Wireless powered federated learning method, device and electronic equipment

Abstract

The application discloses a wireless energy-supplied federated learning method and device and electronic equipment. First, the channel response of each user to the base station in the current communication round is acquired, a first constraint equation of the local power of each user and a second constraint equation of the signal quality are preset, and a target resource allocation scheme of the current communication round is determined: a transmission beamforming scheme of each user and a reception beamforming scheme of the base station. The user and the base station are controlled to allocate corresponding resources for themselves according to the target resource allocation scheme, so that the base station converges the uploaded local model parameters of each user and acquires original data, thereby completing the aggregation of the global model of this time. The application can save the energy consumption of the user under the premise of ensuring the accuracy of the model convergence process, ensure the accuracy of the acquired local original data, realize the global optimization of the data convergence process of the federated learning model, reduce the signal distortion of the user to the base station, and realize the rapid convergence and higher precision of the federated learning.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a wirelessly powered federated learning method, apparatus, and electronic device. Background Technology

[0002] With the continuous development of distributed learning technology, federated learning technology enables geographically dispersed devices to collaboratively perform model training while users process raw data. Currently, federated learning based on over-the-air computing can reduce communication overhead and latency. In existing federated learning scenarios, the data processed by users may contain many types, mainly divided into model parameters that need to be aggregated at base stations and raw data that needs to be collected by base stations. Currently, both types of data can be uploaded simultaneously to improve spectrum utilization.

[0003] In real-world federated learning scenarios, users are distributed widely and may be deployed in various locations. Some deployment locations, such as underwater, high in buildings, or carried by drones, may make it difficult to frequently charge or replace batteries. Due to the lack of a stable power supply, some users will find it difficult to continuously participate in federated learning, thus affecting the accuracy and convergence of the process.

[0004] Currently, a feasible approach is to maintain a stable power supply for users through wireless charging from base stations. However, wireless charging from base stations also presents resource allocation challenges. Therefore, a resource allocation method is urgently needed to enable the deployment of federated learning in scenarios without a stable power supply. Summary of the Invention

[0005] The purpose of this invention is to provide a wirelessly powered federated learning method, apparatus, and electronic device to support the deployment of wireless federated learning in scenarios without a stable power supply. This invention is used for uplink information transmission by local users and received signal processing by base stations. It can support low-power users to perform data acquisition and federated learning in the absence of a stable local power supply, and can achieve concurrent transmission of raw data and model parameters even with limited spectrum resources.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a federated learning method for wireless power supply, the method comprising: in each communication round, in the first time slot, the base station first acquires the channel response coefficient of each user, determines the target resource allocation scheme, and then uses wireless power transmission technology to wirelessly charge all local users; in the second time slot, the base station broadcasts a global model, and simultaneously, users use the energy collected in the first time slot to perform local data collection; in the third time slot, while local users perform local learning using the collected environmental data, they also continue to collect new environmental data for wireless transmission; in the fourth time slot, local users allocate corresponding resources to themselves according to the resource allocation scheme. The resources are allocated by the base station, which simultaneously uploads the model parameters and environmental data obtained in the third time slot to the base station. In the fifth time slot, based on the received uplink superimposed signal, the base station simultaneously performs model aggregation and data decoding, and allocates corresponding resources to itself according to the resource allocation scheme. The method for determining the target resource allocation scheme is as follows: based on the first constraint equation of each channel response coefficient, each user's local power, and the second constraint equation of signal quality, the target resource allocation scheme for the current training period is determined. The target resource allocation scheme includes: a transmit beamforming scheme for each user in the fourth time slot and a receive beamforming scheme for the base station in the fifth time slot.

[0008] In a second aspect, the present invention provides a wirelessly powered resource allocation device for a federated learning system, the device comprising:

[0009] The information acquisition module is used to acquire the current channel response between each user and the base station during each communication round.

[0010] The scheme determination module is used to determine the target resource allocation scheme for the current communication round based on the channel response of each user and the preset first constraint of local power and second constraint of signal quality for each user in each communication round; wherein the target constraint scheme includes: a transmit beamforming scheme for each user and a receive beamforming scheme for the base station.

