Multi-level federated edge learning methods, systems, and devices based on vehicle-to-everything (V2X) networks

Abstract

This application provides a multi-level federated edge learning method, system, and device based on the Internet of Vehicles (IoV). The method includes: constructing a multi-level federated edge learning framework; obtaining the target task; determining the parameter requirements for the global model to be constructed; distributing the parameter requirements to the vehicle terminal device layer hierarchically based on the federated edge learning framework; obtaining the state information of the vehicle terminal device layer; selecting the vehicle terminal device layer using a scheduling optimization algorithm based on multi-knapsack constraints to obtain target vehicle terminal devices; training the local model corresponding to each target vehicle terminal device according to the parameter requirements to obtain the training parameters of each local model; aggregating the training parameters to obtain the parameter information for the global model to be constructed; and constructing the global model corresponding to the target task based on the parameter information. This method solves the problem of uneven resource allocation among devices in the IoV environment, leading to low model training efficiency in the IoV.

Description

Technical Field

[0001] This application relates to the field of multi-level federated edge learning technology based on vehicle-to-everything (V2X) networks, and more specifically, to a multi-level federated edge learning method, a multi-level federated edge learning system, a multi-level federated edge learning device, and an electronic device based on V2X networks. Background Technology

[0002] With the development of big data technology, more and more applications are using data collected by smart devices for analysis. However, this also brings the risk of data privacy breaches. To address this issue, federated learning, a novel computing paradigm, has been proposed. Unlike traditional machine learning methods, federated learning distributes shared models between cloud-based servers and user devices, allowing users to train models locally using their own data and then upload the updated model parameters to the cloud. Since no raw data needs to be transmitted throughout the process, only the model parameters are transferred, thus effectively improving data privacy and security.

[0003] While federated learning offers significant protection for data privacy, it still faces challenges related to communication resources and latency. Due to varying communication distances between devices and unpredictable network conditions, the uploading of model parameters can suffer from high latency and low efficiency. This is especially true in dynamic environments like Vehicle-to-Everything (V2X), which involve numerous mobile devices (such as smart vehicles and roadside units), where network connections are more unstable, leading to more significant communication latency. Furthermore, in scenarios requiring more user participation in training, device scalability becomes a major bottleneck. The transmission of large amounts of model information also increases the burden on cloud server centers, impacting the overall system efficiency and responsiveness.

[0004] To address this issue, edge computing, as an emerging computing paradigm, is gradually being introduced into federated learning. Edge computing distributes the burden on cloud central servers by deploying computing tasks on edge servers close to the data source, shortening communication paths, reducing latency, and optimizing data processing efficiency. In the connected vehicle environment, edge computing can perform preliminary data processing and model update aggregation on distributed devices such as vehicles and roadside units before uploading the results to the cloud, thereby improving the real-time performance and accuracy of data processing in connected vehicles.

[0005] Currently, federated edge computing typically employs a three-layer architecture: a cloud central server, multiple edge servers, and user terminal devices. In the context of connected vehicles (V2V), terminal devices refer to in-vehicle equipment, roadside sensors, intelligent traffic lights, and other similar devices. However, the number of connected terminal devices remains limited by factors such as network bandwidth, device computing power, and signal coverage. Traditional federated learning algorithms, such as FedAvg, usually select a subset of devices for training, neglecting the devices' computing power and communication quality. In a V2V environment, some devices may be unable to upload model parameters in a timely manner due to poor signal quality or limited transmission capabilities. This not only prolongs the training cycle but also significantly reduces overall efficiency. Therefore, effectively selecting devices for training and optimizing resource allocation has become a crucial issue in V2V. Summary of the Invention

[0006] The main objective of this application is to provide a multi-level federated edge learning method, a multi-level federated edge learning system, a multi-level federated edge learning device, and an electronic device based on vehicle-to-everything (V2X) communication, so as to at least solve the problem of uneven resource allocation in the V2X environment, which leads to low efficiency of model training in V2X.

[0007] To achieve the above objectives, according to one aspect of this application, a multi-level federated edge learning method based on vehicle-to-everything (V2X) is provided, comprising: constructing a multi-level federated edge learning framework, wherein the federated edge learning framework is divided into a cloud center server layer, a regional edge server layer, an edge node layer, and an in-vehicle terminal device layer according to computing power from high to low; obtaining a target task; determining the parameter requirement information of the global model to be constructed corresponding to the target task; and, based on the federated edge learning framework, distributing the parameter requirement information hierarchically to the in-vehicle terminal device layers to obtain the status information of each in-vehicle terminal device layer. In this process, the state information characterizes the communication quality and computing power of the vehicle-mounted terminal device layer. Based on the state information, a scheduling optimization algorithm based on multi-knapsack constraints is used to select multiple target vehicle-mounted terminal devices. Local models corresponding to each target vehicle-mounted terminal device are trained according to the parameter requirement information to obtain training parameters for each local model. The training parameters are aggregated to obtain parameter information for the global model to be constructed. A global model corresponding to the target task is constructed based on the parameter information, wherein the global model is used at least for the autonomous driving control of the vehicle.

[0008] Optionally, the aggregation processing of the training parameters to obtain the parameter information of the global model to be built includes: uploading each of the training parameters to the corresponding edge server in the regional edge server layer, and determining the first latency information of each edge server, wherein the first latency information includes the time for the edge server to aggregate multiple training parameters and the time for the edge aggregated parameters obtained from the aggregation of multiple training parameters to be uploaded to the cloud center server layer; determining the target edge server whose first latency information is less than a set latency time, and controlling each target edge server to aggregate the training parameters to obtain the corresponding edge aggregated parameters; uploading each edge aggregated parameter to the cloud center server layer, and aggregating each edge aggregated parameter to obtain the parameter information of the global model to be built.

[0009] Optionally, uploading each of the training parameters to the corresponding edge server in the regional edge server layer includes: uploading each of the training parameters to the corresponding edge server in the regional edge server layer based on the orthogonal frequency division multiple access protocol.

[0010] Optionally, based on the state information, a scheduling optimization algorithm based on multi-knapsack constraints is used to select the vehicle terminal device layer to obtain multiple target vehicle terminal devices. This includes: determining second latency information for each vehicle terminal device in each vehicle terminal device layer based on the state information, wherein the second latency information includes the training time of the vehicle terminal device for the local model and the time for uploading the training parameters to the regional edge server layer and waiting for the training parameters to be aggregated; obtaining a set time threshold for each edge server in the regional edge server layer; and selecting the vehicle terminal device layer based on the scheduling optimization algorithm based on multi-knapsack constraints according to the set time threshold and the second latency information to obtain multiple target vehicle terminal devices.

[0011] Optionally, determining the time for uploading the training parameters to the regional edge server layer from the second delay information of each of the vehicle-mounted terminal devices in each of the vehicle-mounted terminal device layers includes: according to the first formula: Determine the time to upload the training parameters to the regional edge server layer. ,in, The data size of the training parameters, , For upload speed, Total available bandwidth for edge servers The bandwidth allocation ratio for the training parameters. For the channel gain of the vehicle terminal equipment, For transmission power, This represents the background noise power.

[0012] Optionally, determining the training time of the local model for each vehicle terminal device in the second delay information of each of the vehicle terminal devices in each of the vehicle terminal device layers includes: according to the second formula: The training time of the local model by the vehicle-mounted terminal device is determined, wherein, , The number of iterations for the local model, parameter It is related to the data scale and task complexity. The target accuracy of the local model. The CPU frequency allocated to the local model. The number of CPU cycles required to process one training sample for the local model.

[0013] Optionally, during the process of training the local model corresponding to each of the target vehicle terminal devices according to the parameter requirement information, the method further includes: training the local model with the goal of minimizing the loss function, and updating the model parameters of the local model using a stochastic gradient descent algorithm.

