Edge computing and resource optimization method based on federated learning

Abstract

The present invention discloses an edge computing and resource optimization method based on federated learning, deploys antenna arrays in small base stations, obtains channel information of downlink and channel information after being processed by precoding technology, forms training data with channel and precoding as input and output pairs, and performs federated learning under the support of this data, that is, trains the model at the node end, and finally achieves the purpose of inputting channel information to obtain corresponding precoding information. In this process, in order to obtain a relatively stable learning alliance and control the system energy consumption to the lowest state, user selection is performed, that is, through the physical characteristics of each node, users with stable computing power and communication capabilities are selected from many users to participate in the training; in order to encourage users to actively participate in the training, a contract mechanism is introduced to reward users participating in the training, and the income of each user and the training cost paid are calculated to obtain the utility function, and resources are allocated to users so that the utility of the entire system is maximized.

Description

Technical Field

[0001] The present invention belongs to the field of communication and federated learning. It combines the precoding technology in MIMO with machine learning and applies it to edge computing. It makes the federated learning structure more stable and efficient through selection and resource optimization. Background Art

[0002] Massive Multiple-Input Multiple-Output (MIMO) technology deploys a large-scale array at the base station end. Compared with traditional MIMO, it can effectively resist interference between different users and significantly improve the system capacity. In a massive MIMO system, precoding technology is a crucial signal processing technology in the downlink. It uses the channel state information (CSI) at the transmitter to transform the modulated symbol stream into a data stream that adapts to the current channel, concentrates the signal energy near the target user, effectively resists attenuation and loss, and improves system performance. Specifically, the CSI is used at the transmitter end of the downlink to preprocess the transmitted signal, minimize the interference between different users and antennas, and concentrate the signal energy near the target user, so that the receiver obtains a better signal-to-noise ratio (SNR) and improves the system channel capacity [1]. In the context of machine learning, traditional federated learning is often used. The data from the edge node is sent to the server for processing. The server then distributes the data to the node for data training. This greatly reduces the privacy of the data. In order to protect the data, decentralized federated learning is proposed and well used in the intelligent field. The node does not need to share the data, but only needs to be trained locally. In order to improve the feasibility of the data, it is necessary to train a weight from the data and send it to the server. At this time, the server plays a role of collection. It collects the data weights trained by each node and processes them accordingly, and then broadcasts the processed weights to each node. After the node receives the latest data weight, it updates the weight and applies it to the next round of iteration. This cycle continues until the data is trained to the best model. The traditional system architecture consists of mobile nodes and base stations [2]. The limited computing and traffic capabilities of mobile devices cause machine learning to encounter two major bottlenecks, namely communication overhead and energy overhead. The powerful function of mobile edge computing (MEC) is considered to be a promising distributed computing paradigm. It supports many emerging intelligent applications in the 5G era, such as video streaming, smart cities, and augmented reality. Edge learning allows delay-sensitive and computationally intensive tasks to be offloaded from distributed mobile devices to close-range edge servers, providing real-time response and high energy efficiency. Based on this idea, a hierarchical federated edge learning framework is introduced, in which edge servers are usually deployed fixedly with base stations as intermediaries between mobile devices and the cloud. They can execute local models transmitted from nearby devices for edge aggregation. When each local device reaches a given learning accuracy, it transmits it to the edge for global aggregation to achieve model update. In this way, the communication and energy overhead can be reduced [3], thus entering a three-layer communication system structure.

[0003] [1] Zhang Yu, Zhao Xiongwen. Precoding technology in millimeter wave massive MIMO system[J]. ZTE Communications Technology, 2018, 024(003): 26-31

[0004] [2]Elbir AM,Coleri S. Federated Learning for Hybrid Beamforming inmm-Wave Massive MIMO[J]. IEEE Communications Letters, 2020, PP(99):1-1.

