A device management method based on edge device contribution in federated learning

Abstract

The application provides a device management method based on edge device contribution degree in federated learning, and belongs to the technical field of Internet of Things edge computing. Through data preprocessing on the local training parameters and residual power information of each edge device in federated learning, the local training parameter difference degree between devices is used to group all devices participating in training in this round; each group includes a center device and a common device; the Shapley value of each group center device is obtained as the approximate contribution degree of each group device; the joint contribution degree of all edge devices is obtained by using the residual power information and the approximate contribution degree of each group device; based on the joint contribution degree and the residual power information of each edge device, the device ID participating in training in the next round is selected. The device contribution degree evaluation is more accurate, and the device management is combined with the device energy consumption; the online time of the device is prolonged, the device effectively supports multi-round local training, and the local parameters with high contribution are provided to the parameter server.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology for the Internet of Things, and in particular to a device management method based on the contribution of edge devices in federated learning. Background Technology

[0002] Federated learning in the Internet of Things (IoT) is a distributed machine learning method that generates and collects vast amounts of local data through numerous mobile terminals and sensors, allowing model training to be performed on the devices themselves, eliminating the need to centralize the data on a central server. It leverages the local computing power of the terminals to process the data, transmitting only model parameters to ensure data privacy and security. Currently, with the ever-increasing volume of local data, mobile terminals must dedicate more power to training local models. Since mobile terminals are typically battery-powered, energy consumption is a major obstacle to the long-term, large-scale deployment of federated learning in practical applications. Furthermore, the persistent heterogeneity of devices and data leads to inconsistent contributions from different edge devices to the overall model, resulting in some terminals consuming more energy than they contribute. To address these issues, it is necessary to resolve the problem of inconsistent datasets in terms of sample size, data quality, and sample categories among the edge devices participating in federated learning in real-world deployment scenarios, making it difficult to evaluate device contributions. The goal is to achieve accurate evaluation of device contributions and effectively manage terminal energy consumption based on these contributions.

[0003] In the prior art, Chinese patent document CN116502736A discloses a method, apparatus, medium, and device for determining the contribution of federated learning. This invention determines the data contribution of a device by comparing the differences between different subsets of data samples from different devices. Although the device contribution can be obtained through data interaction, reflecting the impact of data differences on training results to some extent, this method has the risk of data leakage due to local data transmission. Chinese patent document CN112949865A discloses a federated learning contribution evaluation method based on the SIGMA protocol. This invention extracts the gradients of participants through a trusted execution program, performs global aggregation, and runs the test module in the trusted execution program to calculate the contribution of device gradients. Although the contribution of a device can be evaluated through its uploaded gradients, it relies excessively on the test dataset in the test module. This method can only be used in scenarios where all data is known, and it does not select subsequent devices based on gradient contribution, thus failing to meet the application requirements of the Internet of Things (IoT) environment. Chinese patent document CN116009677A discloses a device-side power consumption optimization method based on Cell-Free mMIMO networks for federated learning. This invention reduces the local training power consumption of the user device by iteratively optimizing the CPU operating frequency and training accuracy of the local model during the user's local model training phase. During the user device's model parameter uploading phase, power allocation is performed with the goal of minimizing the user device's maximum latency, thus suppressing the laggard effect and saving uplink power consumption. While this method can reduce device power consumption by limiting the hardware operating frequency of the device terminal, limiting device power consumption leads to a prolonged local training phase and cannot fully support power management in heterogeneous device environments.

[0004] Therefore, a method is needed to effectively manage the energy consumption of federated learning terminals based on accurate device contribution. Summary of the Invention

[0005] In view of this, the present invention provides a device management method based on the contribution of edge devices in federated learning, so as to achieve accurate evaluation of the contribution of edge devices and effective management of device energy consumption based on this, thereby extending the online time of devices and providing high-contribution local parameters to the parameter server.

[0006] Therefore, the present invention provides the following technical solution:

[0007] A device management method based on edge device contributions in federated learning includes:

[0008] The local training parameters and remaining power information of each edge device in federated learning are preprocessed; the differences in local training parameters between devices are used to group all devices participating in the training in this round; each group includes central devices and ordinary devices; the Shapley value of the central device in each group is obtained as the approximate contribution of each group; the joint contribution of all edge devices is obtained using the remaining power information and approximate contribution of each group; based on the joint contribution and remaining power information of each edge device, the device ID to participate in the training in the next round is selected.