[0014] According to another aspect of this application, a multi-level federated edge learning system based on vehicle-to-everything (V2X) is provided, comprising: a controller for executing any of the aforementioned multi-level federated edge learning methods based on V2X; an in-vehicle terminal device layer including multiple in-vehicle intelligent devices and in-vehicle sensors for real-time acquisition of vehicle driving data; an edge node layer including multiple edge nodes, each edge node connecting to multiple in-vehicle intelligent devices and / or in-vehicle sensors, the edge node layer being used to acquire state information, train a local model, and upload training parameters; a regional edge server layer including multiple edge servers, the regional edge server layer being used to receive the training parameters uploaded by the edge node layer, and to aggregate and upload the training parameters; and a cloud center server layer for receiving edge aggregation parameters uploaded by the regional edge server layer, aggregating and processing the edge aggregation parameters to construct a global model, and publishing parameter requirement information for constructing a global model corresponding to a target task.

[0015] According to another aspect of this application, a multi-level federated edge learning device based on vehicle-to-everything (V2X) is provided, comprising: a first construction unit for constructing a multi-level federated edge learning framework, wherein the federated edge learning framework is divided into a cloud center server layer, a regional edge server layer, an edge node layer, and an in-vehicle terminal device layer according to computing power from high to low; and an acquisition unit for acquiring a target task, determining the parameter requirement information of the global model to be constructed corresponding to the target task, and distributing the parameter requirement information to the in-vehicle terminal device layer hierarchically according to the federated edge learning framework to acquire the status information of each in-vehicle terminal device layer, wherein the status information is... The state information characterizes the communication quality and computing power of the vehicle-mounted terminal device layer; the training unit is used to select multiple target vehicle-mounted terminal devices by using a scheduling optimization algorithm based on multiple knapsack constraints based on the state information, and to train the local model corresponding to each target vehicle-mounted terminal device according to the parameter requirement information to obtain the training parameters of each local model; the second construction unit is used to aggregate and process the training parameters to obtain the parameter information of the global model to be constructed, and to construct a global model corresponding to the target task according to the parameter information, wherein the global model is at least used for the autonomous driving control of the vehicle.

[0016] According to another aspect of this application, an electronic device is provided, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing any of the described multi-level federated edge learning methods based on vehicle networking.

[0017] This application employs a multi-level federated edge learning framework, which is divided into five layers based on computing power: a cloud center server layer, a regional edge server layer, an edge node layer, and an in-vehicle terminal device layer. The framework acquires the target task and determines the parameter requirements for the global model to be built corresponding to that task. Based on the federated edge learning framework, these parameter requirements are distributed hierarchically to the in-vehicle terminal device layer to obtain the status information of each layer. This status information represents the communication quality and computing power of each in-vehicle terminal device layer. Based on this status information, a multi-knapsack constraint-based scheduling optimization algorithm is used to select multiple target in-vehicle terminal devices. Local models corresponding to each target in-vehicle terminal device are trained according to the parameter requirements, yielding training parameters for each local model. The training parameters are aggregated to obtain the parameter information for the global model to be built. A global model corresponding to the target task is then constructed based on this parameter information, and this global model is used at least for the autonomous driving control of the vehicle. Through this hierarchical division, each layer of servers or devices processes its corresponding tasks according to its capabilities, thereby achieving efficient data processing and model training. Furthermore, the scheduling optimization algorithm based on multi-knapsack constraints can rationally select devices to participate in training according to the computing power and communication quality of the on-board terminal equipment, ensuring optimal training results with limited resources. This significantly improves the training speed and accuracy of the model, as well as the scalability and responsiveness of the system in large-scale dynamic geographical areas, providing strong support for autonomous driving control of vehicles. It also solves the problem of uneven resource allocation among devices in the connected vehicle environment, which leads to low model training efficiency in connected vehicle networks. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 A hardware structure block diagram of a mobile terminal implementing a multi-level federated edge learning method based on vehicle networking is shown in an embodiment of this application.

[0020] Figure 2 A flowchart illustrating a multi-level federated edge learning method based on vehicle networking provided in an embodiment of this application is shown.

[0021] Figure 3 A schematic diagram illustrating the hierarchical and distributional relationships of a multi-level federated edge learning system based on the Internet of Vehicles provided in an embodiment of this application is shown.

[0022] Figure 4A schematic diagram of an in-vehicle terminal device selection strategy provided according to an embodiment of this application is shown;

[0023] Figure 5 A comparison chart showing the total training time of the local model provided according to embodiments of this application is shown;

[0024] Figure 6 A comparison chart of the training accuracy of DR-Fed and TH-Fed provided according to embodiments of this application is shown;

[0025] Figure 7 A comparison chart of the training losses of DR-Fed and TH-Fed provided according to embodiments of this application is shown;

[0026] Figure 8 A structural block diagram of a multi-level federated edge learning device based on vehicle networking provided according to an embodiment of this application is shown. Detailed Implementation

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] As described in the background section, existing federated learning devices in the Internet of Vehicles (IoV) environment suffer from uneven resource allocation, resulting in low model training efficiency in IoV. To address this issue, embodiments of this application provide a multi-level federated edge learning method, a multi-level federated edge learning system, a multi-level federated edge learning device, and an electronic device based on the Internet of Vehicles.

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0032] The methods and embodiments provided in this application can be executed on a mobile terminal, computer terminal, or similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal based on a multi-level federated edge learning method for vehicle networking, according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0033] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the multi-level federated edge learning method based on vehicle networking in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the above-described networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the above-described networks may include wireless networks provided by the mobile terminal's communication provider. In one example, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0034] This embodiment provides a multi-level federated edge learning method based on vehicle networking that runs on a mobile terminal, computer terminal or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although the logical order is shown in the flowchart, in some cases the steps shown or described can be executed in a different order than that shown here.

[0035] Figure 2 This is a flowchart of a multi-level federated edge learning method based on vehicle-to-everything (V2X) according to an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0036] Step S201: Construct a multi-level federated edge learning framework, wherein the federated edge learning framework is divided into cloud center server layer, regional edge server layer, edge node layer and vehicle terminal device layer according to computing power from high to low.

[0037] The vehicle terminal equipment layer includes a large number of vehicle intelligent devices and vehicle sensors, such as intelligent driving systems, vehicle cameras, millimeter-wave radar, and driving recorders.

[0038] The edge node layer mainly consists of fixed equipment deployed in transportation infrastructure, such as roadside units (RSUs), street light gateways, and traffic signal controllers, and has a medium level of computing and communication capabilities. Each edge node connects to several vehicle-mounted terminal devices through V2I (vehicle-to-infrastructure) communication, and is responsible for coordinating the model uploading and preliminary aggregation tasks of the lower-level devices, and transmitting the results to higher levels.

[0039] The regional edge server layer, typically deployed in city-edge data centers or communication base stations, possesses strong computing power and wider-area connectivity. Nodes in this layer are responsible for integrating model information from multiple edge nodes and performing further model aggregation and intermediate updates, reducing the load on cloud center servers and improving system response speed.

[0040] The cloud central server layer, with nodes possessing the strongest computing and storage capabilities, is typically deployed in remote data centers. The cloud servers are responsible for receiving all aggregated model information from the intermediate layers, performing the final aggregation, optimization, and update of the global model, and distributing the new round of model parameters to the lower-layer edge nodes, forming a closed loop.

[0041] Step S202: Obtain the target task, determine the parameter requirement information of the global model to be built corresponding to the target task, and distribute the parameter requirement information to the vehicle terminal device layer in a hierarchical manner based on the federated edge learning framework to obtain the status information of each vehicle terminal device layer. The status information represents the communication quality and computing power of the vehicle terminal device layer.

[0042] The status information of the vehicle-mounted terminal equipment includes wireless communication channel quality (such as signal-to-noise ratio and bandwidth), computing power (such as CPU / GPU availability), and the scale and type of local data resources (such as the number of vehicle images and traffic scene categories).

[0043] Specifically, edge nodes perform preliminary aggregation of the received information and include their own resource information when uploading it to the upper-layer edge server. Finally, resource information from all levels is uniformly integrated and analyzed by the cloud center server to obtain a comprehensive view of the network's computing and communication resources.

[0044] Step S203: Based on the above state information, a scheduling optimization algorithm based on multi-knapsack constraints is used to select the above vehicle terminal equipment layer to obtain multiple target vehicle terminal equipment. The local model corresponding to each of the above target vehicle terminal equipment is trained according to the above parameter requirement information to obtain the training parameters of each of the above local models.

