An end-to-end cloud collaborative layered federated learning training method, device and storage medium
By employing dynamic clustering and knowledge distillation mechanisms, the problems of synchronization blockage and low resource utilization in heterogeneous resource environments of edge, cloud, and device are solved, enabling efficient and adaptive collaborative training and improving training efficiency and model performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG POLYTECHNIC NORMAL UNIV
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-05
AI Technical Summary
Existing hierarchical federated learning methods suffer from problems such as synchronization blocking, low resource utilization, and cross-layer parameter propagation bias in heterogeneous resource environments such as edge-cloud, making it difficult to achieve efficient and adaptive collaborative training.
By using dynamic clustering terminal devices, we design differentiated computation and communication compression rates, and combine them with teacherless online knowledge distillation and dynamic time threshold scheduling to achieve edge-cloud collaborative training.
It significantly reduces training latency, improves resource utilization and model performance, enhances training stability and consistency, and is suitable for heterogeneous resource environments.
Smart Images
Figure CN122154974A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of distributed machine learning and edge computing technology, and more specifically, to a collaborative hierarchical federated learning training method between edge, cloud, and end devices. Background Technology
[0002] Federated learning, as an emerging distributed machine learning paradigm, is based on the core idea of multiple participants, or clients, collaboratively training a global machine learning model without exchanging raw local data. Participants only upload model updates, such as gradients or parameters, to a central server, such as a cloud server, thereby effectively protecting data privacy and reducing the need for large-scale data transfer.
[0003] To further reduce the communication pressure on the core network and utilize the computing resources at the network edge, hierarchical federated learning has been proposed. This architecture introduces an edge server layer between the terminal device and the cloud server, forming a three-layer structure of end-edge-cloud. The terminal device first uploads its local model updates to its associated edge server for intermediate aggregation, generating edge sub-models; subsequently, the edge server uploads the aggregated model to the cloud server for global aggregation. This mechanism significantly reduces the amount of direct communication between the terminal and the cloud, alleviating network bandwidth pressure.
[0004] However, existing hierarchical federated learning methods still face many challenges in practical deployment. First, due to the significant heterogeneity in computing power, communication capabilities, and resource stability among edge, cloud, and end-device devices, traditional synchronous training mechanisms are limited by the slowest device or node—the "fallback problem"—leading to increased overall system latency and low training efficiency. Second, existing research often treats edge servers as passive model aggregation and forwarding nodes, failing to fully utilize their considerable computing power to assist model optimization, thus limiting the overall computing power of the system. Third, existing lightweight methods for resource-constrained devices, such as model pruning and quantization, mostly fail to differentiate and model the computing and communication capabilities of devices, instead employing a uniform compression strategy, making it difficult to achieve an optimal balance between latency and accuracy. Finally, the lack of an efficient collaborative training mechanism between the edge and cloud layers can lead to deviations in model parameters during cross-layer transmission, affecting the convergence speed and final performance of the global model.
[0005] Existing technologies, for example, primarily focus on reducing communication rounds by introducing an edge layer, but have not effectively addressed the synchronization blocking and resource utilization issues caused by heterogeneity. Other works attempt to cluster clients, but this is typically done in a single-layer client-cloud architecture, and the clustering criteria are mostly based on data distribution rather than device resource status, and are not deeply integrated with model lightweighting strategies. Regarding edge-cloud collaboration, there is a lack of efficient knowledge exchange mechanisms that can effectively reduce communication volume while maintaining model consistency.
[0006] Therefore, there is an urgent need for a new hierarchical federated learning method that can achieve efficient and adaptive collaborative training in heterogeneous resource environments such as edge, cloud, and endpoint. Summary of the Invention
[0007] In view of this, this application provides an edge-cloud collaborative hierarchical federated learning training method to solve or alleviate the technical problems existing in the prior art, and at least provides a beneficial option.
