Vehicle networking-based split federated learning method, system and computing device

Abstract

The application relates to the technical field of Internet of Vehicles, and discloses a splitting federated learning method and system based on Internet of Vehicles and a computing device. A vehicle end uses a local model distributed by a cloud end, extracts data features, and uploads the data features to the cloud end. The cloud end clusters all vehicle ends by using the data features, captures the non-independent and identically distributed characteristics of the vehicle end data, performs intra-cluster training on a cluster model of each cluster, calculates a first gradient update parameter for updating the cluster model, and issues the first gradient update parameter to the corresponding vehicle end. The vehicle end does not need local training, can calculate a second gradient update parameter by combining the first gradient update parameter and the data features, and realizes the updating of the local model, reduces the local calculation and communication requirements of the vehicle end, realizes the collaborative training of the global model, and has a wider application scene. Finally, the cloud end aggregates the updated local models of all vehicle ends, generates an aggregated model, and deploys the aggregated model to the vehicle end after the training is completed, so that the vehicle end can perform data acquisition work.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Vehicles, in particular to a split federated learning method and system based on Internet of Vehicles and a computing device. BACKGROUND

[0002] The rapid development of Internet of Vehicles has given rise to a large amount of diversified data, which has become the key basis for supporting intelligent transportation, autonomous driving, vehicle-road cooperation and remote monitoring and other applications. However, vehicles are widely distributed, data sources are scattered and diversified, and how to efficiently utilize these data faces many challenges.

[0003] The traditional centralized training method collects all data to a unified location for processing, but has a higher requirement for resource configuration and is difficult to deploy effectively. Federated learning allows vehicles distributed in various places to train models locally and share model updates through encryption or other means to achieve collaborative construction of global models, but the computing resources of vehicle-mounted systems are limited and cannot run complete machine learning models locally. Therefore, split federated learning divides the complete model into levels, so that vehicles only need to train the divided model locally, and finally the cloud aggregates the training models of vehicles to realize collaborative training of the complete model.

[0004] That is, the existing federated learning methods all require vehicles to train models locally, but not all vehicle computing resources are sufficient to support the local model training process, and the above methods have certain limitations in actual application. SUMMARY

[0005] Therefore, the present application provides a split federated learning method and system based on Internet of Vehicles and a computing device to solve the problem that the existing federated learning method is limited by vehicle computing resources and has limited application scenarios.

[0006] In a first aspect, the present application provides a split federated learning method based on Internet of Vehicles, applied to the cloud, which includes multiple rounds of training, and each round of training includes:

[0007] sending a local model suitable for the vehicle end to all vehicle ends, so that each vehicle end extracts features from local data based on the local model, and uploads the extracted data features to the cloud;

[0008] clustering all vehicle ends based on data features to obtain multiple clustering clusters;

[0009] performing intra-cluster training according to the cluster model corresponding to each clustering cluster and the data features to obtain a first gradient update parameter; the cluster model in the first round of training and the local model are divided based on a preset complete model;

[0010] updating the cluster model by using the first gradient update parameter, and sending the first gradient update parameter to the corresponding vehicle end, so that the corresponding vehicle end calculates a second gradient update parameter based on the first gradient update parameter and the data feature, updates the local model by using the second gradient update parameter, and uploads the updated local model to the cloud end;

[0011] aggregating all the updated local models uploaded by the vehicle ends to obtain an aggregated model, and deploying the aggregated model obtained in the last round of training to all the vehicle ends.

[0012] The application divides the preset complete model by using the cloud end to obtain a local model suitable for the vehicle end and a cluster model suitable for the cloud end. The vehicle end extracts data features in local data by using the local model distributed by the cloud end, and uploads the data features to the cloud end. The cloud end first clusters all the vehicle ends by using the data features to capture the non-independent and identically distributed characteristics of the vehicle end data, and then performs intra-cluster training on the cluster model associated with each cluster to calculate a first gradient update parameter for updating the cluster model and distribute the first gradient update parameter to the vehicle end included in the corresponding cluster. The vehicle end does not need local training, but can calculate a second gradient update parameter by combining the first gradient update parameter and the local data feature to update the local model, thereby reducing the local calculation and communication requirements of the vehicle end, realizing collaborative training of the global model, and making the application scenario more extensive. Finally, the cloud end aggregates all the updated local models of the vehicle ends to generate an aggregated model and deploys the aggregated model to the vehicle end after the training is completed, so that the vehicle end performs data acquisition work.

[0013] In an optional implementation, the data feature includes a low-dimensional data feature, a decoder, and a label of the local data, wherein the low-dimensional data feature is obtained by compressing the high-dimensional data feature by using the corresponding encoder of the decoder after the local model converts the local data into the high-dimensional data feature by the vehicle end;

[0014] clustering all the vehicle ends based on the data feature to obtain a plurality of cluster groups, including:

[0015] reconstructing the low-dimensional data feature by using the decoder to obtain a reconstructed feature;

[0016] calculating the similarity between all the vehicle ends in pairs based on the reconstructed feature to obtain a similarity matrix, and clustering all the vehicle ends based on the similarity matrix to obtain a plurality of cluster groups; wherein each cluster group includes at least one vehicle end.

[0017] The application compresses high-dimensional data features extracted from local data into low-dimensional data features at the vehicle end, and greatly reduces the communication burden between the vehicle end and the cloud by uploading the compressed low-dimensional data features to the cloud. And the vehicle end uploads the decoder to the cloud to ensure that the cloud can decode and reconstruct the low-dimensional data features to obtain reconstructed features similar to the extracted features of the vehicle end, thereby performing clustering analysis on all vehicle ends and capturing the non-independent and identically distributed characteristics of vehicle data.

[0018] In an optional implementation, the method further comprises:

[0019] When detecting the offline first vehicle end, merging or splitting the cluster cluster to which the first vehicle end belongs;

[0020] When detecting the newly online second vehicle end, determining the cluster cluster to which the second vehicle end belongs.

[0021] After clustering all vehicle ends, the application considers the offline situation of the first vehicle end, and merges or splits the affected cluster cluster. And considering the online situation of the second vehicle end, the second vehicle end is divided into an existing cluster cluster or a new cluster cluster is created. Therefore, when the vehicle end frequently switches between online and offline states, the stability and adaptive ability of the training process are improved by dynamically adjusting the cluster cluster, and the applicable scene range is more extensive.

[0022] In an optional implementation, the cluster-in training is performed according to the cluster model and the data features corresponding to each cluster cluster to obtain the first gradient update parameter, comprising:

[0023] According to the cluster model and the reconstructed features corresponding to the cluster cluster, a predicted value is obtained, and a deviation value between the predicted value and the label of the local data is calculated;

[0024] According to the deviation value, the predicted value, and the cluster model in the last round of training, back propagation is performed to calculate the first gradient update parameter.

[0025] In the application, the cluster model of each cluster cluster at the cloud is trained using the reconstructed features similar to the extracted features of the vehicle end. Based on the predicted value of the cluster model in the current round, the deviation value of the prediction, and the cluster model in the last round of training, back propagation is performed to calculate the first gradient update parameter of the cluster model, so as to update the cluster model at the cloud and the local model at the vehicle end. Therefore, the clustering and training steps are performed using the computing resources at the cloud, so that the vehicle end can update the local model using the first gradient update parameter calculated by the cloud without training, thereby reducing the computing pressure of the vehicle end.