[0011] The resource allocation module is used to control users and base stations to allocate corresponding resources to themselves according to the target resource allocation scheme, so that the base station can aggregate the local model parameters uploaded by each user and obtain the raw data uploaded by the user.

[0012] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0013] Memory, used to store computer programs;

[0014] When a processor executes a program stored in memory, it implements the steps of the method provided in the first aspect above.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method provided in the first aspect.

[0016] Fifthly, the present invention provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps of the method provided in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] The wireless power federated learning method, apparatus, and electronic device of the present invention, during federated learning training, can first acquire the channel response from each user to the base station within the current communication round. Then, based on each acquired channel response, a first constraint equation for each user's local power and a second constraint equation for signal quality are preset to determine the target resource allocation scheme for the current communication round. The target constraint scheme includes: a transmit beamforming scheme for each user and a receive beamforming scheme for the base station. After determining the target resource allocation scheme, the user and the base station are controlled to allocate corresponding resources according to the target resource allocation scheme, so that the base station can aggregate the uploaded local model parameters of each user and acquire raw data, thereby completing the aggregation of the global model. The preset transmit power constraint can save user energy consumption while ensuring the accuracy of the model aggregation process. Furthermore, the preset signal quality constraint can also ensure the accuracy of the acquired local raw data. Further, by jointly optimizing the user's transmit power and the base station's receive beamforming scheme, global optimization of the federated learning model data aggregation process can be achieved, thereby reducing signal distortion from the user to the base station and achieving rapid convergence and higher accuracy in federated learning. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 A schematic diagram illustrating a scenario of wirelessly powered federated learning provided in an embodiment of the present invention;

[0021] Figure 2 A schematic diagram illustrating the operation of the wireless power federated learning method provided in an embodiment of the present invention;

[0022] Figure 3A flowchart illustrating the federated learning method for wireless power supply provided in an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of the structure of a wirelessly powered federated learning system resource allocation device provided in an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In real-world federated learning scenarios, users are distributed widely and may be located in various locations. Some of these locations, such as underwater, high in buildings, or carried by drones, may make frequent charging or battery replacement difficult. Because of the lack of a stable power supply, some users will find it difficult to participate continuously in federated learning, thus affecting the accuracy and convergence of the learning process. Currently, a feasible approach is to maintain a stable power supply for users through wireless charging from base stations.

[0027] Therefore, there is an urgent need for a resource allocation method to enable the deployment of federated learning in scenarios without a stable power supply.

[0028] To address the above technical problems, embodiments of the present invention propose a wirelessly powered federated learning method, apparatus, and electronic device.

[0029] The wireless-powered federated learning method provided in this embodiment of the invention can be applied to any existing federated learning system that requires local model parameter aggregation and raw data transmission. Furthermore, the federated learning system may include at least one user and one base station.

[0030] Each user is located within a designated federated learning system. The base station uses beamforming technology to provide energy to each user via radio electromagnetic waves to meet the energy consumption of the user's subsequent participation in federated learning.

[0031] In this system, each user is located within a designated federated learning system. While training local model parameters, they also collect raw data and upload both the local model parameters and raw data to a base station via wireless signal for aggregation. It should be noted that the local model parameters uploaded by each user can include model information such as local model weights and gradients. The uploaded raw data is the data used in machine learning and can be images, text, audio, etc. This embodiment of the invention does not impose any restrictions on the local model parameters and raw data uploaded by each user.

[0032] It should be noted that each user has their own local original dataset. The datasets of different users may have overlapping parts or may be completely different. Therefore, this embodiment of the invention does not limit the content of each user's local data.

[0033] Before each user uploads their local model parameters and raw data via wireless signal, preprocessing is required. After the base station aggregates the wireless signals uploaded by each user and obtains the uploaded raw data, post-processing is required on the aggregated local model parameters from each user. It should be noted that preprocessing methods for user local model parameters and raw data can include encoding, modulation, etc. This embodiment of the invention does not limit the preprocessing methods for user local models. Similarly, post-processing methods at the base station can include demodulation, decoding, etc. This embodiment of the invention does not limit the post-processing methods used by the base station.

