Wireless energy supply federated learning method, device and equipment based on simultaneous wireless energy transfer, and medium

Abstract

The invention discloses a wireless energy supply federated learning method, device and equipment based on wireless energy-carrying communication, and a medium, relates to the field of federated learning, is applied to a server, and comprises the following steps: jointly optimizing sampling parameters of Internet of Things equipment based on task parameters corresponding to a federated learning task, and distributing a joint optimization result to the Internet of Things equipment; broadcasting the current global model based on the wireless energy-carrying communication technology, so that the Internet of Things equipment judges whether the Internet of Things equipment participates in the model training of the current round based on the target sampling probability, and if so, determines a trained model based on the wireless power transmission technology, the target power segmentation factor and the current global model; receiving the trained model, and determining an updated current global model in combination with a preset adaptive aggregation rule; and skipping to the step of broadcasting the current global model based on the simultaneous wireless information and power transfer technology until a target global model is determined. According to the application, wireless energy supply federal learning can be realized, and the sustainability of the whole federal learning process is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of federated learning, in particular to a wireless powered federated learning method and device based on wireless simultaneous wireless information and power transfer (SWIPT). BACKGROUND

[0002] In the federated learning in the Internet of Things (IoT) scenario, the IoT devices as the federated learning clients usually have limited energy storage capacity, which is difficult to support multiple rounds of local model training and uplink communication.

[0003] To solve the problem, in the existing related solutions: 1) wireless power transfer (WPT) is used to realize wireless power supply, but in this scheme, model broadcasting and energy collection are carried out separately, and energy collection needs to be carried out in a separate time period, which significantly prolongs the duration of each round. 2) wireless simultaneous wireless information and power transfer (SWIPT) is used to realize wireless power supply, but on the one hand, it only considers the case where all clients participate, which leads to the problem of stragglers, i.e. faster devices need to wait for slower devices to complete the energy collection and local calculation / upload phase, slowing down the training process; on the other hand, it cannot ensure the sustainability of the federated learning system, and does not consider how the client scheduling affects the convergence behavior, which leads to the inability to minimize the convergence error within a given system time budget. SUMMARY

[0004] Therefore, the present application aims to provide a wireless powered federated learning method and device based on wireless simultaneous wireless information and power transfer (SWIPT), which can effectively realize wireless powered federated learning, and by jointly optimizing the IoT device sampling probability and power splitting factor, it ensures the sustainability of the entire wireless powered federated learning process and improves the learning efficiency, thereby minimizing the convergence error within a given time budget. The specific scheme is as follows:

[0005] In a first aspect, the present application provides a wireless powered federated learning method based on wireless simultaneous wireless information and power transfer (SWIPT), applied to a server, comprising:

[0006] Based on the task parameters corresponding to the federated learning task, the joint optimization of the IoT device sampling parameters is carried out to determine the joint optimization result; the task parameters include a total federated learning time threshold; the joint optimization result includes target sampling probabilities and target power splitting factors corresponding to each IoT device respectively, the target sampling probability is used to control the training participation frequency, and the target power splitting factor is used to control the energy collection amount;

[0007] The joint optimization result is respectively sent to the corresponding Internet of Things device;

[0008] The current global model is broadcast based on wireless power transfer technology, so that the Internet of Things device determines whether to participate in the current round of model training based on the target sampling probability, and when yes, determines the trained model based on wireless power transfer technology, the target power division factor and the current global model;

[0009] The trained model uploaded by each Internet of Things device is received, and an updated current global model is determined in combination with a preset adaptive aggregation rule;

[0010] Based on the updated current global model, the step of broadcasting the current global model based on wireless power transfer technology is re-executed until a preset task termination condition is met, and a target global model corresponding to the federated learning task is determined.

[0011] Optionally, the joint optimization of the Internet of Things device sampling parameters based on the task parameters corresponding to the federated learning task is performed to determine a joint optimization result, comprising:

[0012] Based on wireless power transfer technology and task parameters corresponding to the federated learning task, an initial model and a preset parameter are sent to each Internet of Things device;

[0013] Parameter information reported by each Internet of Things device is received and saved; the parameter information includes calculation task time consumption, communication rate, communication bandwidth, data volume and wireless channel gain;

[0014] Based on the parameter information, an Internet of Things device energy collection model, an Internet of Things device communication model, an Internet of Things device energy consumption model and an Internet of Things device round time model are constructed to determine a model construction result;

[0015] Based on the parameter information, a constraint relationship between the Internet of Things device sampling probability and the power division factor is constructed;

[0016] A convergence analysis model is constructed based on the constraint relationship; the convergence analysis model is used for convergence analysis of the global model;

[0017] Based on the convergence analysis model, the constraint relationship, the total federated learning time threshold, the approximate algorithm and the model construction result, a joint approximate optimization problem of the Internet of Things device sampling probability and the power division factor is constructed;

[0018] Based on the joint approximate optimization problem, a preset iterative optimization algorithm and a convex optimization toolbox, a joint optimization result corresponding to each Internet of Things device is determined.

