Heterogeneous Internet of Vehicles personalized federal learning method

Abstract

The invention relates to a heterogeneous Internet of Vehicles personalized federal learning method, and belongs to the field of Internet of Vehicles data analysis. According to the method, federated learning is carried out based on a cloud-edge-end three-layer architecture, and a user layer is responsible for data collection, local model training and personalized model parameter decomposition and construction; the edge layer is responsible for aggregating a basic knowledge part of a local model of the vehicle client within the coverage of the edge layer; the cloud server layer is responsible for aggregating the edge basic knowledge of all the edge layers. According to parameter decomposition and construction of the personalized model, basic knowledge and personalized knowledge of a local model of the vehicle client are distinguished by adopting two-dimensional discrete cosine transform, and fine-grained personalized knowledge of the vehicle client is reserved while basic knowledge of other vehicle clients is fused. Compared with other personalized federated learning methods, the method has the advantages that the model training time is shortened while the model performance is guaranteed, and the low-complexity requirement in a networking scene is met.

Description

Technical Field

[0001] The present invention belongs to the field of vehicle networking data analysis and relates to a personalized federated learning method for heterogeneous vehicle networking. Background Art

[0002] With the promotion and application of 5G technology and the development of V2X technology, in the Internet of Vehicles (IoV) scenario, cars can directly share the trajectory, traffic conditions, driver information and other data collected by the vehicle with other vehicles, roadside units, traffic infrastructure, etc. to train machine learning models and provide intelligent services such as recognition, search, parking, environmental perception, and map construction. However, although the data sharing between vehicles and these infrastructures can improve the intelligent decision-making ability of vehicles, it brings serious data privacy leakage problems. Therefore, it is crucial to explore solutions for intelligent applications in the IoV that can both improve intelligent decision-making capabilities and protect data privacy. In order to solve this problem, the federated learning proposed by Google in 2016 brought new ideas. The idea of ​​federated learning is to allow multiple vehicle clients to collaboratively use the collected local data sets to train a shared global model without uploading local data directly to the cloud server for data analysis and model training. It can both improve intelligent decision-making capabilities and protect data privacy. Although federated learning has many advantages in solving data privacy, it still faces many challenges in the current IoV scenario where resources are unbalanced between vehicle clients and vehicles have high-speed mobility.

[0003] Affected by the data dimensions collected by individual vehicles and the differences in the spatiotemporal distribution of vehicles, the data collected between vehicles have serious deviations in data distribution and data volume, resulting in significant non-independent and identically distributed characteristics of data between vehicles, which seriously affects the performance of the global model of federated learning for a single vehicle client. In addition, the spatiotemporal dynamics of vehicle location make the communication resources and network status connected to the roadside unit dynamically time-varying, which easily leads to the incompleteness of the federated training process. Therefore, how to consider the high heterogeneity of Internet of Vehicles data and the high mobility of vehicles to design personalized federated learning methods needs to be studied urgently. At present, the existing personalized federated learning methods in heterogeneous Internet of Vehicles mainly use methods such as client selection, clustering, and asynchronous aggregation to alleviate the impact of data heterogeneity and vehicle mobility on model performance. However, the current methods are insufficient in the granularity of personalized analysis of non-independent and identically distributed data. In addition, for Internet of Vehicles scenarios, due to the high mobility of vehicles, the federated learning model has a requirement for fast convergence, which puts forward low complexity requirements for personalized federated learning methods. Therefore, the existing personalized federated learning methods are not suitable for Internet of Vehicles scenarios. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a personalized federated learning method for heterogeneous Internet of Vehicles, so as to solve the problem that the existing personalized federated learning method has insufficient granularity for personalized analysis of non-independent and identically distributed data and the existing personalized federated learning method does not meet the low complexity requirements in the Internet of Vehicles scenario. First, the method of the present invention adopts a cloud-edge-end three-layer architecture to solve the problems of frequent communication and high overhead in the traditional two-layer architecture. Secondly, in view of the high-speed mobility problem of vehicles in the Internet of Vehicles scenario, the present invention selects a set of vehicle clients participating in the federated training according to the computing power and communication capacity of the vehicle, and ensures that the vehicle can complete the entire federated learning process including the frequency domain parameter decomposition of the model before leaving the current roadside unit, so as to reduce the negative impact of vehicle mobility and channel state time-varying on model training, and accelerate the model convergence speed. In addition, different from the existing personalized federated learning method, the present invention introduces the frequency domain analysis idea to design a personalized federated learning method for heterogeneous Internet of Vehicles, improves the performance of federated learning under heterogeneous Internet of Vehicles, and at the same time, the discrete cosine transform does not involve the calculation of the imaginary part, and the calculation speed is faster. Compared with other personalized federated learning methods, it effectively reduces the model training time and meets the low complexity requirements in the networking scenario.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A personalized federated learning method for heterogeneous Internet of Vehicles is based on a three-layer cloud-edge-end architecture for federated learning. The user layer is responsible for data collection, local model training, and personalized model parameter decomposition and construction; the edge layer is responsible for aggregating the basic knowledge part of the local model of the vehicle client within its coverage; the cloud server layer is responsible for aggregating the edge basic knowledge of all edge layers. For the decomposition and construction of personalized model parameters, a two-dimensional discrete cosine transform is used to distinguish the basic knowledge and personalized knowledge of the local model of the vehicle client, while integrating the basic knowledge of other vehicle clients, retaining its own fine-grained personalized knowledge.