[0009] The process of obtaining the joint contribution of all edge devices by utilizing the remaining power information and approximate contribution of each group of devices includes: based on local parameters uploaded by the devices. Group the devices and receive their remaining battery power information. Obtain the remaining power in each group Where k is the number of devices in each group; based on the approximate contribution of each group of devices and the remaining power of each device in the group, the joint contribution of each device participating in this round of training is obtained.

[0010]

[0011] Through the joint contribution of equipment By matching the device ID number, the joint contribution of the device in this round can be obtained. Among them, the contribution of equipment that did not participate in this round of training is 0;

[0012] The selection of device IDs for the next training round based on the joint contribution and remaining battery power information of each edge device includes:

[0013] The parameter server counts the number of times the device has been trained and obtains the remaining battery power matrix.

[0014] Each device uses a reader to determine if there are empty elements in the remaining power matrix;

[0015] The average energy consumption of each device is obtained by using remaining battery power, index set, and device training count.

[0016] The relative joint contribution of all edge devices is obtained by analyzing communication data with the device parameter server.

[0017] Will The elements are sorted from highest to lowest to obtain C′(n), and then combined with the remaining battery information. Does it support alpha energy consumption? The training process selects the top k devices for the next round of training; the selected device ID set SD is broadcast by the parameter server to all edge devices.

[0018] Furthermore, the data preprocessing of the local training parameters and remaining battery information of each edge device in federated learning includes:

[0019] Use the parameter server to obtain the IDs, local training parameters, and remaining battery information of all edge devices;

[0020] The data is then transmitted to the contribution assessment module for data alignment.

[0021] Remove unresponsive edge device information to obtain device information participating in this round of training.

[0022] Furthermore, the method of grouping all devices participating in the training in this round based on the differences in local training parameters between devices includes:

[0023] Using local training parameters The cosine similarity cs(p,q) between the devices represents the degree of difference in the local training parameters of each device:

[0024]

[0025] Where p and q represent the different device IDs participating in the upload in this round, the device IDs {1,2,…,n}, and the local training parameters.

[0026] Using the difference in local training parameters of each device, cs(p,q), the devices participating in this round of training are divided into m groups, with k devices in each group, through a classification method. The device corresponding to the group center in the classification method is the center device of each group, and the local parameters of the center device in each group are used as the parameters of that group.

[0027] Furthermore, obtaining the Shapley value of each group of central devices as an approximate contribution of each group of devices includes:

[0028] The Shapley values ​​of the central equipment parameters for each of the m groups are calculated as approximate contributions of the equipment.

[0029]

[0030] Among them, the local parameters w are obtained by training m central devices locally. m The set N of participants evaluating the SV value is set N = {w1, w2, ..., w...} m A subset of multiple central equipment parameters is The utility function V(·) evaluates the joint utility of S, which represents the total value of a subset S, where m is the group number of the device.

[0031] Furthermore, the utility function V(·) evaluates the joint utility of S, including:

[0032]

[0033] Among them, the federated learning aggregation method is used. The accuracy of the aggregated global model is used as a utility function. N = {w1, w2, ..., w m} represents the participant, φ m Approximate contribution of the central equipment.

[0034] Furthermore, each device determines whether there are empty elements in the remaining power matrix using a reader, including:

[0035] Each device is accessed via a reader (Counter). n (i) Determine if there are empty elements in the remaining power matrix:

[0036]

[0037] This means that the device did not communicate with the parameter server in a certain round, the bit of the empty element is 0, and the bit of the non-empty element is 1.

[0038] Furthermore, the average energy consumption of each device is obtained by using remaining battery power, index set, and device training count.

[0039]

[0040] Among them, J n (q)={q|Counter n {i, 1} is the index set of all non-empty elements, and Battery(i, n) is the remaining battery power matrix. This means counting the number of communications by devices {1, 2, ..., n} up to the end of the current round, and obtaining the training counts {t1, t2, ..., tn} for n devices. n The parameter server will update the remaining battery information uploaded by the device up to the end of this round. The remaining battery capacity matrix Battery(i,n) is formed.

[0041] Furthermore, based on the communication status of the device parameter server, the relative joint contribution of all edge devices is obtained. include:

[0042] If the device participates in training and completes communication with the parameter server in round i. Equals the joint contribution of the devices in round i.

[0043] If the device did not participate in training or communicate with the parameter server in this round. This is equivalent to the device in the last three valid communication rounds (J n (q-2),J n (q-1),J n The average joint contribution of (q)).