[0019] Optionally, the constructing the Internet of Things device energy collection model, the Internet of Things device communication model, the Internet of Things device energy consumption model and the Internet of Things device round time model based on the parameter information comprises:

[0020] constructing an Internet of Things device energy collection model based on the wireless channel gain, the power split factor, the average transmission power of the server, the energy conversion efficiency and the energy collection time, wherein the energy conversion efficiency is an efficiency of converting collected signal energy into stored energy;

[0021] constructing an Internet of Things device receiving rate model based on the power split factor, the wireless channel gain, the noise power and the Internet of Things device bandwidth allocation information, wherein the Internet of Things device receiving rate model is used to represent a data rate of the Internet of Things device receiving global model;

[0022] constructing a packet error rate model based on the wireless channel gain, the noise power and the average transmission power of the server, wherein the packet error rate model is used to represent a packet error rate from the server to the Internet of Things device;

[0023] constructing a global model decoding variable based on the packet error rate model;

[0024] determining an Internet of Things device communication model based on the Internet of Things device receiving rate model and the global model decoding variable;

[0025] constructing an Internet of Things device energy consumption model based on the Internet of Things device transmission power, the Internet of Things device transmission time and model training energy consumption;

[0026] constructing an Internet of Things device round time model based on the Internet of Things device sampling probability and the packet error rate model.

[0027] Optionally, the constructing a constraint relationship between the Internet of Things device sampling probability and the power split factor based on the parameter information comprises:

[0028] constructing a constraint relationship between the Internet of Things device sampling probability and the power split factor based on the parameter information, the packet error rate model, the energy conversion efficiency, the average transmission power of the server and the Internet of Things device energy consumption model.

[0029] Optionally, the Internet of Things device determines whether to participate in the current round of model training based on the target sampling probability, comprising:

[0030] the Internet of Things device performs a Bernoulli trial based on the target sampling probability and a sampling indicator of the current round to determine a trial result;

[0031] determine whether to participate in the model training of the current round based on the test result.

[0032] Optionally, when the determination is yes, the trained model is determined based on the wireless power transmission technology, the target power split factor and the current global model, including:

[0033] When the IoT device determines to participate in the model training of the current round, it is determined that the IoT device belongs to the sampled IoT devices of the current round;

[0034] The sampled IoT device receives the current global model based on the wireless power transmission technology;

[0035] During the reception of the current global model, the sampled IoT device collects energy based on the wireless power transmission technology and the target power split factor until all the sampled IoT devices of the current round complete the global model reception operation, and then the energy collection result is determined;

[0036] The sampled IoT device determines whether the energy constraint condition is met based on the energy collection result to determine a constraint judgment result;

[0037] If the constraint judgment result is yes, the sampled IoT device decodes the received current global model, and when the decoding is successful, the corresponding model decoding result and the energy collection result are used for model training and gradient updating to determine a trained model and an updated local random gradient.

[0038] Optionally, the receiving of the trained model uploaded by each IoT device and the determination of the updated current global model based on a preset adaptive aggregation rule, including:

[0039] The trained model uploaded by each IoT device and the updated local random gradient are received.

[0040] The updated local random gradient, the trained model, the target sampling probability, the IoT device data distribution weight and the learning rate are used for model aggregation to determine the updated current global model.

[0041] In a second aspect, the present application provides a wireless power-enabled federated learning device based on wireless power communication, applied to a server, including:

[0042] The parameter joint optimization module is configured to perform joint optimization of the sampling parameters of the Internet of Things devices based on task parameters corresponding to the federated learning task to determine a joint optimization result, wherein the task parameters include a total federated learning time threshold, and the joint optimization result includes target sampling probabilities and target power split factors corresponding to the Internet of Things devices respectively, the target sampling probabilities are used to control training participation frequencies, and the target power split factors are used to control energy collection amounts.

[0043] The parameter issuing module is configured to issue the joint optimization result to the corresponding Internet of Things devices respectively.

[0044] The model broadcasting module is configured to broadcast a current global model based on the wireless power communication technology, so that the Internet of Things devices judge whether to participate in model training of a current round based on the target sampling probabilities, and when yes, determine a post-training model based on the wireless power transmission technology, the target power split factor and the current global model.

[0045] The model aggregation module is configured to receive the post-training models uploaded by the Internet of Things devices, and determine an updated current global model in combination with a preset adaptive aggregation rule.

[0046] The step jumping module is configured to jump back to the step of broadcasting the current global model based on the wireless power communication technology based on the updated current global model, and determine a target global model corresponding to the federated learning task when a preset task termination condition is met.

[0047] In a third aspect, the present application provides an electronic device, comprising:

[0048] A memory configured to save a computer program.

[0049] A processor configured to execute the computer program to implement the steps of the wireless power-enabled federated learning method based on the wireless power communication.

[0050] In a fourth aspect, the present application provides a computer readable storage medium configured to save a computer program, wherein the computer program is executed by a processor to implement the steps of the wireless power-enabled federated learning method based on the wireless power communication.

