Vehicle dynamic cooperative control method and system based on opportunistic federated learning

Abstract

The application provides a vehicle dynamic cooperative control method and system based on opportunity federal learning, wherein the method comprises the following steps: acquiring resource information, processing density and position information of a fog node and data of a vehicle node in real time; determining a processing score according to the resource information; calculating a distance score and a data volume from the vehicle node to the fog node; allocating the vehicle node to each fog node based on these factors; the fog node performs model training by using the received data and uploads gradients; finally, the vehicle node inputs real-time data into a global model to obtain a target decision. The application has the beneficial effects that the training speed of the model is improved, the adaptability of the model to diversified driving environments is enhanced, the limitations of insufficient generalization ability and poor training effect of a traditional training method are overcome, the vehicle can quickly adjust the strategy according to the current traffic condition, and thus the overall traffic flow and safety of the intelligent traffic system are optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle joint control, in particular to a vehicle dynamic cooperative control method and system based on opportunistic federated learning. BACKGROUND

[0002] In modern intelligent transportation systems, vehicle dynamic cooperative control technology is increasingly concerned about research and application. As an emerging distributed learning method, opportunistic federated learning (OFL) allows vehicles to improve model performance through collaborative training while protecting data privacy. However, the current OFL framework mainly relies on locally stored models for training, but the parameter update of the local model is only based on a single local data source, resulting in insufficient model generalization ability and suboptimal training results. SUMMARY

[0003] The main purpose of the present application is to provide a vehicle dynamic cooperative control method and system based on opportunistic federated learning, which aims to solve the problem that the parameter update of the local model is only based on a single local data source, resulting in insufficient model generalization ability.

[0004] The present application provides a vehicle dynamic cooperative control method based on opportunistic federated learning, which is realized through an opportunistic federated learning architecture. The architecture includes a cloud, multiple fog nodes, and vehicle nodes. Each fog node is connected to at least one vehicle node, and each fog node is connected to the cloud. The fog node is used to receive data information collected by each vehicle node and train an intermediate model to obtain training gradients, which are uploaded to the cloud. The cloud trains a global model based on the training gradients uploaded by each fog node and sends the global model to each vehicle node. The method includes:

[0005] Real-time acquisition of available resource information, processing density, and first location information of each fog node, as well as data information collected by each vehicle node and second location information of each vehicle node;

[0006] Determine the processing score of the corresponding fog node based on the available resource information;

[0007] According to the first location information and the second location information, calculate the distance score of each vehicle node to each fog node, and calculate the data volume according to the data information;

[0008] Based on the processing score, the distance score, the processing density, and the data volume, assign at least one vehicle node to each fog node to receive data information from the assigned vehicle node;

[0009] Each of the fog nodes trains the intermediate model according to the received data information to obtain a corresponding training gradient, and uploads the training gradient to the cloud;

[0010] The cloud trains a global model based on each of the fog nodes, and sends the global model to each of the vehicle nodes;

[0011] Each of the vehicle nodes obtains real-time data, and inputs the real-time data into the global model to obtain a target decision.

[0012] Further, the step of allocating at least one vehicle node to each fog node to receive data information of the allocated vehicle node based on the processing score, the distance score, the processing density, and the data volume comprises:

[0013] According to the distance score, a maximum end-to-end delay is calculated;

[0014] Each of the vehicle nodes is pre-allocated to each of the fog nodes according to a preset scheme;

[0015] According to the formula The resource requirement of each vehicle node is calculated; wherein, , , represents the resource requirement of the i-th vehicle node, represents the data volume of the i-th vehicle node, represents the processing density of the j-th fog node, represents the preset maximum allowable delay, represents the maximum end-to-end delay of the i-th vehicle node to the j-th fog node, represents a preset error parameter, represents the processing score of the j-th fog node, represents the total number of vehicle nodes;

[0016] According to the formula The preset scheme is optimized to obtain an optimized allocation scheme; wherein, , represents the benefit of the i-th vehicle node, represents maximizing the total benefit, represents the number of fog nodes;

[0017] Each fog node is allocated a vehicle node based on the optimized allocation scheme, so that the fog node receives data information of the allocated vehicle node.