[0044] Furthermore, the selected set of device IDs SD is broadcast by the parameter server to all edge devices:

[0045]

[0046] Advantages and positive effects of the present invention:

[0047] 1) This invention employs a sophisticated and concise contribution evaluation method, enabling rapid assessment of device contributions and supporting evaluation across various federated learning aggregation methods. The contribution evaluation metrics not only encompass the accuracy of locally trained parameters but also incorporate remaining battery power information, resulting in a more comprehensive assessment of device information. Furthermore, compared to traditional single contribution evaluation methods, this approach offers advantages such as simplified computation and faster response time, significantly reducing the resource consumption of the parameter server in contribution evaluation.

[0048] 2) This invention selects devices by comprehensively considering both device contribution and remaining power. While maintaining the accuracy of global parameters, it allows each device to balance local training and energy consumption based on its remaining power, achieving a longer average online time. This is particularly beneficial for mobile edge devices, enabling them to complete the global model training task with minimal power consumption within limited battery capacity.

[0049] 3) This invention has good adaptability and scalability. Federated learning device management is a distributed device selection method in various environments. Whether in heterogeneous edge device scenarios with different device hardware or in various emerging federated learning methods with different global aggregation methods, the device management method based on the contribution of edge devices in federated learning can adapt to various differentiated edge devices and global aggregation methods, and provide relatively efficient and convenient device selection results. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the data processing procedure of the contribution evaluation module receiving device parameters and remaining power information in an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram illustrating the specific calculation process of grouping devices and calculating the approximate contribution of each group using the Shapley value in an embodiment of the present invention.

[0054] Figure 4 This is a schematic diagram illustrating the calculation process of the joint contribution of the device combining power information and approximate contribution calculation in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram illustrating the specific device selection process in this embodiment of the invention, which combines device contribution and remaining battery power to select the device ID to participate in the next training round. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] This invention provides a device management method based on the contribution of edge devices in federated learning. The inventive concept is as follows: Addressing the characteristics of edge devices in federated learning within an IoT environment, this method combines parameters uploaded by the devices with remaining battery power information to perform a simple and rapid evaluation of their contributions, thus solving the problems of complex contribution evaluation and long computation cycles in current methods. Simultaneously, based on the contribution evaluation results, the local computation frequency of some devices is reduced, decreasing device power consumption and solving the problem that mobile edge devices are limited by battery capacity and cannot perform long-term federated learning. The evaluation phase is performed on a parameter server, having no impact on the power consumption of edge devices. It can be performed under various aggregation models, exhibiting strong applicability and scalability for heterogeneous edge devices, allowing the evaluation to obtain the contribution of heterogeneous devices. Device management involves selecting devices based on uploaded parameters and remaining battery power information, making it highly adaptable to various edge devices.

[0059] Combination Figure 1 The flowchart illustrates the steps of this invention in detail:

[0060] The S1 contribution assessment module is used to receive equipment parameters and remaining power information;

[0061] S2 groups the equipment and calculates the approximate contribution of each group using the Shapley value;

[0062] S3 calculates the joint contribution of the devices based on the received power information and the approximate contribution.

[0063] S4 selects the device ID to participate in the next round of training based on the device's contribution and remaining battery power.

[0064] Combination Figure 2 Further explanation is provided regarding S1: the contribution assessment module's receipt of device parameters and remaining power information, specifically including:

[0065] In the i-th round, n edge devices will store device IDs {1,2,…,n} and local training parameters. Information on remaining battery power of the device Transmitted to the parameter server.

[0066] The parameter server will receive the ID, and The data is transferred to the built-in contribution assessment module and aligned.

[0067] The contribution assessment module removes information about edge devices that did not respond in the current round, reducing the amount of data.

[0068] Combination Figure 3 For S2: The equipment is grouped, and the approximate contribution of each group is calculated using the Shapley value. Further explanation includes:

[0069] The contribution evaluation module calculates the received local training parameters. The cosine similarity cs(p,q) between the devices is used to determine the degree of difference in the parameters of each device.

[0070]

[0071] Based on the difference cs(p,q) between the local parameters of the devices, a classification method is used, preferably K-means clustering in this embodiment; the devices are divided into m groups, with k devices in each group, and the local parameters of the central device in each group are denoted as the group parameters.

