A wireless federated learning method for large-scale internet of things collaborative intelligence

Abstract

The application discloses a wireless federated learning method for large-scale Internet of Things cooperative intelligence, aiming at the problem of heterogeneous device computing capability in a large-scale Internet of Things scene, and integrating centralized learning and federated learning to form a unified architecture, so that devices with weak computing capability can participate in global model training. On the one hand, the application determines the data sample selection strategy by the data importance of the centralized learning user, which can reduce the communication overhead and transmission time of data uploading; on the other hand, the application prunes the model of the federated learning user, which can effectively reduce the local computing time under the premise of ensuring the learning performance. The federated learning method provided by the application can realize data sample selection, model pruning and user scheduling of different types of users, which is helpful to improve the utilization rate of wireless network resources and alleviate the problem of limited resources of the Internet of Things.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a wireless federated learning method for large-scale Internet of Things collaborative intelligence. Background Technology

[0002] Leveraging machine learning-driven intelligent solutions, the communication paradigm of future wireless networks is shifting from the Internet of Things (IoT) to interconnected intelligence. While ubiquitous connectivity of massive IoT devices facilitates the provision of massive amounts of real-time data for various intelligent IoT services, how to achieve collaborative intelligence in large-scale IoT networks remains an open question.

[0003] By collecting data samples from massive numbers of IoT devices, edge platforms can leverage their powerful computing resources for centralized learning (CL) to train high-performance models. However, centralized learning frameworks suffer from high communication costs and privacy breaches in vulnerable wireless edge networks.

[0004] In the Federated Learning (FL) paradigm, edge devices do not send raw data to the edge platform; instead, they process the data locally, only transmitting model parameters (e.g., weights or gradients) to the edge platform. However, traditional federated learning frameworks require all edge devices to have sufficient computing resources to support local model training, which is unrealistic. In practice, the datasets possessed by devices with limited computing power are unlikely to participate in the training process of FL models, potentially leading to degraded model performance and low resource utilization.

[0005] Therefore, this invention aims to develop a wireless federated learning framework for collaborative intelligence in large-scale Internet of Things (IoT). By integrating centralized learning and federated learning paradigms and simultaneously considering data uploading and gradient updates, it comprehensively utilizes the advantages of both centralized and federated learning to solve the problem that heterogeneous data and device computing capabilities in large-scale IoT scenarios hinder the improvement of collaborative intelligent learning performance. Summary of the Invention

[0006] The purpose of this invention is to propose a wireless federated learning method for large-scale IoT collaborative intelligence. This method considers both data uploading and gradient updates, combining the advantages of centralized and federated learning. It enables computationally limited devices to participate in the global model training process, improving the overall network scalability. Furthermore, by incorporating learning and scheduling strategies such as network pruning, sample selection, and user selection, computational costs can be reduced while maintaining learning performance, conserving wireless resources, and better utilizing distributed data and computing resources.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This invention provides a wireless federated learning method for large-scale IoT collaborative intelligence, comprising the following steps:

[0009] S1: The edge platform requests all devices to report their information, including data volume and distribution, channel status, and computing resources; the devices report the required information to the edge platform; the edge platform generates a learning strategy based on the received information and notifies all devices of the result; the learning strategy includes device classification, sample selection, model pruning, and device scheduling.

[0010] S2: According to the learning strategy, among the centralized learning users with limited computing power, the top N' users with the best channel conditions are selected, and they are asked to select some data samples based on sample importance and upload them to the edge platform for centralized training; federated learning users with strong computing power perform network pruning and local training, obtain local gradients and upload gradient norms. The edge platform judges the importance of the model based on the gradient norms and selects the top K' federated learning users with the best models, and asks them to upload their local gradients.

[0011] S3: The edge platform aggregates centralized learning gradients and federated learning gradients to obtain global gradients, updates the global model using global gradients, and then distributes the updated global model to all devices to start the next round of model training.

[0012] S4: Repeat steps S1 to S3 until the global model converges or the maximum number of communication rounds is reached.

