Internet of vehicles federal semi-supervised learning method and system

Abstract

The invention relates to the technical field of Internet of Vehicles, in particular to an Internet of Vehicles federal semi-supervised learning method and system, and the method comprises the steps: in each iteration, a vehicle client firstly generates a pseudo label of unlabeled sample data, then carries out the training of a local model, and uploads the trained model to a server; and after the server side aggregates the local models of all the vehicle clients, the local models and the supervision model trained by using the labeled sample data are aggregated again to obtain a global model of the current iteration, and the global model continues to be used for training the local model in the next iteration. According to the invention, the model training efficiency is improved, the real-time performance and adaptability of the system are further improved, and the coping ability of the vehicle in a complex and changeable traffic environment is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Vehicles, and in particular to an Internet of Vehicles federated semi-supervised learning method and system. Background Art

[0002] The construction of new infrastructure such as 5G base stations, artificial intelligence, and new energy vehicle charging piles has greatly promoted the development of the Internet of Vehicles industry. In the Internet of Vehicles, intelligent connected vehicles exchange and share information with people, vehicles, roads, and clouds through on-board sensors and control systems and information terminal equipment, thereby realizing intelligent assisted driving. The continuous improvement of the intelligence level of modern automobiles and the addition of high-performance computing and mobile communication modules have enabled the Internet of Vehicles to go a step further in the original traffic management application. A series of Internet of Vehicles applications based on emerging artificial intelligence technologies, such as vehicle-road collaboration, automatic driving, and vehicle-borne logistics, have emerged.

[0003] The basis for the combination of Internet of Vehicles and artificial intelligence is that vehicles, as intelligent terminals, generate data, and then use this data to generate specific artificial intelligence models for practical applications. The data required to create models for autonomous driving is large and scattered. In vehicle networks, this data is unevenly distributed among a wide range of vehicle terminal devices. To address this problem, in addition to the traditional data-centric model deployment architecture, distributed model learning and deployment methods have gradually developed. In addition, with the improvement of data privacy and security awareness in recent years, laws and regulations related to data privacy and security have been gradually improved, and the way to obtain personal data from vehicle intelligent terminals has become an increasingly important consideration for information security.

[0004] In recent years, the privacy-safe distributed Internet of Vehicles artificial intelligence learning framework has gradually attracted the attention of researchers at home and abroad. Relevant research on distributed learning using methods such as federated learning has been carried out. For distributed vehicle terminal data, traditional data centralized learning has high transmission bandwidth costs and data privacy leakage problems. Data distributed learning provides a potential solution for the Internet of Vehicles scenario.

[0005] Federated learning has attracted widespread attention from researchers at home and abroad due to its distributed learning paradigm and the premise of protecting privacy and security, and has gradually been applied to the construction of intelligent learning frameworks for the Internet of Vehicles.

[0006] Zhou et al. from Harbin Institute of Technology proposed using Federated Learning (FL) as a distributed learning framework. For the local data of vehicles in the Internet of Vehicles, each vehicle terminal can learn a local model locally using local data, and then upload the local model to the roadside unit (RSU). The roadside unit acts as a server to aggregate the local model and send the aggregated global model to each vehicle client for the next local model learning.

[0007] Luca et al. from the Polytechnic University of Milan in Italy considered federated learning on deep learning architectures commonly used to solve autonomous driving tasks, and proposed a modular approach in which the federated learning process is enabled on a variable number of machine learning model layers, reducing the complexity of the model to a certain extent and optimizing the overall learning time.

[0008] Manias et al. from the University of Western Ontario, Canada, discussed the applications and advantages of using federated learning in Internet of Vehicles (IoV) and Intelligent Transportation System (ITS). With the development of technologies such as 5G networks, Network Function Virtualization (NFV), and edge computing, intelligent transportation systems are able to achieve low latency, high bandwidth, and real-time data processing. However, there are challenges in terms of data privacy, system complexity, and model performance. To address these issues, researchers proposed the use of federated learning, a distributed machine learning method that can improve the resilience and efficiency of the system while protecting data privacy. Case studies demonstrate the potential applications of federated learning in ITS, such as intelligent roadside units, NFV orchestration, and intelligent predictive maintenance of vehicles, and point out its broad application prospects in future 5G and beyond networks.

