Equipment management method based on contribution degree of edge equipment in federated learning

Abstract

The invention provides a device management method based on the contribution degree of edge devices in federated learning, and belongs to the technical field of edge computing of the Internet of Things. The method comprises the following steps: performing data preprocessing on local training parameters and residual electric quantity information of each edge device in federated learning; grouping all the devices participating in training in the current round by using the local training parameter difference degree between the devices; each group comprises central equipment and common equipment; obtaining a Sharpley value of each group of central equipment as an approximate contribution degree of each group of equipment; obtaining a joint contribution degree of all edge devices by using the residual electric quantity information and the approximate contribution degree of each group of devices; and based on the joint contribution degree and the residual electric quantity information of each edge device, selecting the ID of the device participating in training in the next round. More accurate equipment contribution degree evaluation is realized, and equipment management is carried out in combination with equipment energy consumption; the online time of the equipment is prolonged, the equipment is effectively enabled to support multi-round local training, and high-contribution local parameters are provided for the parameter server.

Description

Technical Field

[0001] The present 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 Art

[0002] Federated learning in the Internet of Things is a distributed machine learning method that generates and collects a large amount of local data through a large number of mobile terminals and sensors, and performs model training on the device terminal without centralizing the data on a central server. The local computing power of the terminal is used to process the data, and only the model parameters are transmitted to ensure data privacy and security. At present, the amount of local data is increasing. Mobile terminals have to use more power to train local models when performing local training. Mobile terminals are usually powered by batteries. Energy consumption is one of the main obstacles that limit the long-term large-scale deployment of federated learning in practical applications. In addition, due to the existence of device heterogeneity and data heterogeneity, the training results of each edge device contribute differently to the overall model, resulting in the problem that the energy consumption of individual terminals is greater than their contribution. To solve the above problems, it is necessary to solve the problem that in actual deployment scenarios, the data sets of each edge device participating in federated learning are uneven in sample size, data quality, and sample categories, which makes it difficult to evaluate the contribution of the device, to achieve accurate evaluation of the contribution of the device, and to effectively manage the terminal energy consumption based on the contribution of the device.

[0003] In the prior art, Chinese patent document CN116502736A discloses a method, device, medium and device for determining the contribution of federated learning. The invention determines the data contribution of the device by comparing the differences between different data sample subsets of the device. Although the device contribution can be obtained through data interaction, which reflects the impact of data differences on training results to a certain extent, this method has the situation that the data is transmitted out of the local area, and there is a risk of data leakage in some cases. Chinese patent document CN112949865A discloses a method for evaluating the contribution of federated learning based on the SIGMA protocol. The invention extracts the gradients of the participants through a trusted execution program for global aggregation, and runs the test module in the trusted execution program to calculate the contribution of the device gradient. Although the contribution of the device can be evaluated by the uploaded gradient of the device, it is overly dependent on the test data set in the test module. This method can only be used in scenarios where all data is known, and the subsequent devices are not selected by the contribution of the gradient, which cannot meet the application in the actual environment of the Internet of Things. Chinese patent document CN116009677A discloses a method for optimizing energy consumption of federated learning devices based on Cell-FreemMIMO network. In the stage of user training local models, the invention reduces the local training power consumption of user devices by iteratively optimizing the operating frequency of the user device CPU and the training accuracy of the local model; in the stage of uploading model parameters of the user device, power allocation is performed by minimizing the maximum delay of the user device, which not only suppresses the laggard effect, but also saves the uplink power consumption of the device. This method can reduce the energy consumption of the device by limiting the hardware operating frequency of the device terminal, but limiting the power consumption of the device will lead to the extension of the local training stage, and it cannot fully support energy consumption management in the case of heterogeneous devices.

[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 effectively manage the energy consumption of devices based on this, thereby extending the online time of devices while providing high-contribution local parameters to the parameter server.