[0045] Specifically, based on aggregated multi-dimensional resource information and current model training requirements, the cloud center server employs a top-down, layer-by-layer scheduling approach for device selection. At the edge server layer, each node selects a group of edge nodes to participate in the model aggregation task based on the availability, network conditions, and data characteristics of the lower-level devices it manages. At the edge node layer closer to the vehicle-mounted terminal, vehicle-mounted devices that meet requirements such as latency, data volume, and model training capabilities are further selected to participate in the local training task of the current round. Device selection uses a scheduling optimization algorithm under multi-knapsack constraints to ensure that training and uploading are completed within the predetermined time, maximizing training efficiency and system resource utilization. The cloud center server distributes the initialized global model parameters from top to bottom to the selected edge nodes at each level and vehicle-mounted terminal devices through the multi-level edge transmission network, for use as the local model for the current training round.

[0046] Step S204: The above training parameters are aggregated to obtain the parameter information of the global model to be constructed, and a global model corresponding to the target task is constructed based on the above parameter information. The global model is used at least for the autonomous driving control of the vehicle.

[0047] In this embodiment, by applying steps S201, S202, S203, and S204, this hierarchical division allows each server or device at each level to handle corresponding tasks based on its own capabilities, thereby achieving efficient data processing and model training. Furthermore, the use of a multi-knapsack constraint-based scheduling optimization algorithm can rationally select devices participating in training based on the computing power and communication quality of the onboard terminal equipment, ensuring optimal training results with limited resources. This significantly improves the training speed and accuracy of the model, as well as the scalability and responsiveness of the system in large-scale dynamic geographical areas, providing strong support for autonomous driving control of vehicles. This solves the problem of uneven resource allocation among devices in the vehicle-to-everything (V2X) environment, which leads to low model training efficiency in V2X.

[0048] In the specific implementation process, the above-mentioned training parameters are aggregated and processed to obtain the parameter information of the global model to be constructed, including: uploading each of the above-mentioned training parameters to the corresponding edge server in the above-mentioned regional edge server layer, and determining the first latency information of each of the above-mentioned edge servers, wherein the first latency information includes the time for the edge server to aggregate multiple of the above-mentioned training parameters, and the time for the edge aggregated parameters obtained from the aggregation of multiple of the above-mentioned training parameters to be uploaded to the above-mentioned cloud center server layer; determining the target edge server whose first latency information is less than the latency setting time, and controlling each of the above-mentioned target edge servers to aggregate and process the above-mentioned training parameters to obtain the corresponding edge aggregated parameters; uploading each of the above-mentioned edge aggregated parameters to the cloud center server layer, and aggregating and processing each of the above-mentioned edge aggregated parameters to obtain the parameter information of the global model to be constructed.

[0049] The time setting can be configured according to specific upload requirements, such as 5 minutes, 10 minutes, etc.

[0050] This method effectively reduces the communication latency of the entire system and enhances its real-time performance and reliability by optimizing the aggregation process of training parameters. In federated learning, uploading and aggregating training parameters are critical steps, often accompanied by high communication costs and time consumption. To address this, a mechanism is designed to prioritize uploading training parameters to the regional edge server layer. These edge servers are closer to the data generation point, enabling them to complete parameter aggregation faster and upload the results to the cloud center. This strategy not only reduces the load on the cloud center servers but also ensures, by setting a latency threshold, that only edge servers that can complete aggregation within a specified time participate in the construction of the global model, thus avoiding overall efficiency degradation caused by a few slow servers. This method is particularly suitable for vehicle-to-everything (V2X) environments because network conditions may be unstable while vehicles are in motion. The intervention of regional edge servers can significantly improve data transmission efficiency and model training speed, creating favorable conditions for real-time autonomous driving control of vehicles.

[0051] Specifically, uploading the aforementioned training parameters to the corresponding edge server in the aforementioned regional edge server layer includes: uploading the aforementioned training parameters to the corresponding edge server in the aforementioned regional edge server layer based on the orthogonal frequency division multiple access protocol.

[0052] This method employs Orthogonal Frequency Division Multiple Access (OFDMA) protocol to upload training parameters, a key optimization measure in this technical solution. OFDMA is a highly efficient wireless communication technology that divides available spectrum resources into multiple orthogonal subcarriers, allowing multiple users to simultaneously use these subcarriers for data transmission, thereby significantly improving spectrum utilization and data transmission efficiency. In the multi-level federated edge learning framework of vehicle-to-everything (V2X) networks, the onboard terminal device layer needs to upload the training parameters of its local model to the regional edge server layer. Due to the large number and wide distribution of vehicles, traditional single-carrier communication methods may lead to severe spectrum congestion and communication delays. By introducing the OFDMA protocol, parallel uploading of training parameters can be achieved, maintaining good communication quality even in high-density vehicle networks. This ensures that training parameters arrive at the edge server in a timely and accurate manner, thereby accelerating the global model construction process and providing more accurate model support for autonomous driving control of vehicles.

[0053] More specifically, based on the aforementioned state information, a scheduling optimization algorithm based on multi-knapsack constraints is used to select the aforementioned vehicle-mounted terminal device layer to obtain multiple target vehicle-mounted terminal devices. This includes: determining second latency information for each of the aforementioned vehicle-mounted terminal devices in each of the aforementioned vehicle-mounted terminal device layers based on the aforementioned state information, wherein the second latency information includes the training time of the aforementioned vehicle-mounted terminal device for the aforementioned local model, and the time for uploading the aforementioned training parameters to the regional edge server layer and waiting for the training parameters to be aggregated; obtaining a set time threshold for each edge server in the aforementioned regional edge server layer; and, based on the aforementioned set time threshold and the aforementioned second latency information, using the aforementioned scheduling optimization algorithm based on multi-knapsack constraints to select the aforementioned vehicle-mounted terminal device layer to obtain multiple target vehicle-mounted terminal devices.

[0054] This method, based on a multi-knapsack constraint scheduling optimization algorithm, is a key technology in this technical solution for optimizing the selection of in-vehicle terminal devices. In the vehicle-to-everything (V2X) environment, there are significant differences in the computing power and communication quality of in-vehicle terminal devices, which directly affects the efficiency of local model training and the timeliness of parameter upload. To maximize the utilization of these resources within a limited time, a multi-knapsack constraint scheduling algorithm is designed, considering the computing power, communication quality, and processing capacity of the edge server. By dynamically adjusting the participation of in-vehicle terminal devices, the efficiency and fairness of the overall training process are ensured. Specifically, the algorithm intelligently selects in-vehicle terminal devices that can quickly complete local model training and upload training parameters in a timely manner based on the second latency information of the in-vehicle terminal devices (including local model training time and parameter upload waiting time) and the set time threshold of the edge server. This method not only improves the accuracy and speed of model training but also effectively balances the use of network resources, avoiding the problem of overload or idleness of individual devices, and providing a more stable and reliable model foundation for autonomous driving control of vehicles.

[0055] Further, determining the time for uploading the training parameters to the regional edge server layer in the second delay information of each of the aforementioned vehicle-mounted terminal devices in each of the aforementioned vehicle-mounted terminal device layers includes: according to the first formula: Determine the time to upload the above training parameters to the regional edge server layer. ,in, The data size for the above training parameters, , For upload speed, Total available bandwidth for edge servers The bandwidth allocation ratio for the above training parameters. For the channel gain of the vehicle terminal equipment, For transmission power, This represents the background noise power.

[0056] This method accurately assesses the time required for each onboard terminal device to upload training parameters, which is crucial for optimizing the entire training process. In a vehicle-to-everything (V2X) environment, communication conditions between vehicles are constantly changing. Employing precise mathematical models helps us dynamically adjust bandwidth allocation and transmission power, ensuring efficient and stable data transmission even under complex and variable network conditions. This accelerates the construction of the global model and provides more accurate and real-time decision-making support for autonomous vehicle control.

