Vehicle Task Offloading System and Method Based on Federated Learning

Abstract

The present invention discloses a vehicle task offloading system and method based on federated learning, which relates to the technical field of mobile edge computing in the Internet of Vehicles. The method includes dividing a city into different service areas, and each service area contains fixed fog nodes; the regional head in the service area detects vehicles entering or leaving the service area in real time through the existing cellular network registration mechanism, and obtains their location and available computing resource information; the regional head is responsible for distributing the initial model to the fog nodes in the service area and maintaining the blockchain ledger; the client vehicle performs model training locally according to the hardware configuration and computing tasks, and aggregates them into a global model through federated learning, which is recorded in the blockchain ledger. Therefore, by adopting the vehicle task offloading system and method based on federated learning, problems such as uneven distribution of computing resources, resource competition, and task processing delay in the task offloading process in the existing vehicle Internet of Things environment can be overcome, and the effect of model training can be improved and the performance of the system can be enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle networking mobile edge computing, and particularly to a vehicle task offloading system and method based on federated learning. Background Art

[0002] With the development of intelligent driving technology, modern vehicles are equipped with more and more high-performance sensors and computing devices, such as lidar, high-definition cameras, radar sensors, and advanced driver assistance systems (ADAS). The amount of data generated in real time is extremely large, but the computing power and storage space of the vehicle itself are limited, and it is unable to efficiently process and store this large amount of data, resulting in the computing tasks often exceeding the vehicle's load capacity. If these tasks are uploaded to the cloud for processing, although the powerful resources of cloud computing can be utilized, problems such as limited network bandwidth, high communication latency, and network reliability will be faced, and the requirements of real-time and low latency cannot be met. These problems seriously restrict the further development and application of intelligent driving technology.

[0003] Vehicle Fog Computing (VFC) has gradually become an effective solution. The core idea of VFC is to move the computing and storage resources to the network edge, that is, the vehicle can offload the computing tasks to the nearby fog nodes for processing. These fog nodes are usually deployed on infrastructure such as roadside units (RSUs), traffic lights, intelligent road signs, billboards, etc., and have certain computing and storage capabilities. By offloading tasks to the nearby fog nodes, the vehicle can significantly reduce its dependence on cloud computing resources, reduce the demand for network communication bandwidth, and improve the real-time performance and response speed of data processing. However, when a large number of vehicles offload tasks to the fog nodes at the same time and in the same area, the computing and storage resources of the fog nodes may be insufficient, the task processing latency increases, and even the tasks may not be completed in time, affecting the normal operation of the vehicle and the user experience.

[0004] Federated Learning (FL) is an emerging distributed machine learning method aimed at fully utilizing the computing power of distributed devices for model training while ensuring data privacy. In FL, vehicles do not need to upload local data to a central server. Instead, they use local data to train models on the device and then only transfer model parameters or gradients to fog nodes or cloud servers for aggregation and update. In the field of the Internet of Vehicles (IoV), since vehicles need to share a large amount of data and models to achieve functions such as cooperative driving and traffic optimization, FL can achieve distributed model training and optimization without revealing the privacy data of vehicles. This not only improves the efficiency of task processing but also optimizes the data training process to adapt to the characteristics of high-speed vehicle movement and dynamic network changes. In addition, FL can also reduce the communication burden on the network and avoid network congestion and latency caused by uploading a large amount of data.

[0005] However, traditional federated learning schemes usually do not consider the task heterogeneity and dynamic changes among vehicles. Simple aggregation of models may lead to uneven utilization of computing resources and affect the overall performance of the system. Especially when the tasks generated by vehicles have non-independent and identically distributed characteristics, direct aggregation not only consumes more computing resources but also results in a decline in training effects.

[0006] Therefore, it is necessary to improve the traditional federated learning scheme according to the characteristics of the IoV environment, considering the task heterogeneity and dynamic changes among vehicles, in order to improve the effect of model training and the performance of the system. Summary of the Invention

[0007] The object of the present invention is to provide a vehicle task offloading system and method based on federated learning, which can optimize the vehicle task offloading strategy, reduce latency, and improve the overall computing efficiency of the system while ensuring data privacy, and overcome problems such as uneven distribution of computing resources, resource competition, and task processing latency in the task offloading process in the existing vehicle Internet of Things environment.

