coefficient of satellite a;
[0025] The construction of the federated meta-learning transmission model specifically includes: based on a satellite set of a preset target low earth orbit satellite network and an orbit set of the preset target low earth orbit satellite network and N satellites are uniformly distributed in M orbits, the distance between two satellites on the same orbit, the distance between two satellites on adjacent orbits, the maximum communication distance between two satellites on adjacent orbits, the signal-to-noise ratio of two satellites on the same orbit, the signal-to-noise ratio of two satellites on adjacent orbits, and the transmission delay of satellite a in the kth round of iterative training are as follows:
[0026]
[0027] where s i is the distance between two satellites on the i-th orbit, R is the radius of the earth, h ih
[0028]
[0029] wherein s ij is the distance of two satellites in the i-th orbit and the j-th orbit respectively, h i is the altitude of the i-th orbit, h j is the altitude of the j-th orbit, R is the radius of the earth, M is the number of orbits of the target LEO satellite network, and N is the number of satellites of the target LEO satellite network, θ i is the dimension of the i-th orbit, θ j is the dimension of the j-th orbit;
[0030]
[0031] wherein s m is the maximum communication distance of two satellites in the i-th orbit and the j-th orbit respectively, h i is the altitude of the i-th orbit, h j is the altitude of the j-th orbit, R is the radius of the earth;
[0032]
[0033] wherein, is the signal-to-noise ratio of two satellites in the i-th orbit, is the signal-to-noise ratio of two satellites in the i-th orbit and the j-th orbit respectively, is the transmission power of satellite a in the k-th iteration training, G t is the antenna gain of the transmitter, G r is the antenna gain of the receiver, is the ratio of energy per bit to noise density, k B is the Boltzmann constant, r is the total system noise temperature, Q is the inter-satellite link margin, L i is the free space path loss of two satellites in the i-th orbit, L ij is the free space path loss of two satellites in the i-th orbit and the j-th orbit respectively, ω is the carrier frequency, c is the speed of light, s i is the distance of two satellites in the i-th orbit, s ij is the distance of two satellites in the i-th orbit and the j-th orbit respectively;
[0034]
[0035] wherein, is the transmission delay of satellite a in the k-th iteration training, Z(w k) is the bit size of the model parameters, B is the transmission bandwidth, is the signal-to-noise ratio of the two satellites in the ith orbit, is the signal-to-noise ratio of the two satellites in the ith orbit and the jth orbit, respectively;
[0036] The energy consumption model is constructed specifically including:
[0037]
[0038] wherein e k,a is the energy consumption, and κ is the effective switching capacitance, is the computing delay of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, is the satellite computing power of satellite a in the kth round of iterative training.
[0039] Further, the total time required for K rounds of training of the decentralized federated meta-learning system model is:
[0040]
[0041] wherein, is a satellite set of a target low-orbit satellite network, is the computing delay of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training.
[0042] Further, the super parameter and resource allocation joint optimization problem is specifically:
[0043]
[0044] wherein, denotes the mathematical expectation, K is the global training round, w k is the global model parameter in the kth round, ∈ is the preset model accuracy, e k,a denotes the energy consumption of satellite a in the kth round of iterative training, E amax denotes the preset energy consumption threshold of satellite a, is the satellite computing power of satellite a in the kth round of iterative training, is the maximum computing power of satellite a, is the transmission power of satellite a in the kth round of iterative training, is the maximum transmission power of satellite a, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, i.e., the sampling data amount of satellite a in the kth round of iterative training; D a is the local data set of satellite a, i.e., the learning sample of satellite a; s iThe distance between two satellites in the same orbit, s ij The distance between two satellites in adjacent orbits, s m The maximum communication distance between two satellites in adjacent orbits.