[0008] The technical solution of this application embodiment is a layered federated learning training method for edge-cloud collaboration, applied to a system including a cloud server, multiple edge servers, and multiple terminal devices under each edge server. The method includes: Step S1: For each round of global training, the cloud server generates a global model and distributes it to each edge server; Step S2: Each edge server performs dynamic clustering based on the computing and communication capabilities of its subordinate terminal devices, dividing the terminal devices into different clusters, and determining the computing compression ratio and communication compression ratio for each cluster; based on the computing compression ratio and communication compression ratio, the global model is processed to obtain a lightweight model adapted to different clusters and distributed to the terminal devices in the corresponding clusters. Step S3: Each terminal device uses its local dataset to train the received lightweight model locally and uploads the updated model parameters to its respective edge server. Step S4: Online knowledge distillation without a teacher is performed between the edge server and the cloud server based on a shared dataset to update the edge model and the cloud model; the edge server performs local aggregation of the model parameters received from its subordinate terminal devices to update the edge model. Step S5: The cloud server periodically triggers a global aggregation operation based on a dynamic time threshold scheduling strategy to aggregate the cloud model and each edge model, generating a new global model.
[0009] A further preferred embodiment: In step S2, the dynamic clustering process includes: constructing resource status characteristics from the computing power and communication capabilities of the terminal devices; performing clustering analysis on the resource status characteristics using a clustering algorithm, and grouping terminal devices with similar resource statuses into the same cluster.
[0010] Further preferred: In step S2, the process of determining the computational compression ratio and communication compression ratio for each cluster is as follows: with the goal of minimizing the difference between the estimated training latency and the baseline latency of the terminal devices within the cluster, and with the constraint that the model accuracy loss does not exceed a preset threshold, the optimal computational compression ratio and communication compression ratio are solved for each cluster through an optimization algorithm; wherein, the estimated training latency is calculated based on the computational compression ratio, communication compression ratio, and the computing and communication capabilities of the terminal devices, and the baseline latency is the training latency of the terminal device with the most abundant resources when no model compression is performed.
[0011] Further preferred: The optimization algorithm is a multi-objective optimization algorithm, used to solve the computational compression ratio and communication compression ratio under the constraints, and to map the model pruning ratio and quantization bit width to the implementation methods of the computational compression ratio and communication compression ratio.
[0012] A further preferred embodiment: In step S4, the process of performing teacherless online knowledge distillation based on the shared dataset includes: The cloud server uses a shared dataset to train a cloud model and generate a first prediction distribution, which is then sent to the edge server. The edge server trains an edge model using a shared dataset and generates a second prediction distribution, which is then sent to the cloud server. The cloud server updates the cloud model based on its own cross-entropy loss and the KL divergence loss between the first and second prediction distributions. The edge server updates the edge model based on its own cross-entropy loss and the KL divergence loss between the second prediction distribution and the first prediction distribution.
[0013] A further preferred embodiment: In step S5, the dynamic time threshold scheduling strategy determines the global aggregation time interval by solving the following optimization problem: Minimize the total resource waste rate caused by incomplete training across all edge servers; The constraint is that the global aggregation time interval is not less than the maximum latency required to complete one round of local training on any edge server.
[0014] A further preferred embodiment: In step S3, before uploading the updated model parameters, the terminal device compresses the model parameters according to the quantization strategy corresponding to the communication compression rate of its cluster.
[0015] A further preferred embodiment: In step S5, the global aggregation operation specifically involves: performing a weighted average of the parameters of the cloud model and all edge models to generate a new global model.
[0016] Based on the same concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the edge-cloud collaborative hierarchical federated learning training method as described above.
[0017] Based on the same concept, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the edge-cloud collaborative hierarchical federated learning training method as described above.
[0018] Compared with the prior art, this application has the following significant advantages: First, by dynamically clustering based on the computing and communication capabilities of terminal devices and implementing differentiated model lightweighting strategies, the synchronous blocking problem caused by "falling behind" in heterogeneous environments is effectively alleviated, significantly reducing single-round training latency and improving overall training efficiency. Second, by introducing a bidirectional knowledge distillation mechanism at the edge-cloud level, the computing power of edge servers is fully utilized for collaborative training and knowledge transfer, improving the overall computing power utilization and model performance of the system. Third, the dynamic compression strategy can adaptively balance computing and communication overhead while ensuring model accuracy, and the dynamic time threshold aggregation mechanism based on resource status further enhances the robustness and convergence stability of the system.