[0008] To achieve the above object, the present invention provides a vehicle task offloading method based on federated learning, including the following steps:

[0009] S1. Divide the city into different service areas, and each service area contains fixed fog nodes that coexist with roadside infrastructure;

[0010] S2. The area head in the service area uses the existing cellular network registration mechanism to detect vehicles entering or leaving the service area in real time. When a vehicle enters the service area, it reports its location and available computing resources to the area head;

[0011] S3. The regional head distributes the initial concurrent overhead model to the fog nodes within the service area and is responsible for maintaining the blockchain ledger to share computing resources and task offloading information;

[0012] S4. The client vehicle locally trains the initial concurrent overhead model according to its hardware configuration and computing tasks, and aggregates it into a global concurrent overhead model through federated learning, which is recorded in the blockchain ledger.

[0013] Preferably, S4 includes:

[0014] S41. The client vehicle registers with the regional head and receives the initial concurrent overhead model;

[0015] S42. The client vehicle locally trains the initial concurrent overhead model based on historical task data and current task concurrency;

[0016] S43. The client vehicle uploads the locally trained concurrent overhead model to the corresponding fog node and aggregates the concurrent overhead models of different fog nodes using federated learning;

[0017] S44. Distribute the aggregated global concurrent overhead model to all client vehicles and repeat the training until the set number of training rounds is reached.

[0018] Preferably, the local training includes the client vehicle generating new tasks and deciding whether to train locally or offload to the fog node through a task offloading strategy.

[0019] Preferably, the task offloading strategy is optimized by the service delay of task execution;

[0020] Among them, the objective function of each task is defined as:

[0021] ;

[0022] In the formula, is the service delay of task is the remaining computing delay of the currently executing task, is the remaining computing delay of the currently executing task, and are both weight parameters;

[0023] The optimization problem of the task offloading strategy is:

[0024] ;

[0025] In the formula, represents the task offloading strategy, represents the total number of tasks.

[0026] Preferably, when the client vehicle offloads tasks to the fog node, according to the resource competition situation between the new task and the existing tasks, the computing latency of concurrent tasks is optimized through federated learning.

[0027] Preferably, when the task is offloaded to the fog node, including the communication latency, it depends on the network bandwidth, transmission power, and channel gain, and the expression is:

[0028] ;

[0029] In the formula, represents the communication latency when the task is offloaded to the fog node, represents the communication rate from the client vehicle to the fog node, represents the new task data size.

[0030] Preferably, when the task is locally trained, it includes local computing latency, which is related to the computing power of the client vehicle, and the expression is:

[0031] ;

[0032] In the formula, represents the local computing latency of the client vehicle, represents the computing resources required for the new task , represents the computing power of the client vehicle.

[0033] Preferably, when aggregating using federated learning, the formula for updating the concurrent overhead model is:

[0034] ;

[0035] In the formula, represents the updated global concurrent overhead model, represents the current global concurrent overhead model, represents the learning rate, represents the gradient of the loss function with respect to the local dataset .

[0036] A vehicle task offloading system based on federated learning includes:

[0037] A client, used to report the location and available computing resources of the vehicle, and determine the task offloading strategy;

[0038] A regional head, used to distribute the concurrent overhead model to different fog nodes;

[0039] A blockchain ledger, recording the global concurrent overhead model and historical task information;

[0040] A fog node for aggregating the concurrent overhead models after local training to obtain a global concurrent overhead model.

[0041] A computer device includes: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, any of the above vehicle task offloading methods based on federated learning is implemented.

[0042] Therefore, the present invention adopts the above vehicle task offloading system and method based on federated learning, and has the following technical effects: offloading tasks to nearby fog nodes for processing, and performing distributed training and model aggregation through the method of federated learning, which can reduce latency, improve the overall computing efficiency of the system, and ensure data privacy at the same time.