[0013] Furthermore, in step S1, the edge platform, based on the different hardware performance of the local devices, aggregates the entire heterogeneous user set. Divided into a computationally constrained set of centralized learning users and a set of federated learning users with strong computing power

[0014] Furthermore, in step S1, the edge platform uses the information uploaded by the device to determine the pruning rate of the federated learning user, enabling the federated learning user to prune the model while maintaining learning performance. The first optimization problem regarding the pruning rate of the federated learning user is:

[0015]

[0016]

[0017]

[0018] Where, [ρ1,ρ2,...,ρ K ]T This is the model pruning rate vector, and m is a constant related to the convergence analysis. It is the maximum pruning rate allowed for federated learning user k. This is the dataset of FL user k. It's local training latency. It is the uplink transmission delay of the local gradient. This is the maximum allowable delay;

[0019] Solve the first optimization problem regarding the pruning rate of federated learning users to obtain the optimal pruning rate for each federated learning user. The optimal pruning rate for each FL user is:

[0020]

[0021] Among them, D M It is the size of the global model. It is a dataset of federated learning user k.

[0022] It is the number of samples used by user k in the federated learning local update. It is the user's uplink transmission rate, p k This is the user's maximum transmission power. It is the uplink gain, B k It is the user's uplink transmission bandwidth, d c This represents the CPU cycles required to compute one sample on a local device, f. k It is the CPU's rotation speed per second.

[0023] Furthermore, in step S1, the process by which the edge platform uses the information uploaded by the devices to generate a centralized learning user device scheduling strategy is as follows:

[0024] The edge platform uses the channel state information uploaded by centralized learning users to calculate the channel gain of centralized learning user n:

[0025]

[0026] Among them, Λ n The link distance between the edge platform and the centralized learning user n. The path loss is given when the link reference distance is 1m, where α is the path loss exponent and κ is the Rice factor. It is a deterministic line-of-sight channel component. For Rayleigh fading channel components;

[0027] Then, the edge platform sorts the l2 norms of the channel gains of the centralized learning users in descending order, and selects the top N' users with the largest channel gains to participate in this round of global training. The descending sort of the l2 norms of the channel gains of the centralized learning users is as follows:

[0028]

[0029] Where, ||r (n) ||2 is the l2 norm of the channel gain for centralized learning users (n).

[0030] Furthermore, in step S2, during model training, the selected centralized learning user determines the sample selection strategy based on data importance as follows:

[0031] The selected centralized learning users estimate local samples based on the latest global model received at the end of the previous round. The importance is expressed as:

[0032]

[0033] Where s is the number of the current round, Sample data The input and output of the last layer model L after inputting the latest global model satisfy the following conditions:

[0034] It is a loss function;

[0035] The selected centralized learning users then sort the calculated sample importance in descending order, and based on the number of samples determined by the sample selection rate set in the learning strategy, select the data with high importance to upload to the edge platform for centralized training.

[0036] Furthermore, in step S2, during model training, the edge platform learns from the user-uploaded dataset in a centralized manner. The average gradient obtained after training for:

[0037]

[0038] in, For edge platforms, learn from datasets uploaded by user n in a centralized manner. The calculated gradient, It is used to train model parameters The objective function, The model is based on the i-th sample uploaded by user n in centralized learning. The loss function.

[0039] Furthermore, in step S2, during model training, the federated learning user prunes the local model according to the pruning rate specified by the learning strategy.

[0040] Furthermore, in step S2, during model training, the edge platform uses the gradient norm uploaded by each federated learning user to determine the importance of the model, and then determines the federated learning user device scheduling strategy as follows:

[0041] After performing model pruning, federated learning users train the model on their local dataset and obtain the locally updated model gradients as follows:

[0042]

[0043] Edge platforms utilize the l2 norm ||g of local gradients k (s)||2 measures the importance of the local model and selects the top K' federated learning users with the largest l2 norm of local gradients to participate in this round of global training;

[0044] Then, the selected federated learning users upload the updated gradients to the edge platform, which then aggregates the gradients from the federated learning users.

[0045] Furthermore, the aggregated federated learning gradient is represented as:

[0046]

[0047] in, It represents the number of samples used by user k in local updates during federated learning.

