Retransmission count and transmitting power allocation method for federated learning system, and gradient retransmission method

Abstract

Disclosed in the present application are a retransmission count and transmitting power allocation method for a federated learning system, and a gradient retransmission method. The retransmission count and transmitting power allocation method comprises: S1, a user sends a training sequence to an edge server; S2, the edge server estimates an uplink channel gain of the user according to the training sequence and stores same; S3, the edge server solves an optimization problem by using the channel gain of the user, a user energy consumption limit, and a total retransmission count limit, and sets retransmission parameters for the user, i.e., a maximum retransmission count and a transmitting power of each retransmission round; S4, the edge server broadcasts retransmission parameter information of all users; and S5, the user acquires the allocated retransmission parameters from the broadcast information. Thus, the present application can be used for improving the reliability of users uploading local gradients in a federated learning system, thereby ensuring the training effect of federated learning in a wireless communication environment, and ensuring that the energy consumption of users is within a certain range.

Description

Retransmission times and transmission power allocation method and gradient retransmission method of federated learning system Technical Field

[0001] The present invention belongs to the technical field of federated learning transmission optimization, and specifically relates to a method for allocating the number of retransmissions and transmission power of a federated learning system and a gradient retransmission method. Background Art

[0002] With the popularity of the integrated framework of telepathy and computing, federated learning has received increasing attention from academia and industry. The federated learning framework can prevent the leakage of user raw data, thereby protecting the privacy of user terminals, and fully utilizing the user's computing power and storage space, which is in line with the development trend of integrated telepathy and computing. However, the complex and changeable wireless communication environment reduces user reliability. When the communication environment deteriorates, the quality of the local gradients uploaded by users is severely reduced. When they reach the edge server, they can no longer be decoded and are discarded, and cannot be used by the edge server. The more local gradients of users are discarded, the worse the training effect of the global model, which manifests as difficult or non-convergence of training. Therefore, it is necessary to improve the reliability of users uploading local gradients in wireless communication environments. To address the above problems, existing solutions mostly adopt methods such as optimizing transmission bandwidth allocation and power allocation, optimizing user scheduling, that is, selecting users with good channel conditions. In the few solutions that use retransmission to improve user reliability, all users use the same fixed number of retransmissions. However, the following problems still exist in existing solutions:

[0003] 1. Solutions for optimizing transmission bandwidth and power allocation often consider transmission energy consumption and transmission latency, but lack consideration of the effectiveness of federated learning training.

[0004] 2. Although selecting users with good channel conditions through user scheduling optimization can make local gradient transmission more reliable, it will cause the global model update to favor the local models of these users. Users with poor channel conditions cannot contribute to the global model update, causing the global model to be biased and unable to represent all users in the federated learning system.

[0005] 3. The scarcity of wireless communication resources means that users cannot occupy communication time slots indefinitely. The total number of retransmission time slots available to a user cluster is limited. When the number of retransmissions is fixed, users with good channel conditions use too many time slots, resulting in a waste of communication resources. Users with poor channel conditions use too few time slots, which has a limited effect on improving reliability.

[0006] Summary of the Invention

[0007] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and propose a method for allocating the number of retransmissions and transmission power of a federated learning system and a gradient retransmission method.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for allocating the number of retransmissions and transmission power of a federated learning system includes the following steps:

[0010] S1. Each user sends a known training sequence to the edge server using the uplink channel;

[0011] S2. The edge server estimates the uplink channel gain of each user based on the received training sequence sent by each user and stores the uplink channel gain of each user;

[0012] S3. The edge server uses the user's channel gain, the given user energy consumption limit, and the given total retransmission limit to solve the optimization problem and allocate the user's maximum retransmission number and the transmission power for each retransmission round.

[0013] S4. The edge server broadcasts retransmission parameter information including the maximum number of retransmissions for all users and the transmit power for each retransmission round;

[0014] S5. The user obtains the maximum number of retransmissions allowed and the transmission power of each retransmission round from the broadcast information.