[0045] Further, the relationship between the preset accuracy and the hyperparameters in step B is as follows:
[0046]
[0047] wherein, denotes the mathematical expectation, K is the global training round, w k is the global model parameter in the kth round, ∈ is the preset model accuracy, F(·) is the loss function of satellite a, is the gradient of the loss function of satellite a, w o is the initial model parameter, w ∈ is the global optimal model parameter, α is the local learning rate of satellite a, β is the model aggregation learning rate, D a is the local data set of satellite a, i.e., the learning sample of satellite a; D is the total sample number, λ and λ F is the Lipschitz continuous constant, and is used to quantify the difference between the gradient of the loss function and its estimate, is the deviation of the local and global loss function gradients, D in is a sample subset sampled from the local data set, used to calculate the estimated gradient; ∈ is the preset model accuracy.
[0048] Further, based on the relationship between the hyperparameters and the preset model accuracy, the hyperparameter and resource allocation joint optimization problem is decomposed into a global round optimization subproblem, and the optimal global round is calculated, and the specific expression is as follows:
[0049]
[0050] wherein, T is the total learning time of K rounds, F(w0) is the initial global model loss function, F(w ∈ is the global optimal loss function, β is the model aggregation learning rate, ∈ is the preset model accuracy, Ν + represents a set of positive integers, K * is the optimal global round.
[0051] Further, based on the optimal global round, the hyperparameter and resource allocation joint optimization problem is decomposed into a sampling data amount optimization subproblem, and the optimal sampling data amount is calculated, and the specific expression is as follows:
[0052]
[0053] wherein T k is the kth round learning time, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, i.e., the sampling data amount of satellite a in the kth round of iterative training; κ is the effective switch capacitor, τ represents the local iteration number of the satellite, and C is the CPU round number required for a unit bit of sample data, is the satellite computing power of satellite a in the kth round of iterative training, is the transmission power of satellite a in the kth round of iterative training, ζ a is the effective capacitor coefficient of satellite a, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, D a is the local data set of satellite a, i.e., the learning sample of satellite a; is the satellite set, is the upper bound expressed in the second term in the constraint.
[0054] Further, based on the optimal global round number, the hyperparameter and resource allocation joint optimization problem is decomposed into a transmission power optimization subproblem, and the optimal transmission power is calculated, and the specific expression is as follows:
[0055]
[0056] wherein T k is the kth round learning time, is the transmission power of satellite a in the kth round of iterative training, is the maximum transmission power of satellite a, and κ is the effective switch capacitor, is the computing delay of satellite a in the kth round of iterative training, q k,a is the CPU frequency of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, is the satellite set, is the optimal transmission power, is the upper bound expressed in the second term in the constraint.
[0057] Further, based on the optimal global round number, the hyperparameter and resource allocation joint optimization problem is decomposed into a computing power optimization subproblem, and the optimal computing power is calculated, and the specific expression is as follows:
[0058]
[0059]
[0060] wherein T kis the learning time of the kth round, is the satellite computing power of satellite a in the kth round of iterative training, is the maximum computing power of satellite a, κ is the effective switch capacitor, τ represents the local iteration number of the satellite, C is the CPU round required for a unit of bit sample data, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, that is, the sampling data amount of satellite a in the kth round of iterative training; ζ a is the effective capacitor coefficient of satellite a, is the transmission power of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, is a satellite set, is an upper bound expressed by the second term in the constraint.
[0061] The beneficial effects brought by the above technical solutions are:
[0062] The application provides a low-orbit satellite network federal meta-learning hyperparameter optimization and resource allocation method, establishes a federal meta-learning process, and constructs a decentralized federal meta-learning system model; taking minimizing the total learning time of the system as the goal, a hyperparameter optimization and resource allocation problem is established, a joint optimization problem is decomposed into a global round optimization sub-problem and an optimal global round is calculated, a joint optimization problem is decomposed into a sampling data amount optimization sub-problem and an optimal sampling data amount is calculated, a joint optimization problem is decomposed into a transmission power optimization sub-problem and an optimal transmission power is calculated, and a joint optimization problem is decomposed into a computing power optimization sub-problem and an optimal computing power is calculated. The application jointly optimizes multiple variables of wireless communication and model training, reduces the total learning time under the condition that the energy of the satellite is limited and the hardware resources are heterogeneous, and improves the global model precision. BRIEF DESCRIPTION OF DRAWINGS
[0063] Figure 1 is a flowchart of the application;
[0064] Figure 2 is a decentralized federal meta-learning system diagram of the low-orbit satellite network of the application;
[0065] Figure 3 is a schematic diagram of the decentralized federal meta-learning model aggregation method;
[0066] Figure 4 is a comparison diagram of the system model precision change trend with the training round number in different schemes in the embodiment of the application;
[0067] Figure 5A comparison chart of the system learning time trends with model accuracy under different schemes in the embodiments of the present application. DETAILED DESCRIPTION
[0068] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.