[0034] For example, such as Figure 1 The diagram illustrates a scenario of a wirelessly powered federated learning system. It includes K users and their local raw datasets and model parameter sets; and one base station, abbreviated as BS, which stands for BaseStation.

[0035] For example, such as Figure 2 The diagram shown is an operation flowchart of a wirelessly powered federated learning system in one communication round.

[0036] In the first time slot, the base station first acquires the channel response coefficients of the users, and then uses wireless power transmission technology to wirelessly charge all local users. In the second time slot, the base station broadcasts the global model, and at the same time, the users also use the energy collected in the first time slot to collect local data. In the third time slot, while using the collected environmental data for local learning, the local users also continue to collect new environmental data for wireless transmission. In the fourth time slot, the local users simultaneously upload the model parameters and environmental data obtained in the third time slot to the base station. In the fifth time slot, based on the received uplink superimposed signal, the base station simultaneously performs model aggregation and data decoding.

[0037] Furthermore, the wirelessly powered federated learning system resource allocation scheme provided in this embodiment of the invention can be applied to control devices in battery-free wireless federated learning systems. This control device can be installed within a base station or deployed as a standalone device outside the base station. Further, the control device can be any type of electronic device, such as a microcomputer or a microcontroller. Therefore, this embodiment of the invention does not limit the installation location or device type of the control device.

[0038] The following is a detailed description of a wirelessly powered federated learning system resource allocation scheme provided by an embodiment of the present invention.

[0039] Figure 3 This is a flowchart illustrating a resource allocation scheme for a wirelessly powered federated learning system, as provided in an embodiment of the present invention. Figure 3 As shown, the resource allocation method may include the following steps:

[0040] Step 301: Upon entering each communication round, in the first time slot, obtain the channel response coefficient from each user to the base station to determine the resource allocation scheme for the current round. Then, use wireless charging technology to wirelessly charge all local users.

[0041] In this embodiment of the invention, the base station obtains the system's global channel state information (CSI) through a channel estimation method, that is, the base station has the channel response from each user to the base station in the current communication round.

[0042] The channel estimation method can be any method that can obtain the system's global CSI; this embodiment of the invention does not impose specific restrictions on the channel estimation method. Specifically, the channel response from the k-th user to the base station is H. k .

[0043] In this embodiment of the invention, each user obtains the channel response coefficient from the base station by estimating the downlink pilot.

[0044] Specifically, the downlink pilot is a wireless signal sent by the base station to each user before each user transmits local model data uplink, which includes the user's own CSI.

[0045] The downlink pilot may include, but is not limited to, downlink multicast pilots. This embodiment of the invention does not impose specific restrictions on the downlink pilot channel.

[0046] At the start of each communication round, the control device can acquire the channel response from each user to the base station. The control device can acquire the channel response from each user to the base station through various methods. This embodiment of the invention does not specifically limit the methods used.

[0047] The method for determining the resource allocation scheme in the current communication round is as follows: based on the acquired channel response, the first constraint equation of local power for each user and the second constraint equation of signal quality, the target resource allocation scheme in the current training period is determined.

[0048] The target resource allocation scheme includes: a transmit beamforming scheme for each user and a receive beamforming scheme for the base station.

[0049] After obtaining the channel response from each user to the base station, the base station can, based on each obtained channel response, the first constraint equation for the user's local power and the second constraint equation for signal quality, allocate a transmit beamforming scheme for each user and a receive beamforming scheme for the base station within the current communication round, thereby obtaining the target resource allocation scheme.

[0050] Equivalently, the resulting target resource allocation scheme includes a transmit beamforming scheme for each user and a receive beamforming scheme for the base station.

[0051] Step 302: In the second time slot, the base station broadcasts the global model, and at the same time, the user also uses the energy collected in the first time slot to perform local data collection.

[0052] The raw data collected by each user is the data used for machine learning, and can be images, text, audio, etc. This embodiment of the invention does not limit this.

[0053] Step 303: In the third time slot, while the local user is using the collected environmental data for local learning, the local user also continues to collect new environmental data for wireless transmission.