[0005] [3] S.Luo, Communications, vol.19, no.10, pp.6535-6548, Oct.2020, doi:10.1109 / TWC.2020.3003744. Summary of the invention

[0006] The purpose of the present invention is to introduce MEC into the communication system structure of federated learning to form a three-layer structure, namely, macro base station (cloud server), small base station (edge ​​server) and node user, to allocate communication resources to users under the application background of federated learning so that the whole system can achieve maximum utility, and at the same time, to link MIMO precoding technology with federated learning training data, so as to achieve the prediction of corresponding precoding when the channel data is known. First, an antenna array is deployed in a small base station to obtain the channel information of the downlink and the channel information after processing by the precoding technology, to form training data with the channel and precoding as input and output pairs, and to carry out federated learning under the support of this data, that is, to train the model at the node end, and finally to achieve the purpose of obtaining the corresponding precoding information by inputting the channel information. During this process, in order to obtain a more stable learning alliance and control the system energy consumption to the lowest state, user selection is carried out. That is, users with stable computing and communication capabilities are selected from many users to participate in the training through the physical characteristics of each node. At the same time, in order to encourage users to actively participate in the training, a contract mechanism is introduced to reward users who participate in the training, and the income and training cost of each user are calculated to obtain the utility function (utility = income - cost), and resources are allocated to users to maximize the utility of the entire system.

[0007] To solve the above problems, the present invention adopts the following technical solution: an edge computing and resource optimization method based on federated learning, which includes the following steps:

[0008] Step 1: Build a system model based on users, edge servers, and base stations and deploy the array

[0009] The present invention is a multi-user millimeter wave MIMO three-layer structure system structure. The first layer is the user layer, which consists of N single-antenna user devices. The second layer is the ES layer (edge ​​server layer), which includes K with S T The third layer is the CS layer (cloud server layer), which has a cloud server S with a single antenna. Among them, the user layer and the ES layer communicate through wireless channels, and the ES layer and the CS layer are connected through a backhaul link with sufficient capacity.

[0010] Step 2: First, select the set of users that communicate with the edge server based on the distance and represent it as Where N is the number of users, It indicates the number of users selected after distance selection.

[0011] That is, assuming that the user equipment follows a Poisson distribution, the communication range of each edge server is a radius of d k In this three-level structure, it is assumed that there is no overlap between the circular communication areas of the edge servers. Then, users with closer distances are selected based on the distance from each user to the edge server, because users within the communication range have better communication quality with the server, and stable communication is a necessary condition to ensure the smooth progress of federated learning.

[0012] Step 3: Acquisition of training data.

[0013] The channel model is an LOS (line of sight) receiving route with an L path. The millimeter wave channel matrix of the nth user is expressed as:

[0014]

[0015] in, α n,l is the complex channel gain of the lth path, is the direction angle of the downlink transmission route, YesS T ×1 direction vector, whose mth element is expressed as:

[0016]

[0017] λ is the wavelength, σ is the distance between the two antennas, and σ=λ / 2.

[0018] First in The user transmission direction angle is generated within the range. The L-path channel transmission direction angle is randomly generated within the fluctuation range of 5°, and then h is calculated according to the channel matrix formula.n , the angle interval Divide into Q non-overlapping parts equally, each sub-interval Each interval range is marked starting from 1. The interval where the user transmission direction angle is located is marked as q, and the q value is the interval mark value, thus generating (h n ,q) data pair. The final RF precoding f RF,n It can be expressed as in is the midpoint of the interval.

[0019] Step 4: Perform training based on the selected users and the generated MIMO data.

[0020] The training process can be divided into edge aggregation and cloud aggregation

[0021] 1) Edge aggregation: Edge aggregation is divided into local model calculation, local model transmission and edge model aggregation. The details are as follows:

[0022] Local model calculation: After generating MIMO data locally, user n uses local data to train a linear regression model based on gradient descent. When the local model accuracy θ is reached, n ∈[0,1], that is, the number of iterations is L(θ n )=μlog(1 / θ n ), generate the model parameters w t And calculate the model's loss function f i (w t ). In this process, computational delay is generated and computing energy The computing delay and the computing energy consumed can be derived based on the amount of data the user is training and the frequency of data processing.

[0023] Local model transmission: After the local calculation is completed, user n transfers the model parameters w t Uploaded to the edge server k with which it communicates, during the transmission process, the OFDMA (Orthogonal Frequency Division Multiple Access) protocol is used, which generates a transmission delay and transfer energy The transmission delay and transmission energy can be calculated by the Shannon formula according to the transmission rate between user n and edge server k and the bandwidth allocated by server k to user n.