[0051] As can be seen, in this application, applied to the server side, the method includes: jointly optimizing the sampling parameters of IoT devices based on the task parameters corresponding to the federated learning task to determine the joint optimization result; the task parameters include a total federated learning time threshold; the joint optimization result includes a target sampling probability and a target power segmentation factor corresponding to each IoT device, wherein the target sampling probability is used to control the training participation frequency and the target power segmentation factor is used to control the energy harvesting amount; distributing the joint optimization result to the corresponding IoT devices; broadcasting the current global model based on wireless power transmission technology so that the IoT devices can determine whether they participate in the current round of model training based on the target sampling probability, and if so, determining the trained model based on the wireless power transmission technology, the target power segmentation factor, and the current global model; receiving the trained model uploaded by each IoT device and determining the updated current global model in combination with a preset adaptive aggregation rule; based on the updated current global model, re-jumping to the step of broadcasting the current global model based on wireless power transmission technology until a preset task termination condition is met, and determining the target global model corresponding to the federated learning task. In other words, this application first jointly optimizes the sampling probability and power segmentation factor of IoT devices through the server, and then distributes the obtained target sampling probability and target power segmentation factor to the corresponding IoT devices. Then, based on wireless power transmission technology, the current global model is broadcast so that the IoT devices can determine whether to participate in the current round of model training based on the target sampling probability. If so, the trained model is determined based on the wireless power transmission technology, the target power segmentation factor, and the current global model. Afterwards, the received trained models are adaptively aggregated to determine the updated current global model. Then, based on the updated current global model, the model is distributed again until a preset task termination condition is met, at which point the target global model corresponding to the federated learning task is determined. This effectively realizes wireless power federated learning, and by jointly optimizing the sampling probability and power segmentation factor of IoT devices, the sustainability of the entire wireless power federated learning process is ensured, and learning efficiency is improved, thereby minimizing convergence error within a given time budget. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0053] Figure 1A flowchart of a wireless power supply federated learning method based on wireless power-carrying communication provided for this application;

[0054] Figure 2 A schematic diagram of a specific architecture for wireless power federated learning based on wireless power-carrying communication provided for this application;

[0055] Figure 3 A schematic diagram of a wireless-powered federated learning device based on wireless power-carrying communication provided for this application;

[0056] Figure 4 This application provides a structural diagram of an electronic device. Detailed Implementation

[0057] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] In federated learning within IoT scenarios, IoT devices acting as clients typically have limited energy storage capacity, making it difficult to support multiple rounds of local model training and uplink communication. To address this issue, existing solutions include: 1) using wireless power transfer technology for wireless power supply. However, this approach separates model broadcasting and energy harvesting, requiring dedicated time slots for energy harvesting, significantly extending the duration of each round. 2) using wireless power-carrying communication technology for wireless power supply. However, this approach largely considers the scenario of full client participation, leading to a "laggard problem"—faster devices must wait for slower devices to complete energy harvesting and local computation / upload phases, slowing down the training process. Furthermore, it fails to ensure the sustainability of the federated learning system and does not consider how client scheduling affects convergence behavior, resulting in the inability to minimize convergence error within a given system time budget.

[0059] To this end, this application provides a wireless power federated learning scheme based on wireless power-carrying communication, which can effectively realize wireless power federated learning. By jointly optimizing the sampling probability and power division factor of IoT devices, it ensures the sustainability of the entire wireless power federated learning process and improves learning efficiency, thereby minimizing the convergence error within a given time budget.

[0060] See Figure 1 As shown, this embodiment of the invention discloses a wireless power supply federated learning method based on wireless power-carrying communication, applied to the server side, including:

[0061] Step S11: Based on the task parameters corresponding to the federated learning task, perform joint optimization of the sampling parameters of IoT devices to determine the joint optimization result; the task parameters include the total federated learning time threshold; the joint optimization result includes the target sampling probability and target power segmentation factor corresponding to each IoT device, the target sampling probability is used to control the training participation frequency, and the target power segmentation factor is used to control the energy harvesting amount.

[0062] In this embodiment, for the federated learning task, the target sampling probability and power segmentation factor of IoT devices need to be jointly optimized within a given system time budget to ensure that the convergence error is minimized while ensuring the sustainability of the system. Specifically: Based on the wireless energy-carrying communication technology and the task parameters corresponding to the federated learning task, an initial model and preset parameters are sent to each IoT device; the parameter information reported by each IoT device is received and saved; the parameter information includes the calculation task time, communication rate, communication bandwidth, data volume, and wireless channel gain; based on the parameter information, an IoT device energy harvesting model, an IoT device communication model, an IoT device energy consumption model, and an IoT device round-time model are constructed to determine the model construction results; based on the parameter information, a constraint relationship between the IoT device sampling probability and the power segmentation factor is constructed; based on the constraint relationship, a convergence analysis model is constructed; the convergence analysis model is used to perform convergence analysis of the global model; based on the convergence analysis model, the constraint relationship, the total federated learning time threshold, the approximation algorithm, and the model construction results, a joint approximate optimization problem of the IoT device sampling probability and the power segmentation factor is constructed; based on the joint approximate optimization problem, the preset iterative optimization algorithm, and the convex optimization toolbox, the joint optimization results corresponding to each IoT device are determined. Understandably, after obtaining the parameter information reported by the IoT devices acting as clients, the server saves and calculates the known parameters, including the parameters reported by the clients, the total number of clients, the server's total bandwidth, the server's average transmit power, the total data volume of all clients, thermal noise power, circuit noise power, energy conversion efficiency, etc.

[0063] It is important to understand that, in this embodiment, the construction of relevant models for IoT devices includes: constructing an IoT device energy harvesting model based on the wireless channel gain, power segmentation factor, server average transmit power, energy conversion efficiency, and energy harvesting time; the energy conversion efficiency is the efficiency of converting collected signal energy into stored energy; constructing an IoT device receiving rate model based on the power segmentation factor, the wireless channel gain, noise power, and IoT device bandwidth allocation information; the IoT device receiving rate model is used to characterize the data rate at which the IoT device receives the global model; constructing a packet error rate model based on the wireless channel gain, noise power, and server average transmit power; the packet error rate model is used to characterize the packet error rate from the server to the IoT device; constructing global model decoding variables based on the packet error rate model; determining the IoT device communication model based on the IoT device receiving rate model and the global model decoding variables; constructing an IoT device energy consumption model based on the IoT device transmit power, IoT device transmit time, and model training energy consumption; and constructing an IoT device round-time model based on the IoT device sampling probability and the packet error rate model.