[0069] A low-orbit satellite network federated meta-learning hyperparameter optimization and resource allocation method, comprising the following steps:
[0070] Reference Figure 1 And Figure 2 Step A, constructing a federated meta-learning process containing M orbits and N satellites running on the M orbits, and constructing a federated meta-learning system model based on the federated meta-learning process;
[0071] Step B, for the federated meta-learning process, establish a hyperparameter and resource allocation joint optimization problem with the goal of minimizing the overall delay, and based on the relationship between the global model preset accuracy and the hyperparameters, decompose the hyperparameter and resource allocation joint optimization problem into a global round optimization subproblem, a sampling data volume optimization subproblem, a transmission power optimization subproblem, and a computing power optimization subproblem, and obtain the optimal hyperparameters and the optimal resource allocation scheme through solving calculation.
[0072] Further, referring to Figure 3 , the step A establishes the federated meta-learning process based on the decentralized federated meta-learning system model, specifically comprising the following steps:
[0073] Step A1, using a preset ground station to broadcast an initial preset global model and its model parameters w o related to satellite communication transmission to the N satellites running on the M orbits;
[0074] Step A2, initialize k=1, then each satellite respectively uses its local data set to iteratively train the received initial preset global model to obtain the corresponding local model and its model parameters w 1,a of each satellite in the first round of iterative training;
[0075] Step A3, based on the respective local model and its model parameters corresponding to each satellite, the local models on the same orbit are aggregated by single-orbit aggregation to obtain the single-orbit model and its model parameters corresponding to the M orbits respectively in the first round of iterative training, wherein the single-orbit aggregation is specifically: each satellite divides its model parameters into N / M blocks, starts N / M-1 times of scatter-reduce, each satellite accumulates the received data block with its own data block, and sends the updated data block to the next satellite; then the satellite starts N / M-1 times of allgather, each satellite covers its own data block with the received data block, and sends the covered data block to the next satellite; finally each satellite combines all the data blocks to obtain the first round of single-orbit model parameters wherein D m is the sample data amount of the orbit where satellite a is located, and D a is the local data set of satellite a, i.e. the learning sample of satellite a; represents a single-orbit satellite set;
[0076] Step A4, based on the respective single-orbit model corresponding to each orbit, the M single-orbit models are aggregated by orbit aggregation to obtain the global model and its model parameters in the first round of iterative training, i.e. wherein D is the total sample amount, and D m is the sample data amount of the orbit where satellite a is located, and the global model and its model parameters in the first round of iterative training are used as the initial preset global model in the second round of iterative training;
[0077] Step A5, iteratively execute steps A1 to A4, i.e. k=k+1, until the global model reaches the preset number of iterations K, or the preset model accuracy is met.
