A semi-asynchronous federated learning method for mobility-aware clustering based on vehicle networks

Abstract

The present invention belongs to the field of vehicle network communication technology, and specifically relates to a mobile perception clustering semi-asynchronous federated learning method based on vehicle network, comprising the following steps: dividing the vehicle network into a multi-layer network, clustering the vehicles in each layer according to the model training time and the vehicle upload time, and all vehicles are clustered into the corresponding network layer to participate in the federated training; assigning different weights to the models of different network layers according to the influence of vehicles in different network layers on the model; in the semi-asynchronous federated learning MSAFL framework, the central server immediately performs the aggregation process after the model is successfully uploaded, without waiting for all vehicles to complete the upload in a given iteration; collecting perception data inside and outside the vehicle, and processing the data using a local computing unit to achieve data sharing and joint learning. The present invention can improve the efficiency and effect of distributed learning between vehicles under the premise of protecting data privacy.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle network communication, and in particular relates to a mobile-aware clustering semi-asynchronous federated learning method based on vehicle networks. Background Art

[0002] With the advancement of intelligent technology, vehicles now have the ability to perceive the driving environment and make intelligent decisions. Vehicles perceive the surrounding environment through various on-board sensors. These sensors collect internal and external perception data, which is processed by local computing units, and then the vehicle uses machine learning algorithms to learn knowledge from the perception data. However, the perception capabilities and local computing resources of a single vehicle are limited.

[0003] To extend the sensing range and enhance the sensing capability, vehicles can be interconnected to form vehicle networks (VNs) to share sensing data and upload the data to roadside units (RSUs) for more extensive computation. RSUs equipped with powerful computing resources use advanced machine learning algorithms to process the accumulated raw data and transmit the trained learning models or generated decisions back to the vehicles. However, transmitting raw data to RSUs raises significant privacy issues and consumes a large amount of wireless bandwidth. Therefore, federated learning (FL) has emerged as a solution to facilitate distributed learning tasks. Federated learning enables vehicles to contribute to global learning goals by uploading the parameters of local training models (rather than raw data) to a central server. Afterwards, the RSU trains a global machine learning model and transmits it back to the vehicle, which then updates the local model and returns the uploaded parameters to the RSU. The combination of federated learning and VNs effectively protects data privacy while enhancing the understanding of driving behavior and road conditions using distributed sensor data and supporting the development of autonomous driving applications such as traffic sign recognition, adaptive cruise control, and blind spot monitoring.

[0004] Considering the uploading of local model parameters and the transmission of global models, federated learning frameworks are generally divided into two categories: synchronous and asynchronous. FedAvg is an example of a synchronous federated learning framework, in which the central server must wait for all vehicles to upload their models before starting the aggregation process. However, the limited range of the central server's wireless link limits the number of training iterations that highly mobile vehicles can effectively train their local models. In contrast, the asynchronous federated learning framework performs the aggregation process immediately after the model is successfully uploaded to the central server, without requiring all vehicles to complete the upload in a given iteration. Nevertheless, frequent parameter uploads lead to increased communication overhead and may have an adverse impact on algorithm convergence. Therefore, the challenges of applying federated learning to the Internet of Vehicles can be summarized as follows: First, the high mobility of vehicles in the Internet of Vehicles and the limited range of wireless links make it difficult to achieve good local and global machine learning models in time-sensitive situations. Second, the communication overhead introduced by the local parameter upload and / or global model download process complicates the optimization of federated learning, making it challenging to balance the communication overhead with the benefits of federated learning.