[0007] The method specifically comprises the following steps:

[0008] S1: The cloud server initializes model parameters and other basic parameters;

[0009] S2: Select appropriate vehicle clients to participate in federated model training based on the vehicle’s computing and communication performance;

[0010] S3: Use the data collected by each vehicle client to perform local model training;

[0011] S4: Build a personalized model based on the personalized frequency domain parameter decomposition method;

[0012] S5: Repeated iterative training until the predetermined training rounds are reached or the model converges.

[0013] Furthermore, step S1 specifically includes: the cloud server customizes the training target based on user needs, and initializes the local iterator, learning rate size, batch data size, local iteration number, training rounds and other training information according to the training target, and generates an initial global federated learning model; the cloud server initializes the vehicle client participating in this federated training through the roadside unit.

[0014] Further, step S2 specifically includes the following steps:

[0015] S21: The cloud server notifies each roadside unit to evaluate the device performance and communication status of the vehicle client;

[0016] S22: The training delay of the roadside unit when evaluating the vehicle client's local training sample data;

[0017] S23: The roadside unit evaluates the communication delay of the vehicle client;

[0018] S24: Calculate the residence time of the vehicle client within the coverage area of ​​the current roadside unit;

[0019] S25: According to the sum of the calculation delay, communication delay, model aggregation time and frequency domain analysis time of the vehicle client, a vehicle client that can complete a round of model training before moving to the next roadside unit is selected for model training; as shown in the following formula:

[0020]

[0021] In the formula, represents the training time of the vehicle local model parameters, Indicates the time consumed by model uploading, t agg represents the time taken for model aggregation at the roadside unit, δ represents the time taken for frequency domain analysis and decomposition of local model parameters, τ k Indicates vehicle V k The dwell time within the current roadside unit coverage area.

[0022] Further, in step S22, the training delay of the vehicle client when training the sample data locally is Defined as:

[0023]

[0024] in, is the vehicle V in the tth round of federated training k performance, is the vehicle V k The computing resources of E k is the number of local iterations, is the number of training samples.

[0025] Further, step S23 specifically includes: if the vehicle V k The transmission power is p k , the path loss index is γ, and the noise power spectrum density is N0. Under the current roadside unit coverage, the data transmission rate of the vehicle client can be expressed as:

[0026]

[0027] In the formula, For vehicle client V k Channel gain with the roadside unit, is the small-scale channel coefficient, c k For vehicle client V k distance from roadside units;

[0028] If the size of the model parameters sent to the roadside unit is A, then the vehicle client V k Required communication delay for:

[0029]

[0030] Further, step S24 specifically includes: assuming that the vehicle client keeps its speed unchanged when passing the edge server, through the coverage range of the current roadside unit Vehicle Client V k The driving speed v k , its stay time within the coverage of the current roadside unit can be calculated:

[0031]

[0032] In the formula, l k Indicates that from the vehicle client V k The distance from the location to the entrance of the current roadside unit coverage area.

[0033] Further, step S3 specifically includes: according to the quality of the data set collected by the current vehicle client and the adopted loss function, training a local model of the vehicle client, which is expressed as:

[0034]

[0035] Where, d k represents the size of the dataset collected by the vehicle client, represents the jth data sample in the dataset collected by the kth vehicle client, Indicates the corresponding true label of the sample; Represents the difference between the sample prediction value and the sample's true label, that is, the cross entropy loss function; represents the vehicle client V in the tth round of federated training k Personalized model.