[0018] Further, after the step of assigning each fog node with a vehicle node based on the optimized distribution scheme, so that the fog node receives data information of the assigned vehicle node, the method further comprises:

[0019] updating second location information and data volume of each vehicle node, and recalculating maximum end-to-end delay of each vehicle node and each fog node;

[0020] re-optimizing the distribution scheme according to the recalculated maximum end-to-end delay of each vehicle node and each fog node, and the data volume of each vehicle node.

[0021] Further, the step of optimizing the preset scheme to obtain the optimized distribution scheme further comprises:

[0022] judging whether the maximized total revenue is less than a preset revenue threshold;

[0023] if the maximized total revenue is less than the preset revenue threshold, deleting a plurality of vehicle nodes in a preset number one by one until the maximized total revenue is greater than or equal to the preset revenue threshold.

[0024] Further, after the step of the cloud training a global model based on each fog node and sending the global model to each vehicle node, the method further comprises:

[0025] the vehicle node extracts global model parameters in the global model;

[0026] inputting the global model parameters into a local initial model to obtain a local intermediate model;

[0027] training the local intermediate model based on local data stored by the vehicle node to obtain an updated global model.

[0028] The application also provides a vehicle dynamic cooperative control system based on opportunity federated learning, which is realized by an opportunity federated learning architecture, the opportunity federated learning architecture comprising a cloud, a plurality of fog nodes and a plurality of vehicle nodes, one fog node being connected with at least one vehicle node, each fog node being connected with the cloud, the fog node being used for receiving data information collected by each vehicle node and training an intermediate model to obtain a training gradient and uploading the training gradient to the cloud, the cloud training a global model based on the training gradient uploaded by each fog node and sending the global model to each vehicle node, the system comprising:

[0029] ​The acquisition module is configured to acquire available resource information, processing density of each fog node, and first position information of each fog node in real time, and data information collected by each vehicle node and second position information of each vehicle node;

[0030] The determination module is configured to determine a processing score of the fog node according to the available resource information;

[0031] The calculation module is configured to calculate a distance score of each vehicle node to each fog node according to the first position information and the second position information, and calculate a data amount according to the data information;

[0032] The allocation module is configured to allocate at least one vehicle node to each fog node to receive data information of the allocated vehicle node based on the processing score, the distance score, the processing density, and the data amount.

[0033] The training module is configured to train an intermediate model according to the received data information by each fog node to obtain a corresponding training gradient, and upload the training gradient to a cloud side.

[0034] The generation module is configured to generate a global model based on training of each fog node by the cloud side, and send the global model to each vehicle node.

[0035] The input module is configured to acquire real-time data by each vehicle node, and input the real-time data into the global model to obtain a target decision.

[0036] Further, the allocation module comprises:

[0037] The delay calculation sub-module is configured to calculate a maximum end-to-end delay according to the distance score.

[0038] The pre-allocation sub-module is configured to pre-allocate each vehicle node to each fog node according to a preset scheme.

[0039] The resource demand calculation sub-module is configured to calculate resource demand of each vehicle node according to a formula ; wherein, , , represents resource demand of the i-th vehicle node, represents a data amount of the i-th vehicle node, represents processing density of the j-th fog node, represents a preset maximum allowable delay, represents a maximum end-to-end delay of the i-th vehicle node to the j-th fog node, represents a preset error parameter, a processing score of the jth fog node, a total number of vehicle nodes;

[0040] an optimization sub-module configured to optimize the preset scheme according to a formula obtain an optimized allocation scheme; wherein, , a profit of the ith vehicle node, maximize the total profit, a number of fog nodes;

[0041] an allocation sub-module configured to allocate a vehicle node to each fog node based on the optimized allocation scheme, so that the fog node receives data information of the allocated vehicle node.

[0042] Further, the allocation module further comprises:

[0043] a position updating sub-module configured to update second position information and data volume of each vehicle node, and recalculate maximum end-to-end delay of each vehicle node and each fog node;

[0044] a re-optimization sub-module configured to re-optimize the allocation scheme according to the recalculated maximum end-to-end delay of each vehicle node and each fog node, and the data volume of each vehicle node.