[0026] In an optional implementation, the method further comprises:

[0027] determining an average computing load of all vehicles, dividing a preset complete model into a first model applicable to the cloud and a second model applicable to the vehicle end based on the average computing load;

[0028] taking the first model as a cluster model corresponding to each cluster in the first round of training, and taking the second model as a local model applicable to the vehicle end in the first round of training.

[0029] The cloud of the present application dynamically divides the preset complete model into the first model applicable to the cloud and the second model applicable to the vehicle end based on the average computing capacity of all vehicle ends, reduces the computing pressure of the vehicle end, and ensures that the computing resources of the vehicle end are sufficient for collaborative training with the cloud.

[0030] In an optional implementation, the updated local models uploaded by all vehicle ends are aggregated to obtain an aggregated model, including:

[0031] The updated local models uploaded by all vehicle ends are federatedly averaged to obtain an aggregated model; wherein the aggregated model is a local model applicable to the vehicle end in the next round of training.

[0032] The present application federatedly averages the updated local models uploaded by all vehicle ends to obtain an aggregated model, realizes collaborative training of the cloud and the vehicle end, and sends the aggregated model to all vehicle ends for the next round of training.

[0033] In a second aspect, the present application provides a split federated learning method based on the Internet of Vehicles, which is applied to any vehicle end and includes multiple rounds of training. Each round of training includes:

[0034] receiving a local model applicable to the vehicle end sent by the cloud;

[0035] extracting features from local data based on the local model, uploading the extracted data features to the cloud, so that the cloud clusters all vehicle ends based on the data features to obtain multiple cluster clusters, and performs intra-cluster training based on the cluster model corresponding to each cluster cluster and the data features to obtain a first gradient update parameter, and updates the cluster model using the first gradient update parameter; wherein the cluster model and the local model in the first round of training are obtained by dividing the preset complete model;

[0036] receiving the first gradient update parameter sent by the cloud, calculating a second gradient update parameter based on the first gradient update parameter and the data features, updating the local model using the second gradient update parameter, and uploading the updated local model to the cloud, so that the cloud aggregates the updated local models uploaded by all vehicle ends to obtain an aggregated model; wherein the aggregated model obtained in the last round of training is used for deployment to all vehicle ends.

[0037] The application divides the preset complete model by using the cloud end to obtain a local model suitable for the vehicle end and a cluster model suitable for the cloud end. The vehicle end extracts data features in the local data by using the local model distributed by the cloud end and uploads the data features to the cloud end. The cloud end first clusters all vehicle ends by using the data features to capture the non-independent and identically distributed characteristics of the vehicle end data, and then performs intra-cluster training on the cluster model associated with each cluster to calculate a first gradient update parameter for updating the cluster model and distribute the first gradient update parameter to the vehicle ends included in the corresponding cluster. The vehicle end does not need local training, but can calculate a second gradient update parameter by combining the first gradient update parameter and the local data features to update the local model, thereby reducing the local calculation and communication requirements of the vehicle end and realizing collaborative training of the global model. Finally, the cloud end aggregates all the updated local models of the vehicle ends to generate an aggregated model and deploys the aggregated model to the vehicle ends after the training ends, so that the vehicle ends can perform data collection.

[0038] In an optional implementation, the feature extraction is performed on the local data based on the local model, and the extracted data features are uploaded to the cloud end, including:

[0039] The local data is converted into high-dimensional data features by using the local model, and the high-dimensional data features are compressed into low-dimensional data features by using an encoder;

[0040] The low-dimensional data features, a decoder corresponding to the encoder, and labels of the local data are uploaded to the cloud end, so that the cloud end reconstructs the low-dimensional data features by using the decoder to obtain reconstructed features, calculates the similarity between all vehicle ends in pairs based on the reconstructed features to obtain a similarity matrix, and clusters all vehicle ends based on the similarity matrix to obtain a plurality of cluster groups; each cluster group includes at least one vehicle end.

[0041] The vehicle end of the application compresses the high-dimensional data features extracted from the local data into low-dimensional data features, and uploads the compressed low-dimensional data features to the cloud end, thereby greatly reducing the communication burden between the vehicle end and the cloud end. The vehicle end uploads the decoder to the cloud end, so that the cloud end can decode and reconstruct the low-dimensional data features to obtain reconstructed features similar to the extracted features of the vehicle end, thereby performing cluster analysis.

[0042] In an optional implementation, the method further includes:

[0043] The low-dimensional data features are reconstructed by using the decoder to obtain reconstructed features;

[0044] The second gradient update parameter is calculated based on the first gradient update parameter and the data features, including:

[0045] The second gradient update parameters are calculated by backpropagation based on the first gradient update parameters, reconstructed features, high-dimensional data features, and the local model from the previous training round.

[0046] This application splits a pre-defined complete model, allowing the vehicle to extract features using the local model while the cloud handles the inference and training of the cluster model. This enables the vehicle to update the local model using the first gradient update parameters computed by the cloud without needing to train the model itself, effectively reducing the computational load on the vehicle.

[0047] Thirdly, this invention provides a vehicle-to-everything (V2X) based split federated learning system, which includes a cloud and multiple vehicle endpoints. In each round of interaction, the cloud and vehicle endpoints are used for:

[0048] The cloud will send the local model suitable for the vehicle to all vehicles;

[0049] The vehicle receives the local model sent from the cloud; based on the local model, it extracts features from the local data and uploads the extracted data features to the cloud;

[0050] The cloud clusters all vehicles based on data features to obtain multiple clusters; intra-cluster training is performed based on the cluster model and data features corresponding to each cluster to obtain the first gradient update parameters; the cluster model and local model in the first round of interaction are obtained based on the preset complete model partition; the cluster model is updated using the first gradient update parameters;

[0051] The vehicle receives the first gradient update parameters sent from the cloud, calculates the second gradient update parameters based on the first gradient update parameters and data features, updates the local model using the second gradient update parameters, and uploads the updated local model to the cloud.

[0052] The cloud aggregates the updated local models uploaded by all vehicles to obtain an aggregated model; the aggregated model obtained in the last round of interaction is used to deploy to all vehicles.

[0053] Fourthly, the present invention provides a computing device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the vehicle-to-everything (V2X) based split federated learning method of the first aspect or any corresponding embodiment thereof, or to perform the vehicle-to-everything (V2X) based split federated learning method of the second aspect or any corresponding embodiment thereof.

[0054] The beneficial effects of this invention are as follows:

[0055] The application divides the preset complete model by using the cloud end to obtain a local model suitable for the vehicle end and a cluster model suitable for the cloud end. The vehicle end extracts data features in local data by using the local model distributed by the cloud end and uploads the data features to the cloud end. The cloud end first clusters all vehicle ends by using the data features to capture the non-independent and identically distributed characteristics of the vehicle end data, and then performs intra-cluster training on the cluster model associated with each cluster to calculate a first gradient update parameter for updating the cluster model and distribute the first gradient update parameter to the vehicle ends included in the corresponding cluster. The vehicle end does not need local training, but can calculate a second gradient update parameter by combining the first gradient update parameter and the local data features to update the local model, thereby reducing the local computing and communication requirements of the vehicle end, realizing collaborative training of the global model, and having a wider application scenario. Finally, the cloud end aggregates all the updated local models of the vehicle ends to generate an aggregated model and deploy the aggregated model to the vehicle end after the training is completed, so that the vehicle end can perform data acquisition.