[0009] Jin et al. from the Hong Kong University of Science and Technology reviewed the current status and potential of using unlabeled data in FL, pointing out that although FL, as a privacy-preserving distributed learning method, has been widely used in some fields, many existing applications rely on automatically labeled data and are difficult to apply to large-scale data sets that require manual labeling. Researchers believe that FL urgently needs to develop technologies to utilize unlabeled data to address the problem of lack of labeled data in cross-device and cross-organization scenarios. At the same time, Jeong et al. from the Korea Institute of Science and Technology proposed two federated semi-supervised scenarios, one with labels on the client and the other on the server, so that federated learning can better fit the data distribution characteristics in real scenarios and further utilize unlabeled data.

[0010] Zhang et al. from the University of Technology Sydney proposed the problem of using federated semi-supervised learning for traffic mode identification (TMI), paying special attention to privacy protection and efficient crowdsourcing data training under insufficient data conditions. They proposed a novel semi-supervised learning framework that combines deep learning and federated learning techniques to improve model performance through consistent updates and average teacher models without relying on a large amount of labeled data, and effectively protect user privacy.

[0011] However, the existing technology of the Internet of Vehicles federated semi-supervised learning model based on unlabeled data requires a lot of computing resources and time during training, which is not conducive to the rapid training and updating of the model, and cannot meet the real-time and adaptability requirements of the intelligent driving system. Summary of the invention

[0012] To this end, the technical problem to be solved by the present invention is to overcome the problem that the vehicle network federated semi-supervised learning model with unlabeled data in the prior art requires more computing resources and time during training, which is not conducive to the rapid training and updating of the model and cannot meet the real-time and adaptability requirements of the intelligent driving system.

[0013] In order to solve the above technical problems, the present invention provides a federated semi-supervised learning method for Internet of Vehicles, comprising:

[0014] The server uses labeled data to train the supervision model and uses the trained supervision model as the initial global model.

[0015] In each iteration, each vehicle client downloads the current global model as the initial local model, uses local unlabeled sample data to train the initial local model, and uses the pseudo-label selection method to optimize the training of the initial local model to obtain a trained local model; all vehicle clients upload the trained local models to the server; the server aggregates all trained local models to obtain an aggregated local model; the server trains the supervised model with labeled data to obtain a trained supervised model; the server aggregates the aggregated local model and the trained supervised model into a global model for the next iteration;

[0016] After reaching the preset maximum number of iterations, the target global model is output;

[0017] The server publishes the target global model to the vehicle client.

[0018] Preferably, the objective function of the vehicle client training the local model is:

[0019]

[0020] in, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model The target loss function is Represents the initial local model The output of i (x u ) and π j (x u ) represent the unlabeled sample data x u Perform the i-th and j-th data enhancement processing, λ1 and λ2 both represent weight coefficients, Represents unlabeled sample data x u , CE(·) represents the cross entropy function.

[0021] Preferably, a pseudo-label selection method based on Softmax layer confidence and uncertainty is used to optimize the training of the local model, including:

[0022] After multiple data enhancements, the unlabeled sample data is input into the initial local model, and then the multiple confidences output by the Softmax layer of the initial local model are accumulated, and the preliminary pseudo-label is obtained based on the Softmax layer confidence threshold strategy;

[0023] Calculating the confidence uncertainty of the unlabeled sample data based on multiple confidences of the unlabeled sample data;

[0024] When the confidence uncertainty of the unlabeled sample data is less than or equal to the certainty threshold, its preliminary pseudo-label is used as the pseudo-label.