[0054] Step 304: In the fourth time slot, the local user uploads the model parameters and raw data obtained in the third time slot to the base station simultaneously.

[0055] In this process, the base station broadcasts the transmission beamforming scheme to each user, and each user adjusts its own transmission power according to the assigned scheme. Subsequently, users allocate corresponding resources to themselves according to the assigned scheme to complete the wireless transmission task.

[0056] Specifically, all user signals received by the base station can be represented as:

[0057]

[0058] Among them, F k Let g represent the transmit beamforming matrix for the k-th user model parameters. k,l Let n0 represent the transmit beamforming vector for the l-th symbol of the k-th user, and n0 be the Jiaxing Gaussian white noise vector.

[0059] Step 305: In the fifth time slot, based on the received uplink superimposed signal, the base station simultaneously performs model aggregation and data decoding.

[0060] The base station uses the aforementioned receiving beamforming scheme to aggregate and process the local model parameters and raw data uploaded by each user.

[0061] Specifically, the model parameters after base station processing are as follows: The raw data obtained by the base station is At the base station, mean square error (MSE) and signal-to-interference-plus-noise ratio (SINNR) are used respectively. This is used to measure the degree of distortion of model parameters and the signal quality of the original data.

[0062] Where M represents the receive beamforming matrix used by the base station for the model parameters, c k,l This represents the receive beamforming vector for the l-th symbol of the k-th user.

[0063] The expression for the MSE is:

[0064]

[0065] The expression for the signal-to-interference-plus-noise ratio is:

[0066]

[0067] Optionally, in one specific implementation, the determination of the target resource allocation scheme within the current communication round based on each acquired channel response, preset local power constraints, and signal quality constraints in step 301 may include the following steps 11-15:

[0068] Step 11: Using the preset signal distortion optimization equation and the first constraint formula, determine the target resource allocation scheme within the current communication round; wherein, the signal distortion optimization equation is:

[0069]

[0070] in, For the base station's estimation of the global model, The specific expression is:

[0071]

[0072] in, For the channel response from the k-th user to the base station, σ 2 For the power of additive white Gaussian noise, N t N represents the number of antennas for the user. rThis represents the number of antennas at the base station. This represents the transmit beamforming matrix for the parameters of the k-th user model. Let represent the transmit beamforming vector for the l-th symbol of the k-th user, where . This represents the receive beamforming matrix used by the base station for the model parameters.

[0073] The first constraint equation for user local power is:

[0074]

[0075] The second constraint equation for signal quality is:

[0076]

[0077] Among them, each formula in the first constraint equation and the second constraint condition equation can be used as the various constraints that the target resource allocation scheme needs to satisfy when using the above signal distortion degree optimization equation to determine the target resource allocation scheme in the current time slot.

[0078] Specifically:

[0079] As the first constraint, it means that in the target resource allocation scheme, the energy consumption allocated to each user is lower than the energy supplied by the base station in the first time slot.

[0080] As a second constraint, it means that in the target resource allocation scheme, the signal-to-interference-plus-noise ratio (SIR) of each raw data should be higher than the preset minimum SIR to support subsequent demodulation and other processing.

[0081] Step 12: Based on the signal distortion optimization equation, split it into two sub-optimization equations; wherein, the first sub-optimization equation uses the base station's receiving beamforming scheme as the optimization variable, and the second sub-optimization equation uses the user's transmitting beamforming scheme as the optimization variable.

[0082] For the first optimization subproblem, the minimum mean square error acceptance is adopted as the optimization result;

[0083] For the second optimization subproblem, it is transformed into a semi-positive definite programming problem, and then the penalty function method and the continuous convex approximation method are used to determine its optimization result.

[0084] Step 13: Solve the first optimization subproblem to determine the optimal scheme for receiving beamforming.

[0085] The first optimization sub-problem includes a first signal distortion degree sub-optimization equation.

[0086] The sub-optimization equation for the first signal distortion level is:

[0087]

[0088] The beamforming scheme with minimum mean square error used by the base station is as follows:

[0089]

[0090]

[0091] in, This represents the received beamforming vector for the l-th symbol of the k-th user.