[0015] The present invention also includes a gradient retransmission method for a federated learning system, comprising:

[0016] The method for allocating the number of retransmissions and the transmission power provided by the present invention is used to allocate the maximum number of retransmissions and the transmission power of each retransmission round to the user;

[0017] The user reads the maximum number of retransmissions and the transmission power of each retransmission round stored in the user storage unit, and uploads the local gradient to the edge server through an independent and irrelevant uplink channel according to the set transmission power;

[0018] If the edge server successfully receives the user's local gradient, it will send a transmission success signal ACK back to the user, indicating that the user does not need to retransmit;

[0019] If the edge server fails to successfully receive the user's local gradient, the edge server returns a transmission failure signal NACK to the user, instructing the user to re-upload the local gradient to the edge server according to the set transmission power until the user retransmits the gradient to the maximum number of times or the edge server successfully receives the user's local gradient.

[0020] If the edge server still fails to successfully receive the user's local gradient after the user's retransmission times reaches the maximum number of retransmissions, the edge server will aggregate the local gradient that the user successfully transmitted last time.

[0021] The present invention also includes a federated learning system, comprising: an edge server and a plurality of users, wherein an uplink channel between the users and the edge server is a wireless communication channel;

[0022] The edge server is deployed in the base station and includes a computing unit, a storage unit, a training unit, and a sending unit;

[0023] The user includes a storage unit, a training unit, and a sending unit, and is a mobile terminal in a wireless communication environment;

[0024] Edge servers and users collaborate to complete designated tasks according to the federated learning framework. The original data required for the task is stored in the user, and the edge server and other users cannot access the user's original data.

[0025] The edge server and the user exchange local gradients and global model updates. The user updates the local model using local raw data and the global model broadcast by the edge server, and calculates the stochastic gradient of the local model. The edge server then updates the global model using the local gradients uploaded by the user.

[0026] The operation of the federated learning system includes the following steps:

[0027] S101: The edge server randomly selects a number of users to participate in this iteration according to a set ratio;

[0028] S102: If it is the first global iteration, the edge server and the user are initialized;

[0029] S103: The selected user sends a known training sequence to the edge server via an uplink channel;

[0030] S104. The edge server receives a training sequence sent by the selected user, estimates and stores the user channel gain;

[0031] S105. The edge server solves the optimization problem based on the user's channel gain and other constraints, and allocates the user's maximum number of retransmissions and the transmit power for each retransmission round;

[0032] S106. The edge server broadcasts a retransmission parameter information block including the maximum number of retransmissions for the user and the transmit power for each retransmission round, and global model parameters.

[0033] S107. The user queries and records the maximum number of retransmissions allocated to it and the transmission power of each retransmission round from the broadcast information block;

[0034] S108. The user uses the global model parameters and local data set broadcast by the edge server to perform local training and calculate the stochastic gradient of the local model;

[0035] S109. The user reads the allocated maximum number of retransmissions and the transmission power of each retransmission round;

[0036] S110. The user determines whether the current number of transmissions exceeds the maximum number of retransmissions. If the number of retransmissions does not exceed the maximum number of retransmissions, the user uploads the same local gradient according to the transmit power allocated to it, and increases the number of transmissions by one. Otherwise, the retransmission is terminated.

[0037] S111, the edge server determines whether the user's local gradient upload is successful; if the user upload fails, the edge server returns a NACK signal, and the user repeats step S110; if the user uploads successfully, the edge server returns an ACK signal, and the retransmission ends;

[0038] S112, the edge server aggregates all local gradients and updates the global model parameters;

[0039] S113. If the global model converges, the federated learning ends. If the global model does not converge, repeat steps S101-S112 until the global model converges.