[0056] Compared with the traditional centralized data processing and training method, the application does not need to collect all vehicle data to a unified location for feature extraction and training, is less limited by the network bandwidth, privacy security requirements and massive data scale of the Internet of Vehicles, and is easier to deploy and implement.

[0057] Compared with most federal learning schemes that run a complete machine learning model locally on a vehicle, the application has lower requirements for real-time data processing, delay, security and computing resources of the vehicle end.

[0058] Compared with the scheme of training part of the model by the vehicle and aggregating the model by the cloud end, the application does not need to train the model by the vehicle end, but only needs to extract features based on the local model by the vehicle end. The extracted data features are uploaded to the cloud end for training of the cluster model, thereby reducing the local computing and communication requirements of the vehicle end and realizing collaborative training of the global model. Moreover, the application considers the non-independent and identically distributed characteristics of data caused by different environments and driving behaviors of different vehicles in the Internet of Vehicles, clusters the data features of the vehicle end by using the cloud end, and fully captures the data diversity. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the specific embodiments of the application or the technical solutions in the prior art, the drawings needed in the following specific embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0060] Figure 1 is a structural block diagram of a split federal learning system based on the Internet of Vehicles according to an embodiment of the application;

[0061] Figure 2 is an interaction process schematic diagram of a split federated learning system according to an embodiment of the present application;

[0062] Figure 3 is an interaction process schematic diagram of another split federated learning system according to an embodiment of the present application;

[0063] Figure 4 is a flow schematic diagram of a split federated learning method according to an embodiment of the present application;

[0064] Figure 5 is a hardware structure schematic diagram of a computing device according to an embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0066] According to an embodiment of the present application, a split federated learning system based on Internet of Vehicles is provided, as shown in the figure, the split federated learning system based on Internet of Vehicles comprises a cloud end 101 and a plurality of vehicle ends 102, the cloud end 101 and the plurality of vehicle ends 102 perform a plurality of rounds of interaction, wherein the cloud end 101 can be a cloud server, and in each round of interaction, the cloud end 101 and the vehicle end 102 are specifically used for: Figure 1

[0067] The cloud end 101 sends a local model applicable to the vehicle end 102 to all vehicle ends 102.

[0068] The vehicle end 102 receives the local model sent by the cloud end 101; extracts features from local data based on the local model, and uploads the extracted data features to the cloud end 101.

[0069] The cloud end 101 clusters all vehicle ends 102 based on the data features, obtains a plurality of clustering clusters, performs intra-cluster training according to the cluster model and the data features corresponding to each clustering cluster, obtains a first gradient update parameter, and updates the cluster model using the first gradient update parameter. The cluster model and the local model in the first round of interaction are obtained based on a preset complete model.

[0070] ​The vehicle end 102 receives the first gradient update parameter sent by the cloud end 101, calculates a second gradient update parameter based on the first gradient update parameter and the data feature, updates the local model by using the second gradient update parameter, and uploads the updated local model to the cloud end 101.

[0071] The cloud end 101 aggregates all the updated local models uploaded by the vehicle ends 102 to obtain an aggregated model; and the aggregated model obtained in the last round of interaction is used for deployment to all the vehicle ends 102.

[0072] The split federated learning system based on the Internet of Vehicles provided in the embodiments of the present application is used for, in each round of interaction between the cloud end 101 and the vehicle end 102, extracting a data feature in local data by using a local model distributed by the cloud end 101, and uploading the data feature to the cloud end 101. The cloud end 101 first clusters all the vehicle ends 102 by using the data feature, so as to capture the non-independent and identically distributed characteristics of the data of the vehicle ends 102, and then performs intra-cluster training on a cluster model associated with each cluster, calculates a first gradient update parameter for updating the cluster model, and distributes the first gradient update parameter to the vehicle ends 102 included in the corresponding cluster. The vehicle end 102 can calculate a second gradient update parameter by combining the first gradient update parameter and the local data feature, so as to realize the update of the local model, reduce the local calculation and communication requirements of the vehicle end 102, realize the collaborative training of the global model, and have a wider application scenario. Finally, the cloud end 101 aggregates all the updated local models of the vehicle ends 102 to generate an aggregated model, and deploys the aggregated model to the vehicle end after the training is completed, so that the vehicle end can perform data acquisition work.

[0073] The specific working principles and working processes of the cloud end 101 and the vehicle end 102 will be described below in the related description of the method embodiments, and will not be described here in detail.

[0074] According to the embodiments of the present application, a split federated learning method based on the Internet of Vehicles is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0075] In the present embodiment, a split federated learning method based on the Internet of Vehicles is provided, which can be used for the cloud end 101 and the vehicle end 102 as shown in Figure 1 Figure 2 ​is a schematic diagram of an interaction process of a vehicle networking-based split federated learning system according to an embodiment of the present application, wherein the cloud end 101 is configured to perform steps S101-S105, the vehicle end 102 is configured to perform steps S201-S203, and the specific interaction process between the cloud end 101 and the vehicle end 102 is as follows:

[0076] Step S101, sending a local model suitable for the vehicle end to all vehicle ends.

[0077] Specifically, in order to ensure that the computing resources of all vehicle ends are sufficient for federated learning with the cloud end, a preset complete model is dynamically divided to obtain a local model suitable for the vehicle end, and is distributed to all vehicle ends. The local model at least includes a module responsible for feature extraction in the preset complete model, and the cluster model of the cloud end in the first round of training is part of the preset complete model except the feature extraction module. The preset complete model can be a neural network model, a machine learning model, etc.

[0078] Step S201, receiving the local model suitable for the vehicle end sent by the cloud end.

[0079] Specifically, the vehicle end receives the local model uniformly distributed by the cloud end, and all vehicle ends perform feature extraction based on the uniform local model.

[0080] Step S202, performing feature extraction on local data based on the local model, and uploading the extracted data features to the cloud end.

[0081] Specifically, the local data of the vehicle end can be data from vehicle-mounted sensors, intelligent driving assistance systems, in-vehicle entertainment systems, and other vehicle terminal devices. The vehicle end extracts data features from the local data using the local model, and uploads the data features to the cloud end for training.

[0082] Step S102, clustering all vehicle ends based on the data features to obtain a plurality of clustering clusters.

[0083] Specifically, in view of the non-independent and identically distributed problem of vehicle end data, the cloud end clusters all vehicle ends according to the data features uploaded by each vehicle end. Vehicle ends with similar data features are divided into the same clustering cluster, and each clustering cluster is associated with at least one vehicle end.

[0084] Step S103, performing intra-cluster training according to the cluster model corresponding to each clustering cluster and the data features to obtain a first gradient update parameter.

[0085] Specifically, the local model of the vehicle end interacts parameters with the cluster model corresponding to the cluster to which the vehicle end belongs, and the data features extracted by the local model are uploaded to the cluster model corresponding to the cluster in the cloud, forming intra-cluster collaborative training between the local model of the vehicle end and the cluster model of the cluster in the cloud. The cluster model of each cluster in the cloud is trained to obtain the first gradient update parameter required for updating the cluster model.

[0086] In step S104, the cluster model is updated using the first gradient update parameter, and the first gradient update parameter is sent to the corresponding vehicle end.

[0087] Specifically, for the first gradient update parameter corresponding to the cluster model of each cluster, the first gradient update parameter is sent to the vehicle end included in the cluster.