[0057] Furthermore, determining the training time of the local model for each of the aforementioned vehicle-mounted terminal devices in the second latency information of each of the aforementioned vehicle-mounted terminal devices in each of the aforementioned vehicle-mounted terminal device layers includes: according to the second formula: The training time for the aforementioned local model by the aforementioned vehicle-mounted terminal device is determined, wherein, , The number of iterations for the local model described above, parameters It is related to the data scale and task complexity. The target accuracy of the aforementioned local model. The CPU frequency allocated to the above local model, The number of CPU cycles required to process one training sample for the local model described above.

[0058] in, This represents the number of iterations required to achieve the target accuracy θ. This value is affected by the data size and task complexity, reflecting the basic workload of model training.

[0059] This method, by taking these parameters into account, can accurately predict the time required for each on-board terminal device to complete local model training, thereby optimizing device selection and scheduling and ensuring effective resource utilization. In the Internet of Vehicles (IoV), vehicles have varying computing capabilities. By using the second formula, training parameters can be dynamically adjusted according to actual conditions, making the allocation of computing resources more rational. This significantly shortens training time while ensuring model accuracy, providing more agile and intelligent decision support for autonomous driving control of vehicles.

[0060] Specifically, in the process of training the local model corresponding to each of the above target vehicle terminal devices based on the above parameter requirements information, the above method further includes: training the local model with the goal of minimizing the loss function, and updating the model parameters of the local model using a stochastic gradient descent algorithm.

[0061] In the local model training process, this method employs a strategy aimed at minimizing the loss function, combined with stochastic gradient descent (SGD) to update model parameters. This is one of the key technologies for improving model training efficiency and accuracy. The loss function measures the difference between the model's predictions and the actual results. By minimizing the loss function, we can make the model fit the data more accurately and improve its predictive ability. Stochastic gradient descent is an iterative optimization method that estimates the gradient direction by randomly selecting a subset of data, thereby updating the model parameters in each iteration. Compared to traditional batch gradient descent, SGD converges faster, especially on large datasets. In the multi-level federated edge learning framework of vehicle-to-everything (V2X) communication, the in-vehicle terminal device layer collects a large amount of real-time data. The processing of this data and model training need to be completed in a short time to meet the real-time control requirements of vehicles. By adopting the minimization of the loss function and the stochastic gradient descent algorithm, we can effectively accelerate the model training process while ensuring high-quality model output, providing a solid technical guarantee for autonomous driving control of vehicles. In addition, this method also has strong robustness, adapting to the dynamic changes of data in the in-vehicle terminal device layer, ensuring the stability and continuity of model training.

[0062] This embodiment also provides a multi-level federated edge learning system based on vehicle networking, including:

[0063] The controller is used to execute any of the above-mentioned multi-level federated edge learning methods based on vehicle networking;

[0064] The vehicle terminal equipment layer includes multiple vehicle intelligent devices and vehicle sensors for real-time collection of vehicle driving data.

[0065] The edge node layer includes multiple edge nodes, each of which is connected to multiple in-vehicle intelligent devices and / or in-vehicle sensors. The edge node layer is used to acquire status information, train local models, and upload training parameters.

[0066] The regional edge server layer includes multiple edge servers. The regional edge server layer is used to receive the training parameters uploaded by the edge node layer, and to aggregate and upload the training parameters.

[0067] The cloud center server layer is used to receive edge aggregation parameters uploaded by the regional edge server layer, aggregate and process the above edge aggregation parameters to build a global model, and publish parameter requirement information for the global model to be built corresponding to the target task.

[0068] This system achieves efficient data processing and model training through meticulous hierarchical partitioning and intelligent scheduling algorithms. The core of the system is the controller, which not only executes the multi-level federated edge learning method but also coordinates communication and computing resources at each level of the system, ensuring smooth training. The in-vehicle terminal device layer, as the data source, contains various in-vehicle intelligent devices and sensors that collect real-time vehicle driving data, providing a rich source of information for model training. The edge node layer acts as a bridge, connecting the in-vehicle terminal device layer and the regional edge server layer. It can not only obtain the status information of the in-vehicle terminal devices but also perform preliminary training and parameter uploading of the local model, reducing data transmission latency. The regional edge server layer and the cloud center server layer are responsible for data aggregation and processing and global model construction, respectively. Through this multi-level architectural design, efficient data processing and rapid model iteration are achieved. This system design not only solves the problems of inefficiency and uneven resource allocation faced by traditional federated learning in network-dense, heterogeneous vehicle-to-everything (V2X) environments, but also significantly improves the accuracy and speed of model training through collaborative work between different layers. This brings revolutionary changes to fields such as intelligent transportation and autonomous driving, especially in dynamic, high-concurrency scenarios of urban-level V2X networks, where its advantages are particularly evident. Through this system, we can achieve a significant reduction in training time and a substantial improvement in model performance even with large-scale in-vehicle terminal access, providing strong technical support for future intelligent driving.

[0069] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the multi-level federated edge learning method based on vehicle networking will be described in detail below with reference to specific embodiments.

[0070] This embodiment relates to a specific multi-level federated edge learning system based on vehicle-to-everything (V2X) networks. It is divided into layers 1 to Y from bottom to top according to computing power and network roles, realizing collaborative computing and communication optimization of multi-level federated edge learning. It is particularly suitable for dynamic, high-concurrency vehicle-to-everything (V2X) environments.

[0071] The first layer is the vehicle terminal equipment layer, which includes a large number of vehicle intelligent devices and vehicle sensors (such as intelligent driving systems, vehicle cameras, millimeter-wave radar, driving recorders, etc.). These devices collect real-time data on the vehicle's surrounding environment, such as vehicle speed, location, road conditions, traffic signals, etc., and perform local training based on the initial model distributed from the upper-layer nodes.

[0072] The second layer is the edge node layer, which mainly consists of fixed equipment deployed in transportation infrastructure, such as roadside units (RSUs), street light gateways, and traffic signal controllers, and has a medium level of computing and communication capabilities. Each edge node connects to several vehicle-mounted terminal devices through V2I (vehicle-to-infrastructure) communication, and is responsible for coordinating the model uploading and preliminary aggregation tasks of the lower-level devices, and transmitting the results to higher levels.

[0073] The third layer is the regional edge server layer, typically deployed in city edge data centers or communication base stations, possessing strong computing power and wider-area connectivity. Nodes in this layer are responsible for integrating model information from multiple edge nodes and performing further model aggregation and intermediate updates, reducing the load on cloud center servers and improving system response speed.

[0074] Similarly, the fourth layer of the system is the cloud center server layer. The nodes in this layer have the strongest computing and storage capabilities and are usually deployed in remote data centers. The cloud server is responsible for receiving the model information aggregated from all intermediate layers, performing the final aggregation, optimization, and update of the global model, and distributing the new round of model parameters to the lower-layer edge nodes to form a closed loop.

[0075] In the multi-level architecture of the system in this embodiment, considering the unstable communication links and strong node heterogeneity in the vehicle-to-everything (V2X) environment, a time constraint mechanism is introduced during model training and uploading. If a node at a certain level (including vehicles or edge servers) fails to complete the model upload within the predetermined time, its update will not be included in the current aggregation round, thus ensuring the synchronization and efficiency of the overall system training progress. This design effectively addresses issues such as model loss and training delays caused by dynamic communication conditions in V2X, improving the system's stability and convergence speed.

[0076] This embodiment relates to a specific multi-level federated edge learning method based on the Internet of Vehicles (IoV). In the multi-level federated edge learning algorithm framework proposed in this invention, there exists a set of mobile terminal devices in the IoV environment, denoted as […]. Among them, let This indicates that the first layer consists of the first... The set of vehicle client devices connected to each edge node. The first in the layer Each edge server is denoted as (in ), the top layer The layer is the only cloud center server, denoted as Each onboard device can collect and store tagged local traffic data in real time, forming its local dataset. ,in Indicates the first The first terminal device Each sample input (such as road image, vehicle status, location information, etc.) This is assigned to its corresponding target label (e.g., obstacle category, lane type, target behavior, etc.). The overall hierarchy and distribution of the equipment are as follows: Figure 3 As shown.

