An intelligent internet of things federated learning method

Abstract

This invention discloses a federated learning method for intelligent IoT, comprising: dividing the global model into continuous fragments on the server side, configuring lightweight auxiliary head and tail segments to simulate contextual features; introducing an event-triggered resource monitoring and Lyapunov optimization-driven dynamic fragment adjustment mechanism; employing fragment-level knowledge distillation to achieve cross-fragment knowledge transfer, combined with split gradient descent optimization. This invention solves the problems of heavy communication and computational burdens and poor adaptability in resource-constrained scenarios of existing federated learning and fragmented federated learning technologies. This invention can reduce device storage, computation, and communication overhead, adapt to dynamic resource changes, improve model convergence speed and consistency, and performs excellently under both IID and Non-IID data distributions, making it suitable for distributed training scenarios of heterogeneous AIoT devices.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of intelligent Internet of Things (AIoT) and federated learning, specifically involving a dynamic federated learning method for AIoT devices with limited storage, computing, and communication resources, which is suitable for efficient model training in distributed scenarios. Background Technology

[0002] In traditional federated learning frameworks, the server needs to send the complete deep neural network model to the terminal device, and the parameters are aggregated after the terminal trains the model. This mode works well in scenarios with abundant device resources, but it is not suitable for AIoT devices with limited storage, computing, and communication capabilities. The transmission and training of the complete model will bring excessive communication burden and computing pressure. To solve this problem, Partition Federated Learning (SFL) has been proposed. By dividing the global model into multiple smaller pieces, the device only loads and trains a part of the model to reduce resource requirements. However, it still has two major drawbacks: First, traditional SFL adopts a serial piecewise training mode, and there is a sequential dependency between model pieces in forward and backward propagation, which leads to pipeline blockage. Furthermore, device heterogeneity will further aggravate the waiting time, and the overall efficiency is constrained by the slowest link. Second, existing SFL methods use a fixed piecewise strategy, which cannot adapt to the heterogeneity and dynamic changes of device resources, and it is difficult to balance resource utilization and training effect. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent IoT federated learning method to solve the problems of heavy communication and computing burden and poor adaptability of existing federated learning and fragmented federated learning in resource-constrained scenarios. It can reduce device storage, computing and communication overhead, adapt to dynamic resource changes, improve model convergence speed and consistency, and perform well under both IID and Non-IID data distributions. It is suitable for distributed training scenarios of heterogeneous AIoT devices.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions.

[0005] A federated learning method for intelligent IoT includes: the server partitions the global model M into continuous fragments and configures lightweight auxiliary heads and tails to simulate contextual features; an event-triggered resource monitoring and Lyapunov optimization-driven dynamic fragment adjustment mechanism is introduced; fragment-level knowledge distillation is used to achieve cross-fragment knowledge transfer, combined with split gradient descent optimization.

[0006] In the aforementioned intelligent IoT federated learning method, the server divides the global model M into contiguous fragments; the global model M is divided into C contiguous fragments. ,in, For the c-th slice, Each slice is one or more layers in the neural network, and the granularity of the slice can be flexibly adjusted according to the application scenario and device resource constraints.

[0007] The aforementioned intelligent IoT federated learning method configures lightweight auxiliary heads and tails to simulate contextual features; for any segmented region The lightweight auxiliary head and tail include an auxiliary head. and auxiliary tail ,satisfy:

[0008] ;

[0009] ;

[0010] in, For function composition operations; auxiliary header Simulate the output characteristics of the preceding slice, and assist the tail. The input characteristics of subsequent slices are simulated, and both adopt an adapter structure design to compress the number of parameters.

[0011] The aforementioned intelligent IoT federated learning method introduces an event-triggered resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism; it tracks the system state through resource queues, maturity queues, and training progress queues, constructs a weighted Lyapunov function, and minimizes the drift-penalty term to balance resource usage, training balance, and training latency.