[0088] In step S203, the first gradient update parameter sent by the cloud is received, the second gradient update parameter is calculated based on the first gradient update parameter and the data features, the local model is updated using the second gradient update parameter, and the updated local model is uploaded to the cloud.

[0089] Specifically, since the cluster model and the local model are divided based on the preset complete model, the first gradient update parameter for updating the cluster model in the cloud can be combined with the data features of the vehicle end to calculate the second gradient update parameter required for updating the local model.

[0090] In the embodiments of the present application, by dividing the preset complete model into a local model suitable for the vehicle end and a cluster model suitable for the cloud, the vehicle end extracts features using the local model, the cloud clusters all vehicle ends using the data features uploaded by the vehicle ends, and the cluster model corresponding to each cluster is trained within the cluster, thereby reducing the computing pressure of the vehicle end. Moreover, the vehicle end does not need to be trained locally, but can update its local model using the first gradient update parameter sent by the cloud, further reducing the requirement for computing resources required for collaborative training of the vehicle end, and having stronger applicability.

[0091] In step S105, the updated local models uploaded by all vehicle ends are aggregated to obtain an aggregated model; wherein the aggregated model obtained by the last round of training is used to be deployed to all vehicle ends.

[0092] Specifically, the cloud generates an aggregated model by aggregating the updated local models of all vehicle ends, and after the training is completed, the aggregated model is deployed on all vehicle ends for data collection by the vehicle ends.

[0093] Compared with the traditional centralized data processing and training method, the application does not need to collect all vehicle data to a unified location for feature extraction and training, is less limited by the network bandwidth, privacy security requirements and massive data scale of the Internet of Vehicles, and is easier to deploy and implement.

[0094] Compared with most federated learning schemes that run complete machine learning models locally on vehicles, the application has lower requirements for real-time data processing, delay, security and computing resources on the vehicle side.

[0095] Compared with the scheme of training part of the model by the vehicle and finally aggregating the model by the cloud, the application does not need to train the model on the vehicle side, but only needs to extract features based on the local model on the vehicle side. The extracted data features are uploaded to the cloud, and the cloud trains the cluster model, thereby reducing the local computing and communication requirements on the vehicle side and realizing collaborative training of the global model. Moreover, the application takes into account the non-independent and identically distributed characteristics of data caused by different environments and driving behaviors of different vehicles in the Internet of Vehicles, clusters the data features on the vehicle side by the cloud, and fully captures the diversity of data.

[0096] The split federated learning method based on the Internet of Vehicles provided in this embodiment divides the preset complete model by the cloud to obtain a local model suitable for the vehicle side and a cluster model suitable for the cloud side. The vehicle side extracts data features in local data by using the local model distributed by the cloud, and uploads them to the cloud. The cloud first clusters all vehicle sides by using the data features to capture the non-independent and identically distributed characteristics of the vehicle side data, and then trains the cluster model associated with each cluster based on the data features, calculates the first gradient update parameter for updating the cluster model, and distributes it to the vehicle side included in the corresponding cluster. The vehicle side does not need to train locally, but can calculate the second gradient update parameter by combining the first gradient update parameter and the local data features to update the local model, thereby reducing the local computing and communication requirements on the vehicle side, realizing collaborative training of the global model, and having a wider application scenario. Finally, the cloud aggregates all the updated local models of the vehicle side to generate an aggregated model and deploy it to the vehicle side after training is completed, so that the vehicle side can perform data acquisition work.

[0097] In this embodiment, a split federated learning method based on the Internet of Vehicles is provided, which can be used in a system as shown in Figure 1 The cloud side 101 and the vehicle side 102, Figure 3 is an interaction process diagram of a split federated learning system based on the Internet of Vehicles according to an embodiment of the application, wherein the cloud side 101 is used to perform steps S301 to S307, the vehicle side 102 is used to perform steps S401 to S405, and the specific interaction process of the cloud side 101 and the vehicle side 102 is as follows:

[0098] Step S301: Send the local model applicable to the vehicle end to all vehicle ends.

[0099] Specifically, before the first round of interaction between the cloud and the vehicle, the vehicle-to-everything (V2X) based split federated learning system is initialized. The system initialization includes dividing the preset complete model.

[0100] In some alternative implementations, the cloud determines the average computing load of all vehicles, and based on the average computing load, divides the preset complete model into a first model applicable to the cloud. The second model applicable to the vehicle side The cloud dynamically adjusts the segmentation points of the preset complete model based on the average computing load of all vehicles, dynamically segmenting the preset complete model into the first model. Second Model .

[0101] In some embodiments, the second model It should include at least the module responsible for feature extraction in the pre-defined complete model, the first model. This refers to the part of the pre-defined complete model excluding the feature extraction module. The pre-defined complete model can be a neural network model, a machine learning model, etc. At the start of the first round of interaction, the first model is... The second model serves as the cluster model corresponding to each cluster in the first round of training, and is used as the cluster model for each cluster in the first round of training. This serves as the local model applicable to the vehicle during the initial training phase. The cloud then broadcasts the initialized local model to all vehicles, establishing a unified initial training benchmark.

[0102] In this embodiment, based on the average computing power of all vehicles, the cloud dynamically divides the preset complete model into a first model applicable to the cloud and a second model applicable to the vehicles, thereby reducing the computing pressure on the vehicles and ensuring that the computing resources of the vehicles are sufficient to complete collaborative training with the cloud.

[0103] Step S401: Receive the local model for the vehicle sent from the cloud. See details below. Figure 2 The detailed description of step S201 in the illustrated embodiment will not be repeated here.

[0104] Step S402: Use the local model to convert the local data into high-dimensional data features, use the encoder to compress the high-dimensional data features into low-dimensional data features, and use the decoder corresponding to the encoder to reconstruct the low-dimensional data features to obtain the reconstructed features.

[0105] Specifically, regarding the first i individual vehicle end Vehicle side Based on local data Perform forward propagation computation on the local model to compute high-dimensional data features. The specific formula is as follows:

[0106] (1)

[0107] in, For the first t Vehicle end during wheel training The local model, This indicates that forward propagation computation is performed on the local model. For example, ,in, express The corresponding weights express The corresponding bias vector.

[0108] Furthermore, on the vehicle side via encoder High-dimensional data features Mapping to a low-dimensional feature space, thereby transforming high-dimensional data features Compression into low-dimensional data features The specific formula is as follows:

[0109] (2)

[0110] in, This indicates encoding calculation.

[0111] Furthermore, on the vehicle side Using encoder F corresponding decoder Features of compressed low-dimensional data Reconstruction is performed to generate features that approximate high-dimensional data. Reconstruction features The specific formula is as follows:

[0112] (3)

[0113] in, This indicates decoding calculation.

[0114] In some alternative implementations, during each round of training, the vehicle end... Calculate the variance loss function By minimizing the variance loss function in multiple rounds of training Optimize encoder F and decoder G The parameters. Among them, the variance loss function. The specific formula is as follows:

[0115] (4)

[0116] wherein, denotes the number of features, denotes the vehicle end high-dimensional data features the first j feature data, denotes the vehicle end reconstructed features the first j feature data.

[0117] It should be noted that the number of rounds required for the encoder F and the decoder G to converge is much smaller than the maximum number of iterations required for split federated learning T , and in general, the encoder F and the decoder G can converge in the 4th-5th round of training.

[0118] In step S403, the low-dimensional data features, the decoder, and the labels of the local data are uploaded to the cloud.