[0064] In other words, combining Figure 2 As shown, considering Phase 1 in the diagram (used for global model broadcasting and SWIPT), when the server broadcasts the global model, all IoT devices acting as clients can use SWIPT for energy harvesting, with a target power segmentation factor of [missing value]. The client splits the received signal into two circuits: Power is used for energy storage, 1- Power is used for data detection. The average transmit power at the server side is [value missing]. In the case of wireless channel gain The energy collected per second by client n is: ,in This indicates the conversion efficiency of collecting signal energy into stored energy. ∈(0,1). The client will be in the entire phase one time. Energy is harvested internally. Therefore, an IoT device energy harvesting model is used to characterize the energy collected by client n in each round. , can be represented as: .

[0065] Regarding the IoT device reception rate model, in this embodiment, Orthogonal Frequency Division Multiple Access (OFDMA) is used for downlink transmission. The IoT device reception rate model, characterizing the data rate at which client n receives the global model data from the base station (i.e., the server), can be expressed as:

[0066] .

[0067] in, It is Gaussian noise power. This represents the average noise power introduced during the conversion from received signal to baseband signal, and W is the bandwidth allocated by the base station to each IoT mobile device. (Utilizing the waterfall threshold...) Packet error rate from base station to client n It can be represented as:

[0068] .

[0069] The above formula is the grouping error rate model for IoT devices.

[0070] Regarding the global model decoding variables, client n successfully decoded the binary variables of the global model in round r. This can be expressed as: if decoding is successful, =1 (probability is 1-) );otherwise =0 (probability is) ).

[0071] Regarding the energy consumption model for IoT devices, in this embodiment, as shown in... Figure 2 In Phase Two (for local model training and uploading), the sampled client (i.e., the sampled IoT device) that successfully decoded the global model and harvested energy will use the harvested energy for local training and local model uploading. The energy consumed by the model is calculated by client n. The upload energy of client n is ,in It is the transmit power of client n. This refers to the transmission time. The energy consumption model for IoT devices, used to characterize the total energy consumed by client n in one communication round, is... .

[0072] Regarding the IoT device cycle time model, the time of Phase 1 The longest downlink time among the sampled clients in round r is determined by the length of the downlink time. Expected phase one time for:

[0073] .

[0074] In the formula, Let N be the target sampling probability for client N. For the first phase of client N, Let i be the target sampling probability. Let N be the target sampling probability for client n, and N be the total number of clients. Let n be the phase one time for client n.

[0075] Phase Two Time The decision is made by the slowest of the sampled devices that successfully decoded the global model in this round r, i.e. Expected Phase Two Time for:

[0076] .

[0077] In the formula, Let N be the packet error rate for client N. For client N, the second phase time is... Let i be the packet error rate corresponding to client i. This refers to the second phase time corresponding to client n.

[0078] Then we can obtain the expected duration of each round. The IoT device round-trip time model is as follows: .

[0079] Furthermore, in this embodiment, the construction of the constraint relationship between the sampling probability of IoT devices and the power segmentation factor includes: constructing the constraint relationship between the sampling probability of IoT devices and the power segmentation factor based on the parameter information, the grouping error rate model, the energy conversion efficiency, the server's average transmit power, and the IoT device's energy consumption model. Specifically, the client needs sufficient energy to participate in local training and local model uploading. For any total training epoch R, if the following conditions are met: In this way, wireless power federated learning is sustainable, where It is the target sampling probability of client n. This is the expected time for stage one. This constraint indicates that when When fixed, there exists The upper bound of this means that the client's sampling rate should be small enough to achieve self-sufficiency. Conversely, when When fixed, there exists The lower bound indicates It should be large enough to divert more energy from the received signal for harvesting.

[0080] Regarding the convergence analysis model, the server uses adaptive aggregation rules to ensure that the global aggregation model remains unbiased:

[0081] .

[0082] in, It is a global model for round r. It is the data distribution weight. It's the learning rate. The sampling indicator of client n in round r satisfies , It is the local stochastic gradient of client n. The server derives the upper bound of convergence: for the total number of training epochs R, if... and Then the global model sequence satisfy:

[0083] .

[0084] in, It is the minimum loss value, where L is the Lipschitz constant. It is the upper bound of the gradient divergence. It is the stochastic gradient variance. Let be the gradient of the loss function of the global model in the r-th round. This represents the loss function value of the initial model.

[0085] The construction of a joint approximate optimization problem for IoT device sampling probability and power segmentation factor. In this embodiment, the optimization problem is first established. The server's goal is to achieve the desired result within a given total federated learning time threshold. Internally, optimize independent sampling probability. and power division factor To minimize the average expected global gradient The optimization problem can be formulated as: minimizing The constraint is: total time (Time constraints) (Energy balance constraint). , (Numerical range constraints).

[0086] Next, the server approximates the optimization problem. Due to the complexity of the optimization problem, the server approximates the above optimization problem as: minimizing... The constraints are: .in, for - To further improve algorithm efficiency, the server further approximates... , ,therefore , where M is the size of the training model during transmission. This approximation method captures the basic trends while significantly reducing computational complexity.

[0087] Subsequently, on the server side, considering data heterogeneity, communication heterogeneity, and computational heterogeneity, an iterative optimization algorithm is used to solve the approximate optimization problem. The algorithm flow is as follows:

[0088] (1) Initialize sampling probability and power division factor And set the corresponding threshold. and .