[0012] By using model sharding and a lightweight auxiliary head-and-tail structure design, the device only needs to load a portion of the model shards to complete the entire forward inference and backward propagation, which greatly reduces the storage, memory usage and computing pressure on the device side, thus addressing the core pain point of limited resources in AIoT devices.

[0013] The aforementioned intelligent IoT federated learning method, wherein the triggered resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism, after a device is selected to participate in training, checks the difference between the current resource status and the previous upload when uploading model parameters. If the difference exceeds a preset threshold, it triggers resource status reporting, thereby reducing communication overhead while adapting to resource changes.

[0014] For equipment When a device is selected to participate in the t-th round of training and completes model upload, the server compares the current resource status with the status at the time of the last upload; the resource status at the time of n uploads is defined as... The current resource status is ; Calculate the relative variables of each resource:

[0015] ;

[0016] ;

[0017] ;

[0018] The conditions that trigger resource reporting are:

[0019] ;

[0020] in, To store relative thresholds, As a relative threshold for computing power, This is the relative threshold for communication bandwidth.

[0021] If the triggering condition is met The device will display the current resource status. Upload to the server and update the last upload status in local storage. Otherwise, the device only uploads model parameters and does not update resource status.

[0022] The aforementioned intelligent IoT federated learning method includes a dynamic partitioning adjustment mechanism comprising candidate interval generation and online dynamic adjustment steps. During candidate interval generation, all possible start and end indices are traversed to filter feasible intervals that meet storage constraints. During online dynamic adjustment, the optimal partitioning interval is selected from the candidate intervals based on evaluation metrics and allocated to the device.

[0023] By introducing a dynamic sharding adjustment mechanism, and by monitoring the status of device storage, computing power, and communication bandwidth in real time, combined with virtual queue optimization sharding allocation strategies, the heterogeneity and dynamic fluctuations of device resources can be flexibly adapted to, avoiding the limitations of fixed sharding strategies.

[0024] In the aforementioned intelligent IoT federated learning method, during the shard-level knowledge distillation, the server selects the component with the lowest maturity from the candidate auxiliary component set as the student model. The maturity is determined by both training experience and global stability.

[0025] The aforementioned intelligent IoT federated learning method, wherein the fragment-level knowledge distillation includes auxiliary head distillation and auxiliary tail distillation, wherein the auxiliary head distillation uses mean square error to measure the feature distillation loss, and the auxiliary tail distillation uses KL divergence to calculate the prediction label distillation loss.

[0026] By employing a fragmented knowledge distillation method, knowledge is transferred between fragments using an auxiliary structure, and the distillation intensity is dynamically adjusted based on maturity, thereby accelerating model convergence and enhancing global model consistency. At the same time, by minimizing training latency through a hierarchical approximate allocation algorithm, higher model accuracy and faster convergence speed can be achieved under both IID and Non-IID data distributions.

[0027] The segmented knowledge distillation introduces an adaptive distillation intensity coefficient, which is dynamically adjusted according to the maturity differences between teacher and student models and is proportional to the maturity differences.

[0028] In the aforementioned intelligent IoT federated learning method, the separate gradient descent optimization involves optimizing the entire local model using task loss to improve task performance, and optimizing a single-level distillation component and related parameters using distillation loss to enhance knowledge transfer. The two processes perform gradient descent independently.

[0029] The aforementioned federated learning method for intelligent IoT uses a federated averaging algorithm on the server side to perform block-level aggregation of the updated shard parameters uploaded by the client, thereby completing the global model update.

[0030] The beneficial effects of this invention are as follows: This invention provides an intelligent IoT federated learning method that solves the problems of heavy communication and computing burden and poor adaptability of existing federated learning and fragmented federated learning in resource-constrained scenarios.