[0005] In summary, traditional vehicle network sharing of perception data will cause significant privacy issues and consume a large amount of wireless bandwidth; federated learning is a very promising learning method that protects privacy in a distributed environment, allowing multiple independent vehicles to complete a learning goal by uploading the parameters of the local training model (rather than the original data) to the central server, and enhance the understanding of driving behavior and road conditions; however, the existing synchronous and asynchronous federated learning frameworks have problems such as slow convergence, high communication overhead, and model overfitting under highly mobile vehicles and limited wireless link conditions; therefore, the present invention provides a clustering-based semi-asynchronous federated learning method for edge vehicle mobile perception scenarios, aiming to solve the problems of large communication delay, slow convergence, and model overfitting faced by the existing federated learning framework in the Internet of Vehicles. Under the premise of protecting data privacy, the present invention improves the efficiency and effect of distributed learning between vehicles and supports the development of autonomous driving applications. Summary of the invention

[0006] The purpose of the present invention is to provide a mobile-aware clustering semi-asynchronous federated learning method based on vehicle networks, which can solve the problems of large communication delay, slow convergence speed and model overfitting faced by the existing federated learning framework in the Internet of Vehicles, and improve the efficiency and effect of distributed learning between vehicles while protecting data privacy.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A semi-asynchronous federated learning method for mobility-aware clustering based on vehicle networks, comprising the following steps:

[0009] Step 1: Network layering and vehicle clustering unit: Divide the Internet of Vehicles into multiple layers of networks. The vehicles in each layer are clustered according to the model training time and vehicle upload time. All vehicles are clustered into the corresponding network layer to participate in federated training.

[0010] Step 2: Information age and weight assignment: Assign different weights to models at different network layers based on the impact of vehicles at different network layers on the model;

[0011] Step 3, semi-asynchronous federated learning: In the semi-asynchronous federated learning MSAFL framework, the central server performs the aggregation process immediately after the model is successfully uploaded, without waiting for all vehicles to complete the upload in a given iteration;

[0012] Step 4: Mobile perception and data sharing: Collect perception data inside and outside the vehicle and use local computing units to process the data to achieve data sharing and joint learning.

[0013] Further, when participating in federated training in step 1, vehicles in the low-latency network layer participate in the federated training process more frequently than vehicles in the high-latency network layer.

[0014] Furthermore, the impact of vehicles at different network layers on the model in step 2 is the impact of asynchronous data transmission between vehicles at different network layers and vehicle mobility on the model.

[0015] Furthermore, the perception data collected inside and outside the vehicle in step 4 is the vehicle sensing the surrounding environment through on-board sensors.

[0016] Furthermore, in step 4, the local computing unit is used to process data under a semi-asynchronous federated learning framework. The vehicle uploads the parameters of the local training model to the roadside unit. After the roadside unit trains the global model, it transmits it back to the vehicle. The vehicle updates the local model and returns the uploaded parameters to the roadside unit.

[0017] Furthermore, the step 4 comprises the following steps:

[0018] Build an edge vehicle network architecture consisting of a roadside unit at the edge of the network and The vehicle consists of marked vehicles; the vehicles in the area are connected to the roadside unit through wireless communication; each vehicle cannot exchange data with each other, and only interacts with the roadside unit model; each vehicle travels at a constant speed; different vehicles collect different amounts of data, and their computing power is also different;

[0019] Construct a three-dimensional coordinate system, with the bottom position of the roadside unit set as the origin, the x-axis direction set as the east, the y-axis direction set as the south, and the z-axis direction set as the direction along the roadside unit antenna perpendicular to the x-axis and the y-axis;

[0020] Federated learning is used as the main framework. Each has its own data set, represented as:

[0021]

[0022] in, represents the training data, Corresponding labels, where D k is the number of training data points owned by vehicle k, j is the data sample index, w represents the model parameter, and the loss function corresponding to each sample is f(w, x k，j ,y k，j );set up Denotes the loss corresponding to vehicle k. Federated learning obtains a model with parameters by solving the optimization problem:

[0023]

[0024] In each global iteration, federated learning consists of four steps:

[0025] S1. Broadcast global model: In the tth global iteration, the roadside unit broadcasts the global model w generated in the previous global iteration to all selected vehicles t-1 ;

[0026] S2. Local model training: Each vehicle k is trained in its local dataset D k Train its local model on, that is:

[0027]

[0028] Among them, η t is the learning rate;

[0029] S3. Model upload: Vehicle k exports the local model After that, upload its local model to the federated learning server;