[0024] Edge aggregation: In this process, the edge server aggregates the model parameters uploaded by each user and then averages them to update the model parameters (i.e., federated averaging algorithm). After the edge server aggregates the model parameters and performs the averaging algorithm, it sends the updated model parameters to each user by broadcasting (the delay and energy consumption during the broadcast process are negligible). After the user receives the model parameters sent by the server, it uses this parameter for a new round of data training. This cycle repeats until the edge server reaches its model accuracy value ε, and the number of global aggregations is expressed as Where δ is the learning task parameter. When I(ε,θ n The total delay / energy consumption after ) global aggregation is the sum of the calculation delay / energy consumption and the transmission delay / energy consumption.

[0025] 2) Cloud aggregation: Cloud aggregation is divided into two steps: edge model upload and cloud model aggregation.

[0026] Edge model upload: After the edge server k reaches the edge accuracy, the edge model parameters are uploaded to the cloud server, which causes a delay in the process. and energy consumption It can be calculated based on the transmission rate and transmission power.

[0027] Cloud model aggregation: The cloud server aggregation process is similar to the edge aggregation process, that is, the cloud server receives the model parameters uploaded by the edge server and then performs average calculation.

[0028] The key points of the present invention are:

[0029] Step 1: Allocate resources to users

[0030] Secondary selection: In order to further obtain a stable and efficient learning alliance, a second user selection is performed before allocating resources to the user. The basis for this selection is the user's previous training record. That is, before making a user selection, the user must perform local iteration and achieve the corresponding accuracy. Then, the gradient value is calculated based on the locally calculated loss function, and the difference is made with the gradient gathered by the edge server. The user with the smallest gradient difference is selected to form the final learning alliance. in is the set of users after the first selection. is the set of users after two selections.

[0031] Resource allocation: After the second selection, the alliance is successfully formed, and then resources are allocated to the users who finally participate in the training. In this process, the contract theorem is introduced to encourage and reward users who participate in the training. First, the users are classified according to the local accuracy (the smaller the better), that is, Then sort and set the user's evaluation function in It is a reward for users who participate in alliance learning. is the profit function, and as In this way, the user's computational utility function can be obtained, that is,

[0032]

[0033] Edge Server Offer Contract Terms Users choose the terms that suit them according to their actual situation, based on the constraints of the contract theorem, namely IR (individual rationality constraint) and IC (incentive compatibility constraint).

[0034] IR: The user's income and cost loss are at the lowest level, that is, If u' n <0, the user cannot participate in the training and is abandoned.

[0035] IC: Users must match their own contract terms when selecting contract terms, i.e.

[0036]

[0037] According to the above, we can get the user's utility function u n and the utility function U of the edge server k (The difference between the cost function and the revenue function, the utility function of the cloud server is negligible. In summary, the utility function of the entire system can be obtained

[0038]

[0039] Where K is the set of edge servers, It is the set of users after two selections.

[0040] The purpose of resource allocation is to minimize the system communication overhead and maximize the system utility function, so an optimization problem can be proposed, namely

[0041] max U,

[0042] subject to,

[0043] C1:

[0044] C2:

[0045] C3:

[0046] C4:

[0047] C5:

[0048] C6:

[0049] Among them, the constraints C1 and C2 represent the uplink communication resource constraints and computing power constraints respectively, β k,n is the bandwidth allocated by edge server k to user n, f n is the computation frequency of user n and and are the minimum and maximum values ​​of the calculation frequency. C4 and C5 ensure that the devices involved in model training are stable and efficient, and C6 ensures that the users participating in the training are within the service range of the edge server.

[0050] However, the above optimization problem becomes difficult to solve due to constraints C4 and C5. The optimization problem is simplified below based on these two constraints.

[0051] 1. For any valid terms If and only if hour,

[0052] 2. For any valid terms If and only if hour,

[0053] 3. For any valid terms The computational utility function of each type of user must satisfy

[0054] 0≤u'1<… <u' m <… <u' n

[0055] 4. Based on 1 and 2, constraint C5(IC) can be solved as local downward IC (LDIC), that is

[0056]

[0057] and local upward IC (LUIC), i.e.