[0048] Furthermore, in step S3, the edge platform learns gradients centrally. and Federated learning gradients Perform a weighted average to obtain the global gradient. Use global gradient The process of updating the global model w and obtaining the global gradient is represented as follows:

[0049]

[0050] The update process of the global model is represented as follows:

[0051]

[0052] Where λ>0 is the learning rate.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] This invention discloses a wireless federated learning method for collaborative intelligence in large-scale Internet of Things (IoT). Addressing the issue of heterogeneous computing capabilities among devices in large-scale IoT scenarios, it integrates centralized learning and federated learning to form a unified architecture. This allows devices with weaker computing capabilities to participate in global model training, enabling all edge devices in the IoT to participate in the learning process. This enhances the availability of distributed data and mitigates the impact of heterogeneous computing capabilities of edge devices on collaborative intelligence.

[0055] On the one hand, this method determines the data sample selection strategy by centrally learning the data importance of users, which can reduce the communication overhead and transmission time of data upload; on the other hand, this method can effectively reduce local computing time while ensuring learning performance by pruning the models of federated learning users.

[0056] By applying the federated learning method provided by this invention, data sample selection, model pruning, and user scheduling for different types of users can be achieved, which helps to improve the utilization rate of wireless network resources and alleviate the problem of limited IoT resources. Attached Figure Description

[0057] 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.

[0058] Figure 1 A flowchart illustrating a wireless federated learning method for large-scale IoT collaborative intelligence provided in an embodiment of the present invention.

[0059] Figure 2 This is an architecture diagram of a wireless federated learning method for large-scale IoT collaborative intelligence provided in an embodiment of the present invention.

[0060] Figure 3 This is a learning performance curve obtained from computer simulation provided in an embodiment of the present invention. Detailed Implementation

[0061] 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.

[0062] This invention proposes a wireless federated learning (Semi-Federated Learning, Semi-FL) method for collaborative intelligence in large-scale Internet of Things (IoT). By integrating centralized learning (CL) and federated learning (FL), this method enables devices with lower computing power in large-scale wireless IoT scenarios to participate in the training process of the global model. Furthermore, this invention utilizes learning and scheduling strategies such as network pruning, sample selection, and user selection to make fuller use of distributed and wireless resources.

[0063] Please see Figure 1 This invention provides a wireless federated learning method for large-scale IoT collaborative intelligence, which allows all users to participate in global model training by integrating CL and FL. The method includes the following steps:

[0064] S1: The edge platform requests all devices to report information such as their data volume and distribution, channel status, and computing resources. The devices report the required information to the edge platform. Based on the received information, the edge platform generates a learning strategy and notifies all devices of the results for strategy execution during subsequent model training.

[0065] The learning strategies include device classification, sample selection, model pruning, and device scheduling.

[0066] Please see Figure 2 The aforementioned large-scale IoT collaborative intelligence wireless federated learning framework consists of an edge platform, multiple computationally constrained IoT CL users, and multiple computationally sufficient IoT FL users. CL users upload raw data, enabling the edge platform to perform centralized learning on their behalf. FL users compute local gradients and upload them to the edge platform for aggregation. The edge platform includes a base station and an edge server, responsible for computing CL gradients based on the data uploaded by CL users, aggregating the local gradients uploaded by FL users to obtain FL gradients, and updating the global model by aggregating the CL and FL gradients.

[0067] Corresponding to step S1, the learning strategy includes four aspects:

[0068] Firstly, based on the different computing capabilities of local devices, the entire heterogeneous user set is... Divided into a set of CL users with weaker computing power and a set of FL users with strong computing power

[0069] Secondly, the learning strategy includes guiding CL users on the sampling ratio for selecting local samples. By applying this sampling ratio, CL users can select a portion of data from their local dataset based on sample importance to participate in the current training round.

[0070] Thirdly, the learning strategy includes guiding FL users to prune the local model at a specific pruning rate. FL users sort the absolute values ​​of the weights within each layer of the local model in descending order, and then, based on the appropriate pruning rate determined in the learning strategy, set some weights with smaller absolute values ​​to zero. This reduces the network size and computational cost of training the local model while maintaining training performance.