[0045] Further, the allocation module further comprises:

[0046] a maximum total profit judgment sub-module configured to judge whether the maximum total profit is less than a preset profit threshold;

[0047] a vehicle node deletion sub-module configured to, if the maximum total profit is less than the preset profit threshold, sequentially delete a plurality of vehicle nodes according to a preset number, until the maximum total profit is greater than or equal to the preset profit threshold.

[0048] Further, the vehicle dynamic cooperative control system based on the federated learning by opportunity further comprises:

[0049] a global model parameter extraction module configured to extract global model parameters in the global model by the vehicle node;

[0050] a global model parameter input module configured to input the global model parameters into a local initial model to obtain a local intermediate model;

[0051] a local intermediate model training module configured to train the local intermediate model based on local data stored by the vehicle node to obtain an updated global model.

[0052] The beneficial effects of the present application: using multiple fog nodes for intermediate model training enables the system to effectively integrate information and data from different vehicle nodes, improves the training speed of the model, enhances the adaptability of the model to diversified driving environments, overcomes the limitations of insufficient generalization ability and poor training effect of traditional training methods, and enables the vehicle to quickly adjust the strategy according to the current traffic conditions, thereby optimizing the overall traffic flow and safety of the intelligent traffic system. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flow diagram of a vehicle dynamic cooperative control method based on opportunity federated learning according to an embodiment of the present application;

[0054] Figure 2 is a structural schematic block diagram of a vehicle dynamic cooperative control system based on opportunity federated learning according to an embodiment of the present application;

[0055] Figure 3 is a structural schematic block diagram of a computer device according to an embodiment of the present application.

[0056] The implementation, functional features and advantages of the present application will be further described with reference to the accompanying drawings. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0058] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly. The connection can be direct connection or indirect connection.

[0059] The term "and / or" in this paper is only a description of the association relationship between the associated objects, which means that there can be three kinds of relationships, for example, A and B can represent: A exists alone, A and B exist together, and B exists alone.

[0060] In addition, the description such as "first", "second" and the like in the present application is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implying the number of the indicated technical features. Therefore, the features defined as "first", "second" can be explicitly or implicitly included at least one of the features. In addition, the technical solutions of various embodiments can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, also not within the protection scope required by the present application.

[0061] Reference Figure 1 The present application proposes a vehicle dynamic cooperative control method based on opportunity federal learning, which is realized through an opportunity federal learning architecture. The architecture includes a cloud, a plurality of fog nodes and vehicle nodes. One fog node is connected to at least one vehicle node, and each fog node is connected to the cloud. The fog node is used to receive data information collected by each vehicle node and train an intermediate model to obtain a training gradient, and upload the training gradient to the cloud. The cloud trains a global model based on the training gradient uploaded by each fog node, and sends the global model to each vehicle node. The method comprises:

[0062] S1: Real-time acquisition of available resource information, processing density and first position information of each fog node, and data information collected by each vehicle node and second position information of each vehicle node;

[0063] S2: Determine the processing score of the corresponding fog node according to the available resource information;

[0064] S3: Calculate the distance score of each vehicle node to each fog node according to the first position information and the second position information, and calculate the data volume according to the data information;

[0065] S4: Based on the processing score, the distance score, the processing density and the data volume, at least one vehicle node is allocated to each fog node to receive the data information of the allocated vehicle node;

[0066] S5: Each fog node trains an intermediate model according to the received data information to obtain a corresponding training gradient, and uploads the training gradient to the cloud;

[0067] S6: The cloud trains a global model based on each fog node, and sends the global model to each vehicle node;

[0068] S7: Each of the vehicle nodes acquires real-time data and inputs the real-time data into the global model to obtain a target decision.