[0030] S4, model aggregation: The roadside unit aggregates the local models uploaded by the vehicle and updates the global model based on FedAvg, that is:

[0031]

[0032] Define τ as the deadline of a global round; in each global round, the roadside unit will accept all models uploaded from vehicles before the deadline and reject the models with higher latency from the remaining vehicles in each global iteration;

[0033] Therefore, the vehicles are clustered into several layers, using m k，j Indicates whether vehicle k is on the jth floor. k，j= 1, vehicle k is on the jth floor, m k，j = 0, otherwise; if vehicle k can complete its local model training and upload before the deadline τ, that is, t k ≤τ, then vehicle k is on the first floor, i.e., m k，1 =1; Similarly, if vehicle k can complete the training and uploading of its local model between τ and 2×τ, that is, τ <t k ≤2×τ, then vehicle k is on the second layer, i.e., m k，2 = 1; The general mathematical expression for clustering vehicle k into a specific layer is as follows:

[0034]

[0035] Where j represents the index of the layer, t k represents the date when vehicle k completes its local model training and uploading, and τ represents the deadline;

[0036] After clustering vehicles into corresponding layers, the RSU specifies when vehicles can upload their local models based on time thresholds designed for different layers;

[0037] Cluster all vehicles into four layers; vehicles in the first layer can upload their local models in each global iteration, because their calculation and transmission speed is fast, they can participate more actively in the federated training; and vehicles in the fourth layer will participate in the federated training once every four rounds because of their higher calculation and transmission delays; let t be the index of the global iteration, let is a binary variable indicating whether the vehicle at layer j expects to upload its local model at the end of global iteration t, where:

[0038]

[0039] Among them, % is a modulo operation, so t%j=0 means that t is divisible by j; therefore, the vehicle that the roadside unit expects to upload its local model at global iteration t can be expressed as:

[0040]

[0041] in, is a binary variable, indicates that vehicle k can upload its local model at the end of the tth global iteration, Then according to , aggregating all received local models:

[0042]

[0043] in, represents the total number of data samples of all vehicles willing to upload their local models in global iteration t;

[0044] A regularization term is used in the loss function to alleviate the imbalance of the model caused by semi-asynchronous training; in the global iteration t, the regularized local loss at vehicle k is defined as:

[0045]

[0046] Where ξ is the regularization coefficient and r(w) is the deterministic penalty function; for any vehicle k, its local update is calculated as:

[0047]

[0048] in, represents partial derivative calculation;

[0049] Define Π(t) as the local model that is expected to be uploaded at the end of the tth global iteration Due to the semi-asynchronous setting, the most recent global iteration of the global model that vehicle k receives the update is expressed as:

[0050]

[0051] The roadside unit receives and aggregates the models uploaded by the selected vehicles; it then broadcasts the aggregated model as an updated global model to the vehicles that uploaded the models, thus initiating a new round of joint training;

[0052] The federation cycle age is introduced as a key metric in the federated learning system as follows:

[0053] a k (t) = ts k (t)

[0054] Among them, a k (t) represents the time since vehicle k last received an updated global model;

[0055] Vehicle k will be the weighted local model After uploading to the RSU, the RSU aggregates updates from each vehicle in a semi-synchronous manner;

[0056] The age-aware weighting design is as follows:

[0057]

[0058] Among them, γ is a real-valued constant, and its value can be selected in three cases:

[0059] γ>1, the system prefers older local updates;

[0060] γ<1, the system prefers newer local updates;

[0061] γ=1, which is equivalent to the design in FedAvg.