[0071] Specifically, the edge platform's decision regarding the FL user pruning rate aims to maintain learning performance while pruning the network. Define D M It is the size of the global model. It refers to the size of the model data after pruning. It is the pruning rate of FL user k. This is the dataset of FL user k. This represents the number of samples used by FL user k for local updates. Learning latency mainly consists of two parts: local training latency and... Where d c This represents the CPU cycles required to compute one sample on a local device, f. k The first is the CPU's rotational speed per second. The second is the uplink gradient update latency. in It is the user's uplink transmission rate, p k This is the user's maximum transmission power. It is the uplink gain, B k This refers to the user's uplink transmission bandwidth. The proposed framework employs Orthogonal Frequency Division Multiple Access (OFDMA) technology, where each user occupies one radio resource block when uploading data. Therefore, the user's uplink transmission bandwidth B... k They are the same. Based on the above definition, the optimization problem regarding the pruning rate of FL users can be expressed as:

[0072]

[0073]

[0074]

[0075] Where, [ρ1,ρ2,...,ρ K ] T This is the model pruning rate vector, and m is a constant related to the convergence analysis. It is the maximum pruning rate allowed by FL user k. This is the maximum allowable delay.

[0076] The FL user pruning rate optimization problem is a linear programming problem. Solving the first optimization problem regarding the FL user pruning rate yields the optimal pruning rate for each FL user. Specifically, the optimal pruning rate for each FL user is:

[0077]

[0078] Then, the edge platform informs the user of the optimal pruning rate to guide them in pruning the model before local training, thereby reducing the model size.

[0079] Fourthly, the learning strategy includes a user scheduling strategy for CL users. The edge platform uses the channel information uploaded by each CL user to calculate the channel gain of each CL user, and then determines the scheduling strategy based on the l2 norm of the channel gain.

[0080] Specifically, the channel gain of a wireless link can be obtained by multiplying the path loss, which depends on the transmission distance, by small-scale fading. The link distance between the edge platform and the CL user is expressed as... Large-scale fading Give, Let α be the path loss at a reference distance of 1m, and α be the path loss exponent. For small-scale fading, it is assumed that the link between the user and the edge platform follows Ricean fading. The channel gain of all links between CL users and the edge platform is expressed as:

[0081]

[0082] The edge platform will use the l2 norm ||r of the CL user channel gain. n ||2 sorted in descending order is:

[0083]

[0084] Then, the edge platform selects the top N' users with the largest channel gain to participate in this round of global training and informs all CL users.

[0085] S2: According to the learning strategy, among the computationally limited Centralized Learning (CL) users, the top N' users with the best channel conditions are selected, and they upload a portion of the data samples to the edge platform for centralized training based on sample importance; the computationally powerful Federated Learning (FL) users perform network pruning and local training, obtain local gradients, and upload the gradient norm. The edge platform judges the importance of the model based on the gradient norm, and selects the top K' FL users with the best model, and has them upload their local gradients.

[0086] During the training phase, selected CL users use the gradient upper bound norm to measure the samples based on the latest global model received at the end of the previous round of global training. The importance of the global model.

[0087] Specifically, the model has a total of L layers. Let m represent the weight matrix of the l-th layer of the model. l σ is the number of neurons in the l-th layer of the model; σ is the Lipschitz continuous activation function, β 0 =x,α l =w l (β l-1 ),β l =σ l (α l The selected CL users estimate local samples based on the latest global model received at the end of the previous round. The importance is expressed as:

[0088]

[0089] Where s represents the current round number, These are the inputs and outputs of the last layer of the network after the sample data is input into the global model. This is the loss function. As can be seen from the expression, samples with a larger upper bound norm of the gradient of the global loss with respect to the pre-activation output will have greater importance. Selected CL users, based on the calculated sample importance and the sample selection rate set by the learning strategy, select highly important data to upload to the edge platform for global training.

[0090] The edge platform receives the CL dataset The average gradient obtained after training the model for:

[0091]

[0092] in This represents the gradient of CL user n, which is computed by the edge platform. It is used to train model parameters The objective function, It is the model for the i-th sample of CL user n. The loss function.