[0031] This invention employs model sharding and a lightweight auxiliary head-and-tail structure design, allowing devices to complete forward inference and backward propagation by loading only a portion of the model shards. This significantly reduces storage, memory usage, and computational pressure on the device, addressing the core pain point of resource-constrained AIoT devices. The invention introduces a dynamic sharding adjustment mechanism. By monitoring the device's storage, computing power, and communication bandwidth in real time, and combining this with a virtual queue to optimize sharding allocation strategies, it can flexibly adapt to the heterogeneity and dynamic fluctuations of device resources, avoiding the limitations of fixed sharding strategies. This invention uses a shard-level knowledge distillation method, utilizing an auxiliary structure to transfer knowledge between shards and dynamically adjusting the distillation intensity based on maturity, accelerating model convergence and enhancing global model consistency. Simultaneously, it minimizes training latency through a hierarchical approximation allocation algorithm, achieving higher model accuracy and faster convergence speeds under both IID and Non-IID data distributions. This invention significantly reduces the amount of data transmitted by the model, lowering the communication burden. Under the same training effect, the required communication volume is significantly reduced, improving the overall efficiency of the distributed training system. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall architecture and workflow of the present invention in Example 2;

[0033] Figure 2 This is a schematic diagram showing the experimental performance comparison between the method of the present invention and the baseline method in Example 3 under the IID data distribution;

[0034] Figure 3 This is a schematic diagram showing the experimental performance comparison between the method of the present invention and the baseline method in Example 3 under the Non-IID data distribution;

[0035] Figure 4This is a schematic diagram of the results of the ablation experiment of the present invention under the IID data distribution in Example 3;

[0036] Figure 5 This is a schematic diagram of the results of the ablation experiment of the present invention in Example 3 under the Non-IID data distribution. Detailed Implementation

[0037] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0038] Example 1:

[0039] This embodiment provides an intelligent IoT federated learning method, including: the server divides the global model M into continuous fragments and configures lightweight auxiliary heads and tails to simulate context features; it introduces an event-triggered resource monitoring and Lyapunov optimization-driven dynamic fragment adjustment mechanism; it uses fragment-level knowledge distillation to achieve cross-fragment knowledge transfer, combined with split gradient descent optimization.

[0040] In the aforementioned intelligent IoT federated learning method, the server divides the global model M into contiguous fragments; the global model M is divided into C contiguous fragments. ,in, For the c-th slice, Each slice is one or more layers in the neural network, and the granularity of the slice can be flexibly adjusted according to the application scenario and device resource constraints.

[0041] The aforementioned intelligent IoT federated learning method configures lightweight auxiliary heads and tails to simulate contextual features; for any segmented region The lightweight auxiliary head and tail include an auxiliary head. and auxiliary tail ,satisfy:

[0042] ;

[0043] ;

[0044] in, For function composition operations; auxiliary header Simulate the output characteristics of the preceding slice, and assist the tail. The input characteristics of subsequent slices are simulated, and both adopt an adapter structure design to compress the number of parameters.

[0045] The aforementioned intelligent IoT federated learning method introduces an event-triggered resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism; it tracks the system state through resource queues, maturity queues, and training progress queues, constructs a weighted Lyapunov function, and minimizes the drift-penalty term to balance resource usage, training balance, and training latency.

[0046] The aforementioned intelligent IoT federated learning method, wherein the triggered resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism, after a device is selected to participate in training, checks the difference between the current resource status and the previous upload when uploading model parameters. If the difference exceeds a preset threshold, it triggers resource status reporting, thereby reducing communication overhead while adapting to resource changes.

[0047] For equipment When a device is selected to participate in the t-th round of training and completes model upload, the server compares the current resource status with the status at the time of the last upload; the resource status at the time of n uploads is defined as... The current resource status is ; Calculate the relative variables of each resource:

[0048] ;

[0049] ;

[0050] ;

[0051] The conditions that trigger resource reporting are:

[0052] ;

[0053] in, To store relative thresholds, As a relative threshold for computing power, This is the relative threshold for communication bandwidth.

[0054] If the triggering condition is met The device will display the current resource status. Upload to the server and update the last upload status in local storage. Otherwise, the device only uploads model parameters and does not update resource status.