[0006] To this end, the present invention provides the following technical solutions:

[0007] A device management method based on edge device contribution in federated learning, comprising:

[0008] Perform data preprocessing on the local training parameters and remaining power information of each edge device in federated learning; group all devices participating in the training in this round by using the differences in local training parameters between devices; each group includes central devices and ordinary devices; obtain the Shapley value of each group of central devices as the approximate contribution of each group of devices; use the remaining power information and approximate contribution of each group of devices to obtain the joint contribution of all edge devices; select the device ID participating in the training in the next round based on the joint contribution and remaining power information of each edge device;

[0009] The method uses the remaining power information and approximate contribution of each group of devices to obtain the joint contribution of all edge devices, including: according to the local parameters uploaded by the device Group and receive device remaining power information Get 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] Joint contribution by equipment Match with the device ID number to get the device's joint contribution in this round Among them, the contribution of the equipment that did not participate in this round of training is 0;

[0012] The device IDs for the next round of training are selected based on the joint contribution and remaining power information of each edge device, including:

[0013] Use the parameter server to count the number of device training times and obtain the remaining power matrix;

[0014] Each device determines whether there is an empty element in the remaining power matrix through the reader;

[0015] Get the average energy consumption of each device through the remaining power, index set, and device training times

[0016] Obtain the relative joint contribution of all edge devices through device parameter server communication

[0017] Will The elements in the formula are sorted from high to low to get C′(n), combined with the remaining power information Whether to support α times energy consumption The first k devices are selected for the next round of training; the selected device ID set SD is broadcasted to all edge devices by the parameter server.

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

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

[0020] And transmit it to the contribution evaluation module for data alignment;

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

[0022] Furthermore, the method of utilizing the differences in local training parameters between devices to group all devices participating in the training in this round includes:

[0023] Using local training parameters The cosine similarity cs(p,q) between them indicates the difference of local training parameters of each device:

[0024]

[0025] Among them, p,q are the different device IDs participating in the upload in this round, device ID {1,2,…,n}, local training parameters

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

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

[0028] By calculating the Shapley value of the central device parameter of the m groups, the approximate contribution of the device is obtained:

[0029]

[0030] Among them, the local parameters w obtained by the local training of m central devices m As the set of participants N evaluating SV values, set N = {w1, w2, …, w m}、The subset of multiple central device parameters is The utility function V(·) evaluates the joint utility of S, which means the total value of a subset S, and 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 global model after aggregation as a utility function N={w1,w2,…,w m} is a participant, φ m is the approximate contribution of the central device.

[0034] Furthermore, each device determines whether there is an empty element in the remaining power matrix through a reader, including:

[0035] Each device passes the reader Counter n (i) Determine whether there are empty elements in the remaining power matrix:

[0036]

[0037] That is, the device does 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 through the remaining power, index set, and device training times.

[0039]

[0040] Among them, J n (q) = {q|Counter n (i)=1} is the index set of all non-empty elements, Battery(i,n) is the remaining power matrix, that is, the communication times of devices {1,2,…,n} up to this round are counted, and the training times of n devices {t1,t2,…,t n}; The parameter server will end the remaining power information uploaded by the device in this round Form the remaining power matrix Battery(i,n).

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

[0042] If the device participates in training in round i and completes communication with the parameter server, Equal to the joint contribution of the equipment in round i

[0043] If the device does not participate in training and does not communicate with the parameter server in this round, Equal to the device in the last three valid communication rounds (J n (q-2),J n (q-1),J n (q))’s joint contribution average.

[0044] Furthermore, the selected device ID set SD is broadcasted by the parameter server to all edge devices:

[0045]

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

[0047] 1) The present invention can quickly evaluate the contribution of a device through a sophisticated and concise contribution evaluation method, and supports evaluation under a variety of federated learning aggregation methods. The contribution evaluation index not only covers the accuracy of local training parameters, but also adds the remaining power information of the device, so that the evaluated contribution covers more comprehensive device information. In addition, compared with the traditional single contribution evaluation method, this method has the advantages of more streamlined calculation and faster response time, which greatly reduces the resource consumption of the parameter server in evaluating the contribution.

[0048] 2) The present invention selects devices by comprehensively considering device contribution and remaining power. Under the premise of taking into account the accuracy of global parameters, each device can balance local training and energy consumption based on its own remaining power to achieve a longer average online time. Especially for mobile edge devices, the global model training task can be completed with the minimum power consumption under limited battery power.