[0092] Step 14: First, use matrix lifting techniques to transform the second sub-optimization problem into a semidefinite programming problem. Then, use the penalty function and continuous convex approximation method to transform the semidefinite programming problem into a convex optimization problem. Solve the convex optimization problem to determine the transmit beamforming scheme for each user in the current communication round.

[0093] The second optimization sub-problem includes a second signal distortion degree sub-optimization equation, a first constraint equation for user local power, and a second constraint equation for signal quality.

[0094] The second signal distortion sub-optimization equation is:

[0095]

[0096] The first constraint equation for the user's local power is:

[0097]

[0098] The second constraint equation for signal quality is:

[0099]

[0100] Optionally, matrix lifting techniques can be used to obtain equivalent second signal distortion sub-optimization equations, first constraint equations for user local power, and second constraint equations for signal quality.

[0101] The equivalent second signal distortion sub-optimization equation is:

[0102]

[0103] in satisfy Represents the trace of a matrix.

[0104] The first sub-constraint equation for the equivalent user local power is:

[0105]

[0106] The equivalent signal quality second sub-constraint equation is as follows:

[0107]

[0108] in, t4 is the duration of the fourth time slot.

[0109]

[0110] Optionally, in one specific implementation, the penalty function method and the continuous convex approximation method can be used to handle the new constraints that arise in matrix lifting. Among them, continuous graph approximation methods include, but are not limited to, obtaining The first-order Taylor expansion, express 2-norm.

[0111] Specifically, the second signal distortion sub-optimization equation can be equivalently transformed into the following form:

[0112]

[0113] in, This represents the result obtained after the τth iteration. Feasible solution express The eigenvector corresponding to the largest eigenvalue, where μ is the penalty factor.

[0114] Optionally, the penalty factor is preset by the base station.

[0115] Specifically, the aforementioned second optimization subproblem can be equivalently transformed into an equivalent second optimization subproblem, including an equivalent second signal distortion degree sub-optimization equation, an equivalent user local power first sub-constraint equation, and an equivalent signal quality second sub-constraint equation.

[0116] The equivalent second signal distortion sub-optimization equation is as follows:

[0117]

[0118] The equivalent user local power first sub-constraint equation is:

[0119]

[0120] The equivalent signal quality second sub-constraint equation is:

[0121]

[0122] Step 15: Iteratively determine the base station's receive beamforming scheme and each user's transmit beamforming scheme in sequence until convergence or the maximum number of iterations is reached.

[0123] Optionally, the maximum number of iterations is limited by a limit set by the base station.

[0124] The iterative process is completed at the base station. After convergence or reaching the maximum number of iterations limit, the base station determines the transmit power allocated to each user, the phase shift matrix set for each smart reflector, and the receive factor allocated to the base station in the current communication round, thus obtaining the target resource allocation scheme.

[0125] Corresponding to the wirelessly powered federated learning method provided in the above embodiments of the present invention, the present invention provides a wirelessly powered federated learning system resource allocation device.

[0126] Figure 4 This is a schematic diagram of a wirelessly powered federated learning system resource allocation device provided in an embodiment of the present invention. Figure 4 As shown, the resource allocation device may include:

[0127] The information acquisition module 410 is used to acquire the current channel response between each user and the base station in each communication round.

[0128] The scheme determination module 420 is used to determine the target resource allocation scheme for the current communication round based on the acquired channel response of each user, the preset local power constraints and signal quality constraints of each user; wherein, the target resource allocation scheme includes: a transmit beamforming scheme for each user and a receive beamforming scheme for the base station.

[0129] The resource allocation module 430 is used to control users and base stations to allocate corresponding resources to themselves according to the target resource allocation scheme, so that the base station can aggregate the local model parameters uploaded by each user and obtain the raw data uploaded by the user.

[0130] Optionally, in one specific implementation, the scheme determination module 420 may include:

[0131] The scheme determination submodule is used to determine the target resource allocation scheme within the current communication round by using a preset signal distortion optimization equation and the first constraint formula.