[0077] The specific multi-level federated edge learning method based on vehicle-to-everything (V2X) includes the following steps:

[0078] Step 1: Global model initialization;

[0079] The cloud server randomly initializes the global model parameters corresponding to the target task, serving as the starting model for this round of federated learning training. This model includes neural network structures for typical vehicle-to-everything (V2X) application tasks such as road recognition, path prediction, and driving behavior evaluation.

[0080] Step Two: Resource Information Collection and Summary;

[0081] The cloud center server distributes the parameter requirements and resource request instructions for the current model training task layer by layer from top to bottom to the edge servers at each level, down to the bottom edge nodes (such as roadside units, RSUs) closest to the vehicle terminals. Upon receiving the request, each edge node proactively sends a resource query request to all connected vehicle terminal devices. Each vehicle device, after responding, returns its current status information, including but not limited to:

[0082] 1. Wireless communication channel quality (such as signal-to-noise ratio, bandwidth);

[0083] 2. Computing power (e.g., CPU / GPU availability);

[0084] 3. Scale and type of local data resources (such as the number of vehicle images, traffic scene categories, etc.).

[0085] Edge nodes perform initial aggregation of the received information and include their own resource information when uploading it to the upper-layer edge server. Finally, resource information from all levels is uniformly integrated and analyzed by the cloud center server to obtain a comprehensive view of the network's computing and communication resources.

[0086] Step 3: Equipment selection and scheduling decisions;

[0087] Based on the multi-dimensional resource information summarized in step two and the current model training requirements, the cloud center server adopts a top-down, layer-by-layer scheduling approach for device selection. At the edge server layer, each node selects a group of edge nodes to participate in the model aggregation task based on the availability, network conditions, and data characteristics of the lower-level devices it manages. At the edge node layer closer to the vehicle terminal, vehicle devices that meet the requirements for latency, data volume, and model training capabilities are further selected to participate in the local training task of the current round. The device selection employs a scheduling optimization algorithm under multi-knapsack constraints to ensure that training and uploading are completed within the predetermined time, maximizing training efficiency and system resource utilization.

[0088] Step 4: Model distribution;

[0089] The cloud center server distributes the initialized global model parameters from top to bottom to the selected edge nodes and vehicle terminal devices at all levels through the edge multi-level transmission network, which are then used as the local initialization model for the current training round.

[0090] Step 5: Local model training and parameter uploading on the vehicle terminal;

[0091] In the federated learning system of this embodiment, after receiving the initial model from the edge server, the vehicle-mounted terminal device performs a local model training task and uploads the updated model parameters to the edge server it is connected to. This step is divided into two sub-stages: local model training and model upload.

[0092] 1) Local model training;

[0093] Suppose a certain vehicle-mounted device Participate in the current round (number) (round) training, its input data is The corresponding target label is The initial model parameters are real vectors. The data is distributed uniformly by the cloud center server. For each sample, its loss function is defined as follows: This represents the prediction error of the current model on this sample. (Device) Local dataset The overall loss function is defined as:

[0094] ;

[0095] The training objective is to minimize the loss function. To optimize the model's generalization performance, stochastic gradient descent is used during the local training process. (SGD) The algorithm updates the model parameters. Let the first... The update rule for the next iteration is:

[0096] ;

[0097] in, The learning rate is used. Due to data privacy protection requirements in the connected vehicle environment, data cannot be directly shared between devices; therefore, the global loss function is used. This cannot be obtained through centralized calculation. This invention employs each onboard device to calculate its local loss function locally. And approximate the global loss through weighted aggregation:

[0098] ;

[0099] ;

[0100] To achieve the common accuracy required by all local users The number of local iterations required for the user is: , where the parameters It is related to the scale of the data and the complexity of the task.

[0101] Furthermore, set up equipment The required processing time for each sample The number of cycles is The local allocation it received Frequency is The total computation delay is then This can be expressed as: .

[0102] 2) Upload local models;

[0103] After completing local training, the vehicle-mounted equipment Its local model parameters Uploaded to the corresponding edge server via wireless network Upload uses Orthogonal Frequency Division Multiple Access (OFDMA). The agreement is being implemented.

[0104] Assume the total available bandwidth of the edge server is ,equipment The bandwidth allocation ratio is Then its available bandwidth is Assume the device channel gain is... Transmission power is Background noise power is Therefore, its achievable upload speed is: ;

[0105] If the model parameter data size is Then the communication transmission delay is: .

[0106] Step Six: Edge Server Model Aggregation and Upload;

[0107] This step is divided into two sub-stages: edge layer model aggregation and model uploading.

[0108] 1) Edge layer model aggregation:

[0109] Second-tier edge server Receives model parameters uploaded by the connected vehicle-mounted device. And a weighted average is calculated based on the amount of data from the devices: ;

[0110] in, This represents the data set of all terminal devices under this edge server. The third-layer edge server further aggregates the data of its subordinate edge node models, according to the following rules: .

[0111] The same applies to the higher levels.

[0112] 2) Edge model upload:

[0113] Set up an edge server The aggregate model data size is Its uplink transmission rate is Then the delay in uploading model parameters is: .

[0114] Step 7: Cloud center server model aggregation;

[0115] Edge servers at each level transmit the aggregated model to upper-layer nodes via wireless communication, until it is aggregated at the cloud center server. The cloud server then performs the final aggregation of all received model parameters. .

[0116] Since the cloud center server has the highest computing power, its aggregation latency is negligible. After aggregation is complete, the cloud server distributes the updated global model from top to bottom to the edge servers and vehicle terminals, initiating the next round of federated training.

[0117] The device selection algorithm based on multi-knapsack constraints specifically includes the following:

[0118] In federated learning systems, increasing the number of in-vehicle terminal devices participating in model training helps improve the convergence speed and generalization ability of the global model. However, increasing the number of participating devices also brings higher communication and aggregation burdens. Especially in the vehicle-to-everything (V2X) environment, there are significant differences in channel status, transmission bandwidth, and physical distance between devices, which may cause some devices to be unable to complete model upload within a limited time, thereby affecting the overall training progress and efficiency.

[0119] In the connected vehicle environment, vehicle terminals exhibit high-frequency mobility and poor network link stability. Furthermore, the system contains a multi-level edge node structure. Therefore, during each round of federated model training and aggregation, not all on-board devices and edge servers can complete their assigned tasks on time. Upload timeouts or aggregation delays can lead to the loss of some model updates, thus limiting the overall performance improvement of the model.

[0120] To address the aforementioned issues, this invention proposes a device selection algorithm based on multi-knapsack constraints suitable for the Internet of Vehicles (IoV) environment. The algorithm aims to maximize the scheduling of available communication and computing resources within a fixed time window, selecting more edge servers and in-vehicle devices that meet the training and communication constraints to participate in model training. This improves training efficiency and model accuracy while ensuring system stability.

[0121] In step two of the above system model in this embodiment: resource information collection and aggregation, the cloud center server has mastered the resource information such as the communication channel status, computing power, and local dataset size of the entire network edge servers and their subordinate vehicle-mounted devices through the edge node hierarchical reporting mechanism, and has completed unified feature extraction and performance evaluation.

[0122] Based on this, the device selection algorithm based on multi-knapsack constraints mainly includes the following two sub-modules:

[0123] 1. Edge server selection:

[0124] In the multi-level federated learning system for vehicle networking proposed in this invention, the edge servers from layer 3 to layer Y act as intermediate aggregation nodes in each round of model training. Their main tasks are: to receive local model parameters from lower-layer edge servers or vehicle-mounted terminal devices, to complete the model aggregation operation, and to upload the aggregation results to the upper-layer server.

[0125] Let the model aggregation time of the edge server be... The required transmission latency for uploading the model is As shown in formula (11). Since the receiving latency of the edge server when receiving model parameters is much smaller than the latency of uploading the model, and is relatively stable in each iteration, it can be ignored. Therefore, the total time required for the edge server to complete one round of training is: ;

[0126] To ensure the system has good real-time performance and robustness in the vehicle-to-everything (V2X) environment, this invention introduces an edge server selection strategy: setting a fixed cutoff time threshold for each candidate edge server. This is used to control eligibility for participation in training. When a certain edge server can participate within the deadline... If the model aggregation and uploading tasks are completed within the specified time, the server will be included in the edge participation set for this round of training; otherwise, if the total time exceeds the threshold, the server will not be included in the aggregation calculation even if the model information has been received, in order to avoid slowing down the overall progress due to inefficient nodes.