[0078] Further, the decentralized federated meta-learning system model further comprises a federated meta-learning computing model constructed based on a federated meta-learning process, a federated meta-learning transmission model, and an energy consumption model, specifically comprising:
[0079] Constructing a federated meta-learning computing model specifically comprises: based on the satellite set of the preset target low-orbit satellite network then satellite After receiving the initial preset global model, a personalized model is obtained by one-step gradient descent on the initial preset global model according to the loss function, and then a local model and its model parameters of satellite a in the kth round of iterative training are obtained by using a stochastic gradient descent, and further the computing time delay of satellite a is obtained:
[0080] F a (w)=f a (φ a (w));
[0081]
[0082] where F a (w) is the personalized model loss function of satellite a, φ a (w) is the personalized model of satellite a, w is the model parameter, and a is the local learning rate of satellite a, denotes the gradient of the loss function of satellite a;
[0083]
[0084] where w k+1,a is the local model parameter of satellite a in the k+1th round of iterative training, w k is the global model parameter in the kth round, and b is the model aggregation learning rate, is the estimated gradient;
[0085]
[0086] where, is the calculation delay of satellite a in the kth round of iterative training, t is the number of local iterations of the satellite, C is the number of CPU rounds required for a unit of bit sample data, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, D a is the local data set of satellite a, i.e., the learning sample of satellite a; is the sample subset sampled by satellite a from its local data set in the kth round of iterative training, is the sample subset sampled by satellite a from its local data set in the kth round of iterative training, is the sample subset sampled by satellite a from its local data set in the kth round of iterative training; q k,a is the CPU frequency of satellite a in the kth round of iterative training, is the satellite computing power of satellite a in the kth round of iterative training, z a is the effective capacitance coefficient of satellite a;
[0087] The construction of the federated meta-learning transmission model specifically includes: based on the satellite set of the preset target low earth orbit satellite network and the orbit set of the preset target low earth orbit satellite network and N satellites are uniformly distributed in M orbits, the distance between two satellites on the same orbit, the distance between two satellites on adjacent orbits, the maximum communication distance between two satellites on adjacent orbits, the signal-to-noise ratio of two satellites on the same orbit, the signal-to-noise ratio of two satellites on adjacent orbits, and the transmission delay of satellite a in the kth round of iterative training are as follows:
[0088]
[0089] where s iis the distance of two satellites in the i th orbit, R is the radius of the earth, h i is the altitude of the i th orbit, M is the number of orbits of the target LEO satellite network, N is the number of satellites of the target LEO satellite network;
[0090]
[0091] wherein s ij is the distance of two satellites in the i th orbit and j th orbit respectively, h i is the altitude of the i th orbit, h j is the altitude of the j th orbit, R is the radius of the earth; i is the dimension of the i th orbit, h j is the dimension of the j th orbit;
[0092]
[0093] wherein s m is the maximum communication distance of two satellites in the i th orbit and j th orbit respectively, h i is the altitude of the i th orbit, h j is the altitude of the j th orbit, R is the radius of the earth;
[0094]
[0095] wherein, is the signal-to-noise ratio of two satellites in the i th orbit, is the signal-to-noise ratio of two satellites in the i th orbit and j th orbit respectively, is the transmission power of satellite a in the k th iteration training, G t is the antenna gain of the transmitter, G r is the antenna gain of the receiver, is the ratio of energy per bit to noise density, k B is the Boltzmann constant, r is the total system noise temperature, Q is the inter-satellite link margin, L i is the free space path loss of two satellites in the i th orbit, L ij is the free space path loss of two satellites in the i th orbit and j th orbit respectively, w is the carrier frequency, c is the speed of light, s i is the distance of two satellites in the i th orbit, s ij is the distance of two satellites in the i th orbit and j th orbit respectively;
[0096]
[0097] wherein, a transmission delay of satellite a in the kth round of iterative training, Z(w k ) is a bit size of a model parameter, B is a transmission bandwidth, a signal-to-noise ratio of two satellites in the ith orbit, a signal-to-noise ratio of two satellites in the ith orbit and the jth orbit, respectively;
[0098] The energy consumption model specifically includes:
[0099]
[0100] wherein e k,a is energy consumption, and K is an effective switching capacitor, a computing delay of satellite a in the kth round of iterative training, a transmission delay of satellite a in the kth round of iterative training, a satellite computing power of satellite a in the kth round of iterative training.
[0101] Further, the total time required for K rounds of training of the decentralized federated meta-learning system model is:
[0102]
[0103] wherein, is a satellite set of a target low-orbit satellite network, a computing delay of satellite a in the kth round of iterative training, a transmission delay of satellite a in the kth round of iterative training.
