A hierarchical federated learning method and system based on split meta-learning

Abstract

This invention discloses a hierarchical federated learning method and system based on split meta-learning. The method specifically includes: a cloud server initializing a global model and distributing it to various edge servers; the edge server layer distributing the global model to its respective clients; each client updating the feature extraction module of the global model locally based on local data, while simultaneously freezing the classifier module of the global model; the edge server layer aggregating the updated parameters of the feature extraction modules from multiple clients and performing meta-optimization on the classifier module using a local validation dataset to generate an edge local model; and the cloud server layer periodically collecting model parameters from multiple edge servers and updating the global model using a gradient-sensitive momentum aggregation strategy. This invention, through a three-layer "vehicle-edge-cloud" architecture and a two-level collaborative optimization mechanism, achieves coordinated optimization between privacy protection, communication efficiency, and model adaptability in a vehicle-to-everything (V2X) environment.

Description

Technical Field

[0001] This invention belongs to the field of distributed machine learning and vehicle-to-everything (V2X) communication technology, and particularly relates to a hierarchical federated learning method and system based on split meta-learning. Background Technology

[0002] With the deep integration of intelligent connected vehicles (ICVs) and 5G communication technology, the Internet of Vehicles (IoV) is rapidly evolving into a key infrastructure supporting autonomous driving, intelligent transportation, and in-vehicle information services. IoV achieves collaborative perception, intelligent decision-making, and precise control through large-scale information interaction (V2X) between vehicles, roadside units (RSUs), pedestrians, and cloud control centers. With the access of hundreds of millions of vehicle terminals, IoV generates massive amounts of multimodal data, including high-definition images, LiDAR point clouds, and driving behavior data. How to safely and efficiently utilize this data to train higher-performance artificial intelligence models has become a core issue driving industry development.

[0003] Traditional centralized machine learning solutions require uploading all raw vehicle data to the cloud for unified training, which faces almost insurmountable obstacles in the connected vehicle environment. First, there are privacy and security red lines: vehicle data contains highly sensitive information such as precise user location trajectories, driving habits, and even in-vehicle audio and video. Its transmission and centralized storage face significant privacy risks and are subject to strict data security regulations worldwide. Second, there are communication bottlenecks and latency challenges: V2X communication, especially for applications like collaborative driving and collision warning to ensure driving safety, requires ultra-low latency down to the millisecond level. Uploading massive amounts of raw data to the cloud in real time would consume enormous network bandwidth, incur high communication costs, and introduce unacceptable communication latency, seriously threatening driving safety.

[0004] Federated Learning (FL), a distributed machine learning paradigm where the data remains stationary while the model moves, offers a novel solution for vehicle-to-everything (V2X) data processing. It allows each vehicle terminal to train its model locally using private data, uploading only encrypted model parameter updates to the server, thus ensuring data privacy at the source and significantly reducing communication load. However, directly applying standard federated learning frameworks (such as FedAvg) to the extremely complex environment of V2X still faces four unique and significant challenges.

[0005] Federated Learning (FL), as a distributed machine learning paradigm where the data remains stationary while the model moves, offers a novel solution for vehicle-to-everything (V2X) data processing. However, existing technologies such as patents CN112487123A (a federated learning method for V2X) and US20210056789A1 (a hierarchical federated learning system), while involving federated learning, have not effectively addressed the issues of asynchronous decomposition meta-learning and highly dynamic topology adaptation in V2X.

[0006] Second, the network topology is extremely dynamic. During high-speed vehicle movement, connections with edge servers (such as RSUs) are transient and unstable, a phenomenon known as the connection opportunity window. Vehicles frequently join and leave the coverage area of ​​an RSU, causing rapid changes in the set of participating nodes in federated learning. Standard synchronous federated learning mechanisms require all participants to complete training and uploading at the same time step, which is almost impossible to achieve in highly dynamic vehicular networks, resulting in significant waiting overhead and training interruptions.

[0007] Third, there is a significant heterogeneity in device resources. The computing capabilities of terminals in the Internet of Vehicles vary greatly, ranging from high-end intelligent vehicles with powerful computing capabilities to traditional vehicles or embedded devices with limited computing power. Required that all vehicles perform the same complex model training tasks (such as the MAML algorithm based on meta-learning, which requires calculating second-order gradients) is impractical. This would lead to a large number of low-computing-power nodes falling behind or failing to train, severely impacting the overall efficiency and fairness of the federated learning system.

[0008] Fourth, stringent real-time requirements. Many connected vehicle applications (such as cooperative lane changing and hazard warning) require models to adapt quickly to changes in the current environment. While existing technologies employ a hierarchical federated learning architecture of "vehicle-roadside unit-cloud" to address these challenges, most fail to effectively solve the asynchronous communication problem under highly dynamic topologies. Furthermore, they have limited effectiveness in balancing the generalization of the global model and the personalization of the local model, resulting in slow model convergence, poor adaptability, and difficulty in meeting the real-time and reliability requirements of connected vehicle application scenarios.

[0009] Therefore, designing a federated learning method for the Internet of Vehicles environment that can collaboratively handle data heterogeneity, adapt to highly dynamic network topologies, be compatible with device resource heterogeneity, and simultaneously meet multiple constraints such as privacy protection, communication efficiency, and real-time model adaptability is a core technical challenge that urgently needs to be overcome. Summary of the Invention

[0010] The purpose of this invention is to provide a hierarchical federated learning method and system based on split meta-learning, so as to overcome the difficulty of the existing technology in meeting the requirements of model adaptability, communication efficiency and privacy protection in the Internet of Vehicles environment, as well as the inability to effectively handle the dual data heterogeneity problem caused by the non-independent and identically distributed data of the vehicle client layer and the geographical location change of the task layer, so as to achieve safe and efficient processing of intelligent data in dynamic Internet of Vehicles, and solve at least one of the above-mentioned problems of the prior art.

[0011] In a first aspect, the present invention provides a hierarchical federated learning method based on split meta-learning, the method specifically comprising:

[0012] The system comprises a client layer, an edge server layer, and a cloud server layer. The client layer includes multiple clients, the edge server layer includes multiple edge servers, and the cloud server layer includes at least one cloud server.

[0013] The global model is initialized on the cloud server and distributed to each edge server. The global model is then distributed to the corresponding client through the edge server layer.

[0014] This allows each client to update the feature extraction module of the global model locally based on local data, while freezing the classifier module of the global model and asynchronously uploading the updated feature extraction module parameters to the associated edge server.

[0015] The edge server layer aggregates the updated parameters of the feature extraction module from multiple clients and performs meta-optimization on the classifier module using the local validation dataset to generate an edge local model.

[0016] Based on the cloud server layer, model parameters are periodically collected from multiple edge servers, and a gradient-sensitive momentum aggregation strategy is used to update the global model;

[0017] Based on the updated global model received at the edge server layer, personalized fusion weights are calculated according to the loss difference between the edge local model and the updated global model on local validation data.

[0018] Based on personalized fusion weights, the edge local model and the global model are weighted and fused to generate a personalized local model, which is then distributed to the corresponding client through the edge server.

[0019] Secondly, this invention provides a hierarchical federated learning system based on split meta-learning, the system specifically comprising:

[0020] The first federated learning module is used to construct a client layer, an edge server layer, and a cloud server layer. The client layer includes multiple clients, the edge server layer includes multiple edge servers, and the cloud server layer includes at least one cloud server.

