High-energy-efficiency federated learning method combining client selection and gradient optimization in heterogeneous mobile edge computing network

Abstract

The invention discloses a high-energy-efficiency federated learning method combining client selection and gradient optimization in a heterogeneous mobile edge computing network. The method comprises the following steps: 1) establishing a heterogeneous mobile edge computing system for federated learning; and 2) the edge server obtains related information of the mobile client participating in federated learning. And 3) the edge server quantifies the calculation quality, the communication quality and the data quality of the client according to the information obtained in the step 2), and calculates the quality score of the mobile client. And 4) using the obtained quality score as an index to select a client to carry out federated learning training, and keeping an effective residual gradient of the mobile client during training to accelerate model convergence. According to the invention, the total quality of the mobile client is quantified, so that the client with better quality can be selected to participate in federal learning training as much as possible. And moreover, the gradient of the mobile client is optimized, so that the energy consumption is effectively reduced while the model performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of mobile edge computing, and in particular to an energy-efficient federated learning method that combines client selection and gradient optimization in a heterogeneous mobile edge computing network. Background Art

[0002] Mobile edge computing emerged alongside the development of mobile communication networks. Smart devices in edge networks continuously generate vast amounts of data, providing a rich set of application scenarios. To better utilize this data while effectively addressing the privacy breaches and network congestion issues associated with centralized cloud computing, federated learning emerged. Federated learning allows mobile clients to collaboratively train a global model by transmitting only locally generated parameters. Combined with mobile edge computing technology, federated learning improves computing efficiency on mobile clients, reduces network latency, and protects client privacy by keeping data local.

[0003] However, there is inherent system and data heterogeneity between different mobile clients. Specifically, different mobile clients possess different data, computing, and communication resources. These resources have different distributions, manifesting as differences in computing power, communication capabilities, data volume, and the diversity of data samples available for training on each mobile client. System heterogeneity refers to differences in computing power and communication capabilities, while data heterogeneity refers to differences in the amount of data available for training and the diversity of data samples. This heterogeneity can reduce the generalization ability of the overall model, leading to problems such as poor model convergence and increased communication overhead and energy consumption. Therefore, it is necessary to design an energy-efficient federated learning method in heterogeneous mobile edge computing networks. Summary of the Invention

[0004] The present invention aims to provide an energy-efficient federated learning method for combining client selection and gradient optimization in a heterogeneous mobile edge computing network, comprising the following steps:

[0005] 1) Build a federated learning framework in a heterogeneous mobile edge computing network, including an edge server located at a base station and N mobile clients;

[0006] 2) Train the federated learning framework to obtain a global model consisting of edge servers and N mobile clients;

[0007] 3) Utilize the global model to execute tasks requested by any mobile client in the mobile edge computing network.

[0008] Furthermore, the CPU clock speed f, transmission speed R, and local dataset D of the N mobile clients are all set independently.

[0009] Furthermore, in step 2), the steps of training the federated learning framework include:

[0010] 2.1) Set up the total training wheel set

[0011] 2.2) In round t=1, the edge server sets the global model parameter ω t Send to all mobile clients;

[0012] The mobile client performs parameter updates locally based on the global model parameters and transmits the trained global model parameters to the edge server;

[0013] The global model parameters after training are as follows:

[0014]

[0015] Where η is the learning rate, Represents the gradient of the local model loss function; The global model parameters after training for mobile client k; is the global model parameter received by mobile client k;

[0016] 2.3) The edge server performs weighted averaging on all received global model parameters and updates the global model parameters, namely:

[0017]

[0018] Among them, γ k is the proportion of the data volume of mobile client k in the total data volume; ω t+1 is the updated global model parameter;

[0019] 2.4) Based on the updated global model parameters, calculate the global loss F(ω); if the global loss F(ω) reaches a predetermined value or the number of training rounds reaches a predetermined value T, then terminate the training of the federated learning framework; otherwise, set t = t + 1 and proceed to step 2.5);

[0020] 2.5) The edge server calculates the quality score of each mobile client based on the data quality, computing quality and communication quality of the mobile client, and selects the top |C t | mobile clients with the highest quality scores participate in the tth round of training;

[0021] In round t, the edge server sets the global model parameter ω t Sent to mobile clients participating in the training;

[0022] The mobile client performs parameter updates locally based on the global model parameters, transmits the trained local model parameters to the edge server, and returns to step 2.3).