[0025] Preferably, multiple confidences output by the Softmax layer of the initial local model are accumulated, and a preliminary pseudo label is obtained based on the Softmax layer confidence threshold strategy, and the formula is:

[0026]

[0027] in, represents the preliminary pseudo-label, I represents the number of data enhancements, π i (x u ) represents the unlabeled sample data x u Perform the i-th data enhancement processing, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model The output of; c represents the predicted category, p represents the confidence, Max(·) represents the function that outputs the sample label category with the maximum cumulative confidence, p crepresents the confidence of the predicted category c, and τ represents the confidence threshold of the Softmax layer.

[0028] Preferably, the confidence uncertainty of the unlabeled sample data is calculated based on multiple confidences of the unlabeled sample data, and the formula is:

[0029]

[0030] Among them, u(x u ) represents the unlabeled sample data x u The confidence uncertainty, I represents the number of data enhancements, p i Represents unlabeled sample data x u The i-th confidence level, E(p i ) indicates p i Expected; π i (x u ) represents the unlabeled sample data x u Perform the i-th data enhancement processing, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model Output.

[0031] Preferably, when the confidence uncertainty of the unlabeled sample data is less than or equal to the certainty threshold, its preliminary pseudo-label is used as the pseudo-label, and the formula is:

[0032]

[0033] in, Represents unlabeled sample data x u The pseudo label of u(x u ) represents the unlabeled sample data x u is the confidence uncertainty, and γ represents the certainty threshold.

[0034] Preferably, the objective function of the server-side training supervision model is:

[0035]

[0036] Among them, σ t represents the supervised model of the tth iteration, L S (σ t ) represents the supervised model σ t The loss, x l and y represent the labeled sample data and their corresponding labels respectively. Represents sample feature data x l In the supervised model σ t The output of , CE(·) represents the cross entropy function.

[0037] Preferably, the server aggregates all local models that have completed training to obtain an aggregated local model, and the formula is:

[0038]

[0039] Among them, φ t+1 represents the local aggregation model at the t+1th iteration, represents the local model trained by the kth vehicle client at the t+1th iteration, |D u,k |

[0040] Preferably, the server aggregates the local model and the trained supervision model into a global model, and the formula is:

[0041] w t+1 =αφ t+1 +(1-α)σ t+1

[0042] Among them, w t+1 represents the global model of the t+1th iteration, α is a hyperparameter, φ t+1 represents the local aggregation model at the t+1th iteration, σ t+1 Indicates the supervised model that completes training at the t+1th iteration.

[0043] The present invention also provides a vehicle network federated semi-supervised learning system, comprising:

[0044] The initial training module is used by the server to train the supervised model using labeled data and use the trained supervised model as the initial global model;

[0045] The iterative training module is used for each vehicle client to download the current global model as the initial local model in each iteration, use the local unlabeled sample data to train the initial local model, and use the pseudo-label selection method to optimize the training of the initial local model to obtain a local model that has been trained; all vehicle clients upload the local models that have been trained to the server; the server aggregates all the local models that have been trained to obtain an aggregated local model; the server trains the supervised model with labeled data to obtain a supervised model that has been trained; the server aggregates the aggregated local model and the supervised model that has been trained into a global model for the next iteration;

[0046] A model output module is used to output the target global model after reaching a preset maximum number of iterations;

[0047] The publishing module is used by the server to publish the target global model to the vehicle client.