[0021] The second federated learning module is used to initialize a global model based on the cloud server and distribute it to various edge servers. The global model is then distributed to the corresponding clients through the edge server layer.

[0022] The third federated learning module enables each client to update the feature extraction module of the global model locally based on local data, while freezing the classifier module of the global model and asynchronously uploading the updated feature extraction module parameters to the associated edge server.

[0023] The fourth federated learning module is used to aggregate the updated parameters of the feature extraction module from multiple clients based on the edge server layer, and perform meta-optimization on the classifier module using the local validation dataset to generate an edge local model.

[0024] The fifth federated learning module is used to periodically collect model parameters from multiple edge servers based on the cloud server layer and update the global model using a gradient-sensitive momentum aggregation strategy.

[0025] The sixth federated learning module is used to receive the updated global model at the edge server layer and calculate personalized fusion weights based on the loss difference between the edge local model and the updated global model on local validation data.

[0026] The seventh federated learning module is used to perform weighted fusion of edge local models and global models based on personalized fusion weights to generate personalized local models, which are then distributed to their respective clients through edge servers.

[0027] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements a hierarchical federated learning method based on split meta-learning as described in any of the above methods.

[0028] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the hierarchical federated learning method based on split meta-learning as described in any of the above methods.

[0029] Compared with the prior art, the present invention has at least one of the following technical effects:

[0030] 1. This invention achieves coordinated optimization between privacy protection, communication efficiency, and model adaptability in the Internet of Vehicles environment through a three-layer architecture design of "vehicle-edge-cloud" and a two-level collaborative optimization mechanism;

[0031] 2. This invention significantly reduces the computational and communication burden of the vehicle terminal through the asynchronous split meta-learning protocol, enabling it to operate efficiently under conditions of limited computing resources and improving the universality of the solution;

[0032] 3. By combining gradient-sensitive momentum aggregation and loss adaptation personalization mechanisms, this invention provides high-precision localized models for vehicles in different regions while ensuring the stability of the global model, thereby improving the scenario adaptability for tasks such as autonomous driving.

[0033] 4. This invention effectively addresses the dual data heterogeneity problem caused by different road environments, driving behaviors, and dynamic traffic flows, significantly improving model convergence speed and accuracy. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments 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.

[0035] Figure 1 This is a flowchart illustrating a hierarchical federated learning method based on split meta-learning provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of a network architecture for a hierarchical federated learning method based on split meta-learning, provided in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the structure of a hierarchical federated learning system based on split meta-learning provided in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0039] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0040] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0041] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0042] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0043] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0044] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0045] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating a hierarchical federated learning method based on split meta-learning, as disclosed in an embodiment of the present invention, is shown below in detail:

[0046] S101, Construct a client layer, an edge server layer, and a cloud server layer. The client layer includes multiple clients, the edge server layer includes multiple edge servers, and the cloud server layer includes at least one cloud server.

[0047] S102 initializes the global model based on the cloud server and distributes it to each edge server. The global model is then distributed to the corresponding client through the edge server layer.

[0048] S103 enables each client to update the feature extraction module of the global model locally based on local data, while freezing the classifier module of the global model and asynchronously uploading the updated feature extraction module parameters to the associated edge server.

[0049] S104 aggregates the updated parameters of the feature extraction module from multiple clients based on the edge server layer, and performs meta-optimization on the classifier module using the local validation dataset to generate an edge local model.

[0050] S105 periodically collects model parameters from multiple edge servers based on the cloud server layer and updates the global model using a gradient-sensitive momentum aggregation strategy.

[0051] S106: Based on the updated global model received by the edge server layer, calculate the personalized fusion weight according to the loss difference between the edge local model and the updated global model on the local validation data.

[0052] S107, based on personalized fusion weights, performs weighted fusion of edge local models and global models to generate personalized local models, which are then distributed to their respective clients via edge servers.

[0053] In this embodiment, firstly, a multi-layered architecture is constructed, consisting of a client layer, an edge server layer (such as a Roadside Unit (RSU), and a cloud server layer. The cloud server initializes the model and distributes it to the edge servers. The edge servers send the model to their respective clients. The client layer performs local updates on the first part of the model (feature extraction module) and asynchronously uploads the update results to its associated edge server. The edge server layer aggregates update results from one or more clients within its coverage area and performs meta-optimization on the second part of the model (classifier module) using its local validation dataset. The cloud server layer periodically collects model parameters from multiple edge servers and updates the global model using a momentum aggregation strategy, where the momentum coefficients are dynamically adjusted based on the magnitude of the validation loss gradient and the dispersion of model parameters. After receiving the global model from the cloud server, the edge server layer performs weighted fusion of the local and global models based on the loss difference between its local model and the global model on the local validation data to generate a personalized local model. The edge servers send the personalized model to their respective clients; this process is repeated until convergence or the training iterations are reached.

[0054] In one possible embodiment, such as Figure 2As shown, this embodiment constructs a three-layer "vehicle-edge-cloud" architecture system, including 200 vehicle clients, 5 edge servers, and 1 cloud server. The vehicle clients adopt a lightweight CNN architecture (containing 3 convolutional layers and 2 fully connected layers) to simulate resource-constrained in-vehicle computing units; both the edge servers and the cloud server use ResNet-18 as the base model to support more complex feature learning.

[0055] like Figure 2 As shown, the specific implementation steps of this embodiment include:

[0056] (1) System initialization phase: The cloud server generates an initial global model and distributes it to each edge server. The edge server then distributes the initial model to the vehicle clients within its coverage area.

[0057] (2) Vehicle client local update stage: The vehicle client executes the asynchronous split meta-learning protocol based on the local private dataset, updates only the feature extraction module, and freezes the classifier parameters;

[0058] (3) Edge layer model aggregation stage: The edge server receives the feature extraction module update uploaded by the vehicle client, performs weighted aggregation, and performs meta-optimization on the classifier parameters based on local validation data;

[0059] (4) Cloud model update stage: The cloud server receives the model parameters uploaded by each edge server and updates the global model using gradient-sensitive momentum aggregation.

[0060] (5) Edge layer personalization stage: The edge server calculates the performance difference between the local model and the global model, generates personalized fusion coefficients based on the loss adaptive mechanism, and performs weighted fusion of the local model and the global model;

[0061] (6) Periodic synchronization phase: with ( (This refers to a global-edge synchronization cycle) cycle. The cloud server distributes the updated global model to the edge server, and the edge server then updates the model of the associated vehicle client.

[0062] In this embodiment, a fixed parameter setting is used to implement the federated learning process: inner loop learning rate. outer loop learning rate Initial global momentum factor Loss adaptive personalized fusion factor Initial warm-up wheel =5 rounds, global-edge synchronization cycle =5, Vehicle movement trigger threshold =10%. These parameter combinations have been validated across multiple datasets and experimental scenarios, achieving an optimal balance between performance and stability. The vehicle client trains locally for each round. Each round randomly selects 50 clients (25% of the total) to participate in the federated learning process. The training process consists of 250 global iterations, using the SGD optimizer (momentum 0.9, batch size 32), and the learning rate is adjusted according to the cosine annealing strategy.