[0089] (2) When or At that time, perform the following steps:

[0090] 1) Record the result of the previous iteration: , .

[0091] 2) Fixed The optimal solution to the P3 problem can be obtained by using a convex optimization toolbox (such as CVX (a convex optimization modeling toolkit)). .

[0092] 3) Fixed For each client n, in Within the range, in step size Perform a linear search to find the optimal solution that satisfies the constraints. .

[0093] (3) Return the optimal and .

[0094] In this way, by using an efficient iterative optimization algorithm, the goal of minimizing the convergence error within a given system time budget is achieved by alternately optimizing the client sampling rate and power segmentation factor. The algorithm has low complexity and high computational efficiency.

[0095] The P3 question is as follows:

[0096] .

[0097] Where I is the average noise power introduced by the conversion of the received signal to the baseband signal. Let n be the phase two time of client n in round r. Furthermore, equation (16) above is the aforementioned sustainable determination condition for wireless power federated learning: .

[0098] Step S12: Send the joint optimization results to the corresponding IoT devices respectively.

[0099] In this embodiment, after the server completes the joint optimization of the parameters, it obtains the sampling probabilities of each target. and each target power segmentation factor The data is then distributed to the corresponding IoT devices. This allows the IoT devices to subsequently participate in federated learning based on their respective received target sampling probabilities and target power segmentation factors.

[0100] Step S13: Broadcast the current global model based on wireless power-carrying communication technology so that the IoT device can determine whether it participates in the current round of model training based on the target sampling probability, and if so, determine the trained model based on wireless power transmission technology, the target power segmentation factor and the current global model.

[0101] In this embodiment, combined with Figure 2 As shown, in Phase 1, the server, based on wireless power-carrying communication technology, transmits the current global model... Broadcast to each IoT device acting as a client, the sampled client The global model is received from the base station via SWIPT. All clients can simultaneously harvest energy using the power division factor until the sampled client completes its transmission. Once the transmission is complete, energy harvesting stops for all devices, requiring no additional energy harvesting time.

[0102] It is important to understand that, in this embodiment, the process by which an IoT device determines whether to participate in the current round of model training based on the target sampling probability includes: the IoT device performing a Bernoulli trial based on the target sampling probability and the sampling indicator of the current round to determine the trial result; and determining whether to participate in the current round of model training based on the trial result. That is, client n determines its participation in the current round of model training based on its own target sampling probability. Conduct Bernoulli's experiment: Each client can independently decide whether to participate in training round r. The client should ensure that its participation rate matches the independent sampling probability provided by the server over a relatively long period.

[0103] Regarding the process of IoT devices collecting energy based on a target power segmentation factor to locally train the current global model, this embodiment includes: when the IoT device determines that it is participating in the model training of the current round, it determines that it belongs to the sampled IoT devices of the current round; it receives the current global model through the sampled IoT devices and based on wireless power transmission technology; during the process of receiving the current global model, it collects energy through the sampled IoT devices and based on wireless power transmission technology and the target power segmentation factor, until all the sampled IoT devices in the current round have completed the global model reception operation, and then determines the energy collection result.

[0104] Subsequently, in Phase Two, the IoT device locally trains a global model based on the energy harvesting results. Specifically, the sampled IoT device determines whether the energy constraints are met based on the energy harvesting results, thus establishing a constraint determination result. If the constraint determination result is yes, the sampled IoT device decodes the received current global model. Upon successful decoding, it uses the corresponding model decoding result and the energy harvesting results to train the model and update the gradient, determining the trained model and the updated local stochastic gradient. In Phase Two, the sampled client that successfully decodes the global model uses the harvested energy (including existing energy) to perform a local model update and uploads it back to the base station for aggregation, while other devices remain idle. If the sampled client n meets the energy constraints... ( (Given the energy consumed by client n in the previous round), then: Update gradient: Client n trains the local model using local data, performs multiple local iterations, and calculates the local stochastic gradient. And send the local gradient to the server: Client n uploads the local gradient to the server.

[0105] In this way, introducing client sampling into wireless power federated learning benefits both phases. In phase one, it is only necessary to ensure that the sampled clients have enough energy to participate in the next round of training, rather than waiting for all devices to collect enough energy as with full client participation. In phase two, client sampling allows faster devices to participate more and slower devices to participate less, effectively solving the lag problem and reducing the clock time required for federated learning convergence.

[0106] Step S14: Receive the trained models uploaded by each of the IoT devices, and determine the updated current global model by combining the preset adaptive aggregation rules.

[0107] In this embodiment, the server receives the trained model and the updated local stochastic gradient uploaded by each of the IoT devices. Based on the updated local stochastic gradient, the trained model, the target sampling probability, the IoT device data distribution weights, and the learning rate, the server performs model aggregation to determine the updated current global model. Specifically, the server uses adaptive aggregation rules to update the global model to determine the updated current global model. :

[0108] .

[0109] This aggregation method ensures that even when different clients are sampled with different probabilities, the aggregation model remains unbiased relative to full client participation.

[0110] Step S15: Based on the updated current global model, jump back to the step of broadcasting the current global model based on wireless power-carrying communication technology until the preset task termination condition is met, and determine the target global model corresponding to the federated learning task.

[0111] In this embodiment, after completing the global model update, the server will aggregate the model. The broadcast is sent to all clients. Phases one and two are repeated until the global model converges to the preset error accuracy or the total system time budget is reached. .