[0036] Further, step S4 specifically includes the following steps:

[0037] S41: Convolutional layer parameters ω′ for the local model of the vehicle client t,k,conv By performing discrete cosine transform, the frequency domain signal F(ω′) after the convolution layer parameters of the vehicle client model are analyzed can be obtained. t,k,conv ), as shown below:

[0038]

[0039] Where x,u=0,1,2,...,s1L-1; y,v=0,1,2,...,s2C-1; x and y are coordinates in the time domain, u and v are coordinates in the frequency domain, L and C are the number of output channels and input channels of the convolution kernel in the local model, s1 and s2 are the spatial shapes of the convolution kernel, ω′ t,k,conv For vehicle V k The convolutional layer parameters of the local model; C(u) is the normalization coefficient used to ensure orthogonality and energy preservation of the transformation; It is a cosine basis function, which mainly performs signal decomposition in the transformation, so as to perform more effective signal analysis and processing;

[0040] S42: Use a low-frequency mask matrix G with all values ​​​​zero except the central area to extract the low-frequency components for model aggregation (basic knowledge), which is expressed as G⊙F(ω′ t,k,conv ), the high-frequency component (personalized knowledge) is expressed as (1-G)⊙F(ω′ t,k,conv ), where F(ω′ t,k,conv ) is the frequency domain signal after the convolution layer parameters of the above-mentioned vehicle client local model are parsed;

[0041] S43: The vehicle client uploads the basic knowledge part of the local model to the roadside unit for preliminary aggregation to obtain the edge basic knowledge

[0042]

[0043] Where t represents the tth round of federated training, l represents the lth roadside unit B l , Indicates roadside unit B l The set of vehicle clients participating in the tth round of federated training within the coverage area, d k represents the kth vehicle client V k The amount of data collected; express The total amount of data from the vehicle clients in the collection;

[0044] S44: Roadside Unit Edge Basics Upload to the cloud server for final aggregation to obtain the aggregated global basic knowledgeω global :

[0045]

[0046] In the formula, represents the set of all roadside units under the coverage of the cloud server, I l Represents roadside unit B l The amount of data you have, Indicates the total data size of all vehicle clients within the coverage area of ​​the cloud service;

[0047] S45: The vehicle client combines the received global basic knowledge with the locally retained personalized knowledge to obtain the representation F(ω' t,k,conv ), defined as:

[0048] F(ω' t,k,conv )=(1-G)⊙F(ω′ t,k,conv )+ω global

[0049] In the formula, ω global is the global basic knowledge, (1-G)⊙F(ω′ t,k,conv ) represents the vehicle client V k personalized knowledge;

[0050] S46: Convert it to the time domain space according to the inverse discrete cosine transform to obtain the time domain representation ω′ of the model convolution layer parameters t,k,conv , and compare it with the fully connected layer parameter ω k,fc Combined, the vehicle client V is obtained k Personalized model

[0051] The beneficial effects of the present invention are as follows: In view of the problem that data heterogeneity and high mobility characteristics of vehicles in the Internet of Vehicles affect the training accuracy of the federated learning model, the present invention proposes a personalized federated learning method for heterogeneous Internet of Vehicles with cloud-edge-end collaboration. By designing a client selection algorithm and a personalized federated learning method, the computing power and communication capabilities of the vehicle are considered at the same time, and the vehicle clients participating in the federated training are dynamically selected; the idea of ​​frequency domain analysis is used to parse the basic knowledge and personalized knowledge of the local model of the vehicle client, and while integrating the basic knowledge of other vehicle clients, the fine-grained personalized knowledge of the vehicle client is retained, thereby training a high-precision personalized federated learning model.