[0063] In some embodiments, step S102 above, which involves initializing the global model based on the cloud server and distributing it to various edge servers, and distributing the global model to the corresponding clients through the edge server layer, specifically includes:

[0064] An initial global model is constructed based on a normal distribution initialization strategy using a cloud server. The initial global model includes initialization parameters containing parameters of the feature extraction module and parameters of the classifier module.

[0065] The cloud server distributes a lightweight diagnostic model to each edge server, collects the real-time computing load and storage status of each edge server, and obtains diagnostic results.

[0066] Based on the diagnostic results, and according to the load of each edge server, the complete initial global model or the simplified initial global model is distributed to each edge server through the cloud server.

[0067] Based on the network status of clients within the coverage area of ​​each edge server, the edge server adaptively distributes the clients to its respective clients.

[0068] In this embodiment, the central server distributes global model parameters to each edge server. ,in These are the parameters for the feature extractor. These are the classifier parameters.

[0069] Specifically, the cloud server first determines the structural framework of the global model, which includes a feature extraction module and a classifier module. Next, the cloud server uses a normal distribution initialization strategy to generate initial values ​​for the parameters of the feature extraction module and the classifier module. The normal distribution initialization strategy is a commonly used parameter initialization method that assigns a reasonable initial range to the model parameters, helping the model converge faster during subsequent training. In this way, the cloud server successfully constructs the initial global model, which provides the foundation for subsequent training and distribution.

[0070] To better understand the real-time status of each edge server, the cloud server distributes a lightweight diagnostic model to each edge server. This lightweight diagnostic model is specifically designed for quickly evaluating edge server performance, featuring low computational load and low resource consumption. Upon receiving the lightweight diagnostic model, each edge server immediately runs it to detect its own real-time computing load and storage status. After the detection is complete, each edge server feeds back the detected real-time computing load and storage status information as diagnostic results to the cloud server. Through this step, the cloud server can promptly grasp the actual operating status of each edge server, providing a basis for subsequent model distribution strategies.

[0071] After receiving diagnostic results from each edge server, the cloud server analyzes and processes these results. Based on the load levels of each edge server, the cloud server formulates different model distribution strategies. For edge servers with high loads, considering their relatively limited computing resources, the cloud server distributes a simplified initial global model to alleviate their computational burden. This simplified initial global model retains core functionality while appropriately simplifying the model structure or parameters, thereby reducing the model's computational resource requirements. For edge servers with low loads, which possess sufficient computing power, the cloud server distributes the complete initial global model to ensure they can fully leverage their performance advantages for more complex model training and processing tasks. This differentiated model distribution approach better adapts to the actual needs of different edge servers, improving the overall system's operational efficiency.

[0072] After receiving the global model (full or simplified version) distributed by the cloud server, each edge server needs to further distribute the model to various clients within its coverage area. To ensure the efficiency and stability of model distribution, the edge server monitors the network status of clients within its coverage area in real time. Network status includes indicators such as network bandwidth, signal strength, and network latency. Based on this network status information, the edge server adopts an adaptive distribution strategy for different clients. For example, for clients with sufficient network bandwidth, good signal strength, and low network latency, the edge server can use a higher data transmission rate to quickly send the complete global model to the client; while for clients with poor network conditions, the edge server may use methods such as batch transmission and data compression to reduce the data transmission volume, ensuring that the model can be successfully transmitted to the client while minimizing the occupation of the client's network resources. Through this adaptive distribution method, the success rate of model distribution can be improved, ensuring that each client in the vehicle network system can obtain the global model in a timely manner, laying the foundation for subsequent local training and data processing.

[0073] In some embodiments, step S103 above, which involves enabling each client to locally update the feature extraction module of the global model based on local data, simultaneously freezing the classifier module of the global model, and asynchronously uploading the updated parameters of the feature extraction module to the associated edge server, specifically includes:

[0074] Configure local training tasks based on the client's local computing resources. Based on the configured local training tasks, sample small batches of data from the local dataset to execute local training cycles. Calculate the loss through forward propagation and backpropagate the gradient, updating only the parameters of the feature extraction module.

[0075] After completing the preset local training cycle, the feature extraction parameter increment is calculated through the client. The feature extraction parameter increment is the difference between the locally updated parameters and the originally received feature extraction module parameters.

[0076] The client asynchronously uploads the incremental feature extraction parameters to the associated edge server.

[0077] In this embodiment, when the vehicle client performs local updates, it adopts a split learning strategy, keeping the parameters of the second part of its model (classifier module) frozen, and only updating the first part of the model (feature extraction module); the specific update formula is as follows:

[0078] ;

[0079] in:

[0080] : No. The vehicle client completes the first The parameters of the feature extraction module after local training (updated state);

[0081] : No. Vehicle client in The parameters of the feature extraction module at the start of each training round (current state);

[0082] The learning rate used during local training on the client side controls the step size for parameter updates.

[0083] : This operator represents the gradient calculation for the parameters of the feature extraction module;

[0084] The client-side local loss function measures the deviation between the model's predictions and the true labels.

[0085] : Global classifier parameters that are frozen and issued by the edge server (remain unchanged during local training);

[0086] : No. Each vehicle client uses a small-batch dataset for local training.

[0087] The technical advantage of this embodiment lies in freezing the classifier parameters. Only update the feature extractor This significantly reduces the computational load on the vehicle terminal, unlike existing methods that update complete model parameters.

[0088] In some embodiments, step S104 above, which involves aggregating updated parameters from multiple clients based on the edge server layer and performing meta-optimization on the classifier module using a local validation dataset to generate an edge local model, specifically includes:

[0089] Based on the updated feature extraction module parameters received from multiple clients at the edge server layer, and combined with the proportion of data volume from each client, the aggregation weight coefficient is calculated, a weighted aggregation operation is performed, and the aggregated feature extraction parameters are generated.

[0090] By using the aggregated feature extraction parameters through the edge server and combining them with the local validation dataset, a meta-optimization process is performed on the classifier module, and the classifier parameters are updated.

[0091] The updated classifier parameters are combined with the aggregated feature extraction parameters to construct an edge-local model.

[0092] In this embodiment, the edge server layer performs joint feature aggregation upon receiving an active subset. Feature increment of contribution Subsequently, RSU reached a consensus through data quality-weighted synthesis, and the aggregation calculation is shown below:

[0093] ;

[0094] The weights are calculated as follows:

[0095] ;

[0096] The above symbols are defined as follows:

[0097] : indicates the first The edge server completed the first After round aggregation, update the parameters of the obtained feature extraction model;

[0098] : indicates the first Edge servers in the The feature extraction model parameters before the previous or initial round of updates;

[0099] : indicates the first During the training round, the set of indexes of active clients participating in the edge server aggregation;

[0100] : indicates the first The weighting coefficient of the client in the model aggregation process (determined by its proportion of data volume).

[0101] : indicates that it belongs to the first The edge server's The client, in the first The increment (i.e., update amount) of the feature extraction model parameters uploaded after local training.

[0102] : indicates that it belongs to the first The edge server's The number of samples (data size) in the local dataset owned by each client.

[0103] : indicates the first In the wheel, the assembly All clients participating in the aggregation (belonging to the edge server) The sum of the number of local dataset samples (i.e., the total amount of data in the current round).