[0127] This selection mechanism takes into account the characteristics of high deployment density, large computational differences, and unstable channel conditions of edge servers (such as base stations, MEC servers, traffic control centers, etc.) in the Internet of Vehicles, and effectively avoids the training bottleneck caused by unreliable nodes.

[0128] 2. Selection strategy for vehicle-mounted terminal equipment;

[0129] After downloading the model, in-vehicle terminal devices (such as intelligent computing modules, camera systems, and sensor processing units in vehicles) need to perform local model training and upload the updated model parameters to the edge nodes they are connected to. Since the latency of the terminal device receiving the downloaded model is much lower than the upload latency and the change is small, it can be ignored in the latency calculation.

[0130] Local model training time of computing terminal devices And the time required to upload the local model. Meanwhile, to facilitate overall system progress control, this embodiment assumes that the edge server still needs to complete one round of standard aggregation operations, which takes time. Therefore, the total time required for the terminal device to complete the upload and wait for aggregation from receiving the model is: .

[0131] This embodiment employs a greedy scheduling algorithm based on multi-knapsack constraints to select as many available terminal devices as possible to participate in the training task within a limited time window, thereby improving the training accuracy of the model.

[0132] The specific process is as follows:

[0133] 1. Set time thresholds for each edge server. This serves as the upper limit for the total training time of the client in this round.

[0134] 2. Iterate through all vehicle-mounted terminal devices connected to the current edge server and calculate the requirements for each candidate device. ;

[0135] 3. Prioritize the device with the shortest processing time, and record its time as [time value missing]. and add it to the current user set. ;

[0136] 4. If the total time consumed by the currently selected equipment Not exceeding the threshold Continue selecting the next optimal device until the accumulated time exceeds the limit.

[0137] The algorithm described above takes into account the realities of the Internet of Vehicles (IoV), such as the high dynamism of nodes, the high heterogeneity of tasks, and the limited channel resources. It can achieve efficient task scheduling and terminal utilization, and improve model convergence efficiency and system stability.

[0138] The execution process of the vehicle terminal equipment selection strategy can be found in [reference needed]. Figure 4 As shown, the edge server selection process and the vehicle terminal device selection strategy have similar control mechanisms.

[0139] To verify the effectiveness of the proposed multi-level federated edge learning system in a vehicle-to-everything (V2X) environment, a city-level V2X simulation environment was built, and image recognition was selected as the test benchmark to simulate the automatic recognition process of road digital signs by vehicles. The publicly available image dataset MNIST was used in the experiment to simulate the image acquisition and recognition tasks of traffic signs and lane numbers by in-vehicle terminal devices. The neural network model used was MobileNetV2, which is lightweight, low-latency, and easy to deploy at the edge, making it suitable for deployment and execution in in-vehicle intelligent terminals with limited computing resources. For each in-vehicle terminal device, Mini-Batch Stochastic Gradient Descent (Mini-Batch SGD) was used for model training.

[0140] This invention sets three key time parameters in the vehicle-to-everything (V2X) scenario to simulate scheduling control under dynamic heterogeneous networks:

[0141] 1. Maximum response time threshold for edge servers 10 minutes;

[0142] 2. Client terminal training deadline 10 minutes;

[0143] 3. Maximum total training time 360 minutes.

[0144] To verify the scalability and training efficiency advantages of the multi-level federated edge learning system proposed in this invention in the vehicle-to-everything (V2X) environment, two system architectures were constructed and comparative experiments were conducted:

[0145] 1) One is a traditional two-tier federated learning architecture (vehicle terminal) (Cloud Center)

[0146] 2) Another type is the multi-layer architecture (vehicle terminal) proposed in this invention. Roadside Unit (RSU) Regional edge server (Cloud Center).

[0147] This experiment deployed different numbers of vehicle-mounted terminal devices (such as moving vehicles) in a simulated urban road network. The simulated task was image classification (such as digit recognition, traffic sign classification, etc.), and the target accuracy was set to an accuracy rate of 0.7.

[0148] To simulate system operation under different client scales, the number of vehicle terminals was set to 300, 450, and 600 respectively, with corresponding server tiers. Specific deployment details are shown in Table 1 (Server Deployment Configuration for Different Numbers of Vehicle Terminals in a Multi-Tier Architecture) and Table 2 (Server Deployment Configuration for Different Numbers of Vehicle Terminals in a Traditional Two-Tier Architecture):

[0149] Table 1

[0150]

[0151] Table 2

[0152]

[0153] In a multi-tiered architecture, the second tier (RSU) is... The sampling rate is selected from the devices participating in the training from the connected vehicle terminals; the third-layer regional edge server and the cloud center server respectively use... The sampling rate is used to select edge servers from its lower-level nodes to participate in aggregation. Traditional two-layer federated learning architectures uniformly use... The ratio of devices selected for training from all vehicle terminals.

[0154] The system operation results are as follows Figure 5 As shown, the horizontal axis represents the number of vehicle-mounted terminals, and the vertical axis represents the total training time required to achieve the set training accuracy (accuracy 0.7).

[0155] The experimental results show that:

[0156] 1. When the number of vehicle terminals is relatively small (300 units), the traditional two-layer architecture has a slightly better training time due to the shorter aggregation path;

[0157] 2. However, as the number of terminals increases, the efficiency of the three-layer, multi-level federated architecture gradually surpasses that of traditional architectures. This is because the intermediate edge layer (RSU and regional edge servers) introduced in this invention effectively shortens the communication path between vehicles and aggregation nodes, reduces single-hop transmission latency, and lowers the risk of communication congestion. Furthermore, under the same terminal scale, the multi-level architecture can alleviate the pressure on cloud servers through distributed aggregation, eliminating the need for cloud servers to process data uploaded by all terminals, effectively improving the overall scalability and scheduling flexibility of the system, making it particularly suitable for model training tasks in large-scale dynamic vehicle networking environments.

[0158] To further verify the adaptability and performance of the proposed "Device Selection Algorithm Based on Multi-Knapsack Constraints" (hereinafter referred to as DR-Fed) in complex vehicle networking environments, a multi-level federated learning system was constructed, and a comparative experiment was conducted with the traditional strategy (hereinafter referred to as TH-Fed) as the benchmark.

[0159] This experiment simulates a typical urban road traffic scenario, deploying a four-level node structure, including vehicle-mounted terminals, roadside edge nodes, regional edge servers, and cloud center servers, as follows:

[0160] First layer: 400 in-vehicle intelligent terminal devices (such as vehicles equipped with intelligent chips);

[0161] The second layer consists of 40 RSUs (Roadside Units), responsible for the initial aggregation of local vehicle data;

[0162] The third layer consists of 20 regional edge servers.

[0163] Fourth layer: 1 cloud center server.

[0164] To be closer to reality, the following assumptions are made:

[0165] 1. Roadside and regional edge servers are fixed devices with continuous power supply and computing capabilities; therefore, their model aggregation time is assumed to be... =0;

[0166] 2. The channel conditions between vehicle terminals are approximately balanced, and the main influencing factor is the equipment's computing power;

[0167] 3. An image classification task is used to reflect application scenarios such as vehicle recognition of road traffic signs and speed markings.

[0168] Comparison algorithm settings:

[0169] 1. DR-Fed (algorithm in this embodiment):

[0170] (1) First-layer vehicle terminal: proportional Select equipment;

[0171] (2) Set the client deadline If the upload fails to complete within a specified time, it will be discarded.

[0172] (3) The second and third layer edge servers are arranged according to Make your selection and set the transmission deadline. =10 minutes;

[0173] (4) The aggregation strategy used is FedProx, which is suitable for non-independent identically distributed (non-IID) data and improves the convergence of the global model;

[0174] (5) The second layer performs 60 aggregations, and the third layer and the cloud center each perform 1 aggregation.