[0052] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0054] Figure 1 A diagram of the personalized federated learning architecture for heterogeneous Internet of Vehicles provided by the present invention;

[0055] Figure 2 A schematic diagram of the process flow of the personalized federated learning method for heterogeneous Internet of Vehicles provided by the present invention. DETAILED DESCRIPTION

[0056] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0057] See also Figure 1-2 , Figure 1 The following is a diagram of the personalized federated learning architecture for heterogeneous Internet of Vehicles. Figure 1 As shown in the figure, the architecture consists of three layers: cloud server layer, edge layer and user layer. The user layer consists of vehicles in motion; the edge layer consists of roadside units; and the cloud server layer consists of central servers connected to macro base stations. The vehicle client is responsible for collecting user data and storing it locally to participate in federated training, and performing frequency domain analysis and personalized parameter decomposition on local model parameters; the cloud server and roadside unit are responsible for aggregating the basic model knowledge uploaded by the vehicle client; assisting the vehicle in completing personalized federated learning training and improving the model accuracy for local vehicle data.

[0058] like Figure 2 As shown, the personalized federated learning method for heterogeneous vehicle networks proposed in the present invention mainly includes the following steps:

[0059] S1: The cloud server initializes the model parameters and other basic parameters.

[0060] The cloud server customizes personalized model training objectives based on user needs, initializes local iterators, learning rate size, batch data size, local iteration times, training rounds and other training information according to the training objectives, and generates an initial global model.

[0061] S2: Select the set of vehicle clients participating in training.

[0062] The cloud server selects the clients to participate in the training by evaluating the computing and communication performance of each vehicle through the roadside unit. k Performance Local iteration number E k , the number of training samples And the computing resources of the current vehicle client By formula To evaluate the training delay of the vehicle client, the transmission power p of the vehicle client is used. k , path loss index γ, noise power spectrum density, through To evaluate the data transmission rate between the vehicle client and the roadside unit, For vehicle client V k The channel gain between the roadside unit and is the small-scale channel coefficient, c k For vehicle client V k The distance to the roadside unit.

[0063] If the model parameter size sent to the roadside unit is A, then the formula The communication delay required by vehicle client k can be calculated. agg represents the time consumed by the aggregation of basic knowledge of the model at the roadside unit, and δ represents the time used for frequency domain analysis and decomposition of local model parameters.

[0064] Since the vehicle client is constantly moving from one roadside unit coverage area to another during the model training process, it is highly mobile. Indicates the coverage of the current roadside unit, l k Indicates that from the vehicle V k The distance from the location to the entrance of the current roadside unit coverage area, v k Represents the vehicle client V k The driving speed can be calculated by the formula To calculate the residence time τ of vehicle client k within the coverage area of ​​the current roadside unit k According to the obtained vehicle stay time τ kAs well as the vehicle's communication and computing delays, the formula To select the client set that participates in model training.

[0065] S3: Local model training.

[0066] The local model training is performed based on the data set size, selected loss function and other information collected by the client. The calculation formula is expressed as:

[0067]

[0068] Where, d k The size of the dataset collected by the representative vehicle client, represents the jth data sample in the dataset collected by the kth vehicle client, Indicates the corresponding sample true label, Represents the difference between the sample prediction value and the sample's true label, which is the cross entropy loss function. represents the vehicle client V in the tth round of federated training k Personalized model.

[0069] S4: Personalized parameter decomposition, distinguishing between the basic knowledge and personalized knowledge of the vehicle client local model.

[0070] Personalized parameter breakdown includes:

[0071] For the tth round of training, the vehicle clients participating in this round of federated learning train the personalized model obtained in the previous round according to their own collected data sets to obtain an updated vehicle local model. k As an example, the convolutional layer parameters of the local model of the vehicle client are Convert to a two-dimensional matrix Where L and C are the number of output channels and input channels of the convolution kernel in the local model, respectively, and s1 and s2 are the spatial shapes of the convolution kernel. t,k,conv By performing discrete cosine transform, the frequency domain signal F(ω′) after the convolution layer parameters of the vehicle client model are analyzed can be obtained. t,k,conv ), the calculation formula is as follows:

[0072]

[0073] Where, x,u=0,1,2,...,s1L-1; y,v=0,1,2,...,s2C-1 (x, y are coordinates in the time domain, u, v are coordinates in the frequency domain), Used to ensure orthogonality and energy preservation of the transformation, It is the cosine basis function, and the signal decomposition is mainly performed in the transformation.

[0074] A low-frequency mask matrix G with all values ​​0 except the central area is used to extract the low-frequency components for model aggregation (basic knowledge), and the model low-frequency components G⊙F(ω′ t,k,conv ), high frequency component (1-G)⊙F(ω′ t,k,conv ).

[0075] S5: The basic knowledge of the model is uploaded, and the personalized knowledge and fully connected layer parameters are retained locally.