[0069] The opportunity federated learning architecture includes a cloud, a plurality of fog nodes, and a plurality of vehicle nodes, one fog node is connected with at least one vehicle node, each fog node is connected with the cloud respectively, the fog node is used for receiving data information collected by each vehicle node and training an intermediate model to obtain a training gradient, and uploading the training gradient to the cloud, the cloud trains a global model based on the training gradient uploaded by each fog node, and sends the global model to each vehicle node. Wherein, the vehicle node only needs to send data to the fog node and will not send to the cloud, the fog node is the intermediate layer of the architecture, used for storage and calculation, and undertakes the preliminary information aggregation and model training task. The fog node reduces the burden of large-scale data processing directly in the cloud by integrating data and computing resources from multiple participants, and improves the data processing efficiency. The cloud is responsible for collecting the training gradient or model parameter uploaded by the fog node, training and updating the global model. The cloud will distribute the trained global model back to each participant, in some embodiments, it can also be sent only to the vehicle node, but in order for the model stored by the fog node to be updated, it can also be sent to the fog node.

[0070] As described in step S1 above, real-time acquisition of available resource information of each fog node, processing density and first position information of each fog node, and data information collected by each vehicle node and second position information of each vehicle node. Real-time acquisition of available resource information, processing density, first position information of each fog node, and data information collected by each vehicle node and second position information. The available resource information includes the current computing power, storage space and network bandwidth of the fog node, which affects the number and speed of tasks that can be processed by the fog node. The processing density reflects the amount of data that can be processed by the fog node per unit time, which determines its processing efficiency and ability. The first position information refers to the physical position of the fog node, while the second position information of the vehicle node also reflects the position of the vehicle. Through these position data, the system can evaluate the distance between the vehicle and the fog node, and prepare for subsequent data allocation. In addition, the data information collected by each vehicle node in this step includes its own perception data, such as speed, position, environmental conditions, etc., for model training.

[0071] As described in step S2, a processing score corresponding to each fog node is determined according to the available resource information. The calculation of the processing score is to evaluate the processing capacity of each fog node, to ensure that the data can be effectively utilized when allocated. This usually involves the fusion and quantification of multiple indicators such as the computing power, storage capacity, network bandwidth of the fog node. Specifically, the relative importance of certain resources can be emphasized by setting weights, for example, it may be considered that the computing power is more important than the storage capacity, and thus a comprehensive processing score is obtained.

[0072] As described in step S3, according to the first position information and the second position information, the distance score of each vehicle node to each fog node is calculated, and the data volume is calculated according to the data information. In order to evaluate the physical distance between the vehicle node and the fog node and the data volume sent to the fog node. The calculation of the distance score can use various methods, such as direct Euclidean distance calculation, the delay of the vehicle transmitting data to the fog node with short distance is usually low, and the analysis of the distance score and the data volume can provide effective basis for subsequent data allocation, so that the data transmission between nodes is more efficient, and the efficiency of model training is significantly improved.

[0073] As described in step S4, based on the processing score, the distance score, the processing density and the data volume, at least one vehicle node is allocated to each fog node to receive the data information of the allocated vehicle node. Based on the processing score, the distance score, the processing density and the data volume obtained in the previous steps, at least one vehicle node is allocated to each fog node to receive the data of the vehicle node. Ensure that the data of the vehicle node can be processed by the most suitable fog node. By considering various factors, the system will preferentially select those fog nodes with high processing score, relatively short distance and high processing density. This allocation strategy not only can maximize the efficiency of data processing, but also can reduce the data transmission delay from the vehicle node. In addition, the system can dynamically adjust the allocation strategy to adapt to the changing network conditions and real-time data volume. For example, when the load of a certain fog node is too large or its processing capacity decreases, the vehicle node can be reallocated through real-time monitoring mechanism, which can ensure the smoothness and reliability of the data processing process.

[0074] As described in step S5 above, each fog node trains an intermediate model based on the received data information to obtain corresponding training gradients, and then uploads these training gradients to the cloud. Each fog node trains an intermediate model based on the received vehicle node data information, thereby obtaining corresponding training gradients, and then uploads these training gradients to the cloud. As a data processing intermediary, the fog node is not only responsible for receiving information uploaded by vehicle nodes, but also undertakes the task of initial model training. Using the received data information, the fog node can perform machine learning training according to a preset model, such as feature extraction and model optimization using deep learning algorithms. After completing the intermediate model training, the fog node generates corresponding training gradients, which are used to update the global model. The training of the intermediate model reduces direct dependence on the central server, lowers the burden on central computing, and accelerates the optimization process of model parameters. Throughout the entire process, the role of the fog node makes data processing more intelligent and efficient, laying a solid foundation for the subsequent training of the global model.