[0132] The optimization equation for the degree of signal distortion is as follows:

[0133]

[0134] in, For the base station's estimation of the global model, The specific expression is:

[0135]

[0136] in, For the channel response from the k-th user to the base station, σ 2 For the power of additive white Gaussian noise, N t N represents the number of antennas for the user. r This represents the number of antennas at the base station. This represents the transmit beamforming matrix for the parameters of the k-th user model. Let represent the transmit beamforming vector for the l-th symbol of the k-th user, where . This represents the receive beamforming matrix used by the base station for the model parameters.

[0137] The first constraint equation for the user's local power is:

[0138]

[0139] The second constraint equation for signal quality is:

[0140]

[0141] Among them, each formula in the first constraint equation and the second constraint condition equation can be used as the various constraints that the target resource allocation scheme needs to satisfy when using the above signal distortion degree optimization equation to determine the target resource allocation scheme in the current time slot.

[0142] Optionally, in one specific implementation, the scheme determining sub-module includes:

[0143] The signal distortion optimization equation splitting unit is used to split the signal distortion optimization equation into two sub-optimization equations; wherein, the first sub-optimization equation uses the base station's receiving beamforming scheme as the optimization variable, and the second sub-optimization equation uses the user's transmitting beamforming scheme as the optimization variable;

[0144] The base station receive beamforming solution unit is used to solve the first sub-optimization problem and determine the optimal solution of the base station receive beamforming scheme in the current communication round.

[0145] The user transmit beamforming solution unit is used to first transform the second sub-optimization problem into a semidefinite programming problem using matrix lifting techniques, and then transform the semidefinite programming problem into a convex optimization problem using a penalty function and a continuous convex approximation method. Solving the convex optimization problem determines the transmit beamforming scheme for each user in the current communication round.

[0146] The iterative unit is used to iteratively determine the base station's receive beamforming scheme and the user's transmit beamforming scheme in sequence until convergence or the maximum number of iterations is reached.

[0147] Corresponding to the resource allocation scheme for a multi-intelligent reflector-assisted federated learning system provided in the above embodiments of the present invention, the present invention also provides an electronic device, such as... Figure 4 As shown, it includes a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504.

[0148] Memory 503 is used to store computer programs;

[0149] When the processor 501 executes the program stored in the memory 503, it implements the steps of the resource allocation method for any of the multi-intelligent reflective surface-assisted federated learning systems provided in the above embodiments of the present invention.

[0150] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0151] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0152] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0153] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0154] A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, it implements the steps of any of the multi-intelligent reflector-assisted federated learning system resource allocation methods provided in the embodiments of the present invention.

[0155] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform the steps of any of the multi-intelligent reflector-assisted federated learning system resource allocation methods provided in the embodiments of the present invention.