[0072] By calculating the Shapley value (SV value) of the central device parameters for m groups, using... This represents the local parameters of the central device, and yields the device's approximate contribution φ in this round. m That is, m is the group number of the device. The formula for calculating the SV value is as follows:

[0073]

[0074] Where N is the set of all participants who contribute to the evaluation, S is a subset of N, and the utility function V(·) can take any form to evaluate the joint utility of S, that is, to represent the total value of a subset S.

[0075] In this embodiment, the specific method is as follows: The local parameters w obtained by training m central devices locally are... m As participants in evaluating the SV value, the set of participants is N, N = {w1, w2, ..., w m A subset of multiple central equipment parameters is... Federated learning aggregation method The accuracy of the aggregated global model is used as a utility function. Participants {w1,w2,…,w m The SV value of} is the approximate contribution φ of the central equipment. m The specific calculation formula is as follows:

[0076]

[0077] Combination Figure 4 Regarding S3: Further explanation is needed regarding the joint contribution of the device calculated by combining power information with approximate contribution rates, specifically including:

[0078] Based on the device grouping in S2, combined with the remaining battery information uploaded by the device. Get the remaining power of k devices in the group

[0079] Based on the approximate contribution φ of the central equipment in the group to which the equipment belongs m Remaining power of equipment within the group Calculate the joint contribution of the equipment The calculation method is as follows:

[0080]

[0081] Through the joint contribution of equipment By matching the device IDs, the joint contribution of the devices in round i can be obtained.

[0082] Combination Figure 5 Regarding S4: The selection of device IDs for the next training round based on device contribution and remaining battery power is further explained, specifically including:

[0083] Step a: Count the number of communications between devices {1,2,…,n} up to the i-th round using the parameter server, and obtain the training counts {t1,t2,…,t} for n devices. n}

[0084] Step b: First, the parameter server will upload the remaining battery information from the device up to round i. Construct a remaining battery power matrix Battery(i,n); then for each device {1,2,…,n}, use a reader Counter. n (i) Determine if there is an empty element in {Battery(1,n),Battery(2,n),…,Battery(i,n)}, i.e., the device did not communicate with the parameter server in a certain round. The bit of the empty element is 0, and the bit of the non-empty element is 1. J n (q)={q|Counter n (i) = 1} is the set of indices for all non-empty elements, and the reader is Counter. n (i) is calculated as follows:

[0085]

[0086] Step c, combine Battery(i,n) and J n (q) and the number of equipment training iterations {t1,t2,…,t} n Calculate the average energy consumption of all devices. The calculation method is as follows:

[0087]

[0088] Step d: Calculate the relative joint contribution of n edge devices based on the device parameters and server communication status. If the device participates in training and completes communication with the parameter server in this round (round i), Equals the combined contribution of the device in this round If the device did not participate in training or communicate with the parameter server in this round. This is equivalent to the device in the last three valid communication rounds (J n (q-2),J n (q-1),J n The average joint contribution of (q)).

[0089] Step e: Obtain the result from step d The elements are sorted from highest to lowest to obtain C′(n), and then combined with the remaining battery information. Does it support alpha energy consumption? For training, the first k devices are selected for the next round of training. The selected device ID set SD is broadcast by the parameter server to all edge devices, and its calculation method is as follows:

[0090]

[0091] The method of the present invention will be further explained with reference to specific application embodiments:

[0092] In this embodiment, federated learning processes a linear regression problem involving 20 edge devices. There are 1100 data samples, each containing 6 features. The parameter server retains 100 data samples, and the remaining 1000 data samples are evenly distributed among the 20 edge devices, with each edge device receiving 50 data samples. After 3 rounds of responses from all devices and 4 rounds of responses from some devices...

[0093] Combination Figure 2 In round 8, some devices (ID: {1, 3, 4, 5, 16, 17, 18, 19, 20}) performed local training and transmitted the training results and remaining battery power to the parameter server, as shown in Table 1:

[0094] Table 1

[0095]

[0096] In the parameter server, the contribution evaluation module removes information about devices that did not participate in the calculation, as shown in Table 2:

[0097] Table 2

[0098]

[0099] Combination Figure 3Nine edge devices (ID: {1, 3, 4, 5, 16, 17, 18, 19, 20}) calculated the parameter difference using the cosine similarity of their locally trained parameters according to formula (1). The specific results are shown in Table 3.

[0100] Table 3

[0101]

[0102]

[0103] The devices were divided into three groups based on the degree of difference in their local parameters using the K-means clustering algorithm. Each group contained three devices. The specific grouping information is shown in Tables 4 and 5.