[0175] 2. TH-Fed (Traditional Three-Tier Federal Strategy):

[0176] (1) All nodes at all levels are randomly selected using a fixed sampling ratio, with no awareness of resource status;

[0177] (2) Without considering client latency or device capabilities, time-out nodes are eliminated solely based on time thresholds;

[0178] (3) The aggregation strategy used is the traditional FedAvg;

[0179] (4) The number of polymerizations was kept consistent with that of DR-Fed for comparative analysis.

[0180] 2. Neural network model configuration:

[0181] To evaluate the adaptability of the algorithm under different model complexities, as shown in Table 3, the following two types of vehicle-mounted terminal-adapted neural networks were selected:

[0182] Table 3

[0183]

[0184] Both models are deployed on in-vehicle terminals, with the input data being the MNIST extended dataset (used for simulating traffic numbers and sign recognition).

[0185] Analysis of experimental results:

[0186] 1. Figure 6 The difference in training accuracy between DR-Fed and TH-Fed under the two model configurations is shown; it can be seen that DR-Fed can dynamically schedule computing and communication resources, allowing more available on-board devices to participate in training in a timely manner, thereby accelerating the convergence of the global model;

[0187] 2. Figure 7 The convergence process of the training loss is demonstrated; DR-Fed shows a more stable loss decrease throughout the training cycle, and the final value is significantly lower than TH-Fed, demonstrating better robustness and convergence performance.

[0188] In summary, the DR-Fed algorithm proposed in this embodiment can not only effectively expand the available training resources, but also has excellent latency tolerance and scheduling flexibility, making it particularly suitable for dynamic, high-density urban-level vehicle-to-everything (V2X) application scenarios.

[0189] This implementation addresses the characteristics of the Vehicle-to-Everything (V2X) environment, including the widespread distribution of vehicle-mounted terminals, heterogeneous computing capabilities, and sensitivity to communication latency. It proposes a multi-level federated edge learning system. This architecture constructs a multi-level collaborative computing system consisting of vehicle-mounted terminal devices, roadside units (RSUs), regional edge servers, and a cloud center. It fully leverages the aggregation and buffering effects of edge computing nodes, effectively reducing the load on the cloud center server and improving the system's scalability and responsiveness across large-scale dynamic geographical areas. Simulation results show that, with an increased number of vehicle-mounted terminals, compared to the traditional two-layer federated learning architecture, this embodiment significantly reduces the training time required to achieve the same model accuracy, demonstrating improved overall system communication efficiency and computational collaboration capabilities.

[0190] Furthermore, based on the aforementioned multi-level federated learning architecture, this embodiment proposes a device selection algorithm based on multi-knapsack constraints. This algorithm, based on a greedy selection principle, comprehensively considers the computing power, communication latency, and task completion time of onboard devices, dynamically selecting onboard terminals and edge nodes that meet the requirements of the current training round to participate in model updates. Compared with traditional random scheduling strategies (such as FedAvg), the device selection algorithm based on multi-knapsack constraints can include more effective devices in training within the same time window, significantly improving the training accuracy, convergence speed, and robustness of the system model in heterogeneous vehicle-to-everything (V2X) scenarios. Experimental results verify that the algorithm exhibits higher stability and adaptability under various neural network models and different data distribution conditions, making it particularly suitable for application in large-scale, high-density deployed intelligent connected vehicle systems.

[0191] This application also provides a multi-level federated edge learning device based on vehicle-to-everything (V2X) communication. It should be noted that this V2X-based multi-level federated edge learning device can be used to execute the multi-level federated edge learning method for V2X communication provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0192] The following describes the multi-level federated edge learning device based on the Internet of Vehicles provided in the embodiments of this application.

[0193] Figure 8 This is a schematic diagram of a multi-level federated edge learning device based on the Internet of Vehicles according to an embodiment of this application. Figure 8 As shown, the device includes:

[0194] The first building unit 81 is used to build a multi-level federated edge learning framework, wherein the federated edge learning framework is divided into cloud center server layer, regional edge server layer, edge node layer and vehicle terminal device layer according to computing power from high to low.

[0195] The acquisition unit 82 is used to acquire the target task, determine the parameter requirement information of the global model to be built corresponding to the target task, and distribute the parameter requirement information to the vehicle terminal device layer in a hierarchical manner based on the federated edge learning framework to acquire the status information of each vehicle terminal device layer. The status information represents the communication quality and computing power of the vehicle terminal device layer.

[0196] Training unit 83 is used to select the vehicle terminal device layer based on the above state information using a scheduling optimization algorithm based on multi-knapsack constraints, to obtain multiple target vehicle terminal devices, and to train the local model corresponding to each of the above target vehicle terminal devices according to the above parameter requirement information, so as to obtain the training parameters of each of the above local models.

[0197] The second construction unit 84 is used to aggregate and process the above training parameters to obtain the parameter information of the global model to be constructed, and to construct a global model corresponding to the above target task based on the above parameter information, wherein the global model is used at least for the autonomous driving control of the vehicle.

[0198] In this embodiment, the first construction unit is used to construct a multi-level federated edge learning framework, wherein the federated edge learning framework is divided into a cloud center server layer, a regional edge server layer, an edge node layer, and an in-vehicle terminal device layer according to computing power from high to low; the acquisition unit is used to acquire the target task, determine the parameter requirement information of the global model to be constructed corresponding to the target task, and distribute the parameter requirement information to the in-vehicle terminal device layer in a hierarchical manner based on the federated edge learning framework to obtain the status information of each in-vehicle terminal device layer, wherein the status information represents the communication quality and computing power of the in-vehicle terminal device layer; the training unit is used to select the in-vehicle terminal device layer based on the status information and using a scheduling optimization algorithm based on multi-knapsack constraints to obtain multiple target in-vehicle terminal devices, and train the local model corresponding to each target in-vehicle terminal device according to the parameter requirement information to obtain the training parameters of each local model; the second construction unit is used to aggregate and process the training parameters to obtain the parameter information of the global model to be constructed, and construct the global model corresponding to the target task according to the parameter information, wherein the global model is used at least for the autonomous driving control of the vehicle.

[0199] As an optional solution, the second construction unit includes a first determining module, a first aggregation processing module, and a second aggregation processing module. The first determining module uploads each of the aforementioned training parameters to the corresponding edge server in the aforementioned regional edge server layer, and determines the first latency information of each of the aforementioned edge servers. The first latency information includes the time for the edge server to aggregate the multiple training parameters, and the time for the edge aggregated parameters obtained from the aggregation of the multiple training parameters to be uploaded to the aforementioned cloud center server layer. The first aggregation processing module is used to determine the target edge server whose first latency information is less than a set latency time, and controls each of the aforementioned target edge servers to perform aggregation processing on the aforementioned training parameters to obtain the corresponding edge aggregated parameters. The second aggregation processing module is used to upload each of the aforementioned edge aggregated parameters to the cloud center server layer, and perform aggregation processing on each of the aforementioned edge aggregated parameters to obtain the aforementioned parameter information of the global model to be constructed.

[0200] In one alternative approach, the first determining module includes an uploading submodule, used to upload the aforementioned training parameters to the corresponding edge server in the aforementioned regional edge server layer based on the orthogonal frequency division multiple access protocol.

[0201] In one optional scheme, the training unit includes a second determining module and a first acquiring module. The second determining module is used to determine the second latency information of each of the aforementioned vehicle terminal devices in each of the aforementioned vehicle terminal device layers based on the aforementioned state information. The second latency information includes the training time of the aforementioned vehicle terminal device on the aforementioned local model and the time for uploading the aforementioned training parameters to the regional edge server layer and waiting for the training parameters to be aggregated. The first acquiring module is used to acquire the set time threshold of each edge server in the aforementioned regional edge server layer. Based on the aforementioned set time threshold and the aforementioned second latency information, the aforementioned scheduling optimization algorithm based on multi-knapsack constraints is used to select the aforementioned vehicle terminal device layers to obtain multiple of the aforementioned target vehicle terminal devices.