[0119] Specifically, the local data refers to the data collected by the vehicle end and stored locally. Under the premise of compliance and privacy protection, the vehicle end sends the stored local data to the data administrator through an encrypted channel, manually labels the local data, sets the labels of the local data, and returns them to the corresponding vehicle end through an encrypted manner. Taking image data collected by a camera as an example, the boundary box region in the image data is labeled manually, thereby setting the label of the image data, which can be “pedestrian”, “vehicle”, “obstacle”, etc.

[0120] In some embodiments, the vehicle end can also use the unsupervised learning model deployed on itself to automatically label the labels of the local data. For details, please refer to the description of related technologies, which will not be repeated here.

[0121] Specifically, the vehicle end stores the local data and the labels of the local data. After the vehicle end converts and compresses the local data into low-dimensional data features, the vehicle end uploads the decoder, the low-dimensional data features, and the labels of the local data to the cloud. G

[0122] In step S302, the low-dimensional data features are reconstructed using the decoder to obtain reconstructed features.

[0123] ​​​​​​Specifically, the cloud uses a decoder. G Features of low-dimensional data Reconstruction is performed to obtain reconstructed features. ,Right now It should be noted that, assuming no data transmission errors or decoding errors occur, the reconstruction characteristics in the cloud are... Reconstruction features of the vehicle end They are consistent.

[0124] Step S303: Calculate the pairwise similarity between all vehicle terminals based on the reconstructed features to obtain a similarity matrix, and cluster all vehicle terminals based on the similarity matrix to obtain multiple clusters; wherein each cluster contains at least one vehicle terminal.

[0125] Specifically, the cloud can build structures of size [size missing]. Similarity matrix , This represents the total number of active vehicles and the similarity matrix. Initialization is performed. The cloud-based system calculates vehicle-side parameters based on reconstructed features. and vehicle end cosine similarity Among them, cosine similarity The calculation formula is as follows:

[0126] (5)

[0127] in, Indicates vehicle end Reconstruction features .

[0128] Furthermore, by calculating the pairwise similarity between vehicles, the similarity matrix is... Maintain and utilize the similarity matrix. All vehicle terminals are clustered to obtain multiple clusters. For a detailed description of the clustering process, please refer to the relevant technical descriptions, which will not be repeated here.

[0129] In some embodiments, the cloud determines the cluster to which each vehicle belongs, and assigns it to a specific vehicle cluster. Assign a unique cluster identifier Determine the first t Clustering in rounds of training Corresponding cluster model and maintain In the first round of training, the cluster model of each cluster is the first model obtained by dividing the preset complete model in the cloud. .

[0130] In some embodiments, the cloud depends on the vehicle end cluster identifier of the cluster , a dynamic binding relationship is established between the local model of the vehicle end and the cluster model of the cluster to which the vehicle end belongs, that is , wherein .

[0131] In the embodiments of the present application, the vehicle end compresses the high-dimensional data features extracted from the local data into low-dimensional data features, and greatly reduces the communication burden between the vehicle end and the cloud end by uploading the compressed low-dimensional data features to the cloud end. And the vehicle end uploads the decoder to the cloud end to ensure that the cloud end can decode and reconstruct the low-dimensional data features to obtain reconstructed features similar to the extracted features of the vehicle end, thereby performing clustering analysis on all vehicle ends and capturing the non-independent and identically distributed characteristics of vehicle data.

[0132] In some optional embodiments, when the first vehicle end is detected to be offline, the cluster cluster to which the first vehicle end belongs is merged or split.

[0133] Specifically, when the vehicle end exits the connection with the cloud end, that is, when the first vehicle end is detected to be offline, the corresponding entry of the first vehicle end in the similarity matrix is deleted, and the affected cluster cluster is re-evaluated to determine whether the merging or splitting condition is met.

[0134] In some embodiments, after the first vehicle end exits, the average cosine distance between the cluster cluster to which the first vehicle end belongs and other cluster clusters is calculated, and when the average cosine distance between two cluster clusters is less than a set merging threshold , it indicates that the cluster cluster to which the first vehicle end belongs can be merged with other cluster clusters. Wherein, the calculation formula of the average cosine distance is as follows:

[0135] (6)

[0136] Wherein, denotes the average cosine distance between the cluster cluster and the cluster cluster .

[0137] In some embodiments, the cloud end generates a cluster model of a new cluster cluster for the cluster clusters that meet the merging condition, that is :

[0138] (7)

[0139] The embodiments of the present application merge the cluster cluster to which the first vehicle end belongs and other cluster clusters when the first vehicle end is offline, and determine a new cluster cluster The cluster model enables dynamic adjustment of clusters.

[0140] In some embodiments, after the first vehicle exits, it is determined whether the cluster to which the first vehicle belongs meets the splitting criteria. When the cluster... The splitting operation is performed when the following splitting conditions are met:

[0141] (8)

[0142] in, This indicates the set splitting threshold. Represents clusters The sum of squares within the cluster, Represents clusters The center of mass.

[0143] Furthermore, the cloud provides clusters that meet the splitting criteria. Calculate the removal of the first vehicle end The average sum of squares within the cluster after that ,like Then the first vehicle end Mark as removed. Repeat the above process for each first vehicle terminal in the offline vehicle terminal set until there are no more first vehicle terminals that can further reduce the average intra-cluster sum of squares.

[0144] In this embodiment, when the first vehicle is offline, the cluster to which the first vehicle belongs is split to achieve dynamic adjustment of the cluster.

[0145] In some alternative implementations, when a newly online second vehicle terminal is detected, the cluster to which the second vehicle terminal belongs is determined.

[0146] Specifically, when the new second vehicle end When joining the vehicle-to-everything (V2X) network, the second vehicle is detected. Upon launch, cloud computing will be used for the second vehicle terminal. With existing vehicle terminals cosine similarity :

[0147] (9)

[0148] Furthermore, if the second vehicle end With existing vehicle terminals cosine similarity Maximum cosine similarity Exceeding the set similarity threshold Then the second vehicle end Distributed to vehicle end If the cluster contains the second vehicle, then create a new cluster containing the second vehicle. New clusters The new cluster The cluster model can be the first model. .

[0149] To address the issue of non-independent and identically distributed data, various personalized strategies have been proposed. For example, multi-task learning methods treat the data collected from each vehicle as multiple interconnected but distinct tasks; model interpolation techniques balance the influence of the global and local models to achieve a trade-off between model generalization ability and personalized specialization; and cluster-based joint learning methods further enhance the model's adaptability to non-independent and identically distributed data environments by grouping vehicles with similar data features.

[0150] Although related technologies have improved the training performance of models in non-independent and identically distributed scenarios to some extent, they often assume that the vehicle group is in a static state or require high computing resources. When faced with the actual dynamic environment of vehicles frequently going online or offline, their scalability and stability are significantly limited.

[0151] In this embodiment, after clustering all vehicle terminals, the offline status of the first vehicle terminal is considered, and the affected clusters are merged or split; and the online status of the second vehicle terminal is considered, and the second vehicle terminal is assigned to an existing cluster or a new cluster is created for it. Thus, when the vehicle terminal frequently switches between online and offline states, the clusters are dynamically adjusted, which improves the stability and adaptability of the training process and makes it applicable to a wider range of scenarios.

[0152] Step S304: Based on the cluster model and reconstruction features corresponding to the clusters, obtain the predicted values ​​and calculate the deviation between the predicted values ​​and the labels of the local data.