[0058]

[0059] According to constraints 1, 2, 3, and 4, C4 and C5 can be solved as:

[0060]

[0061]

[0062] Then the constraints of the optimization problem are simplified to:

[0063] C1:

[0064] C2:

[0065] Assumptions And ∈≠1, from the constraint condition C1, we can solve

[0066] From the constraint C2, we can solve

[0067] In summary, we can get This is a nonlinear optimization problem with constraints. It can be solved by finding the optimal solution f under the condition of maximizing the utility function. n and β k,n , which is the optimal value, thus completing resource allocation.

[0068] Traditional federated learning usually has a two-layer structure, namely the user layer and the server layer. However, the distance between the user layer and the server layer is usually far, which leads to serious transmission delay and transmission energy consumption, greatly increasing the communication overhead. In order to reduce the communication overhead, mobile edge computing, namely the edge server layer, is introduced to offload the computing tasks to the edge server, which can greatly reduce the communication overhead.

[0069] In the present invention, in order to enhance the stability of the training alliance, a user selection link is added to select users with strong computing and communication capabilities to join the training, which greatly reduces computing delay and computing energy consumption, and further reduces communication overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The figure shows the loss function of the linear regression model calculated by a user after reaching local accuracy when training a local MIMO channel data pair using the gradient descent algorithm. It can be clearly seen from the figure that when the number of iterations reaches between 400 and 500, the loss function converges to a fixed minimum value, indicating that the user's training model has reached the optimal state at this time.

[0071] Figure 2 and Figure 3 The resource allocation results after the second user selection are shown, starting from Figure 3 It can be seen from the figure that when the step value is 2, the inverse value F of the utility function of the system reaches the minimum value (that is, the utility function reaches the maximum value). At this time, the frequency resources and bandwidth resources take the optimal value. The optimal value can be obtained from Figure 2 It can be clearly seen in.

[0072] Figure 4 and Figure 5 The figure shows the delay and energy consumption of the traditional two-layer federated learning (UM) and the three-layer federated learning (USM) after the edge server is introduced. When the number of users and the accuracy of the user model are the same, the number of users is 15-20. It can be seen that the delay and energy consumption of the three-layer structure are significantly less than those of the two-layer structure.

[0073] Figure 6 It shows the system delay and energy consumption with and without user selection and resource allocation. It can be seen from the figure that under the same conditions, the system delay and energy consumption after selection are much smaller than those before selection, and the system delay and energy consumption after optimization are smaller than those before optimization. The best state, that is, the situation with the smallest delay and energy consumption is the system after selection and optimization. DETAILED DESCRIPTION

[0074] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0075] The edge computing and resource optimization method based on federated learning includes the following steps:

[0076] Step 1: Build a system model based on users, edge servers, and base stations and deploy the array

[0077] The present invention is a multi-user millimeter wave MIMO three-layer structure system structure. The first layer is the user layer, which consists of N single-antenna user devices. The second layer is the ES layer (edge ​​server layer), which includes K with S T The third layer is the CS layer (cloud server layer), which has a cloud server S with a single antenna. Among them, the user layer and the ES layer communicate through wireless channels, and the ES layer and the CS layer are connected through a backhaul link with sufficient capacity.

[0078] Step 2: First, select the set of users that communicate with the edge server based on the distance and represent it as Where N is the number of users, It indicates the number of users selected after distance selection.

[0079] That is, assuming that the user equipment follows a Poisson distribution, the communication range of each edge server is a radius of d k In this three-level structure, it is assumed that there is no overlap between the circular communication areas of the edge servers. Then, users with closer distances are selected based on the distance from each user to the edge server, because users within the communication range have better communication quality with the server, and stable communication is a necessary condition to ensure the smooth progress of federated learning.

[0080] Step 3: Acquisition of training data.

[0081] The channel model is an LOS (line of sight) receiving route with an L path. The millimeter wave channel matrix of the nth user is expressed as:

[0082]

[0083] in, α n,l is the complex channel gain of the lth path, is the direction angle of the downlink transmission route, YesS T ×1 direction vector, whose mth element is expressed as:

[0084]

[0085] λ is the wavelength, σ is the distance between the two antennas, and σ=λ / 2.