[0023] Furthermore, during the training process, the mobile client k is used to locally update the gradient of the global model parameters. As shown below:

[0024]

[0025] Where, is the residual gradient; ζ is Impact factor in the current round; is the gradient of the mobile client k that was not selected in the previous round.

[0026] Further, the quantity |C t |As shown below:

[0027]

[0028] Where p represents the proportion of mobile clients.

[0029] Furthermore, the quality scores for mobile clients are as follows:

[0030]

[0031] Among them, ρ1, ρ2, and ρ3 represent the weights of data quality, calculation quality, and communication quality, respectively; Respectively represent the data quality, computing quality and communication quality of mobile client k; V k is the quality score of mobile client k.

[0032] Furthermore, the data quality of mobile client k is as follows:

[0033]

[0034] Where α and β are hyperparameters;

[0035] Among them, individual data contribution and overall data value They are as follows:

[0036]

[0037] Among them, λ1, λ2, α, and β are all hyperparameters. is the gradient calculated by mobile client k in round t, y k is the data label of mobile client k. is the maximum gradient; |D k |、|D j | is the data dependency depth.

[0038] Furthermore, the calculated quality of mobile client k is as follows:

[0039]

[0040] Where θ and μ are hyperparameters; is the maximum delay and maximum energy consumption of the calculation model parameters in the client set N;

[0041] Among them, the delay required for mobile client k to update the local model The energy consumed by mobile client k to update the local model As shown below:

[0042]

[0043] Where, κ is the number of local iterations; is the number of CPU cycles required to process one sample; ν is the effective capacitance coefficient. k is the CPU clock frequency.

[0044] Furthermore, the communication quality of mobile client k is as follows:

[0045]

[0046] Where μ is a hyperparameter; Collection for clients The maximum delay and maximum energy consumption of China Mobile clients uploading local model parameters to the edge server;

[0047] Among them, the delay of the mobile client uploading local model parameters to the edge server Energy consumed As shown below:

[0048]

[0049] Among them, q k is the transmission power of mobile client k; S is the size of the model parameters that need to be uploaded. k is the transmission rate.

[0050] Furthermore, the tasks requested by mobile clients include intelligent perception, autonomous driving assistance, and smart device management.

[0051] The technical effects of the present invention are undoubted, and the beneficial effects of the present invention are as follows:

[0052] Considering the system heterogeneity and data heterogeneity in mobile edge computing networks, compared with other mobile edge computing network scenarios that do not consider heterogeneity, this framework is more in line with real life and has a wider range of applications.

[0053] By quantifying the computing, communication, and data quality of mobile clients, we designed a new client selection mechanism that prioritizes mobile clients with better overall quality. This allows the model to learn more data variance in a shorter time, reducing the negative impact of heterogeneity while enhancing the model's generalization capabilities and ensuring the quality of communication between mobile clients and edge servers.

[0054] The valid gradients of the mobile client are retained. These gradients contain information that is beneficial to model convergence. This operation can accelerate the convergence of the global model.

[0055] The energy consumption during the federated learning training process and the loss function of the model are comprehensively considered to reduce energy consumption while ensuring the convergence of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a diagram of the federated learning architecture in mobile edge computing networks.

[0057] Figure 2 Flowchart for energy-efficient federated learning with joint client selection and gradient optimization. DETAILED DESCRIPTION

[0058] The present invention will be further described below with reference to the following examples, but it should not be understood that the scope of the present invention is limited to the following examples. Without departing from the above technical ideas of the present invention, various substitutions and modifications can be made according to common technical knowledge and customary means in the art, and all should be included in the scope of protection of the present invention.

[0059] Example 1:

[0060] See also Figures 1 to 2 , an energy-efficient federated learning method combining client selection and gradient optimization in heterogeneous mobile edge computing networks, comprising the following steps:

[0061] 1) Build a federated learning framework in a heterogeneous mobile edge computing network, including an edge server located at a base station and N mobile clients;

[0062] 2) Train the federated learning framework to obtain a global model consisting of edge servers and N mobile clients;

[0063] 3) Utilize the global model to execute tasks requested by any mobile client in the mobile edge computing network.

[0064] The CPU clock speed f, transmission speed R, and local dataset D of the N mobile clients are all set independently.