[0104] Table 4

[0105] Grouping 3 1 2 3 1 2 2 3 1

[0106] Table 5

[0107]

[0108] Based on the grouping information, we know that in Group 1, the device IDs are 3, 16, and 20, with device ID 16 being the central device in Group 1; in Group 2, the device IDs are 4, 17, and 18, with device ID 17 being the central device in Group 2; and in Group 3, the device IDs are 1, 5, and 19, with device ID 5 being the central device in Group 3. The parameters for the central device are selected as the parameters for this group. Specific information is shown in Table 6.

[0109] Table 6

[0110]

[0111] The approximate contribution of each central device in the three groups (ID: {16, 17, 5}) is represented by the SV value according to formula (3). The specific results are shown in Table 7.

[0112] Table 7

[0113] <![CDATA[φ m ]]> 32.69 55.63 46.88

[0114] refer to Figure 4 Based on the remaining power information uploaded by the devices and the device grouping information, the remaining power of the devices within each group is calculated. Specific information is shown in Table 8.

[0115] Table 8

[0116] 68.88 32.49 65.04 73.81 32.3 44.55 36.09 65.91 56.11

[0117] The above information, combined with the approximate contribution of the central equipment in the group to which the equipment belongs, is used to calculate the joint contribution of the equipment according to formula (4). The specific calculation results are shown in Table 9.

[0118] Table 9

[0119] 37.78 41.49 55.21 40.49 41.24 37.81 19.81 84.16 47.62

[0120] The above results, combined with grouping information and device participation in training, yielded the joint contribution of the devices in round 8. The joint contribution of devices that did not participate in training was 0. Specific joint contribution information is shown in Table 10.

[0121] Table 10

[0122]

[0123] refer to Figure 5 The parameter server obtains the number of local training iterations for each device based on the device upload data from the first 8 rounds. Specific information is shown in Table 11.

[0124] Table 11

[0125] <![CDATA[t * ]]> 5 4 5 8 7 5 5 5 3 6 ID 11 12 13 14 15 16 17 18 19 20 <![CDATA[t * ]]> 5 7 3 3 4 5 7 5 6 7

[0126] The above information, combined with the remaining power information uploaded by the device during its last communication with the parameter server, is used to calculate the average power consumption of each device according to formulas (5) and (6). The specific results are shown in Table 12.

[0127] Table 12

[0128]

[0129]

[0130] The parameter server calculates the relative joint contribution of the device in round 8 by combining the device's training status in round 8 with the joint contribution of the device in the first 7 rounds. The specific results are shown in Table 13.

[0131] Table 13

[0132]

[0133] Combining the results of the two parts above, and based on the conditions in formula (7), the devices to participate in the training in the next round are selected. The specific results are shown in Table 14:

[0134] Table 14

[0135] SD 18 9 2 14 8 1 10 15

[0136] The parameter server broadcasts the global parameters for round 8 and the training devices (IDs: {1, 2, 8, 9, 10, 14, 15, 18}) for round 9 to all 20 edge devices.