[0019] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of this application will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is the overall flowchart of the edge-cloud collaborative hierarchical federated learning training method in this application. Detailed Implementation
[0022] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of this application. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.
[0023] The embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0024] refer to Figure 1 A collaborative hierarchical federated learning training method for edge-cloud systems is proposed, applicable to systems including cloud servers, multiple edge servers, and multiple terminal devices under each edge server. The method includes: Step S1: For each round of global training, the cloud server generates a global model and distributes it to each edge server; Step S2: Each edge server performs dynamic clustering based on the computing and communication capabilities of its subordinate terminal devices, divides the terminal devices into different clusters, and determines the computing compression ratio and communication compression ratio for each cluster; based on the computing compression ratio and communication compression ratio, the global model is processed to obtain a lightweight model adapted to different clusters and distributed to the terminal devices in the corresponding clusters. Step S3: Each terminal device uses its local dataset to train the received lightweight model locally and uploads the updated model parameters to its respective edge server. Step S4: Online knowledge distillation without a teacher is performed between the edge server and the cloud server based on a shared dataset to update the edge model and the cloud model; the edge server performs local aggregation of the model parameters received from its subordinate terminal devices to update the edge model. Step S5: The cloud server periodically triggers a global aggregation operation based on a dynamic time threshold scheduling strategy to aggregate the cloud model and each edge model, generating a new global model.
[0025] The steps described above will be explained in detail below with reference to specific embodiments.
[0026] Further preferred: In step S2, the dynamic clustering process includes: constructing resource status characteristics from the computing power and communication capabilities of the terminal devices; performing clustering analysis on the resource status characteristics using a clustering algorithm, and grouping terminal devices with similar resource statuses into the same cluster.
[0027] Further optimization: In step S2, the process of determining the computational compression ratio and communication compression ratio for each cluster is as follows: with the goal of minimizing the difference between the estimated training latency and the baseline latency of the terminal devices within the cluster, and with the constraint that the model accuracy loss does not exceed a preset threshold, the optimal computational compression ratio and communication compression ratio are solved for each cluster through an optimization algorithm; wherein, the estimated training latency is calculated based on the computational compression ratio, communication compression ratio, and the computing and communication capabilities of the terminal devices, and the baseline latency is the training latency of the terminal device with the most abundant resources when no model compression is performed.
[0028] Further preferred: The optimization algorithm is a multi-objective optimization algorithm, used to solve the computational compression ratio and communication compression ratio under the constraints, and to map the model pruning ratio and quantization bit width to the implementation methods of computational compression ratio and communication compression ratio.
[0029] Further preferred: In step S3, before uploading the updated model parameters, the terminal device compresses the model parameters according to the quantization strategy corresponding to the communication compression rate of its cluster.
[0030] In the specific implementation process, on a cloud server... An edge server and its subordinate servers. The system consists of several terminal devices. At the start of each round of global training, the cloud server generates and distributes the global model to each edge server. The edge servers dynamically cluster the models based on the computing and communication capabilities of the terminal devices, design personalized lightweight model strategies for different clusters, and distribute the original or compressed models to the terminal devices. Using local datasets Perform local training to obtain updated model parameters And upload it to the corresponding edge server either as is or after quantization. .
[0031] Meanwhile, at the edge-cloud layer, a teacherless online knowledge distillation framework is constructed, connecting edge and cloud servers based on a shared dataset. Achieve efficient collaborative training of models and update cloud models. With edge model After the edge server receives the terminal models, it performs local aggregation to obtain the updated edge model. The cloud server periodically performs global aggregation to generate a unified global model, and uses a dynamic time threshold scheduling strategy to adaptively adjust the frequency of global aggregation based on the system resource status, thereby reducing cross-layer communication while maintaining model consistency.