[0104] Further, the super parameter and resource allocation joint optimization problem specifically is:
[0105]
[0106] wherein, denotes a mathematical expectation, K is a global training round, w k is a global model parameter in the kth round, ∈ is a preset model accuracy, e k,a denotes energy consumption of satellite a in the kth round of iterative training, E amax denotes a preset energy consumption threshold of satellite a, a satellite computing power of satellite a in the kth round of iterative training, is a maximum computing power of satellite a, a transmission power of satellite a in the kth round of iterative training, is a maximum transmission power of satellite a, d k,a is a sample quantity of satellite a used for updating a gradient in the kth round of iterative training, i.e., a sampling data quantity of satellite a in the kth round of iterative training; D ais the local dataset of satellite a, i.e., the learning samples of satellite a; D is the total number of samples, λ and λ i is the distance between two satellites in the same orbit, s ij is the distance between two satellites in adjacent orbits, s m is the maximum communication distance between two satellites in adjacent orbits.
[0107] Further, the relationship between the preset accuracy and the hyperparameters in step B is as follows:
[0108]
[0109] wherein, denotes the mathematical expectation, K is the global training round, w k is the global model parameter in the kth round, ∈ is the preset model accuracy, F(·) is the loss function of satellite a, is the gradient of the loss function of satellite a, w o is the initial model parameter, w ∈ is the global optimal model parameter, α is the local learning rate of satellite a, β is the model aggregation learning rate, D a is the local dataset of satellite a, i.e., the learning samples of satellite a; D is the total number of samples, λ and λ F is the Lipschitz continuous constant, and is used to quantify the difference between the gradient of the loss function and its estimate, is the deviation of the local and global loss function gradients, D in is the sample subset sampled from the local dataset, used to calculate the estimated gradient; ∈ is the preset model accuracy.
[0110] Further, based on the relationship between the hyperparameters and the preset model accuracy, the hyperparameter and resource allocation joint optimization problem is decomposed into a global round optimization subproblem, and the optimal global round is calculated, and the specific expression is as follows:
[0111]
[0112] wherein, T is the total learning time of K rounds, F(w0) is the initial global model loss function, F(w ∈ is the global optimal loss function, β is the model aggregation learning rate, ∈ is the preset model accuracy, Ν + represents a set of positive integers, K * is the optimal global round.
[0113] Further, based on the optimal global round, the hyperparameter and resource allocation joint optimization problem is decomposed into a sampling data amount optimization subproblem, and the optimal sampling data amount is calculated, and the specific expression is as follows:
[0114]
[0115] wherein T k is the kth round learning time, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, i.e., the sampling data amount of satellite a in the kth round of iterative training; κ is the effective switch capacitor, τ represents the local iteration number of the satellite, and C is the CPU round number required for a unit bit of sample data, is the satellite computing power of satellite a in the kth round of iterative training, is the transmission power of satellite a in the kth round of iterative training, ζ a is the effective capacitor coefficient of satellite a, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, D a is the local data set of satellite a, i.e., the learning sample of satellite a; is the satellite set, is the upper bound expressed in the second term in the constraint.
[0116] Further, based on the optimal global round number, the hyperparameter and resource allocation joint optimization problem is decomposed into a transmission power optimization subproblem, and the optimal transmission power is calculated, and the specific expression is as follows:
[0117]
[0118] wherein T k is the kth round learning time, is the transmission power of satellite a in the kth round of iterative training, is the maximum transmission power of satellite a, and κ is the effective switch capacitor, is the computing delay of satellite a in the kth round of iterative training, q k,a is the CPU frequency of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, is the satellite set, is the optimal transmission power, is the upper bound expressed in the second term in the constraint.
[0119] Further, based on the optimal global round number, the hyperparameter and resource allocation joint optimization problem is decomposed into a computing power optimization subproblem, and the optimal computing power is calculated, and the specific expression is as follows:
[0120]
[0121] wherein T kis the learning time of the kth round, is the satellite computing power of satellite a in the kth round of iterative training, is the maximum computing power of satellite a, κ is the effective switch capacitor, τ represents the local iteration number of the satellite, C is the CPU round required for unit bit sample data, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, that is, the amount of sample data of satellite a in the kth round of iterative training; ζ a is the effective capacitor coefficient of satellite a, is the transmission power of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, is a satellite set, is an upper bound expressed by the second term in the constraint.