[0093] FL user k first performs network pruning on the latest global model received at the end of the previous round based on the optimal pruning rate generated by the edge platform, and then performs local model training:

[0094]

[0095] Where s represents the current global training round. This represents the gradient of FL user k during local training. It is used to train model parameters w k The objective function of (s) It is the model with respect to the i-th sample of FL user k. The loss function, η k Let k be the learning rate of user k. Therefore, the locally updated model gradient is expressed as:

[0096]

[0097] The FL user scheduling strategy generated by the edge platform is based on the importance of the updated model to the global model update, using g k The l2 norm of (s) ||g k (s)||2 is used as the measure. Specifically, FL users will calculate ||g k (s)||2 is sent to the edge platform; the edge platform selects the one with the largest ||g k The top K' devices with a value of (s)||2 participate in this round of updates, denoted as . And inform FL users; selected FL users upload their updated model gradients g. k The edge platform aggregates the gradients received from FL users.

[0098]

[0099] S3: The edge platform aggregates the CL gradient and FL gradient to obtain the global gradient, updates the global model using the global gradient, and then distributes the updated global model to all devices to start the next round of model training.

[0100] Specifically, the edge platform utilizes CL gradients. and FL gradient The global gradient obtained by weighted averaging is:

[0101]

[0102] Accordingly, the obtained global gradient is used pass A global model update is performed, where λ > 0 is the learning rate. The updated model is then distributed to the set. All devices in the system.

[0103] S4: Repeat steps S1 to S3 until the global model converges or the maximum number of communication rounds is reached.

[0104] In computer simulation, an IoT network with 4 computationally constrained devices and 6 computationally powerful devices is considered. The learning performance of the proposed SemiFL framework is evaluated by training a multilayer perceptron on the Fashion-MNIST dataset. To verify the performance of the proposed SemiFL framework, the following three comparison schemes are considered:

[0105] Baseline 1 (CL): All devices send their datasets to the edge platform for centralized training;

[0106] Baseline 2 (FL): The model is trained in a federated manner on all computationally powerful devices;

[0107] Baseline 3 (Semi-FL without sample selection and model pruning): All CL users upload their entire datasets, and all FL users upload their entire models to the server.

[0108] Please see Figure 3 Based on the aforementioned computer simulation settings, the learning performance of various schemes under non-independent and identically distributed conditions is demonstrated. On the one hand, simulation results show that the proposed SemiFL method outperforms FL (i.e., baseline 2, 0% sample selection, 30% network pruning) in the setting without sample selection and network pruning (i.e., baseline 3, 100% sample selection, 0% network pruning), and achieves similar performance to CL (i.e., baseline 1, 100% sample selection, 0% network pruning); however, baseline 1 has higher communication overhead than the proposed SemiFL method because it transmits the local datasets of all users. On the other hand, simulation results show that although adding sample selection and network pruning strategies can further reduce the communication overhead of SemiFL, it also reduces the achievable prediction accuracy.

[0109] This invention proposes a wireless federated learning method for large-scale IoT collaborative intelligence. By integrating federated learning and centralized learning, it enables all edge devices in the IoT to participate in the learning process, enhancing the availability of distributed data and mitigating the impact of heterogeneous computing capabilities of edge devices on collaborative intelligence. Simultaneously, through dataset transmission and collaborative computing between local devices and the edge platform, the utilization of computing and storage resources in the edge network is effectively improved. In the framework proposed in this invention, the edge platform uses massive amounts of data collected from CL users to reconstruct a high-quality dataset, which can improve the heterogeneous statistical characteristics of distributed data, thereby further enhancing model performance. Furthermore, the framework proposed in this invention integrates learning and scheduling strategies such as sample selection, network pruning, and user selection, which can reduce computational costs and conserve wireless resources while maintaining learning performance.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications 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 the present invention.