[0048] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0049] The method for semi-supervised learning of a federated Internet of Vehicles described in the present invention takes into account the characteristics of distributed unlabeled sample data in a real Internet of Vehicles scenario. In each iteration, the vehicle client first generates a pseudo-label for the unlabeled sample data, then trains the local model and uploads it to the server. After the local models of all vehicle clients are aggregated, the server aggregates them again with the supervised model trained with labeled sample data to obtain the global model of the current iteration, which is continued to be used for the training of the local model in the next iteration. The present invention combines the semi-supervised learning method of the vehicle client with the distributed federated learning architecture, making full use of a small amount of labeled sample data of the server and a large amount of unlabeled sample data of the vehicle client. The unlabeled sample data in the Internet of Vehicles can be used for local training while protecting data privacy and security, reducing data communication costs and improving the real-time performance of the system. In addition, the present invention uses a small amount of labeled sample data from the server as an aid, further optimizes the federated semi-supervised learning through the pseudo-label selection method, improves the model training efficiency, further improves the real-time performance and adaptability of the system, enhances the vehicle's ability to cope with complex and changeable traffic environments, and provides more accurate and comprehensive data analysis and decision support for intelligent driving. Furthermore, the present invention does not rely on a specific deep learning network architecture and data category and has a certain degree of versatility.

[0050] Furthermore, the present invention adopts a pseudo-label selection method based on Softmax layer confidence and uncertainty to optimize the training of the local model. Through the Softmax layer confidence threshold and uncertainty threshold, the pseudo-label selection in the local model is reasonably constrained, thereby improving the quality of pseudo-label selection and thus improving the training accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 It is a flow chart of a method for federated semi-supervised learning of Internet of Vehicles of the present invention;

[0053] Figure 2 This is a flowchart of the pseudo-label selection method based on the confidence and uncertainty of the Softmax layer;

[0054] Explanation of the markings in the specification: 1. Local unlabeled sample data training; 2. Server labeled sample data training; 3. Local model upload; 4. Global model download; 5. Local model aggregation; 6. Global model aggregation. DETAILED DESCRIPTION

[0055] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0056] Embodiment 1

[0057] Reference Figure 1 As shown, the present invention proposes a federated semi-supervised learning method for Internet of Vehicles, comprising:

[0058] The server uses labeled data to train the supervision model and uses the trained supervision model as the initial global model.

[0059] In each iteration, each vehicle client downloads the current global model as the initial local model, uses local unlabeled sample data to train the initial local model, and uses the pseudo-label selection method to optimize the training of the initial local model to obtain a trained local model; all vehicle clients upload the trained local models to the server; the server aggregates all trained local models to obtain an aggregated local model; the server trains the supervised model with labeled data to obtain a trained supervised model; the server aggregates the aggregated local model and the trained supervised model into a global model for the next iteration;

[0060] After reaching the preset maximum number of iterations, the target global model is output;

[0061] The server publishes the target global model to the vehicle client.

[0062] The scenario defined in the present invention includes a server and K vehicle clients, where the server only has labeled data, and the labeled data set is D l , the vehicle client only has unlabeled sample data, and the unlabeled dataset of each vehicle client is D u,k .

[0063] The server is responsible for aggregating the local models of the vehicle clients and training the supervised model with labeled data. The client is responsible for training with local unlabeled sample data and optimizing the semi-supervised learning effect of the local model using pseudo-label selection methods.

[0064] The entire learning framework of the present invention includes three parts: training part, communication part and aggregation part. The training part includes local learning of the vehicle client and server, the communication part includes local model uploading and global model downloading, and the aggregation part includes local model aggregation and global model aggregation.

[0065] This embodiment takes the t+1th iteration as an example, and the server uses the labeled data set D lTrain the supervised model σ for the tth iteration t , and obtain the supervised model σ after parameter update t+1 .

[0066] The objective loss function of the supervised model is:

[0067]

[0068] Among them, σ t represents the supervised model of the tth iteration, L S (σ t ) represents the supervised model σ t The loss, x l and y represent the labeled sample data and their corresponding labels in the labeled dataset, respectively. Represents sample feature data x l In the supervised model σ t The output of , CE(·) represents the cross entropy function.

[0069] For vehicle clients, the kth vehicle client receives the global model w of the tth iteration t , which is used as the initial local model for the t+1th iteration

[0070] Reference Figure 2 As shown, the present invention uses a pseudo-label selection method based on Softmax layer confidence and uncertainty on the vehicle client to calculate the pseudo-label of each unlabeled sample data in the local unlabeled data set, including:

[0071] Each unlabeled sample data in the local unlabeled dataset is simultaneously enhanced by I different data. The data enhancement methods include but are not limited to flipping, translation and rotation.