[0032] The client uses a tuple Indicates its computing power With communication resources The set of resource states of all clients is denoted as . This invention employs the Mean-Shift clustering algorithm to... Clients are divided into A cluster. In the cluster The Middle The resource status of each client is represented as follows: The resource status of the cluster core is recorded as follows: For the first Clients in a cluster The total latency of a single training round is defined as follows: in and These are the client and client respectively. Cluster Computational and communication compression ratio; To calculate the delay, For communication delay, respectively defined as (2) (3)
[0033] in, This refers to the size of each batch of datasets. It is the computational cost per data sample. It is a client computing power It is a round calculated locally. Indicates the model size. It is a client The transmission rate to the edge server.
[0034] In synchronous federated learning under an edge-edge architecture, the training latency is determined by the slowest client. Therefore, the local training latency under the edge server is: (4) To ensure model performance, in each round of local training, a resource-rich device trains the complete model. This paper defines the training latency of a resource-rich device as the baseline latency, expressed as: (5) The optimization objective for a single edge server is to minimize the difference between training latency and baseline latency. (6) in, For the first wheel cluster The computational compression ratio of the client-side model in the dataset. For the first wheel cluster Communication compression rate of the client model in the text. This indicates the performance loss caused by model compression. It is the maximum allowable accuracy loss threshold.
[0035] Constraints and Limitations on computational and communication compression rates, constraints Ensure the accuracy of the model.
[0036] To reduce the complexity of problem solving, clustering is used. The computing and communication resources corresponding to the cluster center Substitute these into formulas (2) and (3) for analysis. Meanwhile, to simplify the solution process of the optimization problem, the constraints in formula (6) are... Transformed into computational compression ratio and communication compression ratio Constraints. Before each training round begins, i.e. when the m-th communication round starts, the server customizes a corresponding lightweight strategy based on the current cluster resource status and task requirements. At this time, the original optimization objective function (6) can be updated to: (7) in, and These represent the thresholds for computational compression ratio and communication compression ratio, respectively. By setting differentiated compression ratio thresholds for different models, the system can reduce training latency while maintaining model accuracy within an acceptable range.
[0037] This invention employs the NSGA-II (Nondominated Sorting Genetic) genetic algorithm to solve the above problem. Each individual is represented as... Its appropriate function Defined as: in The penalty factor is used to ensure that the solution selected by the optimization algorithm satisfies the constraints in formula (7). and The algorithm approximates the Pareto optimal front through iterative evolution, thereby finding the optimal solution.
[0038] The specific algorithm flow is shown in Algorithm 1.
[0039]
[0040] The server will provide the parameters of the optimal solution. The mapping is applied to the corresponding pruning rate and quantization bit width. L1 norm pruning is then performed on the global model. Subsequently, the server broadcasts the pruned model and quantization configuration to the corresponding terminal devices within the cluster. Each terminal device in the cluster trains the received lightweight model based on its local dataset and compresses the updated local model using the assigned quantization strategy. Finally, each terminal device uploads its compressed local model to the server, which then performs an aggregation operation to obtain the updated global model.
[0041] Further optimization: In step S4, the process of performing teacherless online knowledge distillation based on the shared dataset includes: The cloud server uses a shared dataset to train a cloud model and generate a first prediction distribution, which is then sent to the edge server. The edge server trains an edge model using a shared dataset and generates a second prediction distribution, which is then sent to the cloud server. The cloud server updates the cloud model based on its own cross-entropy loss and the KL divergence loss between the first and second prediction distributions. The edge server updates the edge model based on its own cross-entropy loss and the KL divergence loss between the second and first prediction distributions.
[0042] Further preferred: In step S5, the dynamic time threshold scheduling strategy determines the global aggregation time interval by solving the following optimization problem: Minimize the total resource waste rate caused by incomplete training across all edge servers; The constraint is that the global aggregation time interval is not less than the maximum latency required to complete one round of local training on any edge server.
[0043] Further optimization: In step S5, the global aggregation operation specifically involves: weighting the parameters of the cloud model and all edge models to generate a new global model.
[0044] In the specific implementation process, the cloud server utilizes a shared dataset. Training cloud models And generate a prediction distribution on a shared dataset. , predict distribution Send to the edge server. The edge server utilizes the shared dataset. Training edge models Generate a prediction distribution on the shared dataset. The predicted distribution is sent to the cloud server. Once the cloud and edge servers receive the predicted distribution from each other, they calculate the loss function: in, and These represent the cross-entropy loss of the cloud server and the edge server, respectively. The weighting coefficient for distillation loss is used to weigh prediction error against knowledge transfer. It's the Kullback-Leibler divergence. Cloud and edge servers are based on the loss function. and Update the cloud model separately and edge model .