[0055] The aforementioned intelligent IoT federated learning method includes a dynamic partitioning adjustment mechanism comprising candidate interval generation and online dynamic adjustment steps. During candidate interval generation, all possible start and end indices are traversed to filter feasible intervals that meet storage constraints. During online dynamic adjustment, the optimal partitioning interval is selected from the candidate intervals based on evaluation metrics and allocated to the device.

[0056] In the aforementioned intelligent IoT federated learning method, during the shard-level knowledge distillation, the server selects the component with the lowest maturity from the candidate auxiliary component set as the student model. The maturity is determined by both training experience and global stability.

[0057] The aforementioned intelligent IoT federated learning method, wherein the fragment-level knowledge distillation includes auxiliary head distillation and auxiliary tail distillation, wherein the auxiliary head distillation uses mean square error to measure the feature distillation loss, and the auxiliary tail distillation uses KL divergence to calculate the prediction label distillation loss.

[0058] The segmented knowledge distillation introduces an adaptive distillation intensity coefficient, which is dynamically adjusted according to the maturity differences between teacher and student models and is proportional to the maturity differences.

[0059] In the aforementioned intelligent IoT federated learning method, the separate gradient descent optimization involves optimizing the entire local model using task loss to improve task performance, and optimizing a single-level distillation component and related parameters using distillation loss to enhance knowledge transfer. The two processes perform gradient descent independently.

[0060] The aforementioned federated learning method for intelligent IoT uses a federated averaging algorithm on the server side to perform block-level aggregation of the updated shard parameters uploaded by the client, thereby completing the global model update.

[0061] Example 2:

[0062] This embodiment provides an overall architecture and workflow of the present invention, such as... Figure 1 As shown, this framework aims to address the challenges of federated learning on storage-constrained devices in AIoT environments. Through key technologies such as model sharding, auxiliary head-and-tail structures, and knowledge distillation, it achieves efficient and collaborative model training. The entire system mainly consists of a server and multiple clients, and its training process can be summarized into the following key steps.

[0063] Before training begins, the server first performs model segmentation and adds auxiliary structures (①), dividing the global model into multiple consecutive slices and adding lightweight auxiliary structures to the head and tail of the model. These auxiliary structures are designed to simulate the contextual information of the model slice sequence so that client devices can perform effective forward inference and backward propagation even if they only load a portion of the model slices. Next, in each global training round, the server performs slice selection (②), dynamically selecting a suitable model slice interval for each client based on the resource status (e.g., storage space, computing power, communication bandwidth, etc.) of the client devices participating in the training. After slice selection, the server performs slice distribution (③), distributing the selected model slices and corresponding auxiliary head and tail structures to the corresponding client devices so that the client devices can train the model locally. After receiving the model components, the client devices begin local training (④), using the local dataset to train the received model slices. During training, the client devices concatenate the model slices with the auxiliary head and tail structures to build a complete local model, approximating the behavior of the global model and ensuring the effectiveness of local training. To improve model performance and training collaboration, the framework introduces a shard-level knowledge distillation method, performing either auxiliary head or auxiliary tail distillation (⑤ and ⑥) based on server selection. Knowledge distillation guides student model learning through the teacher model, improving model performance. In server selection, the server chooses the auxiliary head or tail with the lowest maturity as the student model for distillation. In auxiliary head distillation, the client builds a teacher model, which is a sub-network of the local backbone model. The teacher model transfers knowledge to the auxiliary head student model sent by the server, improving its ability to simulate global model input features. In auxiliary tail distillation, the client similarly builds a teacher model (a sub-network of the local backbone model) and transfers knowledge to the auxiliary tail student model sent by the server, improving its ability to simulate global model output features. The distillation process achieves knowledge transfer by narrowing the output gap between the teacher and student models. Finally, after the client completes local training, shard parameters are collected and aggregated (⑦), and the updated model shard parameters are uploaded back to the server. The server receives updated parameters from each client and performs block-level aggregation using a federated averaging algorithm. After parameter aggregation is completed, the server completes one global training round and begins the next round of training. The above steps are repeated until the model converges or reaches the preset number of training rounds.