[0062] The technical effects achieved by the present invention are:

[0063] The mobile-aware clustering semi-asynchronous federated learning method based on vehicle networks of the present invention can solve the problems of large communication delay, slow convergence speed and model overfitting faced by the existing federated learning framework in the Internet of Vehicles. Under the premise of protecting data privacy, it improves the efficiency and effect of distributed learning between vehicles and supports the development of autonomous driving applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the edge vehicle network framework of the present invention;

[0065] Figure 2 It is a diagram of the federated learning training process in the vehicle network scenario of the present invention;

[0066] Figure 3 is a schematic diagram of a delay-aware vehicle clustering selection scheme of the present invention;

[0067] Figure 4 It is the MSAFL algorithm proposed by the present invention;

[0068] Figure 5 is a probability density diagram of vehicle delay distribution of the present invention;

[0069] Figure 6 is the accuracy and loss of the proposed method and the baseline method on the MNIST dataset;

[0070] Figure 7 is the training time of the proposed method and the baseline method on the MNIST dataset. DETAILED DESCRIPTION

[0071] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0072] like Figure 1-7 As shown, a semi-asynchronous federated learning method for mobile-aware clustering based on vehicle networks includes the following steps;

[0073] Step 1, network stratification and vehicle clustering unit: The Internet of Vehicles is divided into multiple layers of networks. Vehicles in each layer are clustered according to the model training time and vehicle upload time. When participating in federated training, vehicles in the low-latency network layer participate in the federated training process more frequently than vehicles in the high-latency network layer. All vehicles are clustered into the corresponding network layer to participate in federated training, which effectively reduces the slow convergence and model overfitting problems encountered in synchronous federated learning. Through network stratification and vehicle clustering mechanisms, vehicles in the low-latency network layer participate in federated training more frequently, reducing communication delays and convergence time.

[0074] Step 2, information age and weight assignment: Introduce the concept of information age, and assign different weights to models at different network layers according to the impact of asynchronous data transmission and vehicle mobility between vehicles at different network layers on the model. By introducing the concept of information age and assigning different weights according to the impact of asynchronous data transmission and mobility on the model, the problem of model obsolescence caused by lagging vehicles is solved, ensuring the real-time and effectiveness of the model.

[0075] Step 3, semi-asynchronous federated learning: In the semi-asynchronous federated learning MSAFL framework, the central server performs the aggregation process immediately after the model is successfully uploaded, without waiting for all vehicles to complete the upload in a given iteration, thereby reducing communication delays and improving learning efficiency; at the same time, through appropriate network layer stratification and information age weight distribution, the communication overhead and algorithm convergence problems caused by frequent parameter uploads are alleviated.

[0076] Compared with the existing method of transmitting models, the semi-asynchronous federated learning mechanism allows the aggregation process to be performed immediately after the model is successfully uploaded, reducing communication overhead. The vehicle uploads the parameters of the local training model instead of the original data, which effectively protects data privacy and reduces the risk of privacy leakage during data transmission.

[0077] Step 4: Mobile perception and data sharing: The vehicle perceives the surrounding environment through on-board sensors, collects internal and external perception data, and processes the data using local computing units. Under the semi-asynchronous federated learning framework, the vehicle uploads the parameters of the local training model to the roadside unit. After the roadside unit trains the global model, it transmits it back to the vehicle. The vehicle updates the local model and returns the uploaded parameters to the roadside unit to achieve data sharing and joint learning.

[0078] This technical solution can improve the accuracy and reliability of autonomous driving applications, such as traffic sign recognition, adaptive cruise control, and blind spot monitoring, by enhancing the understanding of driving behavior and road conditions. At the same time, the framework can still perform federated learning efficiently in a highly mobile Internet of Vehicles environment, improving the robustness and flexibility of the system.

[0079] In summary, the present invention provides a clustering-based semi-asynchronous federated learning method for edge vehicle mobile perception scenarios, aiming to solve the problems of large communication delay, slow convergence speed and model overfitting faced by the existing federated learning framework in the Internet of Vehicles. Under the premise of protecting data privacy, the present invention improves the efficiency and effect of distributed learning between vehicles and supports the development of autonomous driving applications.