[0112] Furthermore, it is understood that the aforementioned wireless power federated learning process can be applied to IoT scenarios related to industries such as security, industry, and healthcare. After obtaining the target global model, the server distributes the target global model to each IoT device, enabling each IoT device to use the target global model for intelligent analysis, device monitoring, diagnostic assistance, or other operations. For example, in the first specific implementation, for a smart security scenario, the aforementioned target global model can be a facial recognition model or a human abnormal behavior detection model. Each IoT device can use the obtained target global model and combine it with a camera device to perform facial recognition or abnormal behavior detection, without needing to upload video data to the server; real-time alarm requirements can be met locally. In the second specific implementation, for an industrial IoT scenario, the aforementioned target global model can be an industrial equipment fault detection model. Each IoT device can use the obtained target global model to predict industrial equipment faults (such as through data such as temperature and vibration data), achieving predictive maintenance. In the third specific implementation, for the medical IoT scenario, the aforementioned target global model can be a medical diagnostic model. The IoT devices in each hospital can train the diagnostic model locally and combine it with the target global model obtained by federated learning to help improve the accuracy of the diagnostic model's disease prediction while protecting patient privacy.

[0113] In summary, the proposed scheme in this embodiment adopts a two-stage design: Stage 1 (global model broadcasting and SWIPT) and Stage 2 (local model training and uploading). This two-stage design effectively solves the problem of existing WPT schemes requiring additional energy collection time, while also addressing the issue of lagging behind through client sampling. Furthermore, client sampling is achieved by the server providing independent sampling probabilities to the client. This independent sampling mechanism allows the client to choose not to participate in the federated learning round if it has a higher priority task or is disconnected due to poor wireless network signal. This effectively copes with the fluctuating environment of the federated learning system.

[0114] Therefore, in this embodiment, the server first jointly optimizes the sampling probability and power segmentation factor of IoT devices. Then, the obtained target sampling probability and target power segmentation factor are distributed to the corresponding IoT devices. Next, the current global model is broadcast using wireless power transmission technology, allowing IoT devices to determine whether to participate in the current round of model training based on the target sampling probability. If so, the trained model is determined based on the wireless power transmission technology, the target power segmentation factor, and the current global model. Then, the received trained models are adaptively aggregated to determine the updated current global model. This updated current global model is then distributed again until a preset task termination condition is met, at which point the target global model corresponding to the federated learning task is determined. This effectively achieves wireless power federated learning and, by jointly optimizing the sampling probability and power segmentation factor of IoT devices, ensures the sustainability of the entire wireless power federated learning process and improves learning efficiency, thereby minimizing convergence error within a given time budget.

[0115] See Figure 3 As shown in the illustration, this application also discloses a wirelessly powered federated learning device based on wireless power-carrying communication, applied to a server, comprising:

[0116] The parameter joint optimization module 11 is used to jointly optimize the sampling parameters of IoT devices based on the task parameters corresponding to the federated learning task, so as to determine the joint optimization result; the task parameters include the total federated learning time threshold; the joint optimization result includes the target sampling probability and target power segmentation factor corresponding to each IoT device, the target sampling probability is used to control the training participation frequency, and the target power segmentation factor is used to control the energy harvesting amount;

[0117] The parameter distribution module 12 is used to distribute the joint optimization results to the corresponding IoT devices respectively;

[0118] The model broadcasting module 13 is used to broadcast the current global model based on wireless power-carrying communication technology, so that the IoT device can determine whether it participates in the current round of model training based on the target sampling probability, and when it does, determine the trained model based on wireless power transmission technology, the target power segmentation factor and the current global model.

[0119] The model aggregation module 14 is used to receive the trained models uploaded by each of the IoT devices and determine the updated current global model by combining the preset adaptive aggregation rules.

[0120] The step jump module 15 is used to jump back to the step of broadcasting the current global model based on the updated current global model, until the preset task termination condition is met, and then determine the target global model corresponding to the federated learning task.

[0121] In some specific embodiments, the parameter joint optimization module 11 may specifically include:

[0122] The information sending unit is used to send the initial model and preset parameters to each IoT device based on the wireless power-carrying communication technology and the task parameters corresponding to the federated learning task.

[0123] The parameter receiving unit is used to receive and save the parameter information reported by each of the IoT devices; the parameter information includes the calculation task time, communication rate, communication bandwidth, data volume, and wireless channel gain.

[0124] The model building unit is used to build an IoT device energy harvesting model, an IoT device communication model, an IoT device energy consumption model, and an IoT device round-time model based on the parameter information, so as to determine the model building results;

[0125] The constraint construction unit is used to construct the constraint relationship between the sampling probability of the Internet of Things device and the power segmentation factor based on the parameter information.

[0126] A convergence analysis model construction unit is used to construct a convergence analysis model based on the constraint relationship; the convergence analysis model is used to perform convergence analysis of the global model.

[0127] The problem construction unit is used to construct a joint approximate optimization problem of IoT device sampling probability and power segmentation factor based on the convergence analysis model, the constraint relationship, the total federated learning time threshold, the approximation algorithm and the model construction results;

[0128] The joint optimization result determination unit is used to determine the joint optimization result corresponding to each IoT device based on the joint approximate optimization problem, the preset iterative optimization algorithm, and the convex optimization toolbox.