[0086] First in The user transmission direction angle is generated within the range. The L-path channel transmission direction angle is randomly generated within the fluctuation range of 5°, and then h is calculated according to the channel matrix formula. n , the angle interval Divide into Q non-overlapping parts equally, each sub-interval Each interval range is marked starting from 1. The interval where the user transmission direction angle is located is marked as q, and the q value is the interval mark value, thus generating (h n ,q) data pair. The final RF precoding f RF,n It can be expressed as in is the midpoint of the interval.

[0087] Step 4: Perform training based on the selected users and the generated MIMO data.

[0088] The training process can be divided into edge aggregation and cloud aggregation

[0089] 2) Edge aggregation: Edge aggregation is divided into local model calculation, local model transmission and edge model aggregation. The specific contents are as follows:

[0090] Local model calculation: After generating MIMO data locally, user n uses local data to train a linear regression model based on gradient descent. When the local model accuracy θ is reached, n ∈[0,1], that is, the number of iterations is L(θ n )=μlog(1 / θ n ), generate the model parameters w t And calculate the model's loss function f i (w t). In this process, computational delay is generated and computing energy The computing delay and the computing energy consumed can be derived based on the amount of data the user is training and the frequency of data processing.

[0091] Local model transmission: After the local calculation is completed, user n transfers the model parameters w t Uploaded to the edge server k with which it communicates, during the transmission process, the OFDMA (Orthogonal Frequency Division Multiple Access) protocol is used, which generates a transmission delay and transfer energy The transmission delay and transmission energy can be calculated by the Shannon formula according to the transmission rate between user n and edge server k and the bandwidth allocated by server k to user n.

[0092] Edge aggregation: In this process, the edge server aggregates the model parameters uploaded by each user and then averages them to update the model parameters (i.e., federated averaging algorithm). After the edge server aggregates the model parameters and performs the averaging algorithm, it sends the updated model parameters to each user by broadcasting (the delay and energy consumption during the broadcast process are negligible). After the user receives the model parameters sent by the server, it uses this parameter for a new round of data training. This cycle repeats until the edge server reaches its model accuracy value ε, and the number of global aggregations is expressed as Where δ is the learning task parameter. When I(ε,θ n The total delay / energy consumption after ) global aggregation is the sum of the calculation delay / energy consumption and the transmission delay / energy consumption.

[0093] 2) Cloud aggregation: Cloud aggregation is divided into two steps: edge model upload and cloud model aggregation.

[0094] Edge model upload: After the edge server k reaches the edge accuracy, the edge model parameters are uploaded to the cloud server, which causes a delay in the process. and energy consumption It can be calculated based on the transmission rate and transmission power.

[0095] Cloud model aggregation: The cloud server aggregation process is similar to the edge aggregation process, that is, the cloud server receives the model parameters uploaded by the edge server and then performs average calculation.

[0096] The key points of the present invention are:

[0097] Step 1: Allocate resources to users

[0098] Secondary selection: In order to further obtain a stable and efficient learning alliance, a second user selection is performed before allocating resources to the user. The basis for this selection is the user's previous training record. That is, before making a user selection, the user must perform local iteration and achieve the corresponding accuracy. Then, the gradient value is calculated based on the locally calculated loss function, and the difference is made with the gradient gathered by the edge server. The user with the smallest gradient difference is selected to form the final learning alliance. in is the set of users after the first selection. is the set of users after two selections.

[0099] Resource allocation: After the second selection, the alliance is successfully formed, and then resources are allocated to the users who finally participate in the training. In this process, the contract theorem is introduced to encourage and reward users who participate in the training. First, the users are classified according to the local accuracy (the smaller the better), that is, Then sort and set the user's evaluation function in It is a reward for users who participate in alliance learning. is the profit function, and as In this way, the user's computational utility function can be obtained, that is,

[0100]

[0101] Edge Server Offer Contract Terms Users choose the terms that suit them according to their actual situation, based on the constraints of the contract theorem, namely IR (individual rationality constraint) and IC (incentive compatibility constraint).

[0102] IR: The user's income and cost loss are at the lowest level, that is, If u' n <0, the user cannot participate in the training and is abandoned.