[0104] The edge server layer utilizes the aggregated feature states. Meta-adaptation of classifier parameters is performed by second-order gradient feedback generated on the edge validation set.

[0105] The meta-optimization update formula is:

[0106] ;

[0107] The above symbols are defined as follows:

[0108] : No. The parameters of the classifier module on the edge server after meta-optimization and update;

[0109] Initial global classifier parameters (the starting point for meta-optimization) are sent from the cloud.

[0110] The outer learning rate when the edge server performs meta-learning updates.

[0111] : This operator represents the gradient calculation on the classifier parameters;

[0112] Loss function on the local validation set of the edge server;

[0113] The parameters of the feature extraction module after aggregation by the edge server are used as input to calculate the validation loss.

[0114] The technical advantage of this embodiment is that it uses second-order gradient information to simulate the decision boundary after client feature drift, enabling the classifier to quickly adapt to heterogeneous data features in different regions.

[0115] In some embodiments, step S105 above, which involves periodically collecting model parameters from multiple edge servers based on the cloud server layer and updating the global model using a gradient-sensitive momentum aggregation strategy, specifically includes:

[0116] Based on a preset synchronization cycle, the cloud server collects the corresponding model parameter sets from multiple edge servers.

[0117] Based on the model parameter set, the validation loss gradient of each edge server model in the current round is calculated through the cloud server, and the spatial distribution differences between each model parameter in the model parameter set are analyzed to obtain the parameter dispersion index.

[0118] Based on the validation loss gradient and parameter dispersion index, a dynamic momentum coefficient is calculated by the cloud server through a dynamic adjustment mechanism. The dynamic momentum coefficient is negatively correlated with the changing trend of the validation loss gradient and positively correlated with the parameter dispersion index.

[0119] The historical global model parameters and the model parameter set are weighted and fused using dynamic momentum coefficients to perform a global model update and generate a new round of global model.

[0120] In this embodiment, the momentum aggregation strategy of the cloud server layer adopts a gradient-sensitive and discreteness-aware dynamic momentum update mechanism.

[0121] Cloud server for feature parameters and classifier parameters The formula for updating is:

[0122] ;

[0123] The above symbols are defined as follows:

[0124] The updated global model parameter set (including the feature extractor) and classifier );

[0125] The previous set of global model parameters (historical model);

[0126] : Dynamic momentum coefficient, used to balance the weights of historical global model information and current cycle update information;

[0127] Total number of edge servers participating in this round of updates;

[0128] : No. The total amount of data within the coverage area of ​​the edge server;

[0129] The sum of all data volume from all participating edge servers;

[0130] : No. Complete model parameters uploaded from the edge server.

[0131] The dynamic momentum coefficient It is not a static constant, but is dynamically calculated according to the following formula:

[0132] ;

[0133] The above symbols are defined as follows:

[0134] : Preset baseline momentum factor (basic inertia value);

[0135] Gradient-sensitive hyperparameter adjustment controls the model's response to convergence (gradient changes);

[0136] and : The gradient norm of the global validation loss, representing the previous round ( ) and the previous round ( The state of ) is used to determine the convergence trend;

[0137] : Prevents smoothing terms (minimum values) where the denominator is zero;

[0138] Discreteness adjustment hyperparameter controls the model's sensitivity to data heterogeneity;

[0139] The spatial dispersion (variance) of the model parameters submitted by each edge server reflects the differences in data distribution (Non-IID degree).

[0140] While existing technologies (such as CN118900423B) also employ a layered architecture, their cloud aggregation typically uses a weighted average (FedAvg) or fixed momentum aggregation method based on data volume. In the highly dynamic environment of connected vehicles, the distribution of vehicle data (Non-IID) changes extremely rapidly, and a fixed aggregation strategy is prone to causing the global model to oscillate violently between different distributions, making convergence difficult.

[0141] The formula proposed in this invention (the momentum formula above) introduces the gradient norm and parameter dispersion as feedback variables. The physical meaning is that when the system detects significant (increased) differences in the models uploaded by various edge servers, it determines that the current data heterogeneity is high. The algorithm automatically increases the momentum coefficient, forcing the model to rely more on historical experience rather than current drastic fluctuations, thus fundamentally solving the training divergence problem caused by heterogeneous vehicle network data at the mathematical level.

[0142] The technical advantage of this embodiment lies in that, unlike the fixed aggregation weights in existing technologies, this method automatically increases the weights of historical models when it detects large dispersion in model parameters (i.e., severe data heterogeneity). To suppress oscillations; when a large gradient is detected (early stage of convergence), reduce inertia to accelerate convergence.

[0143] In some embodiments, step S106 above, which involves receiving the updated global model at the edge server layer and calculating the personalized fusion weights based on the loss difference between the edge local model and the updated global model on local validation data, specifically includes:

[0144] The edge server receives updated global model parameters from the cloud server and compares them with the parameters of its own maintained edge local model. The local model loss value and the global model loss value are calculated on the local validation dataset of the edge server.

[0145] The loss difference value is obtained by using the calculated local model loss value and global model loss value through interpolation.

[0146] The loss difference value is input into the hyperbolic tangent activation function for nonlinear transformation to generate a weight adjustment factor.

[0147] Based on the weight adjustment factor, combined with the preset baseline retention rate and sensitivity scaling factor, the personalized fusion weight is calculated.

[0148] In this embodiment, in the personalized weighted fusion step of the edge server layer, a loss adaptive fusion mechanism based on the hyperbolic tangent function is constructed;

[0149] First, calculate the relative performance difference between the local model and the global model on the edge validation set. :

[0150] ;

[0151] The above symbols are defined as follows:

[0152] The performance (loss) difference between the local model and the global model on the edge validation set;

[0153] : in the Loss calculation function on the local validation set of the edge server;

[0154] The loss calculation function on the global model;

[0155] : No. Local model parameters of the edge server (aggregated and meta-optimized);

[0156] The latest global model parameters distributed from the cloud.

[0157] Secondly, personalized fusion weights are generated based on this difference. :

[0158] ;

[0159] The above symbols are defined as follows:

[0160] : No. Personalized fusion weights for edge servers;

[0161] : Baseline retention rate, ensuring that local features have a minimum retention rate regardless of differences;

[0162] Sensitivity scaling factor, used to control (scaling range).

[0163] The hyperbolic tangent activation function nonlinearly maps the loss difference to the weight adjustment interval.

[0164] The technical effect of this embodiment is that when (i.e., the local model performs worse than the global model) When the function value is negative, the local weights are automatically reduced, and more global knowledge is introduced; conversely, when the value is positive, more local features are retained, thus achieving adaptive personalization without human intervention.

[0165] This embodiment implements a personalized fusion coefficient generation mechanism with adaptive loss, tailored to the local data distribution characteristics of different edge servers (representing different geographical regions). This mechanism dynamically adjusts the fusion coefficients based on the performance differences between the local model and the global model on the validation data.

[0166] Existing technology CN118900423B primarily focuses on improving performance by 'selecting better clients,' and its edge server distributes a uniform model to all vehicles within the area. This invention, however, focuses on 'localized model adaptation.' Using the formula (the aforementioned tanh fusion formula), this invention constructs an automatic valve by leveraging the nonlinear saturation characteristics of the hyperbolic tangent function. For edge servers located in traffic-congested areas (with a single data pattern), the local model may overfit; in this case, the formula-calculated... For larger edge nodes, the global model automatically incorporates common features; and for newly established edge nodes, the global model enables rapid cold start. This fine-grained fusion based on loss difference is something that existing client selection methods cannot achieve.