[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0041] 1. By combining retransmission technology, which has been widely used in the communications field, with federated learning, the present invention fully improves the reliability of users in the federated learning system participating in training in a wireless communication environment, thereby ensuring that the federated learning system can still achieve better training results in complex and changeable wireless communication environments. In addition, the use of retransmission technology, which has been well-researched, enhances the feasibility of the present invention.

[0042] 2. The present invention fully utilizes the interactivity between users and edge servers in the federated learning system, enabling the server to obtain the latest information on the uplink channel between the user and the edge server in each global iteration round, thereby utilizing the latest channel information to adaptively adjust the user's maximum number of retransmissions and the transmission power of each retransmission round, thereby ensuring the training effect of the federated learning system in a wireless communication environment.

[0043] 3. Based on the limited wireless communication network resources and limited user energy supply, the present invention provides users with a maximum number of channel-adaptive retransmissions and a transmission power allocation for each retransmission round. Retransmission increases the probability of the user's local gradient being successfully uploaded to the edge server in a wireless communication environment. By dynamically adjusting the user's maximum number of retransmissions and the transmission power for each retransmission round to meet the user's limited energy supply, the federated learning system of the present invention takes into account both maintaining the training effect and saving the user's energy consumption in a wireless communication environment.

[0044] 4. The present invention adopts the hybrid automatic repeat request retransmission protocol (CC-HARQ) with appending merging. Users only need to upload the same local gradient in each round. The edge server performs appending merging and merges the gradients. After the retransmission is completed, the merged decoding is performed, making full use of the information uploaded in each retransmission round. There is no need to add additional devices to the user terminal and edge server, effectively reducing the system deployment cost. It is simple, easy to implement, and has strong scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] 1 is a flow chart of a method for allocating the number of retransmissions and transmit power according to the present invention;

[0046] Figure 2 is a schematic diagram of the federated learning system;

[0047] FIG3 is a schematic diagram of the structure of an edge server in the federated learning system of the present invention;

[0048] FIG4 is a schematic diagram of the structure of users in the federated learning system of the present invention;

[0049] FIG5 is a schematic diagram of a training sequence block sent by a user for channel estimation in an embodiment;

[0050] FIG6 is a schematic diagram of a channel gain information block obtained after channel estimation in an embodiment;

[0051] 7 is a schematic diagram of a retransmission parameter information block including the maximum number of retransmissions for all users and the transmit power information for each retransmission round obtained after allocating the number of retransmissions and transmit power in an embodiment;

[0052] FIG8 is a flowchart of the operation of the federated learning system of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0054] Example

[0055] As shown in Figure 1, a method for allocating the number of retransmissions and transmission power of a federated learning system is applied to the federated learning system. In this embodiment, as shown in Figures 2, 3, and 4, the federated learning system includes an edge server and several users, and the uplink channel between the users and the edge server is a wireless communication channel; the edge server is deployed in a base station and includes a computing unit, a storage unit, a training unit, and a sending unit; the user includes a storage unit, a training unit, and a sending unit, and is a mobile terminal in a wireless communication environment.

[0056] Since users are usually distributed in a complex and changeable wireless communication network, uploading local gradients through poor user channels may fail. Since users are usually portable mobile smart devices such as mobile phones and smart watches, the battery size is limited, which limits the user's energy consumption. Communication resources are limited, and the user group cannot perform a large number or unlimited retransmissions. Therefore, the present invention proposes a method for allocating the number of retransmissions and the transmit power. In this embodiment, the method includes the following steps:

[0057] S1. Each user sends a known training sequence to the edge server using the uplink channel. Specifically:

[0058] The user adds the user number to the front of the training sequence to form a user training sequence block. The user participating in federated learning sends the user training sequence block to the edge server, as shown in Figure 5; the same training sequence is also stored in the storage unit of the edge server, which is information known to both the user and the edge server.