[0045] During the training of the edge-cloud collaborative model, the edge server collects local datasets from terminal devices. Training a local model Once the edge server has completed the edge model update and collected all client models, it performs a local aggregation update of the edge model. To ensure model consistency, the cloud server periodically performs global aggregation to generate a global model. This process continues until a set time threshold is reached. Then, the edge server sends the model parameters to the cloud server, which then performs a global aggregation of the cloud model and all edge models, i.e.: in, and The first A global round of cloud and edge models.
[0046] The setting of the global aggregation time interval has a significant impact on system performance. This invention introduces a dynamic time threshold scheduling strategy. If an edge device fails to complete the current round of training, it will result in a waste of training resources for that round. The optimization objective of this paper is to minimize the system's resource waste rate, defined as:
[0047] in It is an edge server The latency for the client to complete local training. This refers to global aggregation time. Constraints. Ensure that the global aggregation interval is no less than the single-round training delay of any edge node, and avoid model oscillation and convergence instability caused by using outdated model parameters in global updates due to some edge nodes not completing training.
[0048] Based on the same concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described edge-cloud collaborative hierarchical federated learning training method.
[0049] Based on the same concept, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described edge-cloud collaborative hierarchical federated learning training method.
[0050] The edge-cloud collaborative hierarchical federated learning training method provided by this invention effectively alleviates the synchronization blocking problem caused by "fallback" devices in heterogeneous environments by dynamically clustering based on the real-time computing and communication capabilities of terminal devices and designing differentiated model lightweighting strategies for clusters with different resource states. This mechanism enables resource-constrained devices to obtain adapted simplified models for training within an acceptable accuracy loss, thereby significantly reducing the overall latency of a single training round and improving the overall training efficiency and throughput of the system in scenarios with uneven resources.
[0051] While achieving efficient training, this invention establishes a mathematical model with minimizing latency difference as the optimization objective and accuracy loss as the constraint, and combines it with a multi-objective optimization algorithm to solve for the optimal compression ratio, thus achieving an adaptive and fine-grained balance of computing and communication resources for terminal devices. The system can intelligently determine the degree of model pruning and quantization based on the dynamic changes in device capabilities, thereby minimizing unnecessary computing and communication overhead while ensuring model performance and optimizing the overall utilization of edge-side resources.
[0052] To further enhance the efficiency of cross-layer collaboration, this invention introduces a teacherless online knowledge distillation mechanism based on a shared dataset between the edge server and the cloud server. This mechanism enables continuous knowledge transfer and collaborative optimization between the edge and cloud models during training by bidirectionally exchanging prediction distributions and fusing KL divergence loss. This not only significantly reduces the number of model parameters that need to be transmitted between the edge and cloud layers, alleviating the burden of cross-layer communication, but also effectively improves the knowledge representation ability and generalization performance of each layer's models, accelerating the convergence process of the global model.
[0053] To ensure the stability of large-scale distributed training, this invention designs a dynamic time threshold scheduling strategy to control the triggering timing of global aggregation. This strategy optimizes the aggregation interval, ensuring that global updates are only performed after the vast majority of edge nodes have completed local training. This avoids model oscillations and convergence instability caused by the use of outdated parameters due to delays in some nodes. This enhances the system's robustness to node performance fluctuations or network instability, making the training process more stable and reliable.
[0054] In summary, the "edge-cloud" three-layer collaborative training architecture constructed in this invention organically integrates dynamic lightweighting, cross-layer knowledge distillation, and intelligent aggregation scheduling to form a complete adaptive hierarchical federated learning solution. This system not only significantly improves training efficiency and resource utilization in heterogeneous resource environments but also enhances model performance and training stability. Furthermore, it naturally aligns with data privacy protection requirements, making it particularly suitable for large-scale distributed scenarios such as smart cities and the Industrial Internet of Things (IIoT) that require efficient collaborative intelligent training of massive numbers of edge devices.