[0122] The technical effects of the present application will be further described in detail in combination with simulation experiments. Figure 4 is a comparison chart of the change trend of system model accuracy with training round number under different schemes. From Figure 4 It can be observed that under the same training round, the model accuracy of the hyperparameter optimization and resource allocation method proposed in the present application is significantly better than that of the traditional meta-learning method and the traditional federated learning method. Figure 5 is a comparison chart of the change trend of system learning time with model accuracy under different schemes. From Figure 5 It can be observed that the hyperparameter optimization and resource allocation method proposed in the present application performs best in the balance between training time and model accuracy, and can achieve high model accuracy in a short time, which has obvious efficiency advantage compared with the traditional method.
[0123] The above is only a preferred embodiment of the present application, and does not limit the present application in any way. Any simple modification, change and equivalent structural change made according to the technical essence of the present application to the above embodiment are still within the protection scope of the technical solution of the present application.
Claims
1. A method for low earth orbit satellite network federated meta-learning hyperparameter optimization and resource allocation, characterized in that, The method comprises the following steps: Step A, constructing a federated meta-learning process comprising M orbits and N satellites running on the M orbits, and constructing a federated meta-learning system model based on the federated meta-learning process; Step B, for the federated meta-learning process, establishing a hyperparameter and resource allocation joint optimization problem with the objective of minimizing the overall time delay, and based on the relationship between the global model preset accuracy and the hyperparameters, decomposing the hyperparameter and resource allocation joint optimization problem into a global round optimization subproblem, a sampling data amount optimization subproblem, a transmission power optimization subproblem, and a computing power optimization subproblem, and obtaining the optimal hyperparameters and the optimal resource allocation scheme through solving calculation.
2. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 1, characterized in that, The step A establishes the federated meta-learning process based on the decentralized federated meta-learning system model, and specifically comprises the following steps: Step A1, using a preset ground station to broadcast an initial preset global model and its model parameters related to satellite communication transmission to the N satellites running on the M orbits; Step A2, each satellite respectively uses its local data set to iteratively train the received initial preset global model to obtain the corresponding local model and its model parameters of each satellite in the kth iteration training; Step A3, based on the corresponding local model and its model parameters of each satellite, the inter-satellite links between satellites in a single orbit are used to aggregate the local models on the same orbit to obtain the corresponding single-orbit models and their model parameters of the M orbits in the kth iteration training; Step A4, based on the corresponding single-orbit models of each orbit, the inter-satellite links between satellites on different orbits are used to aggregate the M single-orbit models to obtain the global model and its model parameters of the kth iteration training, and the global model and its model parameters of the kth iteration training are used as the initial preset global model of the k+1th iteration training; Step A5, iteratively performing steps A1 to A4 until the global model reaches a preset iteration number K or meets a preset model accuracy.
3. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 2, characterized in that, The decentralized federated meta-learning system model further comprises a federated meta-learning computing model, a federated meta-learning transmission model, and an energy consumption model, which are constructed based on the federated meta-learning process, and specifically comprise: The constructing the federated meta-learning computing model specifically comprises: Then, the satellite After receiving the initial preset global model, a personalized model is obtained by performing one-step gradient descent on the initial preset global model according to a loss function, and then a local model of the satellite a and model parameters of the local model are obtained by using a stochastic gradient descent in the kth round of iterative training, and further, a computing time delay of the satellite a is obtained. F a (w) = f a (φ a (w)); φ a (w) = w - a Vf a (w); where F a (w) is the personalized model loss function of satellite a, φ a (w) is the personalized model of satellite a, w is the model parameter, a is the local learning rate of satellite a, and ▽f a (·) represents the gradient of the loss function of satellite a; w k+1,a : = w k - β∇F a (w k ); wherein w k+1,a is the local model parameter of satellite a in the k+1th iteration training, w k is the global model parameter in the kth iteration, β is the model aggregation learning rate, and ▽F a is the estimated gradient; wherein, is the computation latency of satellite a in the kth round of iterative training, τ is the local iteration number of the satellite, C is the number of CPU rounds required for a unit of bit sample data, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, D a is the local dataset of satellite a, i.e., the learning samples of satellite a; is the sample subset sampled by satellite a from its local dataset in the kth round of iterative training, is the sample subset sampled by satellite a from its local dataset in the kth round of iterative training, is the sample subset sampled by satellite a from its local dataset in the kth round of iterative training; q k,a is the CPU frequency of satellite a in the kth round of iterative training, is the satellite computation power of satellite a in the kth round of iterative training, ζ a is the effective capacitance coefficient of satellite a; The constructing the federated meta-learning transmission model specifically comprises: And N satellites are uniformly distributed in M orbits, then the distance of two satellites in the same orbit, the distance of two satellites in adjacent orbits, the maximum communication distance of two satellites in adjacent orbits, the signal-to-noise ratio of two satellites in the same orbit, the signal-to-noise ratio of two satellites in adjacent orbits, and the transmission delay of satellite a in the kth iteration training are as follows: wherein s i is the distance of two satellites on the i th orbit, R is the radius of the Earth, h i is the height of the i th orbit, M is the number of orbits of the target LEO satellite network, and N is the number of satellites of the target LEO satellite network. wherein s ij is the distance of the two satellites in the i-th orbit and the j-th orbit, respectively, h i is the altitude of the i-th orbit, h j is the altitude of the j-th orbit, R is the radius of the earth, M is the number of orbits of the target low earth orbit satellite network, N is the number of satellites of the target low earth orbit satellite network, θ i is the dimension of the i-th orbit, θ j is the dimension of the j-th orbit; where s m is the maximum communication distance of the two satellites in the i-th orbit and the j-th orbit, respectively, h i is the height of the i-th orbit, h j is the height of the j-th orbit, and R is the radius of the Earth; wherein, the signal-to-noise ratio of two satellites in the i-th orbit, the signal-to-noise ratio of two satellites in the i-th and j-th orbits, respectively, the transmit power of satellite a in the k-th iteration training, G t the antenna gain of the transmitter, G r the antenna gain of the receiver, the ratio of energy per bit to noise density, k B the Boltzmann constant, r is the total system noise temperature, Q is the inter-satellite link margin, L i the free space path loss of two satellites in the i-th orbit, L ij the free space path loss of two satellites in the i-th and j-th orbits, respectively, ω is the carrier frequency, c is the speed of light, s i the distance of two satellites in the i-th orbit, s ij the distance of two satellites in the i-th and j-th orbits, respectively; wherein, is the transmission delay of satellite a in the kth round of iterative training, Z(w k is the bit size of the model parameters, B is the transmission bandwidth, is the signal-to-noise ratio of two satellites in the ith orbit, is the signal-to-noise ratio of two satellites in the ith orbit and the jth orbit, respectively; The construction of the energy consumption model specifically comprises: where e k,a is the energy consumption, k is the effective switching capacitance, is the calculated time delay of satellite a in the kth round of iterative training, is the transmission time delay of satellite a in the kth round of iterative training, is the satellite calculation power of satellite a in the kth round of iterative training.
4. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 3, characterized in that, The overall time required for K rounds of training of the decentralized federated meta-learning system model is: wherein, a set of satellites being a target LEO satellite network, a computation delay of satellite a in the kth round of iterative training, a transmission delay of satellite a in the kth round of iterative training.
5. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 4, characterized in that, The hyperparameter and resource allocation joint optimization problem specifically is: 0 < s i ≤ s m 0 < s ij ≤ s m wherein, denotes the mathematical expectation, K is the global training round, w k is the global model parameter of the kth round, ∈ is the preset model accuracy, e k,a denotes the energy consumption of satellite a in the kth iteration training, E amax denotes the preset energy consumption threshold of satellite a, is the satellite computing power of satellite a in the kth iteration training, is the maximum computing power of satellite a, is the transmission power of satellite a in the kth iteration training, is the maximum transmission power of satellite a, d k,a is the number of samples used by satellite a to update the gradient in the kth iteration training, that is, the sampling data amount of satellite a in the kth iteration training; D a is the local data set of satellite a, that is, the learning sample of satellite a; s i is the distance between two satellites in the same orbit, s ij is the distance between two satellites in adjacent orbits, s m is the maximum communication distance between two satellites in adjacent orbits.
6. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 1, characterized in that, The relationship between the global model preset accuracy and the hyperparameters in step B is as follows: wherein, denotes the mathematical expectation, K is the global training round, w k is the global model parameter of the kth round, ∈ is the preset model accuracy, F(·) is the loss function of satellite a, ▽F(·) is the gradient of the loss function of satellite a, w o is the initial model parameter, w ∈ is the global optimal model parameter, α is the local learning rate of satellite a, β is the model aggregation learning rate, D a is the local data set of satellite a, that is, the learning sample of satellite a; D is the total sample number, λ and λ F is the Lipschitz continuous constant, and is used to quantify the difference between the gradient of the loss function and its estimate, is the deviation of the local and global loss function gradients, D in is a sample subset sampled from the local data set, used to calculate the estimated gradient; ∈ is the preset model accuracy.
7. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 6, characterized in that, Based on the relationship between the hyperparameters and the preset model accuracy, the hyperparameter and resource allocation joint optimization problem is decomposed into a global round optimization subproblem, and the optimal global round is calculated, and the specific expression is as follows: wherein T is the total learning time of K rounds, F(w0) is the initial global model loss function, F(w ∈ ) is the global optimal loss function, β is the model aggregation learning rate, ∈ is the preset model accuracy, N + represents a set of positive integers, K * is the optimal global round number.
8. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 7, characterized in that, Based on the optimal global round, the hyperparameter and resource allocation joint optimization problem is decomposed into a sampling data amount optimization subproblem, and the optimal sampling data amount is calculated, and the specific expression is as follows: wherein T k is the kth round of learning time, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, that is, the sampling data amount of satellite a in the kth round of iterative training; κ is the effective switch capacitor, τ represents the local iteration number of the satellite, C is the CPU round required for a unit bit of sample data, is the satellite computing power of satellite a in the kth round of iterative training, is the transmission power of satellite a in the kth round of iterative training, ζ a is the effective capacitor coefficient of satellite a, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, D a is the local data set of satellite a, that is, the learning sample of satellite a; is the satellite set, is the upper bound expressed by the second term in the constraint. 9.The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 7, characterized in that, Based on the optimal global round, the hyperparameter and resource allocation joint optimization problem is decomposed into a transmission power optimization subproblem, and the optimal transmission power is calculated, and the specific expression is as follows: wherein T k is the kth round of learning time, is the transmit power of satellite a in the kth round of iterative training, is the maximum transmission power of satellite a, and κ is the effective switching capacitor, is the calculation delay of satellite a in the kth round of iterative training, and q k,a is the CPU frequency of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, and E amax represents the preset energy consumption threshold of satellite a, is a satellite set, is the optimal transmit power, is the upper bound expressed by the second term in the constraint.
10. The LEO satellite network federated meta-learning hyperparameter optimization and resource allocation method according to claim 7, characterized in that, Based on the optimal global round number, the joint optimization problem of hyperparameters and resource allocation is decomposed into a computing power optimization subproblem, and the optimal computing power is calculated as follows: wherein T k is the kth round of learning time, is the satellite computing power of satellite a in the kth round of iterative training, is the maximum computing power of satellite a, κ is the effective switch capacitor, τ represents the local iteration number of the satellite, C is the CPU round required for unit bit sample data, d k,a is the number of samples used by satellite a to update the gradient in the kth round of iterative training, that is, the sampling data amount of satellite a in the kth round of iterative training; ζ a is the effective capacitor coefficient of satellite a, is the transmission power of satellite a in the kth round of iterative training, is the transmission delay of satellite a in the kth round of iterative training, E amax represents the preset energy consumption threshold of satellite a, is a satellite set, is an upper bound expressed by the second term in the constraint.