[0080] In order to more clearly explain the technical solution, the technical solution discloses the steps of step 4 in detail, which specifically includes the following steps:

[0081] Build an edge vehicle network architecture such as Figure 1 As shown in Figure 2, the architecture consists of a roadside unit (RSU) at the edge of the network and The vehicle is composed of marked vehicles; the vehicles in the area are connected to the RSU through wireless communication; each vehicle cannot exchange data with each other and only interacts with the roadside unit model; each vehicle travels at a constant speed; different vehicles collect different amounts of data, and their computing capabilities may also be different; in order to facilitate real-time determination of the vehicle position, the technical solution constructs a three-dimensional coordinate system, the bottom position of the roadside unit is set as the origin, the x-axis direction is set to the east, the y-axis direction is set to the south, and the z-axis direction is set to the direction along the roadside unit's antenna perpendicular to the x-axis and y-axis.

[0082] Develop a model that is trained across all vehicles without exchanging local data with other vehicles or RSUs. To achieve this, we adopt federated learning as the main framework. Each has its own data set, represented as:

[0083]

[0084] in, represents the training data, Corresponding label. Among them, D k is the number of training data points owned by vehicle k, j is the data sample index, w represents the model parameter, and the loss function corresponding to each sample is f(w, x k，j ,y k，j ).set up Denotes the loss corresponding to vehicle k. Federated learning obtains a model with parameters by solving the following optimization problem:

[0085]

[0086] like Figure 2 As shown, in each global iteration, federated learning FL consists of four steps:

[0087] S1. Broadcast global model: In the tth global iteration, the RSU broadcasts the global model w generated in the previous global iteration to all selected vehicles t-1 .

[0088] S2. Local model training: Each vehicle k is trained in its local dataset D k Train its local model on, that is:

[0089]

[0090] Among them, η t is the learning rate.

[0091] S3. Model upload: Vehicle k exports the local model After that, upload its local model to the federated learning FL server.

[0092] S4, model aggregation: RSU aggregates the local models uploaded from vehicles and updates the global model based on FedAvg, i.e.

[0093]

[0094] This technical solution defines τ as the deadline of a global round; in each global round, the RSU will accept all models uploaded from vehicles before the deadline and reject the models with higher latency from the remaining vehicles in each global iteration; therefore, this technical solution clusters the vehicles into several layers, using m k，j Indicates whether vehicle k is on the jth floor. k，j = 1, vehicle k is on the jth floor, m k，j = 0 otherwise. If vehicle k can complete its local model training and upload before the deadline τ, that is, t k ≤τ, then vehicle k is on the first floor, i.e., m k，1 = 1. Similarly, if vehicle k can complete the training and uploading of its local model between τ and 2×τ, that is, τ <t k ≤2×τ, then vehicle k is on the second layer, i.e., m k,2 = 1. The following equation provides a general mathematical expression for clustering vehicle k into a specific layer:

[0095]

[0096] Where j represents the index of the layer, t k represents the date when vehicle k completes its local model training and uploading, and τ represents the deadline.

[0097] After clustering vehicles into corresponding layers, the RSU can specify when vehicles can upload their local models based on the time thresholds designed for different layers.

[0098] For example, Figure 3 As shown in Figure 2, all vehicles are clustered into four layers; vehicles in the first layer can upload their local models in each global iteration, because their calculation and transmission speeds are faster, and they can participate more actively in the federated training; and vehicles in the fourth layer will participate in the federated training once every four rounds because of their higher calculation and transmission delays; let t be the index of the global iteration, let is a binary variable indicating whether the vehicle at layer j expects to upload its local model at the end of global iteration t, where:

[0099]

[0100] Among them, % is a modulo operation, so t%j=0 means that t is divisible by j; therefore, the vehicle that the roadside unit expects to upload its local model at global iteration t can be expressed as:

[0101]

[0102] in, is a binary variable, indicates that vehicle k can upload its local model at the end of the tth global iteration, Then according to , aggregating all received local models:

[0103]

[0104] in, Represents the total number of data samples of all vehicles willing to upload their local models in global iteration t.

[0105] This technical solution uses a regularization term in the loss function to alleviate the imbalance of the model caused by semi-asynchronous training. In the global iteration t, the regularized local loss at vehicle k is defined as:

[0106]

[0107] Where ξ is the regularization coefficient and r(w) is the deterministic penalty function. For any vehicle k, its local update is calculated as:

[0108]

[0109] in, Represents partial derivative calculation.