[0167] In the parameter sensitivity analysis, a loss-adaptive personalized fusion factor was found. The optimal value range for is [0.5, 0.7], within which the system exhibits the most stable performance. When At times, the model relies too heavily on global knowledge and lacks personalization capabilities; when At times, the local model's dominance is too strong, leading to difficulties in global convergence. Optimal performance occurs when... This is consistent with the default settings of this invention.

[0168] Furthermore, the method of the present invention addresses the Dirichlet distribution parameters. Its adaptability was also evaluated. (Highly non-IID) to The method of this invention performs well across a wide range of (approximate IID) datasets, consistently maintaining an advantage over existing methods on the CIFAR-100 dataset, demonstrating the method's adaptability to varying degrees of data heterogeneity.

[0169] In some embodiments, in S107, the step of weighted fusion of the edge local model and the global model based on personalized fusion weights to generate a personalized local model, and distributing the personalized local model to the corresponding client through the edge server, specifically includes:

[0170] By utilizing personalized fusion weights, a weighted fusion operation is performed on the parameters of the edge local model and the global model to generate fused model parameters;

[0171] Based on the parameters of the fusion model, a personalized local model is constructed according to the preset model architecture;

[0172] The parameters of the personalized local model are distributed to the respective clients via edge servers.

[0173] In this embodiment, the final generated personalized model for:

[0174] ;

[0175] The above symbols are defined as follows:

[0176] : No. The edge server ultimately generates a personalized local model (which will be sent to the client).

[0177] Local model parameters of the edge server;

[0178] : Global model parameters.

[0179] In some embodiments, in steps S101-S107 above, the model synchronization period between the cloud server layer and the edge server layer is adaptively adjusted according to the proportion of vehicle clients switching to different edge servers within a certain time window. When the vehicle client performs local updates, the classifier parameters are frozen, and the feature extraction module is optimized and updated multiple times based only on the local private dataset. The vehicle client and the edge server use asynchronous communication to adapt to the dynamic connection characteristics brought about by the high-speed movement of the vehicle client. The edge server performs weighted aggregation on the received feature extraction module updates, with the weights determined according to the proportion of the local dataset size of each vehicle client. Then, it performs meta-optimization updates on the classifier parameters based on the local validation data of the edge server.

[0180] In some embodiments, in steps S101-S107 above, the method further includes a system robustness improvement mechanism, which fixes the momentum coefficient and personalized fusion coefficient during the initial training phase, and tightens the dynamic adjustment range of the coefficient when the vehicle movement ratio is detected to exceed a preset threshold. When the cloud server updates the global model, it adopts a momentum smoothing fusion method, and its momentum coefficient is dynamically adjusted according to the magnitude of the validation loss gradient and the dispersion of model parameters. When the edge server performs personalized weighted fusion of the local model and the global model, it first calculates the loss difference between the local model and the global model on the local validation data of the edge server, then generates a loss-adaptive personalized fusion coefficient based on this loss difference, and finally performs weighted fusion of the local model and the global model based on this coefficient. In the dynamic adjustment mechanism of the momentum coefficient, the weight of historical model parameters is increased when the validation loss gradient increases, and the weight of historical model parameters is decreased when the dispersion of model parameters increases. The loss difference is calculated by the relative difference between the local model loss and the global model loss. The loss-adaptive personalized fusion coefficient is generated by mapping the loss difference to a preset range; the larger the loss difference, the more the fusion coefficient tends to retain the features of the local model.

[0181] In some embodiments, in steps S101-S107 above, the feature extraction module is used to process the time-series data collected by the client based on a lightweight Transformer architecture using a multi-head self-attention mechanism, and extract a temporal feature vector containing driving behavior intent, including...

[0182] It receives raw time-series data collected by the vehicle-mounted sensors on the client, and performs noise reduction and normalization preprocessing on the raw time-series data to generate standardized sequence data.

[0183] Segment and embedding mapping are performed on standardized sequence data to generate sequence embedding vectors;

[0184] Add positional encoding information to the sequence embedding vector to preserve temporal order dependencies and generate position-aware embedding vectors;

[0185] The location-aware embedding vector is input into the multi-head self-attention mechanism, and the attention distribution is calculated through the query matrix, key matrix, and value matrix. Then, the contextual features are aggregated to generate enhanced feature representations.

[0186] Based on enhanced feature representation, high-order temporal feature vectors are extracted by nonlinear transformation and dimension adjustment through a feedforward neural network.

[0187] Features at key time steps are selected from high-order temporal feature vectors based on attention weights to generate temporal feature vectors that contain driving behavior intentions.

[0188] In this embodiment, the first part of the model deployed on the vehicle client (feature extraction module) is based on a lightweight Transformer architecture and uses a multi-head self-attention mechanism to process the time-series data collected by the vehicle sensors and extract a temporal feature vector containing driving behavior intentions.

[0189] Considering the limited computing power of in-vehicle terminals and the strong time dependence of driving data (such as LiDAR point cloud sequences and IMU data), the vehicle client... Feature extraction module Instead of using standard deep CNNs, a variant of the lightweight Vision Transformer is adopted.

[0190] Specifically, the client will continuously Frame sensor input data The sequence is segmented into patches, and temporal information is embedded through positional encoding. Features are then computed using a multi-head self-attention mechanism.

[0191] ;

[0192] in:

[0193] (Query Matrix): The query vector matrix. It represents the sensor data segment at the current moment or the one of interest, such as the LiDAR point cloud features of the current frame. It searches for the correlation between itself and other parts of the sequence;

[0194] (Key Matrix): A key vector matrix. It represents the index features of data segments at all times in the input sequence. Used in conjunction with... Perform matching and calculate relevance;

[0195] (Value matrix): A value vector matrix. It represents the actual content features contained in the input sequence, such as specific vehicle speeds and obstacle texture information;

[0196] : Dimension of the key vector. Scaling factor, used to adjust the size of the dot product;

[0197] (Scaling item): When When the value is large, the dot product The result will be very large, causing the softmax function to enter the saturation region where the gradient is extremely small (gradient vanishing). Divide by The values ​​can be pulled back to the gradient-sensitive region to ensure the stability of model training;

[0198] (Relevance Score): Calculates the similarity between the query and the key. In autonomous driving, this represents the dependency between the current vehicle state and the state over the past few seconds; for example, determining whether the current deceleration is a continuation of a previously identified obstacle ahead.

[0199] Softmax: A normalization function. It converts the relevance scores into a probability distribution (with weights summing to 1), determining how much attention should be allocated to different parts of the sequence when synthesizing the final features.

[0200] In this way, the features output by the client It not only includes current visual features but also implicitly reveals vehicle motion trends, such as changes in speed and acceleration, which are crucial for predicting collision risk. Because only the extracted feature sequences are uploaded, the original video stream remains within the vehicle, ensuring privacy.