[0065] In step 2), the steps for training the federated learning framework include:

[0066] 2.1) Set up the total training wheel set

[0067] 2.2) In round t=1, the edge server sets the global model parameter ω t Send to all mobile clients;

[0068] The mobile client performs parameter updates locally based on the global model parameters and transmits the trained global model parameters to the edge server;

[0069] The global model parameters after training are as follows:

[0070]

[0071] Where η is the learning rate, Represents the gradient of the local model loss function; The global model parameters after training for mobile client k; is the global model parameter received by mobile client k;

[0072] 2.3) The edge server performs weighted averaging on all received global model parameters and updates the global model parameters, namely:

[0073]

[0074] Among them, γ k is the proportion of the data volume of mobile client k in the total data volume; ω t+1 is the updated global model parameter;

[0075] 2.4) Based on the updated global model parameters, calculate the global loss F(ω); if the global loss F(ω) reaches a predetermined value or the number of training rounds reaches a predetermined value T, then terminate the training of the federated learning framework; otherwise, set t = t + 1 and proceed to step 2.5);

[0076] 2.5) The edge server calculates the quality score of each mobile client based on the data quality, computing quality and communication quality of the mobile client, and selects the top |C t | mobile clients with the highest quality scores participate in the tth round of training;

[0077] In round t, the edge server sets the global model parameter ω t Sent to mobile clients participating in the training;

[0078] The mobile client performs parameter updates locally based on the global model parameters, transmits the trained local model parameters to the edge server, and returns to step 2.3).

[0079] During training, the mobile client k is used to perform the gradient of global model parameter update locally. As shown below:

[0080]

[0081] Where, is the residual gradient; ζ is Impact factor in the current round; is the gradient of the mobile client k that was not selected in the previous round. The gradient generated by client k in this round of training. Plus the gradient generated by k participating in the previous round of training as the gradient of parameter updates

[0082] Quantity|C t |As shown below:

[0083]

[0084] Where p represents the proportion of mobile clients.

[0085] The quality scores for mobile clients are as follows:

[0086]

[0087] Among them, ρ1, ρ2, and ρ3 represent the weights of data quality, calculation quality, and communication quality, respectively; Respectively represent the data quality, computing quality and communication quality of mobile client k; V k is the quality score of mobile client k.

[0088] The data quality of mobile client k is as follows:

[0089]

[0090] Where α and β are hyperparameters;

[0091] Among them, individual data contribution and overall data value They are as follows:

[0092]

[0093] Among them, λ1, λ2, α, and β are all hyperparameters. is the gradient calculated by mobile client k in round t, y k is the data label of mobile client k. is the maximum gradient; |D k |、|Dj | is the data dependency depth.

[0094] The calculated quality of mobile client k is as follows:

[0095]

[0096] Where θ and μ are hyperparameters; Collection for clients The calculation model parameters in the maximum delay and maximum energy consumption;

[0097] Among them, the delay required for mobile client k to update the local model The energy consumed by mobile client k to update the local model As shown below:

[0098]

[0099] Where, κ is the number of local iterations; is the number of CPU cycles required to process one sample; ν is the effective capacitance coefficient. k is the CPU clock frequency.

[0100] The communication quality of mobile client k is as follows:

[0101]

[0102] Where μ is a hyperparameter; Collection for clients The maximum delay and maximum energy consumption of China Mobile clients uploading local model parameters to the edge server;

[0103] Among them, the delay of the mobile client uploading local model parameters to the edge server Energy consumed As shown below:

[0104]

[0105] Among them, q k is the transmission power of mobile client k; S is the size of the model parameters that need to be uploaded. k is the transmission rate.

[0106] The tasks requested by mobile clients include intelligent perception, autonomous driving assistance, and smart device management.

[0107] Example 2:

[0108] An energy-efficient federated learning method for combining client selection and gradient optimization in a heterogeneous mobile edge computing network includes the following steps:

[0109] 1) Build a federated learning framework in a heterogeneous mobile edge computing network, including an edge server located at a base station and N mobile clients;

[0110] 2) Train the federated learning framework to obtain a global model consisting of edge servers and N mobile clients;

[0111] 3) Utilize the global model to execute tasks requested by any mobile client in the mobile edge computing network.