[0110] This technical solution defines Π(t) as the local model that is expected to be uploaded at the end of the tth global iteration. Due to the semi-asynchronous setting, the most recent global iteration of the global model that vehicle k receives the update is expressed as:

[0111]

[0112] RSU receives and aggregates the models uploaded by the selected vehicles. Subsequently, it broadcasts the aggregated model as an updated global model to the vehicles that uploaded the model, thereby starting a new round of joint training; Under the guidance of the theoretical framework of information age, this technical solution introduces the federation cycle age as a key metric in the federated learning system as follows:

[0113] a k (t) = ts k (t)

[0114] Among them, a k (t) represents the time elapsed since vehicle k last received an updated global model.

[0115] Vehicle k will be the weighted local model After uploading to the roadside unit, the roadside unit aggregates updates from each vehicle in a semi-synchronous manner; in order to solve the asynchronous problem in model aggregation, this technical solution not only assigns weights based on the proportion of data carried by the vehicle, but also assigns weights based on its federation cycle age. The following is the age-aware weighting design:

[0116]

[0117] Among them, γ is a real-valued constant, and its value can be selected in three cases:

[0118] γ>1, the system prefers older local updates.

[0119] γ<1, the system prefers newer local updates.

[0120] γ=1, which is equivalent to the design in FedAvg.

[0121] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A semi-asynchronous federated learning method for mobile-aware clustering based on vehicle networks, characterized by: The following steps are involved: Step 1: Network layering and vehicle clustering unit: Divide the Internet of Vehicles into multiple layers of networks. The vehicles in each layer are clustered according to the model training time and vehicle upload time. All vehicles are clustered into the corresponding network layer to participate in federated training. Step 2: Information age and weight assignment: Assign different weights to models at different network layers based on the impact of vehicles at different network layers on the model; The impact of vehicles in different network layers on the model in step 2 is the impact of asynchronous data transmission and vehicle mobility between vehicles in different network layers on the model; Step 3, semi-asynchronous federated learning: In the semi-asynchronous federated learning MSAFL framework, the central server performs the aggregation process immediately after the model is successfully uploaded, without waiting for all vehicles to complete the upload in a given iteration; Step 4: Mobile sensing and data sharing: Collect the sensing data inside and outside the vehicle and process the data using the local computing unit to achieve data sharing and joint learning. Define τ as the deadline of a global round; in each global round, the roadside unit will accept all models uploaded from vehicles before the deadline and reject the models with higher latency from the remaining vehicles in each global iteration; Therefore, the vehicles are clustered into several layers, using m k,j Indicates whether vehicle k is on the jth floor. k,j = 1, vehicle k is on the jth floor, m k,j = 0, otherwise; if vehicle k can complete its local model training and upload before the deadline τ, that is, t k ≤τ, then vehicle k is on the first floor, i.e., m k,1 =1; Similarly, if vehicle k can complete the training and uploading of its local model between τ and 2×τ, that is, τ <t k ≤2×τ, then vehicle k is on the second layer, i.e., m k,2 = 1; The general mathematical expression for clustering vehicle k into a specific layer is as follows: Where j represents the index of the layer, t k represents the date when vehicle k completes its local model training and uploading, and τ represents the deadline; After clustering vehicles into corresponding layers, the RSU specifies when vehicles can upload their local models based on time thresholds designed for different layers.

2. According to claim 1, a semi-asynchronous federated learning method for mobile-aware clustering based on vehicle networks is characterized by: When participating in federated training in step 1, vehicles in the low-latency network layer participate in the federated training process more frequently than vehicles in the high-latency network layer.

3. The method of semi-asynchronous federated learning based on vehicle network mobility-aware clustering according to claim 1 is characterized by: The perception data collected inside and outside the vehicle in step 4 is the vehicle's perception of the surrounding environment through on-board sensors.