[0201] In some embodiments, in steps S101-S107 above, the classifier module is used to extract spatial location relationships and interaction risk features between vehicles by fusing multi-client features through graph convolution operations based on the graph neural network aggregation layer, including:

[0202] Based on the initial feature vectors received from multiple clients by the edge server, and based on the relative positions or communication connection status between the vehicles corresponding to the clients, a dynamic graph structure is constructed, wherein the graph nodes of the dynamic graph structure are clients and the graph edges are connections between vehicles.

[0203] The node features of the dynamic graph structure are aggregated by graph convolution operation, and the initial feature vector is transformed based on the adjacency matrix and degree matrix to generate enhanced node features that fuse multi-client features.

[0204] The spatial positional relationship features and interaction risk features between vehicles are extracted from the enhanced node features. Based on the spatial positional relationship features and interaction risk features, collaborative scenario understanding is performed, and collaborative decision-making results are output.

[0205] In this embodiment, the second part (classifier module) of the model deployed on the edge server uses a front graph neural network (GNN) aggregation layer to treat multiple vehicle clients within the coverage area as graph nodes and the communication links between vehicles as graph edges. It fuses features from multiple clients through graph convolution operations to extract spatial location relationships and interaction risk features between vehicles, thereby enabling collaborative scene understanding and decision-making.

[0206] On the edge server The location received data from multiple vehicles within the area. After uploading features, traditional average aggregation cannot capture the relative positional relationships between vehicles (e.g., the risk of a rear-end collision between a following vehicle and the vehicle in front is much greater than that between parallel vehicles). Therefore, this invention utilizes a classifier on an edge server. The front end incorporates a graph neural network layer.

[0207] Building dynamic graphs , where nodes Representative vehicle client Features, edges Representative vehicle and The relative distance or communication connection between them. The edge aggregation process uses a graph convolutional network (GCN):

[0208] ;

[0209] in:

[0210] : No. The layer's node feature matrix. Each row of the matrix represents a feature vector uploaded by a vehicle (client), and each column represents a dimension of the feature. This is the set of states of all vehicles received by the edge server at the current moment;

[0211] : No. The layer outputs a feature matrix (the aggregated result). After graph convolution processing, the features of each car are fused with information from its neighbors (surrounding vehicles). For example, the features of the following car now include the braking intention of the car in front;

[0212] : Added adjacency matrix with self-loops ( ). It records the connection relationships between vehicles, such as communication links or physical distances that are less than a threshold; It is an identity matrix, which ensures that during aggregation, vehicles not only consider the information of their neighbors, but also retain their own original features;

[0213] : The degree matrix. A diagonal matrix that records how many neighbors each vehicle is connected to;

[0214] A symmetric normalized adjacency matrix. This prevents nodes with high degree (cars with many neighbors) from having excessively large eigenvalues, leading to numerical instability. It implements a weighted average of features, rather than a simple summation.

[0215] : No. The learnable weight matrix of a layer. A core parameter of the neural network, responsible for learning how to extract useful cooperative risk patterns from neighboring features, such as multi-vehicle cooperative lane changing and intersection game theory.

[0216] Nonlinear activation function (usually ReLU). Increases the model's nonlinear expressive power, enabling it to fit complex traffic scenarios.

[0217] By using GNN aggregation, edge servers can understand complex macro-level scenarios such as congested intersections or coordinated overtaking, thereby optimizing the global classifier.

[0218] In some embodiments, the method further includes, in steps S101 to S107 above:

[0219] Based on the historical traffic accident data and traffic flow characteristics of the corresponding area of ​​the edge server, a regional benchmark risk model is trained and maintained for each edge server, and benchmark model parameters are generated.

[0220] Based on the driving behavior features extracted by the client from the local CAN bus data, the local feature extraction module is updated based on the driving behavior features to generate feature increment parameters;

[0221] Based on the feature increment parameters, the classifier boundary of the regional benchmark risk model is adjusted by using a meta-optimization strategy through the edge server to generate optimized local model parameters.

[0222] Based on the first loss difference between the optimized local model parameters and the baseline model parameters calculated by the edge server on the local validation data, a first personalized fusion weight is generated based on the first loss difference.

[0223] The optimized local model parameters and the baseline model parameters are weighted and fused using the first personalized fusion weight to generate a personalized risk assessment model. This personalized risk assessment model is used for risk scoring and real-time adjustment of the vehicle's basic insurance premium rate.

[0224] In this embodiment, in the vehicle insurance risk assessment scenario, the edge server generates a personalized risk model based on dynamic driving behavior and regional environmental context:

[0225] The edge server maintains a regional baseline risk model, which is trained based on historical traffic accident data and traffic flow characteristics of the region;

[0226] 1) The vehicle client extracts driving behavior features such as rapid acceleration, rapid deceleration and frequent lane changes from the local CAN bus data, and uploads the feature increments after local updates;

[0227] 2) After receiving the feature increment, the edge server uses the meta-optimization strategy to adjust the classifier boundary of the regional benchmark risk model, and combines it with the personalized weighted fusion mechanism to generate a personalized risk assessment model for a specific driver.

[0228] The risk score output by the personalized risk assessment model is used to adjust the Use-Based Insurance (UBI) rate for the vehicle in real time.

[0229] Specifically, the method described in this invention has significant application value in the field of auto insurance (especially UBI, Usage-based Insurance). Existing technologies typically rely on centralized cloud analysis, while this invention achieves more accurate personalized risk pricing through edge computing. The specific implementation steps include:

[0230] 1. Data source and feature construction:

[0231] Vehicle client Raw data is read from the OBD (On-Board Diagnostics) interface, including vehicle speed, engine speed, brake pedal depth, and steering wheel angle. Local feature extractor. Trained to identify high-risk events such as hard braking, hard cornering, and speeding;

[0232] 2. Integration of Regional Benchmarks and Personalized Standards

[0233] Edge server (For example, RSUs deployed at accident-prone intersections) Maintain a regional baseline risk model. The model learns the average risk level of a specific road segment by aggregating the features of all vehicles that have passed through the intersection (e.g., the average frequency of emergency braking at the intersection is 5 times per hour).

[0234] 3. Personalized risk model generation:

[0235] When the vehicle When passing through this region, the split meta-learning process of this invention is executed:

[0236] Local update: The vehicle updates the feature extractor based on its own driving habits and calculates its feature deviation relative to the average driver.

[0237] Meta-optimization and fusion: Edge servers utilize personalized fusion formulas:

[0238] ;

[0239] in:

[0240] Vehicle client The final personalized risk assessment model parameters are then used. This model will be deployed directly on the vehicle or used for edge inference, taking real-time driving data as input and outputting the driver's current risk score.

[0241] Feature extractor / model parameters trained on local data. These parameters reflect the driver's individual characteristics. For example, the model parameters remember the driver's habitual braking force, lane change frequency, and other micro-driving behaviors.

[0242] These are the parameters of the regional baseline risk model maintained by the edge server. They reflect common environmental risks. For example, the intersection itself has a complex design, resulting in a high average braking frequency for all passing vehicles. These parameters represent normal public driving levels or road condition benchmarks.

[0243] Personalized fusion weighting coefficient (by Mechanism calculation). Responsibility determination weight;

[0244] like Larger (closer to 1): This indicates that the model trusts local features more. It means that the current risk mainly stems from the driver's personal driving habits (such as sudden braking without reason on a straight road), and insurance pricing should take individual behavior into greater consideration.