[0059] S2. The edge server estimates the uplink channel gain of each user based on the received training sequence blocks sent by each user, and stores the uplink channel gain of each user. Specifically,

[0060] The edge server receives a user training sequence block sent by the user through a wireless communication channel between the user and the edge server, and sequentially stores the received user training sequence blocks in a queue in a storage unit;

[0061] During the upload process, the user training sequence block is affected by the wireless communication channel between the user and the edge server, resulting in a certain degree of deformation and degradation;

[0062] The edge server reads one user training sequence block from the queue each time, compares the training sequence in the user training sequence block with the original training sequence stored in the storage unit, and performs user channel estimation in the calculation unit;

[0063] The edge server copies the user number at the front end of the user training sequence block for which channel estimation has been completed, and adds the channel gain information obtained after channel estimation to the user number to form a user channel gain unit. Each time the edge server obtains a user channel gain unit, it concatenates it with the user channel gain unit that has already been obtained. Finally, a channel gain information block containing the channel information of all users participating in the training is obtained and stored in the storage unit. Figure 6 shows a channel gain information block.

[0064] S3. The edge server uses the user's channel gain, the given user energy consumption limit, and the given total retransmission limit to solve the optimization problem and set the maximum number of retransmissions and the transmission power of each retransmission round for the user. Specifically,

[0065] The edge server reads the preset user energy consumption limit, total retransmission limit, and channel gain information block stored in the storage unit, substitutes them into the optimization problem with unknown parameters and a determined specific model, copies the user number at the front end of the channel gain information unit, solves the maximum number of retransmissions allocated to the user and the transmission power of each retransmission round, and after splicing with the user number, obtains the user retransmission parameter unit. After splicing the user retransmission parameter units of all users, the retransmission parameter information block is obtained and stored in the storage unit of the edge server; as shown in Figure 7, this is the retransmission parameter information block obtained after the retransmission number and transmission power are allocated.

[0066] The optimization problem is specifically:

[0067] Among them, P i Specifically expressed as It is the user's transmission power in different retransmission rounds; M i is the maximum number of retransmissions for user i, and the first transmission round is also included in the retransmission round, that is, user i can retransmit at most M times in addition to the first transmission i -1 time; is the path loss noise ratio, is the length of each retransmission slot, B i is the number of bits of information that user i should transmit, ε is the energy consumption limit of each user, M is the number of time slots shared by all users, that is, the sum of the maximum number of retransmissions of users participating in training, P max is the maximum transmit power of the user, and N is the number of users. The maximum number of user retransmissions is a positive integer greater than or equal to 1. The optimization problem needs to be solved in each round of federated learning to achieve the best improvement in federated learning reliability and energy efficiency.

[0068] Among them, the objective function of the optimization problem P1 is:

[0069] It reflects the training effect of the federated learning system in the wireless communication environment. The smaller the objective function, the better the training effect of the federated learning.

[0070] The first constraint of the optimization problem P1 is:

[0071] It means that for any user i, the energy consumed by the entire local gradient upload and retransmission process needs to be less than or equal to the user's maximum energy consumption limit ε.

[0072] By configuring the federated learning system based on the maximum number of retransmissions per user and the transmit power per retransmission round, obtained after solving optimization problem P1, both training performance and energy efficiency can be balanced. Specifically, to minimize the objective function to maximize the effectiveness of federated learning training, the maximum number of retransmissions per user must be as high as possible. However, the constraints limit the energy consumption of each user's local gradient upload process, thus limiting the maximum number of retransmissions per user. Furthermore, through the mutual constraints of the objective function and the constraints, the optimal transmit power per user per retransmission round can be obtained; therefore, the role of the objective function and the constraints in the optimization problem is to balance the training performance and energy efficiency of federated learning.