4. The method of semi-asynchronous federated learning based on vehicle network mobility-aware clustering according to claim 1 is characterized in that: In step 4, the local computing unit is used to process data under the semi-asynchronous federated learning framework. The vehicle uploads the parameters of the local training model to the roadside unit. After the roadside unit trains the global model, it transmits it back to the vehicle. The vehicle updates the local model and returns the uploaded parameters to the roadside unit.

5. The method of semi-asynchronous federated learning based on vehicle network mobility-aware clustering according to claim 1 is characterized in that: The step 4 comprises the following steps: Build an edge vehicle network architecture consisting of a roadside unit at the edge of the network and The vehicle consists of marked vehicles; the vehicles in the area are connected to the roadside unit through wireless communication; each vehicle cannot exchange data with each other, and only interacts with the roadside unit model; each vehicle travels at a constant speed; different vehicles collect different amounts of data, and their computing power is also different; Construct a three-dimensional coordinate system, with the bottom position of the roadside unit set as the origin, the x-axis direction set as the east, the y-axis direction set as the south, and the z-axis direction set as the direction along the roadside unit antenna perpendicular to the x-axis and the y-axis; Federated learning is used as the main framework. Each has its own data set, represented as: in, represents the training data, Corresponding labels, where D k is the number of training data points of vehicle k, j is the data sample index, w represents the model parameter, and the loss function corresponding to each sample is f(w,x k,j ,y k,j );set up Denotes the loss corresponding to vehicle k. Federated learning obtains a model with parameters by solving the optimization problem: In each global iteration, federated learning consists of four steps: S1. Broadcast global model: In the tth global iteration, the roadside unit broadcasts the global model w generated in the previous global iteration to all selected vehicles t-1 ; S2. Local model training: Each vehicle k is trained in its local dataset D k Train its local model on it, namely: Among them, η t is the learning rate; S3. Model upload: Vehicle k exports the local model After that, upload its local model to the federated learning server; S4, model aggregation: The roadside unit aggregates the local models uploaded by the vehicle and updates the global model based on FedAvg, that is:

6. The method of semi-asynchronous federated learning based on vehicle network mobility-aware clustering according to claim 5 is characterized by: Cluster all vehicles into four layers; vehicles in layer 1 can upload their local models in each global iteration; and vehicles in layer 4 will participate in federated training every four rounds; let t be the index of the global iteration, let is a binary variable indicating whether the vehicle at layer j expects to upload its local model at the end of global iteration t, where: Among them, % is a modulo operation, so t%j=0 means that t is divisible by j; therefore, the vehicle that the roadside unit expects to upload its local model at global iteration t is expressed as: in, is a binary variable, indicates that vehicle k uploads its local model at the end of the tth global iteration, Then according to , aggregating all received local models: in, represents the total number of data samples of all vehicles willing to upload their local models in global iteration t; A regularization term is used in the loss function to alleviate the imbalance of the model caused by semi-asynchronous training; in the global iteration t, the regularized local loss at vehicle k is defined as: Where ξ is the regularization coefficient and r(w) is the deterministic penalty function; for any vehicle k, its local update is calculated as: in, represents partial derivative calculation; Define Π(t) as the local model that is expected to be uploaded at the end of the tth global iteration Due to the semi-asynchronous setting, the most recent global iteration of the global model that vehicle k receives the update is expressed as: The roadside unit receives and aggregates the models uploaded by the selected vehicles; subsequently, it broadcasts the aggregated model as an updated global model to the vehicles that uploaded the models, thus initiating a new round of joint training.

7. The method of semi-asynchronous federated learning based on vehicle network mobility-aware clustering according to claim 6 is characterized by: The federation cycle age is introduced as a key metric in the federated learning system as follows: a k (t)=t-s k (t) Among them, a k (t) represents the time since vehicle k last received an updated global model; Vehicle k will be the weighted local model After uploading to the RSU, the RSU aggregates the updates from each vehicle in a semi-synchronous manner; The age-aware weighting design is as follows: Among them, γ is a real-valued constant, and its value can be selected in three cases: γ>1, the system prefers older local updates; γ<1, the system prefers newer local updates; γ=1, which is equivalent to the design in FedAvg.

Images