[0245] like Smaller (closer to 0): This indicates that the model trusts the regional benchmark more. It means that the current outliers are mainly caused by environmental factors (e.g., all cars slowing down here), and insurance pricing should exclude environmental factors to avoid misjudging driver violations.

[0246] Output: Generated personalized model Output a normalized risk score ;

[0247] (Risk Score): Risk score (the output mentioned in the text). The normalized risk probability is directly used as a multiplier factor for UBI rates, for example... This may result in an increase in insurance premiums for that trip.

[0248] If the model output , If a preset risk score threshold is set, the vehicle is deemed to be engaging in high-risk behavior on the current road segment. The insurance system can adjust premiums in real time accordingly (e.g., increasing the premium by 5% for this trip) or send real-time safety alerts through the vehicle's infotainment system. Compared to traditional methods, this edge-side personalized fusion approach utilizes both environmental big data (edge-side knowledge) and accurately characterizes driver traits (local-side knowledge), without requiring the uploading of specific driving trajectories, thus perfectly resolving privacy compliance issues.

[0249] In some embodiments, the method further includes, in steps S101 to S107 above: a) multiple vehicle clients configured to perform local updates only on the first part of the model and asynchronously upload the update results; b) multiple edge servers configured to aggregate updates from the vehicle clients, perform meta-optimization on the second part of the model, and perform personalized weighted fusion; c) a cloud server configured to periodically update the global model using the momentum aggregation strategy.

[0250] Furthermore, the vehicle client is the on-board computing unit of an intelligent connected vehicle, and the system complies with automotive functional safety standards in its design and operation.

[0251] Those skilled in the art will understand that, without departing from the core principles of this invention, the above parameters can be appropriately adjusted to suit specific application scenarios. For example, in scenarios with extremely poor network communication conditions, the parameters can be appropriately increased. To reduce communication frequency; in scenarios where client data is highly heterogeneous, the frequency can be appropriately increased. To enhance personalization capabilities. These adjustments still fall within the scope of protection of this invention.

[0252] Reference Figure 3 An embodiment of the present invention provides a hierarchical federated learning system 3 based on split meta-learning, wherein the hierarchical federated learning system 3 based on split meta-learning specifically includes:

[0253] The first federated learning module 301 is used to construct a client layer, an edge server layer, and a cloud server layer. The client layer includes multiple clients, the edge server layer includes multiple edge servers, and the cloud server layer includes at least one cloud server.

[0254] The second federated learning module 302 is used to initialize a global model based on the cloud server and distribute it to various edge servers. The global model is then distributed to the corresponding clients through the edge server layer.

[0255] The third federated learning module 303 is used to enable each client to update the feature extraction module of the global model locally based on local data, while freezing the classifier module of the global model and asynchronously uploading the updated feature extraction module parameters to the associated edge server.

[0256] The fourth federated learning module 304 is used to aggregate the updated parameters of the feature extraction module from multiple clients based on the edge server layer, and perform meta-optimization on the classifier module using the local validation dataset to generate an edge local model.

[0257] The fifth federated learning module 305 is used to periodically collect model parameters from multiple edge servers based on the cloud server layer and update the global model using a gradient-sensitive momentum aggregation strategy.

[0258] The sixth federated learning module 306 is used to receive the updated global model based on the edge server layer and calculate personalized fusion weights based on the loss difference between the edge local model and the updated global model on local validation data.

[0259] The seventh federated learning module 307 is used to perform weighted fusion of edge local models and global models based on personalized fusion weights to generate personalized local models, which are then distributed to their respective clients via edge servers.

[0260] It is understandable that, such as Figure 1 The content of the hierarchical federated learning method embodiment based on split meta-learning shown is applicable to this hierarchical federated learning system embodiment based on split meta-learning. The specific functions implemented by this hierarchical federated learning system embodiment based on split meta-learning are the same as those shown below. Figure 1 The illustrated hierarchical federated learning method based on split meta-learning is the same as the one shown, and achieves the same beneficial effects. Figure 1 The beneficial effects achieved by the hierarchical federated learning method based on split meta-learning shown in the example are also the same.

[0261] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0262] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0263] Reference Figure 4 The present invention also provides a computer device 4, including: a memory 402 and a processor 401, and a computer program 403 stored on the memory 402. When the computer program 403 is executed on the processor 401, it implements the hierarchical federated learning method based on split meta-learning as described in any of the above methods.

[0264] The computer device 4 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 4 may include, but is not limited to, a processor 401 and a memory 402. Those skilled in the art will understand that... Figure 4 The computer device 4 is merely an example and does not constitute a limitation on the computer device 4. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0265] The processor 401 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0266] In some embodiments, the memory 402 may be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 402 may be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Furthermore, the memory 402 may include both internal and external storage units of the computer device 4. The memory 402 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 402 can also be used to temporarily store data that has been output or will be output.

[0267] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the hierarchical federated learning method based on split meta-learning as described in any of the above methods.

[0268] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0269] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0270] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0271] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0272] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A hierarchical federated learning method based on split meta-learning, characterized in that... The method specifically includes: The system comprises a client layer, an edge server layer, and a cloud server layer. The client layer includes multiple clients, the edge server layer includes multiple edge servers, and the cloud server layer includes at least one cloud server. A global model is initialized on a cloud server and distributed to various edge servers. The global model is then distributed to the corresponding clients through the edge server layer. This allows each client to update the feature extraction module of the global model locally based on local data, while freezing the classifier module of the global model and asynchronously uploading the updated feature extraction module parameters to the associated edge server. The edge server layer aggregates updated parameters from multiple clients' feature extraction modules and performs meta-optimization on the classifier module using a local validation dataset to generate an edge-local model. The model parameters are periodically collected from multiple edge servers based on the cloud server layer, and the global model is updated using a gradient-sensitive momentum aggregation strategy. Based on the updated global model received at the edge server layer, personalized fusion weights are calculated according to the loss difference between the edge local model and the updated global model on local validation data. Based on personalized fusion weights, the edge local model and the global model are weighted and fused to generate a personalized local model, which is then distributed to the corresponding client through the edge server; Specifically, the step of aggregating updated feature extraction module parameters from multiple clients based on the edge server layer and performing meta-optimization on the classifier module using a local validation dataset to generate an edge local model includes: Based on the updated feature extraction module parameters received from multiple clients at the edge server layer, and combined with the data volume ratio of each client, the aggregation weight coefficient is calculated, a weighted aggregation operation is performed, and the aggregated feature extraction parameters are generated. By using the aggregated feature extraction parameters through the edge server and combining them with the local validation dataset, a meta-optimization process is performed on the classifier module, and the classifier parameters are updated. The updated classifier parameters are combined with the aggregated feature extraction parameters to construct an edge-local model.

2. The method according to claim 1, characterized in that... The process of initializing a global model based on a cloud server and distributing it to various edge servers, and then distributing the global model to its respective clients through the edge server layer, specifically includes: An initial global model is constructed based on a normal distribution initialization strategy using a cloud server. The initial global model includes initialization parameters containing parameters for the feature extraction module and parameters for the classifier module. The cloud server distributes a lightweight diagnostic model to each edge server, collects the real-time computing load and storage status of each edge server, and obtains diagnostic results. Based on the diagnostic results, and according to the load levels of each edge server, the complete initial global model or a simplified initial global model is distributed to each edge server via the cloud server. Based on the network status of clients within the coverage area of ​​each edge server, the edge server adaptively distributes the clients to its respective clients.