[0202] In one alternative approach, the second determining module includes a first determining submodule, used to determine according to a first formula: Determine the time to upload the above training parameters to the regional edge server layer. ,in, The data size for the above training parameters, , For upload speed, Total available bandwidth for edge servers The bandwidth allocation ratio for the above training parameters. For the channel gain of the vehicle terminal equipment, For transmission power, This represents the background noise power.

[0203] In one alternative approach, the second determining module includes a second determining submodule, used to determine according to the second formula: The training time for the aforementioned local model by the aforementioned vehicle-mounted terminal device is determined, wherein, , The number of iterations for the local model described above, parameters It is related to the data scale and task complexity. The target accuracy of the aforementioned local model. The CPU frequency allocated to the above local model, The number of CPU cycles required to process one training sample for the local model described above.

[0204] In an optional embodiment, the apparatus further includes an update unit, which, during the training of the local model corresponding to each of the aforementioned target vehicle-mounted terminal devices based on the aforementioned parameter requirement information, trains the local model with the objective of minimizing the loss function and updates the model parameters of the local model using a stochastic gradient descent algorithm.

[0205] The aforementioned multi-level federated edge learning device based on vehicle-to-everything (V2X) communication includes a processor and a memory. The first construction unit, acquisition unit, training unit, and second construction unit are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules reside in the same processor; alternatively, the modules may be located in different processors in any combination.

[0206] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can address the problem of uneven resource allocation in connected vehicle environments, which leads to low model training efficiency in connected vehicle environments.

[0207] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0208] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the aforementioned multi-level federated edge learning method based on vehicle networking.

[0209] This invention provides a processor for running a program, wherein the program executes the aforementioned multi-level federated edge learning method based on vehicle networking.

[0210] This invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements at least the steps of the multi-level federated edge learning method based on vehicle networking described above. The device in this document can be a server, PC, PAD, mobile phone, etc.

[0211] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform an initialization procedure having at least the steps of the above-described multi-level federated edge learning method based on vehicle networking.

[0212] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0213] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0214] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0215] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0216] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0217] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0218] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0219] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0220] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0221] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-level federated edge learning method based on vehicle-to-everything (V2X) communication, characterized in that, include: A multi-level federated edge learning framework is constructed, wherein the federated edge learning framework is divided into cloud center server layer, regional edge server layer, edge node layer and vehicle terminal device layer according to computing power from high to low. The target task is obtained, and the parameter requirement information of the global model to be built corresponding to the target task is determined. Based on the federated edge learning framework, the parameter requirement information is distributed to the vehicle terminal device layer in a hierarchical manner to obtain the status information of each vehicle terminal device layer. The status information represents the communication quality and computing power of the vehicle terminal device layer. Based on the state information, a scheduling optimization algorithm based on multi-knapsack constraints is used to select the vehicle terminal device layer to obtain multiple target vehicle terminal devices. Then, the local model corresponding to each target vehicle terminal device is trained according to the parameter requirement information to obtain the training parameters of each local model. The training parameters are aggregated to obtain the parameter information of the global model to be built, and a global model corresponding to the target task is constructed based on the parameter information, wherein the global model is used at least for the autonomous driving control of the vehicle.

2. The method according to claim 1, characterized in that, The parameter information of the global model to be built is obtained by aggregating and processing the training parameters, including: Each of the training parameters is uploaded to the corresponding edge server in the regional edge server layer, and the first latency information of each edge server is determined, wherein the first latency information includes the time for the edge server to aggregate multiple training parameters, and the time for the edge aggregated parameters obtained by aggregating multiple training parameters to be uploaded to the cloud center server layer; The target edge server whose first delay information is less than the delay set time is determined, and each target edge server is controlled to perform aggregation processing on the training parameters to obtain the corresponding edge aggregation parameters; Each edge aggregation parameter is uploaded to the cloud center server layer, and the edge aggregation parameters are aggregated and processed to obtain the parameter information of the global model to be built.

3. The method according to claim 2, characterized in that, Uploading each of the training parameters to the corresponding edge server in the region edge server layer includes: Based on the orthogonal frequency division multiple access protocol, each of the training parameters is uploaded to the corresponding edge server in the region edge server layer.

4. The method according to claim 1, characterized in that, Based on the state information, a scheduling optimization algorithm based on multi-knapsack constraints is used to select multiple target vehicle terminal devices, including: Based on the state information, the second latency information of each vehicle terminal device in each vehicle terminal device layer is determined, wherein the second latency information includes the training time of the vehicle terminal device for the local model and the time for uploading the training parameters to the regional edge server layer and waiting for the training parameters to be aggregated. A set time threshold is obtained for each edge server in the regional edge server layer. Based on the set time threshold and the second delay information, the scheduling optimization algorithm based on multi-knapsack constraints is used to select the vehicle terminal device layer to obtain multiple target vehicle terminal devices.

5. The method according to claim 4, characterized in that, Determining the time in the second latency information of each of the aforementioned vehicle-mounted terminal devices in each of the vehicle-mounted terminal device layers for uploading the training parameters to the regional edge server layer includes: According to the first formula: Determine the time to upload the training parameters to the regional edge server layer. ,in, The data size of the training parameters, , For upload speed, Total available bandwidth for edge servers The bandwidth allocation ratio for the training parameters. For the channel gain of the vehicle terminal equipment, For transmission power, This represents the background noise power.

6. The method according to claim 4, characterized in that, Determining the training time of the local model for each of the vehicle terminal devices in the second latency information of each of the vehicle terminal devices in each of the vehicle terminal device layers includes: According to the second formula: The training time of the local model by the vehicle-mounted terminal device is determined, wherein, The training time is mentioned above. , The number of iterations for the local model, parameter It is related to the data scale and task complexity. The target accuracy of the local model. The CPU frequency allocated to the local model. The number of CPU cycles required to process one training sample for the local model.

7. The method according to claim 1, characterized in that, During the training of the local model corresponding to each of the target vehicle-mounted terminal devices based on the parameter requirement information, the method further includes: The local model is trained with the goal of minimizing the loss function, and the model parameters of the local model are updated using the stochastic gradient descent algorithm.

8. A multi-level federated edge learning system based on vehicle-to-everything (V2X) communication, characterized in that, include: A controller, the controller being configured to execute the multi-level federated edge learning method based on vehicle networking as described in any one of claims 1 to 7; The vehicle terminal equipment layer includes multiple vehicle intelligent devices and vehicle sensors for real-time collection of vehicle driving data. An edge node layer, comprising multiple edge nodes, each edge node being connected to multiple in-vehicle intelligent devices and / or in-vehicle sensors, the edge node layer being used to acquire status information, train a local model, and upload training parameters; A regional edge server layer, comprising multiple edge servers, is used to receive the training parameters uploaded by the edge node layer, and to aggregate and upload the training parameters. The cloud center server layer is used to receive edge aggregation parameters uploaded by the regional edge server layer, aggregate and process the edge aggregation parameters to construct a global model, and publish parameter requirement information for the global model to be constructed corresponding to the target task.

9. A multi-level federated edge learning device based on vehicle-to-everything (V2X) communication, characterized in that, include: The first building unit is used to build a multi-level federated edge learning framework, wherein the federated edge learning framework is divided into cloud center server layer, regional edge server layer, edge node layer and vehicle terminal device layer according to computing power from high to low. An acquisition unit is used to acquire a target task, determine the parameter requirement information of the global model to be built corresponding to the target task, and distribute the parameter requirement information to the vehicle terminal device layer in a hierarchical manner based on the federated edge learning framework, so as to obtain the status information of each vehicle terminal device layer, wherein the status information represents the communication quality and computing power of the vehicle terminal device layer. The training unit is used to select multiple target vehicle terminal devices by using a scheduling optimization algorithm based on multiple knapsack constraints based on the state information, and to train the local model corresponding to each target vehicle terminal device according to the parameter requirement information to obtain the training parameters of each local model. The second construction unit is used to aggregate and process the training parameters to obtain the parameter information of the global model to be constructed, and to construct a global model corresponding to the target task based on the parameter information, wherein the global model is used at least for the autonomous driving control of the vehicle.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing the multi-level federated edge learning method based on vehicle networking as described in any one of claims 1 to 6.

Images