[0049] 3) The present 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 a variety of differentiated edge devices and global aggregation methods, and provide more efficient and convenient device selection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0051] Figure 1 is a flow chart of the present invention;

[0052] Figure 2 Schematic diagram of the data processing process of receiving device parameters and remaining power information by the contribution evaluation module in an embodiment of the present invention;

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

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

[0055] Figure 5 This is a schematic diagram of a specific device selection process for selecting the device ID for the next round of training in combination with the device contribution and the remaining power in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work 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 and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0058] The present invention provides a device management method based on the contribution of edge devices in federated learning. The invention concept is: for the edge device characteristics of federated learning in the Internet of Things environment, the contribution of the edge device is simply and quickly evaluated by combining the parameters uploaded by the device with the remaining power information to solve the problems of complex contribution evaluation and long calculation cycle at the current stage. At the same time, according to the contribution evaluation results, the local calculation frequency of some devices is reduced, the energy consumption of the device is reduced, and the problem that the current mobile edge devices are limited by the battery capacity and cannot perform long-term federated learning is solved. The evaluation stage is carried out in the parameter server, which has no effect on the energy consumption of the edge device. It can be evaluated under a variety of aggregation models. The algorithm has strong applicability. For the problem of heterogeneous edge devices, the algorithm has strong scalability, and the contribution of heterogeneous devices can be obtained through evaluation. Device management is to select devices by uploading parameters and remaining power information of the device. It can be applied to a variety of edge devices and has strong adaptability.

[0059] Combination Figure 1 Flow chart specifically describes the steps of the present invention:

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

[0061] S2 groups the devices 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 contribution and remaining power;

[0064] Combination Figure 2 , S1: The contribution evaluation module receives the device parameters and the remaining power information, and further describes, specifically including:

[0065] In the i-th round, n edge devices send device ID {1,2,…,n}, local training parameters Remaining battery information Transmitted to the parameter server.

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

[0067] The contribution evaluation module removes the edge device information that did not respond in this round to reduce the amount of data.

[0068] Combination Figure 3 , for S2: group the devices and calculate the approximate contribution of each group through the Shapley value, further explanation, specifically including:

[0069] The contribution evaluation module calculates the received local training parameters The cosine similarity cs(p,q) between them determines the difference between the parameters of each device:

[0070]

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

[0072] By calculating the Shapley value (SV value, Shapley Value) of the central device parameters of m groups, use Represents the local parameters of the central device, and obtains the approximate contribution of the device in this round φ m , that is, m is the group number of the device. The calculation formula of SV value is as follows:

[0073]

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

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

[0076]

[0077] Combination Figure 4 , S3: Calculating the joint contribution of the device by combining the power information and the approximate contribution, further explained, specifically including:

[0078] According to the grouping of devices in S2, combined with the device uploading the remaining power information Get the remaining power of k devices in the group

[0079] Combine the approximate contribution of the central device in the group to which the device belongs φ m Remaining power of devices in the group Calculate the joint contribution of the equipment It is calculated as follows:

[0080]

[0081] Joint contribution by equipment Match with the device ID number to get the joint contribution of the device in the i-th round

[0082] Combination Figure 5 , for S4: combining the device contribution and the remaining power to select the device ID for the next round of training, further explanation, specifically including:

[0083] Step a: Count the communication times of devices {1,2,…,n} up to the i-th round through the parameter server, and obtain the training times of n devices {t1,t2,…,t n}.

[0084] Step b: First, the parameter server will upload the remaining power information of the device up to the i-th round. The remaining power matrix Battery(i,n) is formed; then for each device {1,2,…,n}, the reader Counter n (i) Determine whether there is an empty element in {Battery(1,n),Battery(2,n),…,Battery(i,n)}, that is, 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 index set of all non-empty elements, and the reader Counter n (i) is calculated as follows:

[0085]

[0086] Step c: Combine Battery(i,n), J n (q) and the number of equipment training {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 communication status of the device parameter server If the device participates in training in this round (round i) and completes communication with the parameter server, Equal to the joint contribution of the equipment in this round If the device does not participate in training and does not communicate with the parameter server in this round, Equal to the device in the last three valid communication rounds (J n (q-2),J n (q-1),J n (q))’s joint contribution average.