[0153] Specifically, the local model on the vehicle interacts with the cluster model corresponding to its cluster. The high-dimensional data features extracted by the local model are compressed and uploaded to the cloud for decoding and reconstruction. The reconstructed features are then input into the corresponding cluster model, forming a collaborative training between the local model on the vehicle and the cluster model of the cloud cluster.

[0154] In some embodiments, the cloud targets each cluster. Perform intra-cluster training for clusters Each vehicle end Based on reconstruction features Cluster model Perform forward propagation calculations to obtain the predicted values. ,in, This indicates that forward propagation computation is performed on the cluster model, for example, ,in, denotes the corresponding weight, denotes the corresponding bias vector. The prediction bias is evaluated by the following loss function L:

[0155] (10)

[0156] wherein, denotes the prediction value and the label of the local data .

[0157] Step S305, according to the bias value, the prediction value and the cluster model in the last round of training , the first gradient update parameter is calculated by back propagation.

[0158] Specifically, the first gradient update parameter can be calculated according to the following formula:

[0159] (11)

[0160] wherein, denotes the cluster in the first gradient update parameter in the first round of training, denotes the cluster in the first gradient update parameter in the first round of training.

[0161] In the embodiment of the present application, the cloud end uses the reconstructed features similar to the features extracted by the vehicle end to train the cluster model of the cluster, and performs back propagation based on the prediction value of the cluster model in the current round, the predicted bias value and the cluster model in the last round of training to calculate the first gradient update parameter of the cluster model, so as to update the cluster model of the cloud end and the local model of the vehicle end. Thus, the clustering and training steps are performed by using the computing resources of the cloud end, so that the vehicle end can update the local model by using the first gradient update parameter calculated by the cloud end without training, thereby reducing the computing pressure of the vehicle end.

[0162] Step S306, the cluster model is updated by using the first gradient update parameter, and the first gradient update parameter is sent to the corresponding vehicle end.

[0163] Specifically, the formula for updating the cluster model by using the first gradient update parameter is as follows:

[0164] (12)

[0165] wherein, denotes the cluster The corresponding updated cluster model, represents the learning rate set.

[0166] In some embodiments, the clustering cluster The first gradient update parameter The relevant information is sent to the vehicle end of the clustering cluster .

[0167] Step S404, receiving the first gradient update parameter sent by the cloud end, based on the first gradient update parameter, the reconstructed feature, the high-dimensional data feature and the local model in the last round of training, performing back propagation to calculate the second gradient update parameter.

[0168] Specifically, the calculation formula of the second gradient update parameter is as follows:

[0169] (13)

[0170] Wherein, represents the second gradient update parameter of the vehicle end in the first round of training, represents the local model of the vehicle end in the first round of training.

[0171] Step S405, updating the local model by using the second gradient update parameter, and uploading the updated local model to the cloud end.

[0172] Specifically, the formula for updating the local model by using the second gradient update parameter is as follows:

[0173] (14)

[0174] Wherein, represents the updated local model of the vehicle end .

[0175] The application splits the preset complete model, and the vehicle end extracts features by using the local model, and the cloud end is responsible for the inference training of the cluster model, so that the vehicle end can update the local model by using the first gradient update parameter calculated by the cloud end without training, thereby effectively reducing the calculation load of the vehicle end.

[0176] Step S307, performing federated averaging on all the updated local models uploaded by the vehicle end to obtain an aggregated model; wherein the aggregated model is a local model suitable for the vehicle end in the next round of training, and the aggregated model obtained in the last round of training is used for deployment to all vehicle ends.

[0177] Specifically, the cloud receives the updated model parameters of all vehicle ends , and generates an aggregated model using a federated averaging algorithm . The calculation formula can be as follows:

[0178] (15)

[0179] wherein, n represents the number of vehicle ends uploading the updated local model.

[0180] Further, the aggregated model is distributed to each vehicle end as a training benchmark for the next round of training, completing the federated learning closed loop. The above interaction process is repeated until the training round reaches the maximum iteration number .

[0181] In the embodiments of the present application, the updated local models uploaded by all vehicle ends are federated averaged to obtain an aggregated model, realizing collaborative training of the cloud and the vehicle end, and the aggregated model is sent to all vehicle ends for the next round of training.

[0182] The split federated learning method based on Internet of Vehicles provided by the embodiments of the present application trains a model based on split federated learning, divides a preset complete model into front-end feature extraction of a vehicle end and back-end reasoning of a cloud, compresses high-dimensional data features extracted by the vehicle end into low-dimensional data features and uploads the low-dimensional data features to the cloud for reasoning training, which can effectively reduce the calculation load of the vehicle end and the system communication pressure, while guaranteeing the data privacy of users.

[0183] The present application is aimed at the non-independent and identically distributed of heterogeneous vehicle data, the dynamic change of online state of the vehicle end and the bandwidth-limited Internet of Vehicles scene, dynamically adjusts the clustering cluster of the cloud in combination with the offline or online situation of the vehicle end, thereby realizing collaborative update of the cluster model and the local model, significantly improving the training effect and speed in a heterogeneous environment, and being applicable to complex application scenarios, and having stronger expansibility and adaptability.

[0184] The split federated learning based on Internet of Vehicles will be described in detail in combination with a specific application example as shown in the following. Figure 4 The application example includes the following steps:

[0185] Step 1, system initialization.

[0186] Step 1.1, the cloud dynamically adjusts the split point of the preset complete model according to the average calculation load of all vehicle ends, dynamically divides the preset complete model into a first model and a second model , and ensures the calculation pressure of the lightweight vehicle end.

[0187] Step 1.2, the cloud distributes the second model​ As the local model applicable to the vehicle side in the first round of training, the initialized local model is broadcast to all vehicles to establish a unified initial training benchmark.

[0188] Step 1.3: A similarity matrix can be constructed in the cloud. And for the similarity matrix Perform initialization.

[0189] Step 2: Autoencoder training.

[0190] Step 2.1, Vehicle side Based on local data Perform forward propagation computation on the local model to compute high-dimensional data features. .

[0191] Step 2.2, Vehicle Side via encoder High-dimensional data features Mapping to a low-dimensional feature space to transform high-dimensional data features Compression into low-dimensional data features .

[0192] Step 2.3, Vehicle Side Using encoder F corresponding decoder Features of compressed low-dimensional data Reconstruction is performed to generate features that approximate high-dimensional data. Reconstruction features .

[0193] Step 2.4, during each round of training, the vehicle end... Calculate the variance loss function By minimizing the variance loss function in multiple rounds of training Optimize encoder F and decoder G The parameters.

[0194] Step 2.5, Vehicle Side decoder G Low-dimensional data characteristics and local data tags Uploaded to the cloud.

[0195] Step 3: Cluster management.

[0196] Step 3.1, use the decoder in the cloud. G Features of low-dimensional data Reconstruction is performed to obtain reconstructed features. .

[0197] Step 3.2, the cloud maintains a similarity matrix by computing the similarity between each pair of vehicle ends, and uses the similarity matrix to cluster all vehicle ends into multiple clusters.

[0198] Specifically, when a new second vehicle end joins, the cloud computes the cosine similarity between the second vehicle end and existing vehicle ends, and assigns the second vehicle end to the cluster that the second vehicle end has the largest cosine similarity with, if the largest cosine similarity exceeds a set similarity threshold; otherwise, a new cluster is created.