[0137] Compared to previous device management methods, this invention comprehensively considers the impact of device-uploaded local parameters and remaining battery power on the global model from the perspectives of device contribution and remaining battery power. This reduces the number of training iterations for some devices and extends the average online time of devices under the same battery power. A contribution evaluation module is added to the parameter server to assess device contribution. First, the uploaded local parameters and remaining battery power are preprocessed to remove empty slots for devices that did not upload in the current round, reducing data length. Then, devices are grouped according to their local parameters, and only the Shapley value of the local parameters of the central device in each group is calculated to obtain the approximate contribution of each group, reducing computational complexity and excessive time caused by calculating Shapley values ​​for a large number of devices. Second, the joint contribution of devices is calculated by combining the remaining battery power information of devices within each group with the approximate contribution of each group. The added remaining battery power information not only provides a more comprehensive understanding of device contribution but also differentiates the approximate contribution of devices within a group. Finally, devices that meet the criteria are selected for the next round of training based on the joint contribution and remaining battery power. This effectively enables devices to support multi-round local training and provides high-contribution local parameters to the parameter server.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A device management method based on the contribution of edge devices in federated learning, characterized in that, include: Data preprocessing is performed on the local training parameters and remaining power information of each edge device in federated learning; the differences in local training parameters between devices are used to group all devices participating in the training in this round; each group includes central devices and ordinary devices; the Shapley value of the central device in each group is obtained as the approximate contribution of each group; the joint contribution of all edge devices is obtained by using the remaining power information and approximate contribution of each group. Based on the joint contribution and remaining power information of each edge device, select the device ID to participate in the next round of training; The process of obtaining the joint contribution of all edge devices by utilizing the remaining power information and approximate contribution of each group of devices includes: based on local parameters uploaded by the devices. Group the devices and receive their remaining battery power information. , obtained the Remaining power of equipment in the group ,in, For edge devices, the device ID number. Number of rounds For the first Number of devices in the group The devices are grouped and numbered; based on the approximate contribution of each group and the remaining power of each device within the group, the joint contribution of each device participating in this round of training is obtained. : in, Indicates the first The first in the group The combined contribution of individual devices Indicates the first The device ID within the group is The corresponding remaining power, Indicates by the first The approximate contribution calculated by the group's central equipment. Indicates the first The total remaining battery power of all devices in the group; Through the joint contribution of equipment By matching the device ID number, the joint contribution of the device in this round can be obtained. Of these, the contribution of equipment that did not participate in this round of training was 0. The selection of device IDs for the next training round based on the joint contribution and remaining battery power information of each edge device includes: The parameter server counts the number of times the device has been trained and obtains the remaining battery power matrix. Each device uses a reader to determine if there are empty elements in the remaining power matrix; The average energy consumption of each device is obtained by using remaining battery power, index set, and device training count. ; The relative joint contribution of all edge devices is obtained by analyzing communication data with the device parameter server. ; Will Sort the elements from highest to lowest to obtain Combined with remaining battery power information Does it support? Secondary energy consumption is The training, among which, Given the set of indices of all non-empty elements output by the reader, select the first... Each device will undergo the next round of training; the selected set of device IDs will be used. The parameter is broadcast to all edge devices by the parameter server.

2. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, The data preprocessing of the local training parameters and remaining battery information of each edge device in federated learning includes: Use the parameter server to obtain the IDs, local training parameters, and remaining battery information of all edge devices; The data is then transmitted to the contribution assessment module for data alignment. Remove unresponsive edge device information to obtain device information participating in this round of training.

3. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, The method of grouping all devices participating in the training in this round based on the differences in local training parameters between devices includes: Using local training parameters Cosine similarity between This indicates the degree of difference in local training parameters across different devices: in, For the different device IDs and device IDs that participated in the upload in this round Local training parameters ; By utilizing the differences in local training parameters across devices, a classification method is used to average the devices participating in this round of training into Groups, each group contains For each device, the device corresponding to the group center in the classification method is the group center device, and the local parameters of the center device within the group are used as the group parameters. .

4. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, The process of obtaining the Shapley value of each group of central devices as an approximate contribution of each group of devices includes: By calculating the first The Shapley value of the central equipment parameter of the group is used as an approximate contribution of the equipment: Among them, the Local parameters obtained by the group center device during local training As a set of participants in evaluating SV value ,gather A subset of multiple central equipment parameters is Utility function Evaluate The joint utility, that is, representing a subset Total value, The group number to which the device belongs. This represents the number of groups.

5. The device management method based on edge device contribution in federated learning according to claim 4, characterized in that, The utility function Evaluate The combined effects include: Among them, the federated learning aggregation method is used. The accuracy of the aggregated global model is used as a utility function. ; As a participant, For the first Approximate contribution of the group's central equipment.

6. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, Each device uses a reader to determine whether there are empty elements in the remaining power matrix, including: Each device is accessed via a reader. Determine if there are any empty elements in the remaining battery power matrix: This means that the device did not communicate with the parameter server in a certain round, the bit of the empty element is 0, and the bit of the non-empty element is 1.

7. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, The average energy consumption of each device is obtained by using remaining battery power, index set, and device training count. : in, The set of indices for all non-empty elements This is a remaining power matrix, which represents the data up to the current cycle for each device. The number of communications, obtain Number of training sessions per device The parameter server will update the remaining battery information uploaded by the device up to the end of this round. Forming a residual power matrix .

8. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, The relative joint contribution of all edge devices is obtained based on the communication status of the device parameter server. ,include: If the device is in the Each round participates in training and communicates with the parameter server. Equal to the device in the first Combined contribution of rounds ; If the device did not participate in training or communicate with the parameter server in this round. This is equivalent to the device in the last three valid communication rounds. The average of their combined contributions.

9. The device management method based on edge device contribution in federated learning according to claim 1, characterized in that, The selected set of device IDs The parameter server broadcasts the information to all edge devices: 。

Images