[0072] The unlabeled sample data x u After I types of data enhancement, input into the initial local model Then accumulate the confidence of the output of the Softmax layer of the initial local model.

[0073] The preliminary pseudo-labels obtained based on the Softmax layer confidence threshold strategy are:

[0074]

[0075] in, represents the preliminary pseudo-label, I represents the number of data enhancements, π i (x u ) represents the unlabeled sample data x u Perform the i-th data enhancement processing, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model The output of; c represents the predicted category, p represents the confidence, Max(·) represents the function that outputs the sample label category with the maximum cumulative confidence, p c represents the confidence of the predicted category c, and τ represents the confidence threshold of the Softmax layer.

[0076] When the confidence level p of category c c ≥τ, category c becomes the preliminary pseudo label.

[0077] After the confidence threshold of the Softmax layer is limited, in order to further ensure the reliability of the confidence pseudo-label selection, it is necessary to u The confidence uncertainty is calculated based on the I confidence levels.

[0078] The confidence uncertainty is calculated as:

[0079]

[0080] Among them, u(x u ) represents the unlabeled sample data x u The confidence uncertainty, p i Represents unlabeled sample data x u The i-th confidence level, E(p i ) indicates p i expectations.

[0081] When the unlabeled sample data x u When the confidence uncertainty is less than or equal to the certainty threshold, the preliminary pseudo label is selected As a pseudo-label, the formula is:

[0082]

[0083] in, Represents unlabeled sample data x u The pseudo-label of , γ represents the certainty threshold.

[0084] For the local unlabeled dataset D u,k Calculate pseudo labels for all unlabeled sample data in the local unlabeled dataset D u,k In the initial local model Train on it and get the local model with updated parameters

[0085] The objective loss function for local model training is:

[0086]

[0087] in, Represents the initial local model The target loss function, π j (x u ) represents the unlabeled sample data x u The jth data enhancement processing is performed, λ1 represents the weight coefficient of consistency regularization, and λ2 represents the weight coefficient of the pseudo-label cross entropy function. Used to limit the difference in the prediction output of the local model for the same sample after different data augmentation processing.

[0088] The kth vehicle client completes the initial local model After training, the local model with updated parameters is obtained

[0089] All vehicle clients will complete the trained local model Upload to the server, and the server receives K local models

[0090] The server aggregates all K local models that have completed training to obtain an aggregated local model. The formula is:

[0091]

[0092] Among them, φ t+1 represents the local aggregation model at the t+1th iteration, represents the local model trained by the kth vehicle client at the t+1th iteration, |D u,k | represents the local unlabeled dataset D of the kth vehicle client u,k The number of sample data.

[0093] The server will aggregate the local model φ t+1 And the supervised model σ that completes the training t+1 Aggregate into a global model w t+1 , the formula is:

[0094] w t+1 =αφ t+1 +(1-α)σ t+1

[0095] Among them, w t+1 represents the global model at the t+1th iteration, φ t+1 represents the local aggregation model at the t+1th iteration, σ t+1 represents the supervised model trained at the t+1th iteration; α is a hyperparameter used to control the aggregation ratio between the server supervised model and the client local aggregation model in aggregation.

[0096] By aggregating the local models of all vehicle clients and the supervised model on the server side, the scattered data can be integrated while reducing data transmission costs. The resulting global model can learn more comprehensive driving scenarios and modes, improving the accuracy and generalization ability of the model.