[0073] S4. The edge server broadcasts a retransmission parameter information block containing the maximum number of retransmissions for all users and the transmission power information for each retransmission round through the sending unit;

[0074] S5. The user obtains the maximum number of retransmissions allowed and the transmit power for each retransmission round from the broadcast information, specifically:

[0075] The user receives a retransmission parameter information block containing the maximum number of retransmissions for all users and the transmit power for each retransmission round. The user finds its own retransmission parameter information unit using the user sequence number at the front end of each retransmission parameter information unit in the information block, obtains its assigned maximum number of retransmissions and transmit power for each retransmission round, and stores them in a storage unit.

[0076] In another embodiment, a gradient retransmission method for a federated learning system is provided, including:

[0077] The method for allocating the number of retransmissions and the transmit power of each retransmission round is used to allocate the maximum number of retransmissions and the transmit power of each retransmission round to the user.

[0078] The user reads the maximum number of retransmissions and the transmission power of each retransmission round stored in the user storage unit, and uploads the local gradient to the edge server through an independent and irrelevant uplink channel according to the set transmission power;

[0079] If the number of information bits obtained by the edge server is greater than the number of bits that the user should send, it means that the edge server has successfully received the user's local gradient. The edge server uses a reliable method, such as increasing the signal transmission power or repeatedly sending the signal, to return a transmission success signal ACK to the user, indicating that the user does not need to retransmit. Otherwise, the edge server has not successfully received the user's local gradient. The edge server uses a reliable method to return a transmission failure signal NACK to the user. After receiving the NACK signal, the user re-uploads the same local gradient to the edge server. If the number of information bits accumulated by the edge server after the retransmission round is less than the number of bits that the user should send, it means that the retransmission was unsuccessful. The user then continues to retransmit until the number of retransmissions exceeds the maximum number of retransmissions or the retransmission is successful.

[0080] If the user fails to successfully transmit the local gradient to the edge server after the number of retransmissions reaches the maximum number of retransmissions, the edge server aggregates the local gradient that the user successfully transmitted last time.

[0081] In this embodiment, the user uses the hybrid automatic repeat request retransmission protocol CC-HARQ with additional combining;

[0082] Based on the CC-HARQ protocol, users upload the same local gradient in each retransmission round. The edge server combines the information received in each retransmission round and decodes the information using maximum ratio combining after the retransmission. The process is shown as follows:

[0083] In the CC-HARQ protocol, the user sends the same information in each retransmission round. represents the length of the retransmission slot. In the mth retransmission round, the data upload rate of user i is:

[0084] The transmission bandwidth is normalized to 1, h i (m) is the channel gain of the mth retransmission round; the amount of data transmitted by user i in the mth retransmission round is:

[0085] In the CC-HARQ protocol, the information transmitted by user i after the retransmission is the superposition of the information in all retransmission rounds. The cumulative amount of data transmitted by user i after the mth retransmission round is:

[0086] C i (m) Decode and obtain the information transmitted by the user after the mth retransmission round;

[0087] By comparing the number of bits of information transmitted by the user after the mth retransmission round with the number of bits of information that the user should transmit, it is determined whether the user has successfully transmitted information to the edge server after the retransmission round, which is expressed as:

[0088] If C i (m)≥B i , the user successfully transmits information to the edge server after the mth round, and the retransmission ends. Otherwise, the transmission fails and the next retransmission round is required. i is the number of bits of information that user i should transmit;

[0089] This leads to the interruption probability of user i in the mth retransmission round, that is, the probability of user i failing to transmit in the mth retransmission round: Q i (m)=Pr[C i (m)≤B i ].

[0090] In another embodiment, a federated learning system is provided, as shown in FIG2 , the system includes: an edge server and a plurality of users, wherein the uplink channel between the users and the edge server is a wireless communication channel;

[0091] The edge server is deployed in the base station and includes a computing unit, a storage unit, a training unit, and a sending unit;

[0092] The user includes a storage unit, a training unit, and a sending unit, and is a mobile terminal in a wireless communication environment;

[0093] Edge servers and users collaborate to complete designated tasks according to the federated learning framework. The original data required for the task is stored in the user, and the edge server and other users cannot access the user's original data.