[0089] Step e: the The elements in the formula are sorted from high to low to get C′(n), combined with the remaining power information Whether to support α times energy consumption The first k devices are selected for the next round of training, and the selected device ID set SD is broadcasted by the parameter server to all edge devices. The calculation method is as follows:

[0090]

[0091] The method of the present invention is further described in conjunction with specific application examples:

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

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

[0094] Table 1

[0095]

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

[0097] Table 2

[0098]

[0099] Combination Figure 3, 9 edge devices (ID: {1, 3, 4, 5, 16, 17, 18, 19, 20}) calculate the difference of parameters through the cosine similarity of local training parameters according to formula (1). The specific results are shown in Table 3:

[0100] Table 3

[0101]

[0102]

[0103] The K-means clustering algorithm is used to divide the devices into three groups according to the differences in the local parameters of the devices. There are three devices in each group. The specific grouping information is shown in Table 4 and Table 5:

[0104] Table 4

[0105] ID 1 3 4 5 16 17 18 19 20 Grouping 3 1 2 3 1 2 2 3 1

[0106] Table 5

[0107]

[0108] According to the grouping information, the device IDs in group 1 are 3, 16 and 20, and the device with ID 16 is the central device in group 1; the device IDs in group 2 are 4, 17 and 18, and the device with ID 17 is the central device in group 2; the device IDs in group 3 are 1, 5 and 19, and the device with ID 5 is the central device in group 3. The central device parameters are selected as the parameters of this group. The specific information is shown in Table 6:

[0109] Table 6

[0110]

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

[0112] Table 7

[0113] Group 1 2 3 <![CDATA[φ m ]]> 32.69 55.63 46.88

[0114] refer to Figure 4 According to the remaining power information uploaded by the device and the device grouping information, the remaining power of the devices in the group after grouping is obtained. The specific information is shown in Table 8:

[0115] Table 8

[0116] 1 group 2 groups 3 groups 68.88 32.49 65.04 73.81 32.3 44.55 36.09 65.91 56.11

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

[0118] Table 9

[0119] 1 group 2 groups 3 groups 37.78 41.49 55.21 40.49 41.24 37.81 19.81 84.16 47.62

[0120] The above results are combined with the grouping information and the equipment participation in training to obtain the joint contribution of the equipment in the 8th round. The joint contribution of the equipment that did not participate in the training is 0. The specific joint contribution information is shown in Table 10:

[0121] Table 10

[0122]

[0123] refer to Figure 5 , the parameter server obtains the local training times of each device based on the device upload status in the first 8 rounds. The specific information is shown in Table 11:

[0124] Table 11

[0125] ID 1 2 3 4 5 6 7 8 9 10 <![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 is combined with the remaining power information uploaded by the device when it last communicated with the parameter server, and the average energy consumption of each device is calculated according to formula (5) and formula (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 the 8th round by the training status of the device in the 8th round and 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 above two results, according to the conditions in formula (7), the equipment participating in the next round of training is 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 8th round global parameters and the 9th round participating training devices (ID: {1, 2, 8, 9, 10, 14, 15, 18}) to all 20 edge devices.