[0129] In some specific embodiments, the model building unit may specifically include:

[0130] The first model construction subunit is used to construct an energy harvesting model for IoT devices based on the wireless channel gain, power division factor, average server transmit power, energy conversion efficiency, and energy harvesting time; the energy conversion efficiency is the efficiency of converting the harvested signal energy into stored energy.

[0131] The second model construction subunit is used to construct an IoT device receiving rate model based on the power division factor, the wireless channel gain, noise power, and IoT device bandwidth allocation information; the IoT device receiving rate model is used to characterize the data rate of the IoT device receiving the global model.

[0132] The third model construction subunit is used to construct a packet error rate model based on the wireless channel gain, noise power, and the server's average transmit power; the packet error rate model is used to characterize the packet error rate from the server to the IoT device;

[0133] The fourth model construction subunit is used to construct global model decoding variables based on the grouping error rate model;

[0134] The fifth model construction subunit is used to determine the IoT device communication model based on the IoT device receiving rate model and the global model decoding variables;

[0135] The sixth model building subunit is used to build an energy consumption model for IoT devices based on the transmission power, transmission time, and energy consumption of model training.

[0136] The seventh model construction subunit is used to construct the IoT device round-time model based on the IoT device sampling probability and the grouping error rate model.

[0137] In some specific embodiments, the constraint construction unit may specifically include:

[0138] The constraint relationship construction subunit is used to construct the constraint relationship between the sampling probability of IoT devices and the power segmentation factor based on the parameter information, the grouping error rate model, the energy conversion efficiency, the average transmit power of the server, and the energy consumption model of the IoT devices.

[0139] In some specific embodiments, the model broadcasting module 13 may specifically include:

[0140] The test result determination unit is used by the Internet of Things device to perform a Bernoulli trial based on the target sampling probability and the sampling indicator of the current round to determine the test result;

[0141] The training participation judgment unit is used to determine whether it participates in the current round of model training based on the test results.

[0142] In some specific embodiments, the model broadcasting module 13 may specifically include:

[0143] The judgment result processing unit is used to determine that the IoT device belongs to the sampled IoT device in the current round when the IoT device determines that it is participating in the current round of model training;

[0144] A global model receiving unit is used to receive the current global model through the sampled IoT device and based on wireless power-carrying communication technology;

[0145] An energy harvesting model unit is used to harvest energy through the sampled IoT devices and based on wireless power transfer technology and the target power division factor during the process of receiving the current global model, until all the sampled IoT devices in the current round have completed the global model reception operation, and then determine the energy harvesting result.

[0146] An energy constraint judgment unit is used to determine whether the energy constraint conditions are met based on the sampled IoT device and the energy harvesting results, so as to determine the constraint judgment result.

[0147] The model training unit is used to decode the received current global model through the sampled IoT device if the constraint judgment result is yes, and when the decoding is successful, to perform model training and gradient update using the corresponding model decoding result and the energy harvesting result, so as to determine the trained model and the updated local stochastic gradient.

[0148] In some specific embodiments, the model aggregation module 14 may specifically include:

[0149] The post-trained model receiving unit is used to receive the post-trained model and the updated local stochastic gradient uploaded by each of the IoT devices;

[0150] The model aggregation unit is used to perform model aggregation based on the updated local stochastic gradient, the trained model, the target sampling probability, the IoT device data distribution weights, and the learning rate, so as to determine the updated current global model.

[0151] Furthermore, embodiments of this application also disclose an electronic device, Figure 4This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0152] Figure 4 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the wireless power supply federated learning method based on wireless power-carrying communication disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0153] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0154] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0155] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the wireless power supply federated learning method based on wireless power-carrying communication executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0156] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed wireless power supply federated learning method based on wireless power-carrying communication. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0157] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0158] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0159] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0160] Finally, 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.

[0161] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A wireless power-enabled federated learning method based on wireless power communication, characterized in that, Applied to a server, comprising: Based on the task parameters corresponding to the federated learning task, the joint optimization of the sampling parameters of the Internet of Things devices is carried out to determine the joint optimization result; the task parameters include a total federated learning time threshold; the joint optimization result includes target sampling probabilities and target power split factors corresponding to each Internet of Things device respectively, the target sampling probability is used to control the training participation frequency, and the target power split factor is used to control the energy collection amount; The joint optimization result is respectively sent to the corresponding Internet of Things device; Based on the wireless power communication technology, the current global model is broadcasted, so that the Internet of Things device judges whether to participate in the current round of model training based on the target sampling probability, and when yes, determines the post-training model based on the wireless power transmission technology, the target power split factor and the current global model; The post-training model uploaded by each Internet of Things device is received, and an updated current global model is determined in combination with a preset adaptive aggregation rule; Based on the updated current global model, the step of broadcasting the current global model based on the wireless power communication technology is re-jumped until a preset task termination condition is met, and a target global model corresponding to the federated learning task is determined.

2. The wireless-power-federated-learning method based on wireless power communication according to claim 1, wherein, The joint optimization of the sampling parameters of the Internet of Things devices based on the task parameters corresponding to the federated learning task to determine the joint optimization result, comprising: Based on the wireless power communication technology and the task parameters corresponding to the federated learning task, an initial model and a preset parameter are sent to each Internet of Things device; Receive and save the parameter information reported by each Internet of Things device; the parameter information includes calculation task time consumption, communication rate, communication bandwidth, data volume and wireless channel gain; Based on the parameter information, the Internet of Things device energy collection model, the Internet of Things device communication model, the Internet of Things device energy consumption model and the Internet of Things device round time model are constructed to determine the model construction result; Based on the parameter information, the constraint relationship between the sampling probability of the Internet of Things device and the power split factor is constructed; Based on the constraint relationship, a convergence analysis model is constructed; the convergence analysis model is used for convergence analysis of the global model; Based on the convergence analysis model, the constraint relationship, the total federated learning time threshold, the approximate algorithm and the model construction result, a joint approximate optimization problem of the sampling probability of the Internet of Things device and the power split factor is constructed; Based on the joint approximate optimization problem, a preset iterative optimization algorithm and a convex optimization toolbox, the joint optimization result corresponding to each Internet of Things device is determined.