[0043] The technical solution of the present invention will be further described in detail below through the drawings and embodiments. Description of the Drawings

[0044] Figure 1 is a flowchart of the vehicle task offloading method based on federated learning;

[0045] Figure 2 is a distribution diagram of errors of various clusters in the image recognition task of the aggregated model in different cluster environments in the embodiments of the vehicle task offloading system and method based on federated learning. Detailed Embodiments

[0046] The present invention can be more detailedly explained through the following embodiments. The purpose of disclosing the present invention is to protect all changes and improvements within the scope of the present invention. The present invention is not limited to the following embodiments.

[0047] As Figure 1 shown, the present invention provides a vehicle task offloading method based on federated learning, including the following steps:

[0048] S1. Divide the city into multiple service areas, each service area contains multiple fixed fog nodes, and the fog nodes coexist with roadside infrastructures, such as traffic signs or billboards.

[0049] S2. The area head in the service area detects the vehicles entering or leaving the area in real time through the existing cellular network registration mechanism, and the vehicle reports its location and available computing resources to the area head when entering.

[0050] S3. The area head distributes the initial concurrent overhead model to the fog nodes in the service area and is responsible for maintaining the blockchain ledger to share computing resources and task offloading information;

[0051] S4. The client vehicle trains the concurrent overhead model through federated learning according to its own hardware configuration and computing tasks. The training process includes the following steps:

[0052] S41. The client vehicle registers with the regional head and receives the initial concurrent overhead model.

[0053] Define the basic parameters of the new task , including the task size, the required computing resources, the task start timestamp, and the fog node that executes the task, etc. Among them, the task size represents the data size of the new task , usually in bytes, which determines the amount of data that needs to be transmitted before the task is executed and affects the communication delay. The required floating-point operation volume of the task represents the computing resources required for the new task , usually expressed as the number of floating-point operations (FLOPS), which is used to measure the computational complexity of the task. The task start timestamp represents the time when the task is generated, which is used to calculate the total processing duration of the task. The fog node that executes the task , and the ID indicating the fog computing node, which receives and processes the new task from the client vehicle .

[0054] S42. The client vehicle trains the initial concurrent overhead model in its local environment based on historical task data and current task concurrency.

[0055] Among them, when the client vehicle generates a new task , it needs to use a binary indicator variable , that is, the task offloading strategy, to determine whether the task is processed locally or offloaded to the nearest fog node. This decision needs to consider multiple factors: First, the communication delay needs to be considered, which is related to the physical distance and network bandwidth between the vehicle and the fog node; second, the local computing delay needs to be considered.

[0056] If the vehicle decides to offload the task to the fog node, the communication delay is calculated by the formula: . Among them, is the communication rate from the vehicle to the fog node, which depends on the network bandwidth, transmission power, and channel gain.

[0057] If the task is processed locally by the vehicle, the delay is related to the computing power of the vehicle , and the formula for its local computing delay is: .

[0058] S43. The client vehicle uploads the trained concurrent overhead model to the corresponding fog node, and aggregates the concurrent overhead models of multiple client vehicles (different fog nodes) through federated learning to form a global resource competition model, that is, the global concurrent overhead model.

[0059] When the customer vehicle decides to offload tasks to the fog node, it needs to share computing resources with other tasks on the fog node. Assuming that the computing resources of the fog node are shared, the computing loads of all concurrent tasks will affect each other, increasing the computing latency. To solve this problem, federated learning is used to optimize the computing latency of concurrent tasks. The formula for the computing overhead is: . Among them, the function describes the computing competition situation between the new task and the currently executing task , mainly considering the resource competition between the new task and the existing tasks.

[0060] S44. Distribute the aggregated global concurrent overhead model to all client vehicles, and enter the next round of training until the set number of training rounds is reached. This process can optimize the computing scheduling and resource allocation of tasks through model aggregation in federated learning.

[0061] During the iterative training process, the formula for updating the global concurrent overhead model is:

[0062] ;

[0063] In the formula, represents the updated global concurrent overhead model, is the current global model, is the learning rate, is the gradient of the loss function with respect to the local dataset .

[0064] After the model aggregation for a certain number of training rounds, the regional head updates the final model to the global concurrent overhead model and records it in the blockchain ledger. And the service area head will regularly destroy the expired blockchain ledger and adjust the training parameters of the client vehicles at the end of each round of training to optimize the resource usage efficiency.