[0064] Through the above process, the overall framework of this embodiment achieves efficient federated learning on storage-constrained AIoT devices. The design of model sharding and auxiliary structures significantly reduces the storage and computational burden on the device side, the dynamic sharding adjustment mechanism enables the system to adapt to dynamic changes in device resources, and the shard-level knowledge distillation method improves model performance and training efficiency. The synergistic effect of these key technologies makes the framework an effective federated learning solution for devices with limited storage and computational resources.

[0065] Example 3:

[0066] This embodiment provides an experimental simulation based on Embodiments 1 and 2.

[0067] Step 1: Setting up the experimental platform and parameters. All experiments were conducted on a server equipped with four NVIDIA RTX 4090 GPUs. The experiment simulated a federated learning system with one server and 100 clients. To simulate data heterogeneity in a real IoT environment, the experiment used a Dirichlet distribution to control the degree of non-identical distribution (Non-IID) of the client data. Specifically, this paper sets the Dirichlet distribution parameter α to 0.5 and 1000 to simulate two data distribution scenarios: Non-IID and IID. When α = 0.5, the data distribution exhibits high heterogeneity; when α = 1000, the data distribution approaches homogeneity. To simulate system heterogeneity, the experiment randomly assigned different CPU frequencies (0.6 to 2.5 GHz), communication capabilities (1 to 50 Mbps), and storage capabilities (20 to 120 MB) to each client device to reflect the differences in computing, communication, and storage resources of real IoT devices. To simulate the dynamic changes in client resources, at the start of each round of global training, the computing power, communication bandwidth, and storage capacity of each client are randomly perturbed based on the values ​​from the previous round. The perturbation amplitude is sampled from a uniform distribution U(−10%, +10%), ensuring that the changed values ​​do not exceed the upper and lower bounds of the initial range. In each round of global training, the server randomly selects 10 client devices to participate in the training.

[0068] The second step involves splitting the dataset and selecting the model, including using the CIFAR-10 dataset for image classification. This dataset contains 60,000 32×32 color images across 10 categories, with 50,000 used for training and 10,000 for testing. The experiment uses ResNet50 as the global model, with a total size of 94.11 MB. To support model partitioning, this paper references the AdaptiveNet partitioning method, dividing the ResNet50 model into multiple consecutive partitions and adding auxiliary head and tail structures to each partition.

[0069] Step 3: To comprehensively evaluate the performance of the method of this invention, the following representative baseline methods were selected for comparison:

[0070] (1) FedAvg-Drop: This method is an improved strategy based on FedAvg. It allocates a complete model for training to each device. When the storage requirements of the complete model exceed the storage capacity of the device, the device does not participate in the training of this round. This baseline represents the performance of traditional federated learning methods in storage-constrained scenarios.

[0071] (2) MultiSplitFed: This method is a multi-shard extension of the original SplitFed method. The original SplitFed divides the model into server-side and device-side fragments. For fair comparison, this paper safely extends SplitFed to a multi-shard version. In each round, a subset of devices is randomly selected, and model fragments are allocated sequentially from front to back. The next device continues allocation from the position where the previous device ended, until model fragments are allocated to all devices participating in training in this round. This baseline represents the performance of traditional sharded federated learning methods in resource-heterogeneous environments.

[0072] (3) FedSHTD-Random: This is a variant of the method that randomly selects sharding intervals for selected devices. When the storage demand corresponding to the randomly selected interval exceeds the storage capacity of the device, the device does not participate in the training round. This baseline aims to evaluate the importance of dynamic sharding selection strategies.

[0073] (4) FedSHTD-Greedy: This method is also a variant. It first randomly selects the starting position of the fragment interval, and then greedily selects the largest fragment interval that meets the device's storage requirements. Compared with FedSHTD-Random, FedSHTD-Greedy attempts to maximize the fragment size while meeting storage constraints, but still lacks consideration for system performance optimization.

[0074] Step 4: Determine the evaluation metrics for the experiment. The classification accuracy on the test set is used as the primary metric to evaluate the model's performance. To comprehensively evaluate the efficiency of different methods, the experiment also recorded the global epochs, clock time, and communication volume during the training process.