[0097] In summary, the semi-supervised learning method of the Internet of Vehicles federation described in the present invention takes into account the characteristics of distributed unlabeled sample data in the real Internet of Vehicles scenario. In each iteration, the vehicle client first generates a pseudo-label for the unlabeled sample data, then trains the local model and uploads it to the server; after the server aggregates the local models of all vehicle clients, it aggregates them again with the supervised model trained with labeled sample data to obtain the global model of the current iteration, which is continued to be used for the training of the local model in the next iteration. The present invention combines the semi-supervised learning method of the vehicle client with the distributed federated learning architecture, making full use of the small amount of labeled sample data of the server and the large amount of unlabeled sample data of the vehicle client, and can use the unlabeled sample data in the Internet of Vehicles for local training while protecting data privacy and security, thereby reducing data communication costs. In addition, the present invention uses a small amount of labeled sample data from the server as an aid, further optimizes the federated semi-supervised learning through the pseudo-label selection method, improves the model training efficiency, further improves the real-time and adaptability of the system, enhances the vehicle's ability to cope with complex and changeable traffic environments, and provides more accurate and comprehensive data analysis and decision support for intelligent driving. Furthermore, the present invention does not rely on a specific deep learning network architecture and data category and has a certain degree of versatility.

[0098] Furthermore, the present invention adopts a pseudo-label selection method based on Softmax layer confidence and uncertainty to optimize the training of the local model. Through the Softmax layer confidence threshold and uncertainty threshold, the pseudo-label selection in the local model is reasonably constrained, thereby improving the quality of pseudo-label selection and thus improving the training accuracy of the model.

[0099] Embodiment 2

[0100] Based on the vehicle network federation semi-supervised learning method described in Example 1, this embodiment provides a vehicle network federation semi-supervised learning system, including:

[0101] The initial training module is used by the server to train the supervised model using labeled data and use the trained supervised model as the initial global model;

[0102] The iterative training module is used for each vehicle client to download the current global model as the initial local model in each iteration, use the local unlabeled sample data to train the initial local model, and use the pseudo-label selection method to optimize the training of the initial local model to obtain a local model that has been trained; all vehicle clients upload the local models that have been trained to the server; the server aggregates all the local models that have been trained to obtain an aggregated local model; the server trains the supervised model with labeled data to obtain a supervised model that has been trained; the server aggregates the aggregated local model and the supervised model that has been trained into a global model for the next iteration;

[0103] A model output module is used to output the target global model after reaching a preset maximum number of iterations;

[0104] The publishing module is used by the server to publish the target global model to the vehicle client.

[0105] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0106] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0107] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0108] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0109] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for semi-supervised learning of Internet of Vehicles federation, characterized in that: include: The server uses labeled data to train the supervision model and uses the trained supervision model as the initial global model. In each iteration, each vehicle client downloads the current global model as the initial local model, uses local unlabeled sample data to train the initial local model, and uses the pseudo-label selection method to optimize the training of the initial local model to obtain a trained local model; all vehicle clients upload the trained local models to the server; The server aggregates all trained local models to obtain an aggregated local model. The server uses labeled data to train the supervised model to obtain a trained supervised model. The server aggregates the aggregated local model and the trained supervised model into a global model for the next iteration. After reaching the preset maximum number of iterations, the target global model is output; The server publishes the target global model to the vehicle client.

2. The method for semi-supervised learning of a federated Internet of Vehicles according to claim 1, wherein: The objective function of the vehicle client training the local model is: in, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model The target loss function is Represents the initial local model The output of i (x u ) and π j (x u ) represent the unlabeled sample data x u Perform the i-th and j-th data enhancement processing, λ1 and λ2 both represent weight coefficients, Represents unlabeled sample data x u , CE(·) represents the cross entropy function.

3. The method for semi-supervised learning of Internet of Vehicles federation according to claim 1, characterized in that: The training of the local model is optimized using a pseudo-label selection method based on the confidence and uncertainty of the Softmax layer, including: After multiple data enhancements, the unlabeled sample data is input into the initial local model, and then the multiple confidences output by the Softmax layer of the initial local model are accumulated, and the preliminary pseudo-label is obtained based on the Softmax layer confidence threshold strategy; Calculating the confidence uncertainty of the unlabeled sample data based on multiple confidences of the unlabeled sample data; When the confidence uncertainty of the unlabeled sample data is less than or equal to the certainty threshold, its preliminary pseudo-label is used as the pseudo-label.