[0075] As described in step S6 above, the cloud trains a global model based on each of the fog nodes and sends the global model to each of the vehicle nodes. The cloud trains the global model based on the training gradients uploaded by each fog node. The function of the cloud is to integrate the gradient information provided by all fog nodes to update and optimize the global model. This process typically uses aggregation methods, such as weighted averaging or the FedAvg algorithm, to fuse gradients from different fog nodes. In this way, the cloud can obtain a more comprehensive and accurate overall performance of the global model. The quality of the global model training directly affects the decision-making ability of the vehicle nodes; therefore, clarifying the contribution of each fog node and evaluating the training quality is crucial. During this process, the cloud can also continuously monitor the training effect of the global model and adjust strategies and algorithms based on the model performance feedback to ensure the effectiveness and universality of the final model in various operating environments. After training is completed, the cloud sends the updated global model back to each vehicle node, providing them with a unified knowledge base to help each vehicle make decisions in real time.

[0076] As described in step S7 above, each vehicle node acquires real-time data and inputs it into the global model to obtain the target decision. After the model is updated, the vehicle nodes will use the global model sent from the cloud to make real-time decisions. For example, the vehicle can use the global model to calculate the optimal driving strategy, including decisions on driving speed and steering angle, based on information such as environmental changes and the dynamics of other traffic participants. The acquisition and processing of real-time data enables vehicles to maintain flexible response capabilities in complex and ever-changing traffic environments. In addition, vehicle nodes can continuously collect and feed back new data during operation, forming a closed-loop learning process that promotes continuous optimization and improvement of the global model. In this way, each vehicle can not only obtain efficient decision support but also provide data support for the learning of the entire system, promoting the further development of vehicle dynamic cooperative control and thus achieving the goals of intelligent transportation systems.

[0077] In one embodiment, step S4, which allocates at least one vehicle node to each fog node based on the processing score, the distance score, the processing density, and the data volume, and receives data information on the allocation of vehicle nodes, includes:

[0078] S401: Calculate the maximum end-to-end delay based on the distance score;

[0079] S402: Pre-assign each of the vehicle nodes to each of the fog nodes according to a preset scheme;

[0080] S403: According to the formula Calculate the resource requirements for each vehicle node; among which, , , This represents the resource requirement of the i-th vehicle node. This represents the amount of data for the i-th vehicle node. This represents the processing density of the j-th fog node. This indicates the preset maximum allowable delay. This represents the maximum end-to-end delay of the transmission value from the i-th vehicle node to the j-th fog node. This indicates the preset error parameter. This represents the processing score of the j-th fog node. This represents the total number of vehicle nodes;

[0081] S404: According to the formula The preset scheme is optimized to obtain an optimized allocation scheme; wherein, , Let represent the revenue of the i-th vehicle node. This means maximizing total revenue. Indicates the number of fog nodes;

[0082] S405: Based on the optimized allocation scheme, allocate vehicle nodes to each fog node so that the fog node receives the data information of the allocated vehicle nodes.