[0076] The vehicle client uploads the basic knowledge of the local model to the roadside unit and cloud server for model aggregation operations, and the personalized knowledge part and fully connected layer parameters are retained locally on the vehicle client.

[0077] S6: Model basic knowledge aggregation, including:

[0078] (1) The roadside unit performs preliminary aggregation of the local model basic knowledge uploaded by the vehicle client. Get the Edge Basics In the formula, { } represents roadside unit B l The set of vehicle clients participating in the tth round of federated training within the coverage area, d k represents the kth vehicle client V k The amount of data collected. express{ The total amount of data of the vehicle clients in the collection.

[0079] (2) Roadside unit will be the basic knowledge of the edge Upload to the cloud server for final aggregation, through the formula Calculation can obtain global basic knowledge ω global .in, represents the set of all roadside units under the coverage of the cloud server, I l Represents roadside unit B l The amount of data you have, Indicates the total data volume of all roadside units within the coverage of the cloud service.

[0080] S7: The cloud server delivers global basic knowledge.

[0081] First, the cloud server sends the aggregated global basic knowledge to each roadside unit, and then the roadside unit sends the global basic knowledge to all vehicle clients within its coverage area.

[0082] S8: Personalized model construction.

[0083] Through step S7, the vehicle client obtains the aggregated global basic knowledge, combines it with the personalized knowledge retained in the vehicle, and obtains the representation of the convolution layer parameters of the personalized model of the vehicle client k in the frequency domain space, as shown in the following formula:

[0084]

[0085] pass Discrete cosine inverse transform represents the convolutional layer parameters of the vehicle client personalized model in the frequency domain space F(ω' t,k,conv ) is converted to the time domain representation ω′ t,k,conv , where x,u=0,1,2,...,s1L-1; y,v=0,1,2,...,s2C-1 (x, y are coordinates in the time domain, u, v are coordinates in the frequency domain, L and C are the number of output channels and input channels of the convolution kernel in the local model, s1 and s2 are the spatial shapes of the convolution kernel). is the normalization coefficient used to ensure orthogonality and energy preservation of the transformation, It is a cosine basis function, which mainly performs signal reconstruction in the transformation, thereby performing more effective signal analysis and processing.

[0086] use The obtained ω′ t,k,conv Combined with the fully connected layer parameters of the model retained locally on the vehicle client, personalized model parameters that are adapted to the data distribution of each vehicle user are customized.

[0087] S9: Repeated iterative training until the predetermined training rounds are reached or the model converges.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A personalized federated learning method for heterogeneous Internet of Vehicles, characterized in that: The method specifically comprises the following steps: S1: The cloud server initializes model parameters and other basic parameters; S2: Select vehicle clients to participate in federated model training based on the vehicle’s computing and communication performance; S3: Use the data collected by each vehicle client to perform local model training; S4: Build a personalized model based on the personalized frequency domain parameter decomposition method; S5: Repeated iterative training until the predetermined training rounds are reached or the model converges.

2. The personalized federated learning method for heterogeneous vehicle networks according to claim 1 is characterized in that: Step S1 specifically includes: the cloud server customizes the training target based on user needs, and initializes the local iterator, learning rate size, batch data size, local iteration number and training round according to the training target, and generates an initial global federated learning model; the cloud server initializes the vehicle client participating in this federated training through the roadside unit.

3. The personalized federated learning method for heterogeneous Internet of Vehicles according to claim 1 is characterized in that: Step S2 specifically includes the following steps: S21: The cloud server notifies each roadside unit to evaluate the device performance and communication status of the vehicle client; S22: The training delay of the roadside unit when evaluating the vehicle client's local training sample data; S23: The roadside unit evaluates the communication delay of the vehicle client; S24: Calculate the residence time of the vehicle client within the coverage area of ​​the current roadside unit; S25: According to the sum of the calculation delay, communication delay, model aggregation time and frequency domain analysis time of the vehicle client, a vehicle client that can complete a round of model training before moving to the next roadside unit is selected for model training; as shown in the following formula: In the formula, represents the training time of the vehicle local model parameters, Indicates the time consumed by model uploading, t agg represents the time taken for model aggregation at the roadside unit, δ represents the time taken for frequency domain analysis and decomposition of local model parameters, τ k Indicates vehicle V k The dwell time within the current roadside unit coverage area.