[0094] The edge server and the user exchange local gradients and global model updates. The user uses the local raw data and the global model broadcast by the edge server to update the local model and calculate the stochastic gradient. The edge server uses the user's updated local gradient to update the global model.

[0095] As shown in Figure 8, the operation of the federated learning system includes the following steps:

[0096] S101: The edge server randomly selects a number of users to participate in this iteration according to a set ratio;

[0097] S102: If it is the first global iteration, the edge server and the user are initialized;

[0098] S103: The selected user sends a known training sequence to the edge server via an uplink channel;

[0099] S104. The edge server receives a training sequence sent by the selected user, estimates and stores the user channel gain;

[0100] S105. The edge server solves the optimization problem based on the user's channel gain and other constraints, and allocates the maximum number of retransmissions and the transmit power for each retransmission round to the user.

[0101] S106. The edge server broadcasts a retransmission parameter information block including the maximum number of retransmissions for the user and the transmit power information for each retransmission round, and global model parameters.

[0102] S107. The user queries and records the maximum number of retransmissions allocated to it and the transmission power of each retransmission round from the broadcast information block;

[0103] S108. The user calculates the stochastic gradient of the local model using the local data set and the global model parameters broadcast by the edge server;

[0104] S109. The user reads the allocated maximum number of retransmissions and the transmission power of each retransmission round;

[0105] S110. The user determines whether the current number of transmissions exceeds the maximum number of retransmissions. If the number of retransmissions does not exceed the maximum number of retransmissions, the user uploads the same local gradient according to the transmit power allocated to it, and increases the number of transmissions by one. Otherwise, the retransmission is terminated.

[0106] S111, the edge server determines whether the user's local gradient upload is successful; if the user upload fails, the edge server returns a NACK signal, and the user repeats step S110; if the user uploads successfully, the edge server returns an ACK signal, and the retransmission ends;

[0107] S112, the edge server aggregates all local gradients and updates the global model parameters;

[0108] S113. If the global model converges, the federated learning ends. If the global model does not converge, repeat steps S101-S112 until the global model converges.

[0109] In this embodiment, the total number of retransmission time slots that can be used by the federated learning system and the energy that can be used by users are limited. Therefore, the retransmission number and transmission power allocation method and the gradient retransmission method for improving the reliability and energy efficiency of federated learning need to set the total retransmission number and user energy consumption limit shared by all users.

[0110] It should also be noted that, in this specification, terms such as "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or apparatus comprising the element.

[0111] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for allocating the number of retransmissions and transmission power of a federated learning system, characterized in that: The following steps are involved: S1. Each user sends a known training sequence to the edge server using the uplink channel; S2. The edge server estimates the uplink channel gain of each user based on the received training sequence sent by each user, and stores the uplink channel gain of each user; S3, the edge server uses the user's channel gain, the given user energy consumption limit and the given total retransmission limit to solve the optimization problem and allocate the user's maximum retransmission number and the transmission power of each retransmission round; S4, the edge server broadcasts retransmission parameter information including the maximum number of retransmissions for all users and the transmission power of each retransmission round; S5. The user obtains the maximum number of retransmissions allowed and the transmission power of each retransmission round from the broadcast information.

2. The method for allocating the number of retransmissions and transmission power of a federated learning system according to claim 1, characterized in that: The federated learning system includes an edge server and several users. The uplink channel between the user and the edge server is a wireless communication channel; The edge server is deployed in the base station and includes a computing unit, a storage unit, a training unit, and a sending unit; The user includes a storage unit, a training unit and a sending unit, and is a mobile terminal in a wireless communication environment.

3. The method for allocating the number of retransmissions and transmission power of a federated learning system according to claim 2, characterized in that: Step S1 is specifically as follows: The user adds the user number at the front end of the training sequence to form a user training sequence block. The users participating in federated learning send the user training sequence block to the edge server; the same training sequence is also stored in the storage unit of the edge server, which is information known to both the user and the edge server.