[0112] Example 3:

[0113] A highly energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network, with the same technical content as in Example 2. Furthermore, the CPU clock speed f, transmission speed R, and local data set D of the N mobile clients are independently set.

[0114] Example 4:

[0115] A high-efficiency federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network, the technical content of which is the same as any one of Examples 2-3, further, in step 2), the step of training the federated learning framework includes:

[0116] 2.1) Set up the total training wheel set

[0117] 2.2) In round t=1, the edge server sets the global model parameter ω t Send to all mobile clients;

[0118] The mobile client performs parameter updates locally based on the global model parameters and transmits the trained global model parameters to the edge server;

[0119] The global model parameters after training are as follows:

[0120]

[0121] Where η is the learning rate, Represents the gradient of the local model loss function; The global model parameters after training for mobile client k; is the global model parameter received by mobile client k;

[0122] 2.3) The edge server performs weighted averaging on all received global model parameters and updates the global model parameters, namely:

[0123]

[0124] Among them, γ kis the proportion of the data volume of mobile client k in the total data volume; ω t+1 is the updated global model parameter;

[0125] 2.4) Based on the updated global model parameters, calculate the global loss F(ω); if the global loss F(ω) reaches a predetermined value or the number of training rounds reaches a predetermined value T, then terminate the training of the federated learning framework; otherwise, set t = t + 1 and proceed to step 2.5);

[0126] 2.5) The edge server calculates the quality score of each mobile client based on the data quality, computing quality and communication quality of the mobile client, and selects the top |C t | mobile clients with the highest quality scores participate in the tth round of training;

[0127] In round t, the edge server sets the global model parameter ω t Sent to mobile clients participating in the training;

[0128] The mobile client performs parameter updates locally based on the global model parameters, transmits the trained local model parameters to the edge server, and returns to step 2.3).

[0129] Example 5:

[0130] A high-efficiency federated learning method for joint client selection and gradient optimization in a heterogeneous mobile edge computing network, the technical content of which is the same as any one of Examples 2-4, further, during the training process, the mobile client k is used to locally update the gradient of the global model parameters As shown below:

[0131]

[0132] Where, is the residual gradient; ζ is Impact factor in the current round; is the gradient of the mobile client k that was not selected in the previous round.

[0133] Example 6:

[0134] A high-efficiency federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network, the technical content of which is the same as any one of Examples 2-5, furthermore, the quantity |C t |As shown below:

[0135]

[0136] Where p represents the proportion of mobile clients.

[0137] Example 7:

[0138] A highly energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network, with the same technical content as any one of Examples 2-6, further, the quality score of the mobile client is as follows:

[0139]

[0140] Among them, ρ1, ρ2, and ρ3 represent the weights of data quality, calculation quality, and communication quality, respectively; Respectively represent the data quality, computing quality and communication quality of mobile client k; V k is the quality score of mobile client k.

[0141] Example 8:

[0142] A highly energy-efficient federated learning method combining client selection and gradient optimization in a heterogeneous mobile edge computing network, with the same technical content as any one of Examples 2-7, further, the data quality of the mobile client k is as follows:

[0143]

[0144] Where α and β are hyperparameters;

[0145] Among them, individual data contribution and overall data value They are as follows:

[0146]

[0147] Among them, λ1, λ2, α, and β are all hyperparameters. is the gradient calculated by mobile client k in round t, y k is the data label of mobile client k.

[0148] Example 9:

[0149] A highly energy-efficient federated learning method combining client selection and gradient optimization in a heterogeneous mobile edge computing network, with the same technical content as any one of Examples 2-8, furthermore, the computational quality of mobile client k is as follows:

[0150]

[0151] Where θ and μ are hyperparameters; Collection for clients The calculation model parameters in the maximum delay and maximum energy consumption;

[0152] Among them, the delay required for mobile client k to update the local model The energy consumed by mobile client k to update the local model As shown below:

[0153]

[0154] Where, κ is the number of local iterations; is the number of CPU cycles required to process one sample; v is the effective capacitance coefficient.