[0055] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A collaborative hierarchical federated learning training method for edge-cloud systems, applied to a system including a cloud server, multiple edge servers, and multiple terminal devices under each edge server, characterized in that, The method includes: Step S1: For each round of global training, the cloud server generates a global model and distributes it to each edge server; Step S2: Each edge server performs dynamic clustering based on the computing and communication capabilities of its subordinate terminal devices, dividing the terminal devices into different clusters, and determining the computing compression ratio and communication compression ratio for each cluster; based on the computing compression ratio and communication compression ratio, the global model is processed to obtain a lightweight model adapted to different clusters and distributed to the terminal devices in the corresponding clusters. Step S3: Each terminal device uses its local dataset to train the received lightweight model locally and uploads the updated model parameters to its respective edge server. Step S4: Online knowledge distillation without a teacher is performed between the edge server and the cloud server based on a shared dataset to update the edge model and the cloud model; the edge server performs local aggregation of the model parameters received from its subordinate terminal devices to update the edge model. Step S5: The cloud server periodically triggers a global aggregation operation based on a dynamic time threshold scheduling strategy to aggregate the cloud model and each edge model, generating a new global model.
2. The edge-cloud collaborative hierarchical federated learning training method according to claim 1, characterized in that, In step S2, the dynamic clustering process includes: using the computing power and communication capabilities of terminal devices to form resource status characteristics; using a clustering algorithm to perform cluster analysis on the resource status characteristics, and grouping terminal devices with similar resource statuses into the same cluster.
3. The edge-cloud collaborative hierarchical federated learning training method according to claim 2, characterized in that, In step S2, the process of determining the computational compression ratio and communication compression ratio for each cluster is as follows: with the goal of minimizing the difference between the estimated training latency and the baseline latency of the terminal devices within the cluster, and with the constraint that the model accuracy loss does not exceed a preset threshold, the optimal computational compression ratio and communication compression ratio are solved for each cluster through an optimization algorithm; wherein, the estimated training latency is calculated based on the computational compression ratio, communication compression ratio, and the computing and communication capabilities of the terminal devices, and the baseline latency is the training latency of the terminal device with the most abundant resources when no model compression is performed.
4. The edge-cloud collaborative hierarchical federated learning training method according to claim 3, characterized in that, The optimization algorithm is a multi-objective optimization algorithm used to solve the computational compression ratio and communication compression ratio under the constraints, and to map the model pruning ratio and quantization bit width to the implementation of the computational compression ratio and communication compression ratio.
5. The edge-cloud collaborative hierarchical federated learning training method according to claim 1, characterized in that, In step S4, the process of performing teacherless online knowledge distillation based on a shared dataset includes: The cloud server uses a shared dataset to train a cloud model and generate a first prediction distribution, which is then sent to the edge server. The edge server trains an edge model using a shared dataset and generates a second prediction distribution, which is then sent to the cloud server. The cloud server updates the cloud model based on its own cross-entropy loss and the KL divergence loss between the first and second prediction distributions. The edge server updates the edge model based on its own cross-entropy loss and the KL divergence loss between the second prediction distribution and the first prediction distribution.
6. The edge-cloud collaborative hierarchical federated learning training method according to claim 1, characterized in that, In step S5, the dynamic time threshold scheduling strategy determines the global aggregation time interval by solving the following optimization problem: Minimize the total resource waste rate caused by incomplete training across all edge servers; The constraint is that the global aggregation time interval is not less than the maximum latency required to complete one round of local training on any edge server.
7. The edge-cloud collaborative hierarchical federated learning training method according to claim 1, characterized in that, In step S3, before uploading the updated model parameters, the terminal device compresses the model parameters according to the quantization strategy corresponding to the communication compression rate of its cluster.
8. The edge-cloud collaborative hierarchical federated learning training method according to any one of claims 1 to 7, characterized in that, In step S5, the global aggregation operation specifically involves: weighting the parameters of the cloud model and all edge models to generate a new global model.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the edge-cloud collaborative hierarchical federated learning training method as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the edge-cloud collaborative hierarchical federated learning training method as described in any one of claims 1 to 8.