4. The method for semi-supervised learning of Internet of Vehicles federation according to claim 3, characterized in that: Accumulate multiple confidences output by the Softmax layer of the initial local model, and obtain preliminary pseudo labels based on the Softmax layer confidence threshold strategy. The formula is: in, represents the preliminary pseudo-label, I represents the number of data enhancements, π i (x u ) represents the unlabeled sample data x u Perform the i-th data enhancement processing, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model The output of; c represents the predicted category, p represents the confidence, Max(·) represents the function that outputs the sample label category with the maximum cumulative confidence, p c represents the confidence of the predicted category c, and τ represents the confidence threshold of the Softmax layer.

5. The method for semi-supervised learning of Internet of Vehicles federation according to claim 4, characterized in that: The confidence uncertainty of the unlabeled sample data is calculated based on multiple confidence levels of the unlabeled sample data. The formula is: Among them, u(x u ) represents the unlabeled sample data x u The confidence uncertainty, I represents the number of data enhancements, p i Represents unlabeled sample data x u The i-th confidence level, E(p i ) indicates p i Expected; π i (x u ) represents the unlabeled sample data x u Perform the i-th data enhancement processing, represents the initial local model of the k-th vehicle client at the t+1th iteration, Represents the initial local model Output.

6. The method for semi-supervised learning of Internet of Vehicles federation according to claim 5, characterized in that: When the confidence uncertainty of the unlabeled sample data is less than or equal to the certainty threshold, its preliminary pseudo-label is used as the pseudo-label, and the formula is: in, Represents unlabeled sample data x u The pseudo label of u(x u ) represents the unlabeled sample data x u is the confidence uncertainty, and γ represents the certainty threshold.

7. The method for semi-supervised learning of Internet of Vehicles federation according to claim 1, characterized in that: The objective function of the server-side training supervision model is: Among them, σ t represents the supervised model of the tth iteration, L S (σ t ) represents the supervised model σ t The loss, x l and y represent the labeled sample data and their corresponding labels respectively. Represents sample feature data x l In the supervised model σ t The output of , CE(·) represents the cross entropy function.

8. The method for semi-supervised learning of Internet of Vehicles federation according to claim 1, characterized in that: The server aggregates all local models that have completed training to obtain an aggregated local model. The formula is: Among them, φ t+1 represents the local aggregation model at the t+1th iteration, represents the local model trained by the kth vehicle client at the t+1th iteration, |D u,k | represents the local unlabeled dataset D of the kth vehicle client u,k The number of sample data.

9. The method for semi-supervised learning of Internet of Vehicles federation according to claim 1, characterized in that: The server aggregates the local model and the trained supervision model into a global model. The formula is: w t+1 =after t+1 +(1-a)s t+1 Among them, w t+1 represents the global model of the t+1th iteration, α is a hyperparameter, φ t+1 represents the local aggregation model at the t+1th iteration, σ t+1 Indicates the supervised model that completes training at the t+1th iteration.

10. A federated semi-supervised learning system for Internet of Vehicles, characterized in that: include: The initial training module is used by the server to train the supervised model using labeled data and use the trained supervised model as the initial global model; Iterative training module, used for each vehicle client to download the current global model as the initial local model in each iteration, use the local unlabeled sample data to train the initial local model, and use the pseudo-label selection method to optimize the training of the initial local model to obtain a trained local model; all vehicle clients upload the trained local model to the server; The server aggregates all trained local models to obtain an aggregated local model. The server uses labeled data to train the supervised model to obtain a trained supervised model. The server aggregates the aggregated local model and the trained supervised model into a global model for the next iteration. A model output module is used to output the target global model after reaching a preset maximum number of iterations; The publishing module is used by the server to publish the target global model to the vehicle client.

Images