[0103] IC: Users must match their own contract terms when selecting contract terms, i.e.

[0104]

[0105] According to the above, we can get the user's utility function u n and the utility function U of the edge server k (The difference between the cost function and the revenue function, the utility function of the cloud server is negligible. In summary, the utility function of the entire system can be obtained

[0106]

[0107] Where K is the set of edge servers, It is the set of users after two selections.

[0108] The purpose of resource allocation is to minimize the system communication overhead and maximize the system utility function, so an optimization problem can be proposed, namely

[0109] max U,

[0110] subject to,

[0111] C1:

[0112] C2:

[0113] C3:

[0114] C4:

[0115] C5:

[0116] C6:

[0117] Among them, the constraints C1 and C2 represent the uplink communication resource constraints and computing power constraints respectively, β k,n is the bandwidth allocated by edge server k to user n, f n is the computation frequency of user n and and are the minimum and maximum values ​​of the calculation frequency. C4 and C5 ensure that the devices involved in model training are stable and efficient, and C6 ensures that the users participating in the training are within the service range of the edge server.

[0118] However, the above optimization problem becomes difficult to solve due to constraints C4 and C5. The optimization problem is simplified below based on these two constraints.

[0119] 3. For any valid terms If and only if hour,

[0120] 4. For any valid terms If and only if hour,

[0121] 3. For any valid terms The computational utility function of each type of user must satisfy

[0122] 0≤u'1<… <u'm <… <u' n

[0123] 4. Based on 1 and 2, constraint C5(IC) can be solved as local downward IC (LDIC), that is

[0124]

[0125] and local upward IC (LUIC), i.e.

[0126]

[0127] According to constraints 1, 2, 3, and 4, C4 and C5 can be solved as:

[0128]

[0129]

[0130] Then the constraints of the optimization problem are simplified to:

[0131] C1:

[0132] C2:

[0133] Assumptions And ∈≠1, from the constraint condition C1, we can solve

[0134] From the constraint C2, we can solve

[0135] In summary, we can get This is a nonlinear optimization problem with constraints. It can be solved by finding the optimal solution f under the condition of maximizing the utility function. n and β k,n , which is the optimal value, thus completing resource allocation.

Claims

1. Edge computing and resource optimization method based on federated learning, characterized by: The method comprises the following steps: Step 1: Build a system model based on users, edge servers, and base stations and deploy the array The multi-user millimeter wave MIMO three-layer system structure, the first layer is the user layer, which consists of N single-antenna user devices; the second layer is the ES layer, i.e. the edge server layer, which contains K S T The third layer is the CS layer, i.e., the cloud server layer, which is a cloud server S with a single antenna; wherein the user layer and the ES layer communicate through a wireless channel, and the ES layer and the CS layer are connected through a backhaul link; Step 2: First, select the set of users that communicate with the edge server based on the distance and represent it as Where N is the number of users, When it represents the number of users selected after distance selection, it is assumed that the user equipment follows the Poisson distribution and the communication range of each edge server is a radius of d. k In this three-level structure, it is assumed that there is no overlap between the circular communication areas of the edge servers; then the users with closer distance are selected according to the distance from each user to the edge server; Step 3: Acquisition of training data; The channel model is an LOS receiving route with an L path. The millimeter wave channel matrix of the nth user is expressed as: in, is the complex channel gain of the lth path, is the direction angle of the downlink transmission route, YesS T ×1 direction vector, whose m*th element is expressed as: λ is the wavelength, σ is the distance between the two antennas, and σ = λ / 2; First in The user transmission direction angle is generated within the range. The L-path channel transmission direction angle is randomly generated within the fluctuation range, and then h is calculated according to the channel matrix formula. n , the angle interval Divide into Q non-overlapping parts equally, each sub-interval Each interval range is marked starting from 1. The interval where the user transmission direction angle is located is marked as q, and the q value is the interval mark value, thus generating (h n ,q) data pair; the final RF precoding f RF,n Expressed as in is the midpoint of the interval; Step 4: Perform training based on the selected users and the generated MIMO data; the training process is divided into edge aggregation and cloud aggregation.