[0075] The results of this experiment are presented in two parts: first, the performance of the present invention is compared with that of baseline methods (FedAvg-Drop and MultiSplitFed) to demonstrate its overall superiority; then, the contribution of the dynamic splitting strategy is verified through ablation experiments and its variants (FedSHTD-Random and FedSHTD-Greedy).

[0076] like Figure 2 and Figure 3 As shown, the performance comparison results of the proposed method with the baseline methods FedAvg-Drop and MultiSplitFed are presented under IID and Non-IID data distributions. Each set of figures contains three sub-figures, showing the curves of test accuracy as a function of global rounds, clock time, and communication volume.

[0077] like Figure 2 As shown, under the IID data distribution, the method significantly outperforms the baseline method on all evaluation metrics. Specifically, the highest accuracy achieved by the method is 79.85%, which is approximately 23.31 and 7.23 percentage points higher than FedAvg-Drop and MultiSplitFed methods, respectively. The training epoch curves show that the method converges the fastest, reaching an accuracy of over 75% after approximately 220 training epochs, while the baseline method converges more slowly and achieves a lower final accuracy. The clock time curves show that achieving higher accuracy within the same time frame indicates a significant advantage in system efficiency. The communication volume curves show that the communication volume is significantly lower than other methods when achieving the same accuracy, demonstrating its superiority in reducing communication burden. FedAvg-Drop performed well initially; however, because it requires the device to store the complete model, many storage-constrained devices cannot participate in training, limiting its application scope. MultiSplitFed's performance is inferior, mainly because its fixed splitting strategy cannot adapt to the heterogeneity and dynamism of device resources, and the transfer of a large number of activation and gradient values ​​between splits reduces training efficiency.

[0078] like Figure 3As shown, under Non-IID data distribution, data heterogeneity poses a greater challenge to training federated learning models, but the method still performs well. Its highest accuracy reaches 72.28%, an improvement of approximately 38.12 and 10.36 percentage points compared to FedAvg-Drop and MultiSplitFed, respectively. Similar to the IID scenario, it outperforms baseline methods in convergence speed, system efficiency, and communication efficiency. Notably, FedAvg-Drop's performance significantly decreases in the Non-IID scenario, indicating a significant impact of data heterogeneity. Dynamic sharding and knowledge distillation techniques effectively mitigate this issue, maintaining model stability and high performance. MultiSplitFed's performance further declines in the Non-IID scenario, demonstrating its difficulty in effectively working when data and system heterogeneity coexist.

[0079] To verify the importance of the dynamic fragmentation selection strategy, this embodiment further compares the performance of this method and its variants FedSHTD-Random and FedSHTD-Greedy through ablation experiments.

[0080] like Figure 4 As shown, under the IID data distribution, the highest accuracy of this invention is 79.85%, significantly better than FedSHTD-Random (50.30%) and FedSHTD-Greedy (69.97%). From the round-count curves, the convergence speed of this invention far exceeds that of the variant. FedSHTD-Random suffers from lower performance due to random sharding, while FedSHTD-Greedy, although showing some improvement, still lags behind due to a lack of system optimization. The time and traffic curves further demonstrate that this invention is more efficient than the variant, verifying the crucial role of the dynamic sharding strategy in performance improvement.

[0081] like Figure 5 As shown, under Non-IID data distribution, the highest accuracy of this invention is 72.28%, which is an improvement of approximately 37.19 and 13.32 percentage points compared to FedSHTD-Random (35.09%) and FedSHTD-Greedy (58.96%), respectively. Data heterogeneity exacerbates the performance degradation of variants, while robustness is maintained through dynamically adjusted sharding and knowledge distillation. Time and traffic results also show that the advantages in efficiency and resource utilization stem from its optimized design.

[0082] Ablation experiments show that while random partitioning (FedSHTD-Random) and greedy partitioning (FedSHTD-Greedy) can alleviate storage pressure, they lack comprehensive optimization of system efficiency and model performance, resulting in performance far lower than that of this invention.