[0155] Example 10:

[0156] A highly energy-efficient federated learning method combining client selection and gradient optimization in a heterogeneous mobile edge computing network, with the same technical content as any one of Examples 2-9, further, the communication quality of mobile client k is as follows:

[0157]

[0158] Where μ is a hyperparameter; Collection for clients The maximum delay and maximum energy consumption of China Mobile clients uploading local model parameters to the edge server;

[0159] Among them, the delay of the mobile client uploading local model parameters to the edge server Energy consumed As shown below:

[0160]

[0161] Among them, q k is the transmission power of mobile client k; S is the size of the model parameters that need to be uploaded.

[0162] Example 11:

[0163] A high-energy-efficiency federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network, the technical content of which is the same as any one of Examples 2-10. Furthermore, the tasks requested by the mobile client include intelligent perception, autonomous driving assistance, and smart device management.

[0164] Example 4:

[0165] An energy-efficient federated learning method combining client selection and gradient optimization in a heterogeneous mobile edge computing network, comprising the following steps:

[0166] (1) Design a federated learning framework in a heterogeneous mobile edge computing network. The framework consists of an edge server located at the base station and N mobile clients (we denote the client set as The edge server and the N mobile clients collaborate to train a global model. The N clients have independent CPU clock speed f, transmission speed R, and local dataset D.

[0167] (2) Training is performed in the federated learning framework to ultimately obtain a global model with optimal performance. This model can be used in mobile edge computing networks to perform various tasks such as intelligent perception, autonomous driving assistance, and smart device management, while protecting data privacy and reducing the burden of data transmission. The specific federated learning training process is as follows:

[0168] 2.1 represents the total set of training rounds required at the end of the federated learning process, In the In the round, the edge server sets the global model parameter ω t Sent to all mobile clients, which will then perform model training locally based on the global model parameters.

[0169] 2.2 Edge Server Based on Mobile Client The quality score of mobile client k is calculated by weighting the data quality, computing quality and communication quality, and then the top |C t |Mobile clients with the highest quality scores participate in training. C t For the selected mobile client set, |C t The calculation formula for | is as follows:

[0170]

[0171] Where p represents the proportion of mobile clients.

[0172] To calculate the quality score of a mobile client, assume that the CPU cycles required to process a sample are Then the delay required for mobile client k to update its local model can be calculated as:

[0173]

[0174] Where κ is the number of local iterations. Therefore, the energy consumed by mobile client k to update its local model is:

[0175]

[0176] Where ν is the effective capacitance coefficient.

[0177] Similarly, we assume that the size of the model parameters to be uploaded is S, then the delay of the mobile client uploading the local model parameters to the edge server is:

[0178]

[0179] The energy consumed is:

[0180]

[0181] Among them, q k is the transmit power of mobile client k.

[0182] According to the delay and energy consumption of client calculation and uploading parameters, the calculation quality and communication quality of mobile client k are quantified as:

[0183]

[0184] Among them, θ and μ are hyperparameters.

[0185] At the same time, the data quality of mobile client k is quantified as individual data contribution and overall data value The weighted sum of

[0186]

[0187]

[0188] Among them, λ1, λ2, α, and β are all hyperparameters. is the gradient calculated by mobile client k in round t, y k is the data label of mobile client k. Therefore, the data quality of mobile client k is calculated as:

[0189]

[0190] Finally, based on the data quality, computing quality, and communication quality indicators of the mobile client, its quality score is quantified as follows:

[0191]

[0192] Among them, ρ1, ρ2, and ρ3 represent the weights of data quality, calculation quality, and communication quality, respectively.

[0193] 2.3 The selected mobile client performs parameter updates locally based on the global model parameters sent by the edge server. The specific update rules are as follows:

[0194]

[0195] Where η is the learning rate, Represents the gradient of the local model loss function.

[0196] In order to make the model converge as quickly as possible, this method selectively retains the gradient of the mobile client k that was not selected in the previous round and denotes it as the residual gradient Specifically,

[0197]

[0198] in, Reflect The influence of the current round. These residual gradients contain information that is beneficial to the magnitude and direction of the model update. When mobile client k is selected in the next training round, the residual gradients will be added to the gradients calculated in this round to update the model parameters. The specific calculation formula is:

[0199]

[0200] final is used for local parameter update.

[0201] 2.4 All selected mobile clients upload the updated local model parameters to the edge server.

[0202] 2.5 The edge server performs weighted averaging on the collected parameters to obtain the new round of global model parameters. The specific formula is:

[0203]

[0204] Among them, γ k It is the proportion of the data volume of mobile client k in the overall data volume.