[0065] This embodiment verifies the image recognition algorithm under different service quantities in the real environment and the simulation environment. As Figure 2 shown, as the number of training rounds increases, the changing trend of the mean absolute error under different service quantities (number of clusters) indicates that the error of the model under the clustering condition is significantly lower than that of the unclustered model, and as the number of clusters increases, the error tends to converge. In addition, the model performances in the real environment and the simulation environment are relatively consistent, and the score gap is the smallest when the service quantity is small, not exceeding 1%, and the maximum gap is 26% when the service quantity is large.

[0066] In another embodiment, a vehicle task offloading system based on federated learning is provided, including a client for reporting the location and available computing resources of a vehicle and determining a task offloading strategy; a regional head for distributing concurrent overhead models to different fog nodes; a blockchain ledger for recording the global combined overhead model and historical task information; and fog nodes for aggregating the locally trained concurrent overhead models to obtain a global concurrent overhead model.

[0067] In yet another embodiment, a computer device is further provided, including: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, it implements any one of the above vehicle task offloading methods based on federated learning.

[0068] Therefore, by adopting the above vehicle task offloading system and method based on federated learning, the present invention can overcome problems such as uneven distribution of computing resources, resource competition, and task processing delay in the task offloading process in the existing vehicle Internet of Things environment through federated learning, and achieve the effects of improving model training and enhancing system performance.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A vehicle task offloading method based on federated learning, characterized in that Including the following steps: S1. Divide the city into different service areas, and each service area contains fixed fog nodes, where the fog nodes coexist with roadside infrastructure; S2. The area head in the service area uses the existing cellular network registration mechanism to detect vehicles entering or leaving the service area in real time. When a vehicle enters the service area, it reports its location and available computing resources to the area head; S3. The area head distributes the initial concurrent overhead model to the fog nodes in the service area and is responsible for maintaining the blockchain ledger to share computing resources and task offloading information; S4. The client vehicle locally trains the initial concurrent overhead model according to its hardware configuration and computing tasks, and aggregates it into a global concurrent overhead model through federated learning, which is recorded in the blockchain ledger, including: S41. The client vehicle registers with the area head and receives the initial concurrent overhead model; S42. The client vehicle locally trains the initial concurrent overhead model based on historical task data and current task concurrency, including the client vehicle generating new tasks and deciding whether to train locally or offload to the fog node through the task offloading strategy; The task offloading strategy is optimized through the service latency of task execution; Among them, the objective function of each task is defined as: ； In the formula, is the service latency of the task , is the remaining computing latency of the task currently being executed, and are both weight parameters; The optimization problem of the task offloading strategy is: ； In the formula, represents the task offloading strategy, represents the total number of tasks; When the client vehicle offloads a task to the fog node, according to the resource competition situation of the new task and the existing tasks, it optimizes the computing latency of concurrent tasks through federated learning; S43. The client vehicle uploads the locally trained concurrent overhead model to the corresponding fog node, and the fog node aggregates the locally trained concurrent overhead models to obtain the global concurrent overhead model; S44. Distribute the aggregated global concurrent overhead model to all client vehicles, and repeat the training until the set number of training rounds is reached.

2. The vehicle task offloading method based on federated learning according to claim 1, wherein When the task is offloaded to the fog node, it includes communication latency, which depends on network bandwidth, transmission power, and channel gain, and the expression is: ； wherein, represents the communication delay when the task is offloaded to the fog node, represents the communication rate from the client vehicle to the fog node, represents the new task data size.

3. The vehicle task offloading method based on federated learning according to claim 1, characterized in that When the task is locally trained, it includes local computing latency, which is related to the computing power of the client vehicle, and the expression is: ； In the formula, represents the local computing latency of the client vehicle, represents the new task required computing resources, represents the computing power of the client vehicle.

4. The vehicle task offloading method based on federated learning according to claim 1, wherein When aggregating using federated learning, the formula for updating the concurrent overhead model is: ； In the formula, represents the updated global concurrency overhead model, represents the current global concurrency overhead model, represents the learning rate, represents with respect to the local dataset the gradient of the loss function.

5. A computer device, comprising: A memory and a processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the federated learning-based vehicle task offloading method described in any one of claims 1-4.

Images