[0199] Specifically, when a first vehicle end exits, the cloud deletes the corresponding entry in the similarity matrix, and re-evaluates the affected cluster to determine whether to merge or split. For clusters that meet the merging condition, the cloud generates a new cluster model using the federated averaging algorithm.

[0200] Specifically, for clusters that meet the splitting condition, the cloud computes the average intra-cluster sum of squares after removing the first vehicle end, and if the average intra-cluster sum of squares decreases, the first vehicle end is marked for removal. The process is repeated until there is no first vehicle end that can further reduce the average intra-cluster sum of squares.

[0201] Step 4: Model binding.

[0202] Step 4.1, the cloud assigns a unique cluster identifier to each vehicle end, determines the cluster model corresponding to the cluster in the nth round of training, and maintains the cluster model. t

[0203] ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​Step 4.2, the cloud establishes a dynamic binding relationship for the local model of the vehicle end and the cluster model of the cluster to which the vehicle end belongs, i.e. , to ensure that the vehicle end only cooperatively trains with the cluster model of the cluster to which it belongs.

[0204] Step 5: Model cooperative training and aggregation.

[0205] Step 5.1, the cloud performs intra-cluster training for each cluster , for each vehicle end in the cluster , performs forward propagation calculation on the cluster model based on the reconstructed features to obtain the predicted value . The prediction deviation is evaluated by the loss function L and the first gradient update parameter is calculated by the back propagation calculation, and the cluster model of the cluster is updated using the first gradient update parameter . The cloud sends the first gradient update parameter to the vehicle end in the cluster .

[0206] Step 5.2, the vehicle end receives the first gradient update parameter sent by the cloud , calculates the second gradient update parameter required for updating the local model, and updates the local model using the second gradient update parameter .

[0207] Step 5.3, the cloud receives the updated model parameters of all vehicle ends , and generates an aggregated model using the federated average algorithm.

[0208] Step 5.4, send the aggregated model to each vehicle end as the initialization reference for the next round of training, complete the federated learning closed loop.

[0209] Repeat steps 2 to 5 above until the maximum number of iterations of training is reached .

[0210] ​​The present application aims at the problem of non-independent and identically distributed data in the Internet of Vehicles environment, clusters the vehicle end, assigns vehicles with similar data characteristics to the same cluster, and customizes different cluster models for each cluster, thereby significantly improving the performance and adaptability of the overall model. At the same time, in order to cope with the dynamic characteristics of frequent online and offline of vehicles in the Internet of Vehicles, a cluster management mechanism is introduced, which can efficiently realize the merging and splitting of clusters. In addition, feature compression is used to reduce communication overhead in the data transmission process.

[0211] The present application also provides a computing device which can be applied to the cloud 101 or the vehicle end 102 shown in the above Figure 1 .

[0212] Please refer to Figure 5 , Figure 5 is a structural schematic diagram of a computing device provided by an optional embodiment of the present application, as shown in the figure, the computing device comprises one or more processors 10, a memory 20, and an interface for connecting various components, including a high-speed interface and a low-speed interface. Various components are communicatively connected to each other by different buses, and can be installed on a common motherboard or in other ways as needed. The processor can process instructions executed within the computing device, including instructions stored in the memory or memory to display GUI on an external input / output device (such as a display device coupled to the interface). In some optional embodiments, multiple processors or buses can be used with multiple memories and multiple memories if needed. Similarly, multiple computing devices can be connected, each providing part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 5 . Figure 5 The processor 10 in the above embodiment is taken as an example.

[0213] The processor 10 can be a central processor, a network processor or a combination thereof. The processor 10 can further include a hardware chip. The hardware chip can be an application specific integrated circuit, a programmable logic device or a combination thereof. The programmable logic device can be a complex programmable logic device, a field programmable logic gate array, a general array logic or any combination thereof.

[0214] The memory 20 stores instructions executable by the at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiments.

[0215] The memory 20 can include a program storage area and a data storage area. The program storage area can store an operating system, application programs required for at least one function, etc. The data storage area can store data created by the computing device, etc. In addition, the memory 20 can include a high-speed random access memory, and can also include a non-transitory memory such as at least one disk memory device, a flash memory device, or other non-transitory solid state memory device. In some alternative embodiments, the memory 20 can optionally include memory that is remotely located with respect to the processor 10, and which can be connected to the computing device through a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communications network, and combinations thereof.

[0216] The memory 20 can include a volatile memory, such as a random access memory, and can also include a non-volatile memory, such as at least one disk memory device, a flash memory device, or other non-transitory solid state memory device. The memory 20 can also include a combination of the above-mentioned types of memory.

[0217] The computing device also includes a communication interface 30 for communicating with other devices or communication networks.

[0218] The embodiments of the present application also provide a computer readable storage medium, and the above-mentioned method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or stored in a remote storage medium or a non-transitory machine readable storage medium and downloaded through network and stored in a local storage medium, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special purpose hardware. The storage medium can be a disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that the computer, processor, microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, processor or hardware, the method shown in the above embodiments is implemented.