[0205] 2.6 Repeat the above iterative process until the optimal global parameter ω is found * Make the global loss F(ω) reach a predetermined value or the number of training rounds reach a predetermined value.

Claims

1. An energy-efficient federated learning method for combined client selection and gradient optimization in heterogeneous mobile edge computing networks, characterized by: The following steps are involved: 1) Build a federated learning framework in a heterogeneous mobile edge computing network, including an edge server located at a base station and N mobile clients; 2) Train the federated learning framework to obtain a global model consisting of edge servers and N mobile clients; 3) Utilize the global model to execute tasks requested by any mobile client in the mobile edge computing network.

2. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 1, characterized in that: The CPU clock speed f, transmission speed R, and local dataset D of the N mobile clients are all set independently.

3. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 1, characterized in that: In step 2), the steps for training the federated learning framework include: 2.1) Set up the total training wheel set 2.2) In round t=1, the edge server sets the global model parameter ω t Send to all mobile clients; The mobile client performs parameter updates locally based on the global model parameters and transmits the trained global model parameters to the edge server; The global model parameters after training are as follows: Where η is the learning rate, Represents the gradient of the local model loss function; The global model parameters after training for mobile client k; is the global model parameter received by mobile client k; 2.3) The edge server performs weighted averaging on all received global model parameters and updates the global model parameters, namely: Among them, γ k is the proportion of the data volume of mobile client k in the total data volume; ω t+1 is the updated global model parameter; 2.4) Calculate the global loss based on the updated global model parameters where γ k The number of samples of client k accounts for C t The ratio of the total number of samples in the training set; if the global loss F(ω) reaches a predetermined value or the number of training rounds reaches a predetermined value T, the training of the federated learning framework is terminated; otherwise, set t = t + 1 and proceed to step 2.5); 2.5) The edge server calculates the quality score of each mobile client based on the data quality, computing quality and communication quality of the mobile client, and selects the top |C t | mobile clients with the highest quality scores participate in the tth round of training; In round t, the edge server sets the global model parameter ω t Sent to mobile clients participating in training; The mobile client performs parameter updates locally based on the global model parameters, transmits the trained local model parameters to the edge server, and returns to step 2.3).

4. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 3, characterized in that: During training, the mobile client k is used to perform the gradient of global model parameter update locally. As shown below: Where, is the residual gradient; for Impact factor in the current round; is the gradient of the mobile client k that was not selected in the previous round; 5. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 3, characterized in that: Quantity|C t |As shown below: Where p represents the proportion of mobile clients.

6. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 3, characterized in that: The quality scores for mobile clients are as follows: Among them, ρ1, ρ2, and ρ3 represent the weights of data quality, calculation quality, and communication quality, respectively; Respectively represent the data quality, computing quality and communication quality of mobile client k; V k is the quality score of mobile client k.

7. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 6, characterized in that: The data quality of mobile client k is as follows: Where α and β are hyperparameters; Among them, individual data contribution and overall data value They are as follows: Among them, λ1 and λ2 are hyperparameters. is the gradient calculated by mobile client k in round t, y k is the data label of mobile client k; is the maximum gradient; |D k |、|D j | is the data dependency depth.

8. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 6, characterized in that: The calculated quality of mobile client k is as follows: Where θ is a hyperparameter; Collection for clients The calculation model parameters in the maximum delay and maximum energy consumption; Among them, the delay required for mobile client k to update the local model The energy consumed by mobile client k to update the local model As shown below: Where, κ is the number of local iterations; is the CPU cycles required to process one sample; v is the effective capacitance coefficient; f k is the CPU clock frequency.

9. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 6, characterized in that: The communication quality of mobile client k is as follows: Where μ is a hyperparameter; Collection for clients The maximum delay and maximum energy consumption of China Mobile clients uploading local model parameters to the edge server; Among them, the delay of the mobile client uploading local model parameters to the edge server Energy consumed As shown below: Among them, q k is the transmission power of mobile client k; S is the size of the model parameters that need to be uploaded; R k is the transmission rate.

10. The energy-efficient federated learning method for combined client selection and gradient optimization in a heterogeneous mobile edge computing network according to claim 1, characterized in that: The tasks requested by mobile clients include intelligent perception, autonomous driving assistance, and smart device management.

Images