[0137] Compared with the previous device management method, the present invention takes into account the contribution of local parameters uploaded by the device and the influence of the remaining power of the device on the global model from the perspective of device contribution and remaining power, reduces the number of training times of some devices, and prolongs the average online time of the device under the same power. A contribution evaluation module is added to the parameter server to evaluate the contribution of the device. First, the local parameters and remaining power uploaded by the device are preprocessed, and the vacancies of the device that have not been uploaded in this round are removed to reduce the data length. Then, the device is grouped according to the local parameters of the device, and only the Shapley value of the local parameters of each group of central devices is calculated to obtain the approximate contribution of each group, thereby reducing the complexity of calculation and the high time caused by the calculation of Shapley values ​​by a large number of devices. Secondly, the joint contribution of the device is calculated by combining the remaining power information of the device in each group with the approximate contribution of each group. The added remaining power information can not only obtain the device contribution more comprehensively, but also differentiate the approximate contribution of the device in the group. Finally, the qualified device is selected for the next round of training according to the joint contribution and remaining power of the device, which can effectively enable the device to support multiple rounds of local training and provide 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, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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: Perform data preprocessing on the local training parameters and remaining power information of each edge device in federated learning; group all devices participating in the training in this round by using the differences in local training parameters between devices; each group includes central devices and ordinary devices; obtain the Shapley value of each group of central devices as the approximate contribution of each group of devices; use the remaining power information and approximate contribution of each group of devices to obtain the joint contribution of all edge devices; Based on the joint contribution and remaining power information of each edge device, select the device ID that will participate in the next round of training; The method uses the remaining power information and approximate contribution of each group of devices to obtain the joint contribution of all edge devices, including: according to the local parameters uploaded by the device Group and receive device remaining power information Get 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 Joint contribution by equipment Match with the device ID number to get the device's joint contribution in this round Among them, the contribution of the equipment that did not participate in this round of training is 0; The device IDs for the next round of training are selected based on the joint contribution and remaining power information of each edge device, including: Use the parameter server to count the number of device training times and obtain the remaining power matrix; Each device determines whether there is an empty element in the remaining power matrix through the reader; Get the average energy consumption of each device through the remaining power, index set, and device training times Obtain the relative joint contribution of all edge devices through device parameter server communication Will Sort the elements in from high to low to get C ′ (n), combined with the remaining power information Whether to support α times energy consumption The first k devices are selected for the next round of training; the selected device ID set SD is broadcasted to all edge devices by the parameter server.

2. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: The data preprocessing of the local training parameters and remaining power information of each edge device in the federated learning includes: Use the parameter server to obtain the ID, local training parameters, and remaining power information of all edge devices; And transmit it to the contribution evaluation module for data alignment; Remove the unresponsive edge device information and obtain the device information participating in this round of training.

3. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: The method of utilizing the differences in local training parameters between devices to group all devices participating in the training in this round includes: Using local training parameters The cosine similarity cs(p,q) between them indicates the difference of local training parameters of each device: Among them, p,q are the different device IDs participating in the upload in this round, device ID {1,2,…,n}, local training parameters Using the difference cs(p,q) of the local training parameters of each device, the devices participating in this round of training are evenly divided into m groups through the classification method, with k devices in each group. The device corresponding to the group center in the classification method is the central device of each group, and the local parameters of the central device in the group are used as the parameters of the group.

4. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: The obtaining of the Shapley value of each group of central devices as the approximate contribution of each group of devices includes: By calculating the Shapley value of the central device parameter of the m groups, the approximate contribution of the device is obtained: Among them, the local parameters w obtained by the local training of m central devices m As the set of participants N evaluating SV values, set N = {w1, w2, …, w m }、The subset of multiple central device parameters is The utility function V(·) evaluates the joint utility of S, which means the total value of a subset S, and m is the group number of the device.

5. According to claim 4, a device management method based on the contribution of edge devices in federated learning is characterized in that: The utility function V(·) evaluates the joint utility of S, including: Among them, the federated learning aggregation method is used The accuracy of the global model after aggregation as a utility function N={w1,w2,…,w m } is a participant, φ m is the approximate contribution of the central device.

6. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: Each device determines whether there is an empty element in the remaining power matrix through a reader, including: Each device passes the reader Counter n (i) Determine whether there are empty elements in the remaining power matrix: That is, the device does 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. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: The average energy consumption of each device is obtained by the remaining power, index set, and device training times. Among them, J n (q) = {q|Counter n (i)=1} is the index set of all non-empty elements, Battery(i,n) is the remaining power matrix, that is, the communication times of devices {1,2,…,n} up to this round are counted, and the training times of n devices {t1,t2,…,t n }; The parameter server will end the remaining power information uploaded by the device in this round Form the remaining power matrix Battery(i,n).

8. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: According to the communication status of the device parameter server, the relative joint contribution of all edge devices is obtained. include: If the device participates in training in round i and completes communication with the parameter server, Equal to the joint contribution of the equipment in round i If the device does not participate in training and does not communicate with the parameter server in this round, Equal to the device in the last three valid communication rounds (J n (q-2),J n (q-1),J n (q))’s joint contribution average.

9. According to claim 1, a device management method based on the contribution of edge devices in federated learning is characterized in that: The selected device ID set SD is broadcasted by the parameter server to all edge devices:

Images