2. The method for edge computing and resource optimization based on federated learning according to claim 1, characterized in that: The edge aggregation is divided into local model calculation, local model transmission and edge model aggregation; the specific contents are as follows: Local model calculation: After generating MIMO data locally, user n uses local data to train a linear regression model based on gradient descent. When the local model accuracy θ is reached, n ∈[0,1], that is, the number of iterations is L(θ n )=μlog(1 / θ n ), generate the model parameters w t And calculate the model's loss function f n (w t ); computational delay is generated in this process and computing energy The computational latency and consumed computational energy are derived based on the amount of data the user trains and the frequency of data processing; Local model transmission: After the local calculation is completed, user n transfers the model parameters w t Uploaded to the edge server k with which it communicates, during the transmission process, the orthogonal frequency division multiple access protocol OFDMA is used, which generates a transmission delay and transfer energy The transmission delay and transmission energy are calculated by the Shannon formula based on the transmission rate between user n and edge server k and the bandwidth allocated by server k to user n; Edge aggregation: In this process, the edge server aggregates the model parameters uploaded by each user and averages them to update the model parameters; After the edge server aggregates the model parameters and performs the averaging algorithm, the updated model parameters are sent to each user by broadcasting. After the user receives the model parameters sent by the server, it uses this model parameter for a new round of data training. This cycle repeats until the edge server reaches its model accuracy value ε, and the number of global aggregations is expressed as where δ is the learning task parameter; when I(ε,θ n ) After the global aggregation, the total delay or energy consumption is generated. The total delay is the sum of the calculation delay and the transmission delay, and the total energy consumption is the sum of the calculation energy consumption and the transmission energy consumption.

3. The method for edge computing and resource optimization based on federated learning according to claim 2, characterized in that: The cloud aggregation described: cloud aggregation is divided into two steps: edge model uploading and cloud model aggregation; Edge model upload: After the edge server k reaches the edge accuracy, the edge model parameters are uploaded to the cloud server, which causes a delay in the process. and energy consumption Calculated based on transmission rate and transmission power; Cloud model aggregation: The cloud server aggregation process is similar to the edge aggregation process, that is, the cloud server receives the model parameters uploaded by the edge server and then performs average calculation.

4. The method for edge computing and resource optimization based on federated learning according to claim 3, characterized in that: In order to obtain a learning alliance, a second user selection is performed before allocating resources to the user. The selection is based on the user's previous training record. That is, before making a user selection, the user must perform local iteration and achieve the corresponding accuracy. Then, the gradient value is calculated based on the locally calculated loss function, and the difference is made with the gradient gathered by the edge server. The user with the smallest gradient difference is selected to form the final learning alliance. in is the set of users after the first selection. is the set of users after two selections.

5. The method for edge computing and resource optimization based on federated learning according to claim 4, characterized in that: After the second selection, the alliance is successfully formed, and then resources are allocated to the users who finally participate in the training. In this process, the contract theorem is introduced to encourage and reward users who participate in the training. First, the users are classified according to the local accuracy, that is, Then sort and set the user's evaluation function in It is a reward for users who participate in alliance learning. is the profit function, and as As it increases, it increases monotonically; the user's computational utility function is obtained, that is, Edge Server Offer Contract Terms Users choose the terms that suit them according to their actual situation, based on the constraints of the contract theorem, which are IR and IC; IR: The user's income and cost loss are at the lowest level, that is, if Then the user cannot participate in the training and is abandoned; IC: Users must match their own contract terms when selecting contract terms, i.e. According to the user's utility function u' n and the utility function U of the edge server k That is, the difference between the cost function and the revenue function. The utility function of the cloud server is negligible. In summary, the utility function of the entire system is: Where K is the set of edge servers, is the set of users after two selections; The purpose of resource allocation is to minimize the system communication overhead and maximize the system utility function, so the optimization problem is proposed, namely max U, subject to, C1: C2: C3: C4: C5: C6: Among them, the constraints C1 and C2 represent the uplink communication resource constraint and computing power constraint respectively, β k,n is the bandwidth allocated by edge server k to user n, f n is the computation frequency of user n and and are the minimum and maximum values ​​of the calculation frequency; C4 and C5 ensure that the devices involved in model training are stable and efficient, and C6 ensures that the users participating in the training are within the service range of the edge server.

Images