[0156] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0157] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0158] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments, electronic device embodiments, computer-readable storage medium embodiments, and computer program product embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0159] The embodiments described above are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in this application, or make equivalent substitutions for some of the specific technologies; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A wirelessly powered federated learning method, characterized in that, In each communication round, in the first time slot, the base station first acquires the channel response coefficient of each user, determines the target resource allocation scheme, and then uses wireless power transmission technology to wirelessly charge all local users. In the second time slot, the base station broadcasts the global model, and simultaneously, users use the energy collected in the first time slot to collect local data. In the third time slot, local users, while using the collected environmental data for local learning, also continue to collect new environmental data for wireless transmission. In the fourth time slot, local users allocate corresponding resources to themselves according to the resource allocation scheme and simultaneously upload the model parameters and environmental data obtained in the third time slot to the base station. In the fifth time slot, based on the received uplink superimposed signal, the base station simultaneously performs model aggregation and data decoding, and allocates corresponding resources to itself according to the resource allocation scheme. The method for determining the target resource allocation scheme is as follows: based on the acquired channel response coefficient, the first constraint equation of each user's local power, and the second constraint equation of signal quality, the target resource allocation scheme for the current training period is determined. The target resource allocation scheme includes: a transmit beamforming scheme for each user in the fourth time slot and a receive beamforming scheme for the base station in the fifth time slot. In the fourth time slot, users will simultaneously upload raw data and model parameters, where the first... User model parameters are used This indicates that the raw data is used express, Represents raw data The In this project, the user will apply different beamforming methods to the raw data and model parameters in the fourth time slot, among which, using Indicates that for the first The transmit beamforming matrix of each user model parameter, in order to Indicates that for the first The first user's The transmit beamforming vector of symbols, where, Given the number of antennas for the user, the signal received by the base station is represented as: ， in, For the first Channel response coefficient from a user to the base station, It is an additive white Gaussian noise vector; In the fifth time slot, the base station uses beamforming to obtain two types of data: global model parameters and raw data. This represents the receive beamforming matrix used by the base station for the model parameters. Indicates that for the first The first user's The received beamforming vector of each symbol, Given the number of antennas at the base station, the global model parameters received by the base station are represented as follows: The raw data obtained by the base station is represented as ; The first constraint equation for the local power of each user is: ， in, For the first Energy consumption required for each user to update the local model For the first The power required for a user to transmit a signal For the first Energy received by each user in the first time slot; in, , The duration of the fourth time slot; The second constraint equation for signal quality is: ， in, This represents the minimum required signal-to-interference-plus-noise ratio. Indicates the first The first user's The signal-to-interference-plus-noise ratio (SIR) of each data point is expressed as: ； Using a pre-defined parameter error optimization equation, the target resource allocation scheme for the current training period is determined. The parameter error optimization equation is as follows: ， in, The power of additive white Gaussian noise; The optimization objective for parameter error is: ， Using a pre-defined error optimization objective equation and the first and second constraint equations, with the goal of minimizing the aggregation error of the federated learning model, the resource allocation scheme for user transmit beamforming and base station receive beamforming in the current communication round is determined, including: The optimization equation is split into two sub-optimization problems, where the first sub-optimization equation takes the receiving beamforming scheme as the variable and the second sub-optimization equation takes the transmitting beamforming scheme as the variable. For the first optimization subproblem, the minimum mean square error acceptance is adopted as the optimization result; For the second optimization subproblem, it is transformed into a semi-positive definite programming problem, and then the penalty function method and the continuous convex approximation method are used to determine its optimization result. For solving the first optimization subproblem, the minimum mean square error is used as the optimization result, including: The optimization equation for the first subproblem is: ， The beamforming scheme with minimum mean square error used by the base station is as follows: ， in, ; To solve the second optimization subproblem, it is transformed into a semidefinite programming problem, and then the penalty function method is used to transform it into an equivalent second optimization subproblem to determine its optimization result: Among them, the equivalent second sub-optimization problem includes: the equivalent second signal distortion degree sub-optimization equation, the equivalent user local power first sub-constraint equation, and the equivalent signal quality second sub-constraint equation; The optimization equation for the equivalent second optimization subproblem is: ， in, , Indicates the first After round of iterations Feasible solution express The eigenvector corresponding to the largest eigenvalue. As a penalty factor; in, , satisfy , Represents the trace of a matrix; The equivalent user local power first sub-constraint equation is: ， in, , To determine the time required for the fourth time slot, solve this problem to obtain the beamforming scheme for the transmitter. The equivalent signal quality second sub-constraint equation is: ， in, .

2. The federated learning method for wireless power supply according to claim 1, characterized in that, In the first time slot, the base station obtains the channel response coefficient for each user through a channel estimation method.

3. A wirelessly powered resource allocation device for a federated learning system, characterized in that, The following modules are included for implementing the method of any one of claims 1-2: The information acquisition module is used to acquire the current channel response between each user and the base station in each communication round; The scheme determination module is used to determine the target resource allocation scheme for the current communication round based on the acquired channel response of each user, the preset local power constraints and signal quality constraints of each user; wherein, the target resource allocation scheme includes: a transmit beamforming scheme for each user and a receive beamforming scheme for the base station; The resource allocation module is used to control users and base stations to allocate corresponding resources to themselves according to the target resource allocation scheme, so that the base station can aggregate the local model parameters uploaded by each user and obtain the raw data uploaded by the user.

4. An electronic device for resource allocation in a wirelessly powered federated learning system, characterized in that, The system includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory is used to store computer programs; and the processor, when executing the program stored in the memory, implements the steps of the method described in any one of claims 1-2.

Images