Claims

1. A wireless federated learning method for large-scale Internet of Things collaborative intelligence, characterized in that, Includes the following steps: S1: The edge platform requests all devices to report their information, including data volume and distribution, channel status, and computing resources; the devices report the required information to the edge platform; the edge platform generates a learning strategy based on the received information and notifies all devices of the result; the learning strategy includes device classification, sample selection, model pruning, and device scheduling; the process by which the edge platform uses the information uploaded by the devices to generate a centralized learning user's device scheduling strategy is as follows: The edge platform uses the channel state information uploaded by centralized learning users to calculate the centralized learning user... The channel gain is: ， in, For edge platform and centralized learning users Link distance between This represents the path loss when the link reference distance is 1m. This is the path loss index. It is the Rice factor. It is a deterministic line-of-sight channel component. For Rayleigh fading channel components; Then, the edge platform will centrally learn the user channel gain. Sort the norms in descending order and select the first one. The user with the highest channel gain participates in this round of global training, centrally learning the user's channel gain. The norm is sorted in descending order as follows: ， in, For centralized learning users Channel gain Norm; S2: Based on the learning strategy, select the top [users] among computationally limited, centralized learning users. Users with optimal channel conditions are selected, and a subset of data samples are uploaded to the edge platform for centralized training based on sample importance. Federated learning users with stronger computational capabilities perform network pruning and local training, obtain local gradients, and upload their gradient norms. The edge platform determines model importance based on the gradient norms and selects the top-performing users. The federated learning user with the best model uploads its local gradients; during model training, the selected centralized learning user determines the sample selection strategy based on data importance as follows: The selected centralized learning users estimate local samples based on the latest global model received at the end of the previous round. The importance is expressed as: ， in, It is the number of the current round. Sample data The last layer of the model after inputting the latest global model The input and output satisfy , It is a loss function; Then, the selected centralized learning users sort the calculated sample importance in descending order, and select the data with high importance based on the number of samples determined by the sample selection rate set in the learning strategy, and upload it to the edge platform for centralized training. S3: The edge platform aggregates centralized learning gradients and federated learning gradients to obtain global gradients, updates the global model using global gradients, and then distributes the updated global model to all devices to start the next round of model training. S4: Repeat steps S1 to S3 until the global model converges or the maximum number of communication rounds is reached.

2. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, In step S1, the edge platform combines the entire heterogeneous user set according to the different hardware performance of the local devices. Divided into a computationally constrained set of centralized learning users and a set of federated learning users with strong computing power .

3. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, In step S1, the edge platform uses the information uploaded by the device to determine the pruning rate for federated learning users, enabling them to prune the model while maintaining learning performance. The first optimization problem regarding the pruning rate for federated learning users is: ， ， ， in, It is the model pruning rate vector. These are the constants related to convergence analysis. Federated Learning User Maximum allowable pruning rate FL user The dataset, It's local training latency. It is the uplink transmission delay of the local gradient. This is the maximum allowable delay; Solve the first optimization problem regarding the pruning rate of federated learning users to obtain the optimal pruning rate for each federated learning user. The optimal pruning rate for each federated learning user is: ， in, It is the size of the global model. Federated Learning User The dataset, Federated Learning User The number of samples used for local updates It is the user's uplink transmission rate. This is the user's maximum transmission power. It is the uplink gain. It is the user's uplink transmission bandwidth. This indicates the CPU cycles required to compute one sample on a local device. It is the CPU's rotation speed per second.

4. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, In step S2, during model training, the edge platform learns from user-uploaded datasets in a centralized manner. The average gradient obtained after training for: CL: , in, For edge platforms based on centralized learning users Uploaded dataset The calculated gradient, It is used to train model parameters The objective function, The model is about centralized learning users. The uploaded number Sample The loss function.

5. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, In step S2, during model training, the federated learning user prunes the local model according to the pruning rate specified by the learning strategy.

6. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, Step S2: During model training, the edge platform uses the gradient norm uploaded by each federated learning user to determine the importance of the model, and then determines the federated learning user device scheduling strategy as follows: After performing model pruning, federated learning users train the model on their local dataset and obtain the locally updated model gradients as follows: ， Edge platforms utilize local gradients norm We measure the importance of the local model and select the one with the largest local gradient. The front of the norm One federated learning user participated in this round of global training; Then, the selected federated learning users upload the updated gradients to the edge platform, which then aggregates the gradients from the federated learning users.

7. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, The aggregated federated learning gradient is represented as: IN: , in, Federated Learning User The number of samples used for local updates.

8. The wireless federated learning method for large-scale IoT collaborative intelligence according to claim 1, characterized in that, In step S3, the edge platform learns gradients in a centralized manner. and Federated learning gradients Perform a weighted average to obtain the global gradient. and using global gradients Update global model ; The process of obtaining the global gradient is represented as: SemiFL: , The update process of the global model is represented as follows: ， in, This is the learning rate.

Images