3. The wireless-power-federated-learning method based on wireless power communication according to claim 2, wherein, Based on the parameter information, the Internet of Things device energy collection model, the Internet of Things device communication model, the Internet of Things device energy consumption model and the Internet of Things device round time model are constructed, comprising: Based on the wireless channel gain, the power split factor, the server average transmission power, the energy conversion efficiency and the energy collection time, the Internet of Things device energy collection model is constructed; the energy conversion efficiency is the efficiency of converting collected signal energy into stored energy; constructing an Internet of Things device receiving rate model based on the power splitting factor, the wireless channel gain, the noise power, and Internet of Things device bandwidth allocation information; the Internet of Things device receiving rate model being used to represent a data rate of the Internet of Things device receiving global model; constructing a packet error rate model based on the wireless channel gain, the noise power, and the average transmission power of the server; the packet error rate model being used to represent a packet error rate from the server to the Internet of Things device; constructing a global model decoding variable based on the packet error rate model; determining an Internet of Things device communication model based on the Internet of Things device receiving rate model and the global model decoding variable; constructing an Internet of Things device energy consumption model based on Internet of Things device transmission power, Internet of Things device transmission time, and model training energy consumption; constructing an Internet of Things device round time model based on Internet of Things device sampling probability and the packet error rate model.

4. The wireless power-based federated learning method of claim 3, wherein, The constraint relationship between the Internet of Things device sampling probability and the power splitting factor is constructed based on the parameter information, the packet error rate model, the energy conversion efficiency, the average transmission power of the server, and the Internet of Things device energy consumption model. The Internet of Things device determines whether to participate in the current round of model training based on the target sampling probability, which includes:

5. The wireless power-based federated learning method of wireless power communication according to claim 1, wherein, The Internet of Things device performs a Bernoulli trial based on the target sampling probability and the sampling indicator of the current round to determine a trial result. The Internet of Things device determines whether to participate in the current round of model training based on the trial result. When the answer is yes, the trained model is determined based on the wireless power transmission technology, the target power splitting factor, and the current global model, which includes:

6. The wireless-power-enabled federated learning method based on wireless power communication according to any one of claims 1 to 5, characterized in that, When the Internet of Things device determines that it participates in the current round of model training, it is determined that the Internet of Things device belongs to the sampled Internet of Things device of the current round. The sampled Internet of Things device receives the current global model based on the wireless power transmission technology and the target power splitting factor. During the reception of the current global model, the sampled Internet of Things device collects energy based on the wireless power transmission technology and the target power splitting factor until all the sampled Internet of Things devices of the current round complete the global model reception operation, and then determines an energy collection result. The sampled Internet of Things device determines whether the energy constraint condition is met based on the energy collection result to determine a constraint judgment result. If the constraint judgment result is yes, the sampled Internet of Things device decodes the received current global model, and when the decoding is successful, the model training and gradient update are performed using the corresponding model decoding result and the energy collection result to determine the trained model and the updated local random gradient. The received trained models of all the Internet of Things devices are combined with a preset adaptive aggregation rule to determine an updated current global model, which includes:

7. The wireless-power-federated-learning method based on wireless power communication according to claim 6, wherein, The received trained models of all the Internet of Things devices and the updated local random gradient are received. ​ Based on the updated local stochastic gradient, the trained model, the target sampling probability, the Internet of Things device data distribution weight, and the learning rate, model aggregation is performed to determine an updated current global model.

8. A wireless power-enabled federated learning device based on wireless power-communication, characterized in that, Applied to a server, comprising: A parameter joint optimization module is configured to perform joint optimization of sampling parameters of Internet of Things devices based on task parameters corresponding to a federated learning task to determine a joint optimization result. The task parameters include a total federated learning time threshold. The joint optimization result includes target sampling probabilities and target power split factors corresponding to the Internet of Things devices, respectively. The target sampling probabilities are used to control training participation frequencies, and the target power split factors are used to control energy collection amounts. A parameter issuing module is configured to issue the joint optimization result to the corresponding Internet of Things devices, respectively. A model broadcasting module is configured to broadcast a current global model based on wireless power communication technology, so that the Internet of Things devices determine whether to participate in model training of a current round based on the target sampling probabilities, and when yes, determine a trained model based on wireless power transmission technology, the target power split factor, and the current global model. A model aggregation module is configured to receive the trained models uploaded by the Internet of Things devices and determine an updated current global model in combination with a preset adaptive aggregation rule. A step jump module is configured to jump back to the step of broadcasting the current global model based on wireless power communication technology based on the updated current global model until a target global model corresponding to the federated learning task is determined when a preset task termination condition is met.

9. An electronic device, comprising: Comprise: A memory for saving a computer program; A processor for executing the computer program to implement the wireless power-enabled federated learning method based on wireless power communication according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program for saving, which is executed by a processor to implement the wireless power-enabled federated learning method based on wireless power communication according to any one of claims 1 to 7.

Images