[0083] As described in steps S401-S405 above, the scheme for allocating vehicle nodes to each fog node is optimized. Using the previously calculated distance scores, the maximum end-to-end latency for data transmission between vehicle nodes and fog nodes is estimated. End-to-end latency refers to the total time required from when a vehicle node sends data to when a fog node receives it. This includes network transmission latency, data processing latency, etc. Minimizing end-to-end latency helps improve real-time performance and response speed, thereby optimizing the effectiveness of control decisions. Vehicle nodes are initially allocated to each fog node according to a preset allocation scheme. This preset scheme can be based on rules of thumb or historical data analysis; for example, vehicle nodes that are close to each other are preferentially allocated to fog nodes that are also close. This stage mainly aims to build a preliminary allocation architecture to ensure the system has a reasonable starting point for subsequent optimization, calculating the resource requirements for each vehicle node to process data at the fog node. The benefits of all vehicle nodes are integrated, and the total benefit is calculated. Theoretically, the higher the total benefit, the more optimized the system's resource allocation. Then, the preset scheme is optimized based on the principle of maximizing total benefit, using algorithms such as genetic algorithms, linear programming, or other optimization techniques to adjust the allocation between each vehicle node and fog node. Through this iterative optimization process, the system can obtain a more efficient allocation scheme, thereby improving overall performance. Finally, based on the optimized allocation scheme, the final vehicle nodes are assigned to each fog node. This allows the fog nodes to receive data transmitted from the vehicle nodes assigned to them. This ensures that fog nodes have access to the latest and most relevant data for intermediate model training, improving the efficiency and response speed of the entire system. Ultimately, this process provides strong support for methods of dynamic cooperative vehicle control, facilitating real-time decision-making and optimizing traffic flow.

[0084] In one embodiment, after step S405, which assigns a vehicle node to each fog node based on the optimized allocation scheme so that the fog node receives the data information of the allocated vehicle node, the method further includes:

[0085] S4061: Update the second location information and data volume of each vehicle node, and recalculate the maximum end-to-end delay between each vehicle node and each fog node.

[0086] S4062: Based on the recalculated maximum end-to-end delay between each vehicle node and each fog node, and the data volume of each vehicle node, the allocation scheme is re-optimized.

[0087] As described in steps S4061-S4062 above, the second location information and data volume of each vehicle node are updated. This is because the location and collected data of vehicles may change over time during operation, affecting the real-time performance and accuracy of the system. Therefore, this information is updated in real time based on the dynamic traffic environment. Furthermore, the updated vehicle node locations and data volumes are used to recalculate the maximum end-to-end delay between the vehicle nodes and each fog node. This recalculation of delay provides the system with an assessment of the current transmission timeliness, providing the most accurate basic data for subsequent optimized allocation schemes. The previous allocation scheme is then re-optimized using the recalculated maximum end-to-end delay and the updated data volume of each vehicle node. The calculation result of the maximum end-to-end delay affects the efficiency and real-time performance of data transmission; shorter delays generally mean faster data processing capabilities. Therefore, the allocation between nodes can prioritize combinations with lower delays. The current data volume of the vehicle nodes is also an important optimization indicator, as a larger data volume means more information value, which may require stronger processing resources. Optimization schemes based on this new data will utilize algorithms such as linear programming, genetic algorithms, or other optimization techniques to reconfigure the connections between vehicle nodes and fog nodes, thereby achieving higher overall benefits and better system performance. This can enhance decision support for dynamic cooperative vehicle control and improve the operational efficiency and responsiveness of the entire intelligent transportation system.

[0088] In one embodiment, the formula After step S404, which optimizes the preset scheme to obtain the optimized allocation scheme, the method further includes:

[0089] S4051: Determine whether the maximized total revenue is less than a preset revenue threshold;

[0090] S4052: If the total revenue is less than the preset revenue threshold, then delete multiple vehicle nodes sequentially according to a preset number until the maximum total revenue is greater than or equal to the preset revenue threshold.

[0091] As described in steps S4051-S4052 above, the maximized total revenue calculated using the optimization algorithm will be evaluated and compared with a pre-set revenue threshold. The revenue threshold is a benchmark value set based on system requirements and the actual application scenario, representing the minimum revenue level the system expects to achieve. If the maximized total revenue is less than the preset revenue threshold, it indicates that the current resource allocation scheme has failed to achieve the expected performance goals, meaning that the data allocation efficiency and resource utilization of the current vehicle nodes and fog nodes are insufficient, potentially leading to slower system response and insufficient real-time performance. A resource reduction strategy needs to be initiated, and the system will sequentially delete multiple vehicle nodes. These nodes can be based on priority or low data contribution, or they can be randomly determined nodes, thereby reducing system complexity and resource consumption, potentially improving the collaboration efficiency between the remaining vehicle nodes and fog nodes. By gradually deleting nodes, the system can test the impact of each adjustment on the overall revenue and update the revenue calculation in real time until the maximized total revenue meets or exceeds the preset revenue threshold. This strategy ensures efficiency and performance while reducing system operating costs. Ultimately, if the revenue reaches the preset target, it indicates that the resource allocation is more reasonable and provides a basis for subsequent dynamic adjustments.