4. The method for allocating the number of retransmissions and transmission power of a federated learning system according to claim 3, characterized in that: Step S2 is specifically as follows: The edge server receives the user's wireless communication channel between the user and the edge server. The received user training sequence blocks are sequentially stored in a storage unit in the form of a queue; The user training sequence block is affected by the wireless communication channel between the user and the edge server during the upload process; The edge server reads a user training sequence block from the queue each time, compares the training sequence in the user training sequence block with the known original training sequence in the storage unit, and performs user channel estimation in the calculation unit; The edge server copies the user number at the front end of the user training sequence block that has completed channel estimation, and adds the channel gain information obtained after channel estimation after the user number to form a user channel gain unit; each time the edge server obtains a user channel gain unit, it splices it with the user channel gain unit that has been obtained, and finally obtains the user channel gain information block participating in the training and stores it in the storage unit.

5. The method for allocating the number of retransmissions and transmission power of a federated learning system according to claim 4, characterized in that: Step S3 is specifically as follows: The edge server reads the preset user energy consumption limit, total retransmission number limit, and user channel gain information block stored in the storage unit, substitutes them into the optimization problem with unknown parameters and a determined specific model, copies the user number at the front end of the channel gain information unit, solves the maximum number of retransmissions allocated to the user and the transmission power of each retransmission round, concatenates the information of the maximum number of retransmissions and the transmission power of each retransmission round with the user number, obtains the user retransmission parameter unit, concatenates the user retransmission parameter units of all users, obtains the retransmission parameter information block, and stores it in the storage unit of the edge server; The optimization problem is specifically: Among them, P i Specifically expressed as That is, the user's transmission power in different retransmission rounds; M i is the maximum number of retransmissions for user i, is the path loss noise ratio, is the length of each retransmission slot, B i is the number of bits of information that user i should transmit, ε is the energy consumption limit of each user, M is the number of time slots shared by all users, that is, the sum of the maximum number of retransmissions of users participating in the training, P max is the maximum value of the user's transmit power, N is the number of users; the maximum number of user retransmissions is a positive integer greater than or equal to 1; the optimization problem needs to be solved in each round of federated learning.

6. The method for allocating the number of retransmissions and transmission power of a federated learning system according to claim 5, characterized in that: Step S4 is specifically as follows: The edge server broadcasts a retransmission parameter information block composed of retransmission parameter units of all users through a sending unit.

7. The method for allocating the number of retransmissions and transmission power of a federated learning system according to claim 6, characterized in that: Step S5 is specifically as follows: The user receives a retransmission parameter information block containing the maximum number of retransmissions for all users and the transmission power for each retransmission round. The user searches for its own retransmission parameter unit through the user serial number at the front end of the user retransmission parameter unit, obtains the allocated maximum number of retransmissions and the transmission power for each retransmission round, and stores them in the storage unit.

8. A gradient retransmission method for a federated learning system, characterized in that: include: The method for allocating the number of retransmissions and the transmission power according to any one of claims 1 to 7 is used to allocate the maximum number of retransmissions and the transmission power of each retransmission round to the user; The user reads the maximum number of retransmissions and the transmission power of each retransmission round stored in the user storage unit, and uploads the local gradient to the edge server through an independent and irrelevant uplink channel according to the set transmission power; If the edge server successfully receives the user's local gradient, the edge server returns a transmission success signal ACK to the user, indicating that the user does not need to retransmit; If the edge server fails to successfully receive the user's local gradient, the edge server returns the transmission to the user. The transmission failure signal NACK indicates that the user re-uploads the local gradient to the edge server according to the set transmission power until the user retransmits the maximum number of times or the edge server successfully receives the user's local gradient; If the edge server still fails to successfully receive the user's local gradient after the user's retransmission times reaches the maximum number of retransmissions, the edge server will aggregate the local gradient that the user successfully transmitted last time.