Claims

1. A federated learning method for intelligent Internet of Things, characterized in that, include: The server divides the global model M into contiguous pieces and configures lightweight auxiliary heads and tails to simulate contextual features; An event-triggered resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism are introduced; shard-level knowledge distillation is used to achieve cross-shard knowledge transfer, combined with decoupled gradient descent optimization.

2. The intelligent IoT federated learning method according to claim 1, characterized in that, The server divides the global model M into contiguous fragments; the global model M is divided into C contiguous fragments. ,in, For the c-th slice, Each slice is one or more layers in the neural network, and the granularity of the slice can be flexibly adjusted according to the application scenario and device resource constraints.

3. The intelligent IoT federated learning method according to claim 1, characterized in that, The configuration includes lightweight auxiliary heads and tails to simulate contextual features; for any segmented region The lightweight auxiliary head and tail include an auxiliary head. and auxiliary tail ,satisfy: ； ； in, For function composition operations; auxiliary header Simulate the output characteristics of the preceding slice, and assist the tail. The input characteristics of subsequent slices are simulated, and both adopt an adapter structure design to compress the number of parameters.

4. The intelligent IoT federated learning method according to claim 1, characterized in that, The system introduces an event-triggered resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism; it tracks the system status through resource queues, maturity queues, and training progress queues, constructs a weighted Lyapunov function, and minimizes the drift-penalty term to balance resource usage, training balance, and training latency.

5. The intelligent IoT federated learning method according to claim 4, characterized in that, The trigger-based resource monitoring and Lyapunov optimization-driven dynamic sharding adjustment mechanism checks the difference between the current resource status and the previous upload when uploading model parameters after a device is selected to participate in training. If the difference exceeds a preset threshold, the resource status is reported to reduce communication overhead and adapt to resource changes. For equipment When a device is selected to participate in the t-th round of training and completes model upload, the server compares the current resource status with the status at the time of the last upload; the resource status at the time of n uploads is defined as... The current resource status is ; Calculate the relative variables of each resource: ； ； ； The conditions that trigger resource reporting are: ； in, To store relative thresholds, As a relative threshold for computing power, This is a relative threshold for communication bandwidth; If the triggering condition is met The device will display the current resource status. Upload to the server and update the last upload status in local storage. Otherwise, the device only uploads model parameters and does not update resource status.

6. The intelligent IoT federated learning method according to claim 4, characterized in that, The dynamic partitioning adjustment mechanism includes candidate interval generation and online dynamic adjustment steps. When generating candidate intervals, all possible start and end indices are traversed to filter feasible intervals that meet storage constraints. During online dynamic adjustment, the optimal segmentation interval is selected from the candidate intervals based on the evaluation indicators and assigned to the device.

7. The intelligent IoT federated learning method according to claim 1, characterized in that, In the segmented knowledge distillation, the server selects the component with the lowest maturity from the candidate auxiliary component set as the student model. The maturity is determined by both training experience and global stability.

8. The intelligent IoT federated learning method according to claim 7, characterized in that, The fragment-level knowledge distillation includes auxiliary head distillation and auxiliary tail distillation. The auxiliary head distillation uses mean square error to measure the feature distillation loss, and the auxiliary tail distillation uses KL divergence to calculate the predicted label distillation loss. The segmented knowledge distillation introduces an adaptive distillation intensity coefficient, which is dynamically adjusted according to the maturity differences between teacher and student models and is proportional to the maturity differences.

9. The intelligent IoT federated learning method according to claim 1, characterized in that, In the described separate gradient descent optimization, the task loss optimizes the entire local model to improve task performance, while the distillation loss specifically optimizes a single-level distillation component and related parameters to enhance knowledge transfer. The two perform gradient descent independently.

10. The intelligent IoT federated learning method according to claim 1, characterized in that, The server uses a federated averaging algorithm to perform block-level aggregation of the updated shard parameters uploaded by the client, thereby completing the global model update.

Images