[0092] In one embodiment, after step S6, where the cloud trains a global model based on each of the fog nodes and sends the global model to each of the vehicle nodes, the method further includes:

[0093] S701: The vehicle node extracts the global model parameters from the global model;

[0094] S702: Input the global model parameters into the local initial model to obtain the local intermediate model;

[0095] S703: Train the local intermediate model based on the local data stored in the vehicle node to obtain the updated global model.

[0096] As described in steps S701-S703 above, the local model is updated. Each vehicle node receives the updated global model from the cloud and extracts its global model parameters, typically including model weights and biases. This provides the basic knowledge input for the new model. The vehicle node then inputs the extracted global model parameters into its local initial model to generate a local intermediate model. This maps the global model parameters to the corresponding structure of the local model, enabling the local model to learn knowledge from the global model. By applying the global model parameters to the local model, the vehicle node exhibits higher adaptability and performance in new environments and tasks, demonstrating superior performance when processing real-time data. The deep-learned features of the local intermediate model greatly improve its ability to understand and analyze local data during subsequent training, providing more accurate and reliable information for vehicle node decision-making in dynamic traffic scenarios. The vehicle node will use the stored local data to further train the local intermediate model to generate an updated global model. This deep optimization of the local intermediate model aims to make it more effectively adaptable to specific environments and scenarios. By training on local data, vehicle nodes can further adjust model parameters based on actual traffic conditions, driving habits, and other key factors to improve model accuracy and response speed. During training, vehicle nodes not only fully utilize the knowledge provided by the global model but also combine it with specific local vehicle data, accelerating model learning. The updated model will provide better decision support, helping vehicles react quickly in complex and dynamically changing road environments. This step ensures the effective implementation of the global model in local environments, improving the overall performance of the intelligent transportation system and enabling more personalized dynamic control based on the specific needs of each vehicle node. Through this training and updating process, the system better demonstrates the advantages of distributed learning, ultimately achieving high efficiency and safety in cooperative driving.

[0097] The beneficial effects of this invention are as follows: By using multiple fog nodes for intermediate model training, the system can effectively integrate information and data from different vehicle nodes, improve the training speed of the model, enhance the model's adaptability to diverse driving environments, overcome the limitations of insufficient generalization ability and poor training effect of traditional training methods, and enable vehicles to quickly adjust their strategies according to the current traffic conditions, thereby optimizing the overall flow and safety of the intelligent transportation system.

[0098] Reference Figure 2The present invention also provides a vehicle dynamic cooperative control system based on opportunistic federated learning, implemented through an opportunistic federated learning architecture. This architecture includes a cloud, multiple fog nodes, and individual vehicle nodes. Each fog node is connected to at least one vehicle node, and each fog node is connected to the cloud. The fog nodes receive data collected by each vehicle node, train intermediate models to obtain training gradients, and upload these gradients to the cloud. The cloud trains a global model based on the training gradients uploaded by each fog node and sends the global model to each vehicle node. The system includes:

[0099] The acquisition module 10 is used to acquire in real time the available resource information, processing density, and first location information of each fog node, as well as the data information collected by each vehicle node and the second location information of each vehicle node.

[0100] The determination module 20 is used to determine the processing score corresponding to the fog node based on the available resource information;

[0101] The calculation module 30 is used to calculate the distance score from each vehicle node to each fog node based on the first location information and the second location information, and to calculate the data volume based on the data information;

[0102] The allocation module 40 is used to allocate at least one vehicle node to each fog node based on the processing score, the distance score, the processing density, and the data volume, and to receive data information on the allocation of vehicle nodes.