3. The method according to claim 1, characterized in that... The process of enabling each client to locally update the feature extraction module of the global model based on local data, while simultaneously freezing the classifier module of the global model, and asynchronously uploading the updated feature extraction module parameters to the associated edge server, specifically includes: Configure local training tasks based on the client's local computing resources. Based on the configured local training tasks, sample small batches of data from the local dataset to execute local training cycles. Calculate the loss through forward propagation and backpropagate the gradient, updating only the parameters of the feature extraction module. After completing the preset local training cycle, the feature extraction parameter increment is calculated through the client. The feature extraction parameter increment is the difference between the locally updated parameters and the originally received feature extraction module parameters. The client asynchronously uploads the incremental feature extraction parameters to the associated edge server.

4. The method according to claim 1, characterized in that... The method of periodically collecting model parameters from multiple edge servers based on the cloud server layer and updating the global model using a gradient-sensitive momentum aggregation strategy specifically includes: Based on a preset synchronization cycle, the cloud server collects the corresponding model parameter sets from multiple edge servers. Based on the model parameter set, the validation loss gradient of each edge server model in the current round is calculated through the cloud server, and the spatial distribution differences among the model parameters in the model parameter set are analyzed to obtain the parameter dispersion index; Based on the validation loss gradient and parameter dispersion index, a dynamic momentum coefficient is calculated by the cloud server through a dynamic adjustment mechanism. This dynamic momentum coefficient is negatively correlated with the changing trend of the validation loss gradient and positively correlated with the parameter dispersion index. The historical global model parameters and the model parameter set are weighted and fused using dynamic momentum coefficients to perform a global model update and generate a new round of global model.

5. The method according to claim 1, characterized in that... The step of receiving the updated global model at the edge server layer and calculating personalized fusion weights based on the loss difference between the edge local model and the updated global model on local validation data specifically includes: The edge server receives updated global model parameters from the cloud server, compares these updated global model parameters with the parameters of its own maintained edge local model, and calculates the local model loss value and the global model loss value on the local validation dataset of the edge server. The difference in loss is obtained by using the calculated local model loss value and global model loss value through interpolation. The loss difference value is input into the hyperbolic tangent activation function for nonlinear transformation to generate a weight adjustment factor; Based on the weight adjustment factor, combined with the preset baseline retention rate and sensitivity scaling factor, the personalized fusion weight is calculated.

6. The method according to any one of claims 1 to 5, characterized in that... The feature extraction module is used to process time-series data collected by the client based on a lightweight Transformer architecture and a multi-head self-attention mechanism, extracting temporal feature vectors containing driving behavior intentions, including... It receives raw time-series data collected by the vehicle's onboard sensors from the client, and performs denoising and normalization preprocessing on the raw time-series data to generate standardized sequence data; The standardized sequence data is segmented and embedded to generate sequence embedding vectors; Positional encoding information is added to the sequence embedding vector to preserve temporal order dependencies, generating position-aware embedding vectors; The location-aware embedding vector is input into a multi-head self-attention mechanism. The attention distribution is calculated using the query matrix, key matrix, and value matrix. Contextual features are then aggregated to generate enhanced feature representations. Based on enhanced feature representation, a high-order temporal feature vector is extracted by performing nonlinear transformation and dimension adjustment through a feedforward neural network. Features at key time steps are selected from high-order temporal feature vectors based on attention weights to generate temporal feature vectors that contain driving behavior intentions.

7. The method according to any one of claims 1 to 5, characterized in that... The classifier module is used to extract spatial relationships and interaction risk features between vehicles by fusing features from multiple clients through graph convolution operations based on a graph neural network aggregation layer, including: Based on the initial feature vectors received from multiple clients by the edge server, and based on the relative positions or communication connection states between the vehicles corresponding to the clients, a dynamic graph structure is constructed. The graph nodes of the dynamic graph structure are clients, and the graph edges are connections between vehicles. The node features of the dynamic graph structure are aggregated by graph convolution operation, and the initial feature vector is transformed based on the adjacency matrix and degree matrix to generate enhanced node features that fuse multi-client features; The spatial positional relationship features and interaction risk features between vehicles are extracted from the enhanced node features. Based on the spatial positional relationship features and interaction risk features, collaborative scenario understanding is performed, and collaborative decision-making results are output.

8. The method according to any one of claims 1 to 5, characterized in that... The method further includes: Based on historical traffic accident data and traffic flow characteristics of the corresponding regions of the edge servers, a regional benchmark risk model is trained and maintained for each edge server, and benchmark model parameters are generated; Based on the driving behavior features extracted by the client from the local CAN bus data, the local feature extraction module is updated to generate incremental feature parameters. Based on the feature increment parameters, the classifier boundary of the regional benchmark risk model is adjusted by using a meta-optimization strategy through the edge server to generate optimized local model parameters; Based on the first loss difference between the optimized local model parameters and the baseline model parameters calculated on the local validation data using the edge server, a first personalized fusion weight is generated based on the first loss difference. The optimized local model parameters and the baseline model parameters are weighted and fused using the first personalized fusion weight to generate a personalized risk assessment model. The personalized risk assessment model is used for risk scoring and real-time adjustment of the vehicle's basic insurance premium rate.

9. A hierarchical federated learning system based on split meta-learning, characterized in that... The system specifically includes: The first federated learning module is used to construct a client layer, an edge server layer, and a cloud server layer. The client layer includes multiple clients, the edge server layer includes multiple edge servers, and the cloud server layer includes at least one cloud server. The second federated learning module is used to initialize a global model based on the cloud server and distribute it to various edge servers. The global model is then distributed to the corresponding clients through the edge server layer. The third federated learning module enables each client to locally update the feature extraction module of the global model based on local data, while freezing the classifier module of the global model and asynchronously uploading the updated feature extraction module parameters to the associated edge server; The fourth federated learning module aggregates updated parameters from multiple clients based on the edge server layer and performs meta-optimization on the classifier module using the local validation dataset to generate an edge-local model. The fifth federated learning module is used to periodically collect model parameters from multiple edge servers based on the cloud server layer and update the global model using a gradient-sensitive momentum aggregation strategy. The sixth federated learning module is used to receive the updated global model at the edge server layer and calculate personalized fusion weights based on the loss difference between the edge local model and the updated global model on local validation data. The seventh federated learning module is used to perform weighted fusion of the edge local model and the global model based on personalized fusion weights to generate a personalized local model, which is then distributed to the corresponding client through the edge server. Specifically, the step of aggregating updated feature extraction module parameters from multiple clients based on the edge server layer and performing meta-optimization on the classifier module using a local validation dataset to generate an edge local model includes: Based on the updated feature extraction module parameters received from multiple clients at the edge server layer, and combined with the data volume ratio of each client, the aggregation weight coefficient is calculated, a weighted aggregation operation is performed, and the aggregated feature extraction parameters are generated. By using the aggregated feature extraction parameters through the edge server and combining them with the local validation dataset, a meta-optimization process is performed on the classifier module, and the classifier parameters are updated. The updated classifier parameters are combined with the aggregated feature extraction parameters to construct an edge-local model.

Images