9. The gradient retransmission method according to claim 8, characterized in that: The user adopts the hybrid automatic repeat request retransmission protocol CC-HARQ with additional combination; Based on the CC-HARQ protocol, the user uploads the same local gradient in each retransmission round. The edge server combines the information received in each retransmission round and decodes the information using the maximum ratio combining method after the retransmission. The process is shown as follows: In the CC-HARQ protocol, the user sends the same information in each retransmission round. represents the length of the retransmission time slot. In the mth retransmission round, the data upload rate of user i is: The transmission bandwidth is normalized to 1, h i (m) is the channel gain of the mth retransmission round; the amount of data transmitted by user i in the mth retransmission round is: In the CC-HARQ protocol, the information transmitted by user i after the retransmission is the superposition of the information of all retransmission rounds. The cumulative amount of data transmitted by user i after the mth retransmission round is: C i (m) Decode to obtain the information transmitted by the user after the mth retransmission round; By comparing the number of bits of information transmitted by the user after the mth round with the number of bits of information that the user should transmit, it is determined whether the user successfully transmits information to the edge server after the retransmission round, which is expressed as: If C i (m)≥B i , the user successfully transmits information to the edge server after the mth round, and the retransmission ends. Otherwise, the transmission fails and the next retransmission round is required. i For user i, The number of bits of information transmitted; This leads to the interruption probability of user i in the mth retransmission round, that is, the probability of user i failing to transmit in the mth retransmission round: Q i (m)=Pr[C i (m)≤B i , 10. A federated learning system, characterized in that: include: The edge server and several users, the uplink channel between the users and the edge server is a wireless communication channel; The edge server is deployed in the base station and includes a computing unit, a storage unit, a training unit, and a sending unit; The user includes a storage unit, a training unit and a sending unit, and is a mobile terminal in a wireless communication environment; The edge server and the user collaborate to complete the specified task according to the federated learning framework. The original data required for the task is stored in the user, and the edge server and other users cannot access the user's original data; The edge server and the user exchange local gradients and global model updates. The user updates the local model using the local original data and the global model broadcast by the edge server, and calculates the stochastic gradient of the local model. The edge server updates the global model using the local gradient uploaded by the user. The operation of the federated learning system includes the following steps: S101, the edge server randomly selects a number of users to participate in this iteration according to a set ratio; S102: If it is the first global iteration, the edge server and the user are initialized; S103, the selected user sends a known training sequence to the edge server via an uplink channel; S104, the edge server receives the training sequence sent by the selected user, estimates and stores the user channel gain; S105, the edge server solves the optimization problem according to the user's channel gain and other constraints, and allocates the user's maximum number of retransmissions and the transmission power of each retransmission round; S106, the edge server broadcasts a retransmission parameter information block including the maximum number of retransmissions of the user and the transmission power information of each retransmission round and global model parameters; S107, the user queries and records the maximum number of retransmissions allocated to it and the transmission power of each retransmission round from the broadcast information block; S108. The user uses the global model parameters and local data set broadcast by the edge server to perform local training and calculate the stochastic gradient of the local model; S109, the user reads the maximum number of retransmissions allocated and the transmission power of each retransmission round; S110, the user determines whether the current number of transmissions exceeds the maximum number of retransmissions; if the number of retransmissions does not exceed the maximum number of retransmissions, the user uploads the same local gradient according to the transmission power allocated to it, and the number of transmissions increases by one, otherwise the retransmission ends; S111, the edge server determines whether the user's local gradient upload is successful; if the user fails to upload, the edge server returns a NACK signal, and the user repeats step S110; if the user uploads successfully, the edge server returns an ACK signal, and the retransmission ends; S112, the edge server aggregates all local gradients and updates the global model parameters; S113. If the global model converges, the current federated learning ends; if the global model does not converge, steps S101-S112 are repeated until the global model converges.

Images