[0219] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, can invoke or provide the method or technical solutions according to the present application. Those skilled in the art should understand that the form of computer program instructions in computer readable medium includes but is not limited to source file, executable file, installation package file and the like, and accordingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0220] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A decomposed federated learning method based on vehicle-to-everything (V2X) communication, characterized in that, The method includes: Step 1.1: The cloud dynamically adjusts the segmentation points of the preset complete model based on the average computing load of all vehicles, dynamically segmenting the preset complete model into the first model. Second Model ; Step 1.2, the cloud will send the second model As a local model applicable to the vehicle during the first round of training, the initialized local model is broadcast to all vehicles. Step 1.3: Construct a similarity matrix in the cloud. and the similarity matrix Perform initialization; Step 2.1, Vehicle side Based on local data Perform forward propagation computation on the local model to compute high-dimensional data features. ; Step 2.2, Vehicle Side via encoder High-dimensional data features Mapping to a low-dimensional feature space to transform high-dimensional data features Compression into low-dimensional data features ; Step 2.3, Vehicle Side Using encoder F corresponding decoder Features of compressed low-dimensional data Reconstruction is performed to generate features that approximate high-dimensional data. Reconstruction features ; Step 2.4, during each round of training, the vehicle end... Calculate the variance loss function By minimizing the variance loss function in multiple rounds of training Optimize encoder F and decoder G Parameters; Step 2.5, Vehicle Side decoder G Low-dimensional data characteristics and local data tags Upload to the cloud; Step 3.1, use the decoder in the cloud. G Features of low-dimensional data Reconstruction is performed to obtain reconstructed features. ; Step 3.2: The cloud calculates the pairwise similarity between vehicles and then applies the similarity matrix. Maintain and utilize the similarity matrix. Clustering is performed on all vehicle terminals to obtain multiple clusters; When there is a new second vehicle end Upon joining, cloud computing is used for the second vehicle terminal. With existing vehicle terminals cosine similarity If the second vehicle end With existing vehicle terminals cosine similarity Maximum cosine similarity Exceeding the set similarity threshold Then the second vehicle end Distributed to vehicle end If the cluster contains the second vehicle, then create a new cluster containing the second vehicle. New clusters ; When the first vehicle exits, its information is removed from the similarity matrix. The corresponding entries are then analyzed, and the affected clusters are re-evaluated to determine whether they meet the conditions for merging or splitting. The cloud platform then considers entries that meet the merging conditions. of Clusters New clusters are generated using the federated averaging algorithm. The cluster model, in which, The set merging threshold; Cloud-based clustering of data that meets the splitting criteria Calculate the removal of the first vehicle end The average sum of squares within the cluster after that ,like Then the first vehicle end Marked as removed, the above process is repeated until no first vehicle end exists that can further reduce the average intra-cluster sum of squares, where, Represents clusters The sum of squares within the cluster; Step 4.1, the cloud is for the vehicle end. Assign a unique cluster identifier Determine the first t Clustering in rounds of training Corresponding cluster model and maintain ; Step 4.2, the cloud platform, based on the vehicle terminal... Cluster identifier This establishes a dynamic binding relationship between the local model on the vehicle side and the cluster model of the cluster to which the vehicle belongs, i.e. This ensures that the vehicle's terminal is only trained in collaboration with the cluster model of its own cluster. Step 5.1, the cloud performs a process for each cluster. Perform intra-cluster training for clusters Each vehicle end Based on reconstruction features Cluster model Perform forward propagation calculations to obtain the predicted values. ,in, This indicates that forward propagation computation is performed on the cluster model; Evaluation of prediction bias using the loss function L : in, Indicates predicted value and local data tags The deviation between; And perform backpropagation to calculate the first gradient and update the parameters. Update parameters using the first gradient Update clusters Cluster model; calculate the first gradient update parameters according to the following formula. : in, Represents clusters In the The first gradient updates the parameters in the first round of training. Indicates the first Clustering in rounds of training The corresponding cluster model; The cloud will update the parameters in the first tier. Distribute to cluster The vehicle end in the middle; Step 5.2, Vehicle Side Receive the first gradient update parameters sent from the cloud. Calculate the second gradient update parameters required for local model update. Update parameters using the second gradient Update the local model; the formula for calculating the second gradient update parameters is as follows: in, Indicates vehicle end In the The second gradient updates the parameters during the training round. Indicates the first Vehicle end during wheel training The local model, For vehicle end High-dimensional data characteristics, For the first t Vehicle end during wheel training The local model, This indicates that forward propagation computation is performed on the local model; Step 5.3: Receive all updated model parameters from the vehicle terminals in the cloud. The aggregate model is generated using the federated average algorithm. ; Step 5.4, aggregate model The data is sent to each vehicle as the initialization baseline for the next round of training, completing the federated learning loop; the aggregated model obtained from the last round of training is used to deploy to all vehicles. Repeat steps 2.1 through 5.4 above until the maximum number of training iterations is reached. .

2. A split-federated learning system based on vehicle networking, characterized in that, The system includes a cloud and multiple vehicle terminals, which are used in each round of interaction for: Step 1.1: The cloud dynamically adjusts the segmentation points of the preset complete model based on the average computing load of all vehicles, dynamically segmenting the preset complete model into the first model. Second Model ; Step 1.2, the cloud will send the second model As a local model applicable to the vehicle during the first round of training, the initialized local model is broadcast to all vehicles. Step 1.3: Construct a similarity matrix in the cloud. and the similarity matrix Perform initialization; Step 2.1, Vehicle side Based on local data Perform forward propagation computation on the local model to compute high-dimensional data features. ; Step 2.2, Vehicle Side via encoder High-dimensional data features Mapping to a low-dimensional feature space to transform high-dimensional data features Compression into low-dimensional data features ; Step 2.3, Vehicle Side Using encoder F corresponding decoder Features of compressed low-dimensional data Reconstruction is performed to generate features that approximate high-dimensional data. Reconstruction features ; Step 2.4, during each round of training, the vehicle end... Calculate the variance loss function By minimizing the variance loss function in multiple rounds of training Optimize encoder F and decoder G Parameters; Step 2.5, Vehicle Side decoder G Low-dimensional data characteristics and local data tags Upload to the cloud; Step 3.1, use the decoder in the cloud. G Features of low-dimensional data Reconstruction is performed to obtain reconstructed features. ; Step 3.2: The cloud calculates the pairwise similarity between vehicles and then applies the similarity matrix. Maintain and utilize the similarity matrix. Clustering is performed on all vehicle terminals to obtain multiple clusters; When there is a new second vehicle end Upon joining, cloud computing is used for the second vehicle terminal. With existing vehicle terminals cosine similarity If the second vehicle end With existing vehicle terminals cosine similarity Maximum cosine similarity Exceeding the set similarity threshold Then the second vehicle end Distributed to vehicle end If the cluster contains the second vehicle, then create a new cluster containing the second vehicle. New clusters ; When the first vehicle exits, its information is removed from the similarity matrix. The corresponding entries are identified, and the affected clusters are re-evaluated to determine whether the conditions for merging or splitting are met. For cloud-based systems that meet the merging conditions, i.e. of Clusters New clusters are generated using the federated averaging algorithm. The cluster model, in which, The set merging threshold; Cloud-based clustering of data that meets the splitting criteria Calculate the removal of the first vehicle end The average sum of squares within the cluster after that ,like Then the first vehicle end Marked as removed, the above process is repeated until no first vehicle end exists that can further reduce the average intra-cluster sum of squares, where, Represents clusters The sum of squares within the cluster; Step 4.1, the cloud is for the vehicle end. Assign a unique cluster identifier Determine the first t Clustering in rounds of training Corresponding cluster model and maintain ; Step 4.2, the cloud platform, based on the vehicle terminal... Cluster identifier This establishes a dynamic binding relationship between the local model on the vehicle side and the cluster model of the cluster to which the vehicle belongs, i.e. This ensures that the vehicle's terminal is only trained in collaboration with the cluster model of its own cluster. Step 5.1, the cloud performs a process for each cluster. Perform intra-cluster training for clusters Each vehicle end Based on reconstruction features Cluster model Perform forward propagation calculations to obtain the predicted values. ,in, This indicates that forward propagation computation is performed on the cluster model; Evaluation of prediction bias using the loss function L : in, Indicates predicted value and local data tags The deviation between; And perform backpropagation to calculate the first gradient and update the parameters. Update parameters using the first gradient Update clusters Cluster model; calculate the first gradient update parameters according to the following formula. : in, Represents clusters In the The first gradient updates the parameters in the first round of training. Indicates the first Clustering in rounds of training The corresponding cluster model; The cloud will update the parameters in the first tier. Distribute to cluster The vehicle end in the middle; Step 5.2, Vehicle Side Receive the first gradient update parameters sent from the cloud. Calculate the second gradient update parameters required for local model update. Update parameters using the second gradient Update the local model; the formula for calculating the second gradient update parameters is as follows: in, Indicates vehicle end In the The second gradient updates the parameters during the training round. Indicates the first Vehicle end during wheel training The local model, For vehicle end High-dimensional data characteristics, For the first t Vehicle end during wheel training The local model, This indicates that forward propagation computation is performed on the local model; Step 5.3: Receive all updated model parameters from the vehicle terminals in the cloud. The aggregate model is generated using the federated average algorithm. ; Step 5.4, aggregate model The data is sent to each vehicle as the initialization baseline for the next round of training, completing the federated learning loop; the aggregated model obtained from the last round of training is used to deploy to all vehicles. Repeat steps 2.1 through 5.4 above until the maximum number of training iterations is reached. .

3. A computing device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to perform the split federated learning method based on the Internet of Vehicles as described in claim 1.

Images