4. The personalized federated learning method for heterogeneous Internet of Vehicles according to claim 3 is characterized in that: In step S22, the training delay of the vehicle client when training sample data locally Defined as: in, is the vehicle V in the tth round of federated training k performance, is the vehicle V k The computing resources of E k is the number of local iterations, is the number of training samples.

5. The personalized federated learning method for heterogeneous Internet of Vehicles according to claim 3 is characterized in that: Step S23 specifically includes: if the vehicle V k The transmission power is p k , the path loss index is γ, the noise power spectrum density is N0, then under the current roadside unit coverage, the data transmission rate of the vehicle client is expressed as: In the formula, For vehicle client V k Channel gain with the roadside unit, is the small-scale channel coefficient, c k For vehicle client V k distance from roadside units; If the size of the model parameters sent to the roadside unit is A, then the vehicle client V k Required communication delay for:

6. The personalized federated learning method for heterogeneous vehicle networks according to claim 3 is characterized in that: Step S24 specifically includes: assuming that the vehicle client keeps its speed unchanged when passing the edge server, passing through the coverage area of ​​the current roadside unit Vehicle Client V k The driving speed v k , calculate its stay time within the coverage of the current roadside unit: In the formula, l k Indicates that from the vehicle client V k The distance from the location to the entrance of the current roadside unit coverage area.

7. The personalized federated learning method for heterogeneous vehicle networks according to claim 3 is characterized in that: Step S3 specifically includes: training a local model of the vehicle client according to the quality of the data set collected by the current vehicle client and the adopted loss function, which is expressed as: Where, d k represents the size of the dataset collected by the vehicle client, represents the jth data sample in the dataset collected by the kth vehicle client, Indicates the corresponding true label of the sample; Represents the difference between the sample prediction value and the sample's true label, that is, the cross entropy loss function; represents the vehicle client V in the tth round of federated training k Personalized model.

8. The personalized federated learning method for heterogeneous vehicle networks according to claim 7 is characterized in that: Step S4 specifically includes the following steps: S41: Convolutional layer parameters ω for the local model of the vehicle client t ' ,k,conv Perform discrete cosine transform to obtain the frequency domain signal F(ω′) after the convolution layer parameters of the vehicle client model are analyzed t,k,conv ), as shown below: Where x,u=0,1,2,...,s1L-1; y,v=0,1,2,...,s2C-1; x and y are coordinates in the time domain, u and v are coordinates in the frequency domain, L and C are the number of output channels and input channels of the convolution kernel in the local model, s1 and s2 are the spatial shapes of the convolution kernel, ω′ t,k,conv For vehicle V k The convolutional layer parameters of the local model; is the normalization coefficient used to ensure orthogonality and energy preservation of the transformation; is the cosine basis function; S42: A low-frequency mask matrix G with all values ​​0 except the central area is used to extract the low-frequency components used for model aggregation, which is expressed as G⊙F(ω t ' ,k,conv ), the high frequency component is expressed as (1-G)⊙F(ω t ' ,k,conv ); S43: The vehicle client uploads the basic knowledge part of the local model to the roadside unit for preliminary aggregation to obtain the edge basic knowledge Where t represents the tth round of federated training, l represents the lth roadside unit B l , Indicates roadside unit B l The set of vehicle clients participating in the tth round of federated training within the coverage area, d k represents the kth vehicle client V k The amount of data collected; express The total amount of data from the vehicle clients in the collection; S44: Roadside Unit Edge Basics Upload to the cloud server for final aggregation to obtain the aggregated global basic knowledgeω global : In the formula, represents the set of all roadside units under the coverage of the cloud server, I l Represents roadside unit B l The amount of data you have, Indicates the total data size of all vehicle clients within the coverage area of ​​the cloud service; S45: The vehicle client combines the received global basic knowledge with the locally retained personalized knowledge to obtain the representation F(ω' t,k,conv ), defined as: F(ω' t,k,conv )=(1-G)⊙F(ω′ t,k,conv )+ω global In the formula, ω global is the global basic knowledge, (1-G)⊙F(ω′ t,k,conv ) represents the vehicle client V k personalized knowledge; S46: Convert it to the time domain space according to the inverse discrete cosine transform to obtain the time domain representation ω′ of the model convolution layer parameters t,k,conv , and compare it with the fully connected layer parameter ω k,fc Combined to obtain the vehicle client V k Personalized model

Images