[0103] The training module 50 is used for each fog node to train an intermediate model based on the received data information to obtain the corresponding training gradient, and to upload the training gradient to the cloud.

[0104] Generation module 60 is used for the cloud to train a global model based on each of the fog nodes and send the global model to each of the vehicle nodes;

[0105] The input module 70 is used for each of the vehicle nodes to acquire real-time data and input the real-time data into the global model to obtain the target decision.

[0106] In one embodiment, the allocation module 40 includes:

[0107] The delay calculation submodule is used to calculate the maximum end-to-end delay based on the distance score;

[0108] The pre-allocation submodule is used to pre-allocate each of the vehicle nodes to each of the fog nodes according to a preset scheme;

[0109] The resource requirement calculation submodule is used to calculate the requirements based on the formula. Calculate the resource requirements for each vehicle node; among which, , , This represents the resource requirement of the i-th vehicle node. This represents the amount of data for the i-th vehicle node. This represents the processing density of the j-th fog node. This indicates the preset maximum allowable delay. This represents the maximum end-to-end delay of the transmission value from the i-th vehicle node to the j-th fog node. This indicates the preset error parameter. This represents the processing score of the j-th fog node. This represents the total number of vehicle nodes;

[0110] The optimization submodule is used to optimize the formula. The preset scheme is optimized to obtain an optimized allocation scheme; wherein, , Let represent the revenue of the i-th vehicle node. This means maximizing total revenue. Indicates the number of fog nodes;

[0111] The allocation submodule is used to allocate vehicle nodes to each fog node based on the optimized allocation scheme, so that the fog node receives the data information of the allocated vehicle nodes.

[0112] In one embodiment, the allocation module 40 further includes:

[0113] The location update submodule is used to update the second location information and data volume of each vehicle node, and recalculate the maximum end-to-end delay between each vehicle node and each fog node.

[0114] The re-optimization submodule is used to re-optimize the allocation scheme based on the recalculated maximum end-to-end latency between each vehicle node and each fog node, as well as the data volume of each vehicle node.

[0115] In one embodiment, the allocation module 40 further includes:

[0116] The maximum total revenue judgment submodule is used to determine whether the maximum total revenue is less than a preset revenue threshold;

[0117] The vehicle node deletion submodule is used to delete multiple vehicle nodes sequentially according to a preset number if the total revenue is less than the preset revenue threshold, until the maximum total revenue is greater than or equal to the preset revenue threshold.

[0118] In one embodiment, the vehicle dynamic cooperative control system based on opportunistic federated learning further includes:

[0119] A global model parameter extraction module is used to extract global model parameters from the global model of the vehicle node;

[0120] A global model parameter input module is used to input the global model parameters into the local initial model to obtain a local intermediate model;

[0121] The local intermediate model training module is used to train the local intermediate model based on the local data stored in the vehicle node to obtain the updated global model.

[0122] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores various data and information. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it can implement the vehicle dynamic cooperative control method based on opportunistic federated learning described in any of the above embodiments.

[0123] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0124] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, can implement the vehicle dynamic cooperative control method based on opportunistic federated learning described in any of the above embodiments.

[0125] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0126] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0127] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A vehicle dynamic cooperative control method based on opportunistic federated learning, characterized in that, This is implemented using an opportunistic federated learning architecture, which includes a cloud, multiple fog nodes, and individual vehicle nodes. Each fog node is connected to at least one vehicle node, and each fog node is connected to the cloud. The fog nodes receive data collected by the vehicle nodes, train intermediate models to obtain training gradients, and upload these gradients to the cloud. The cloud trains a global model based on the training gradients uploaded by each fog node and sends the global model to each vehicle node. The method includes: The system acquires in real time the available resource information, processing density, and first location information of each fog node, as well as the data information collected by each vehicle node and the second location information of each vehicle node. The processing score for the corresponding fog node is determined based on the available resource information; Based on the first location information and the second location information, calculate the distance score from each vehicle node to each fog node, and calculate the data volume based on the data information; Based on the processing score, the distance score, the processing density, and the data volume, at least one vehicle node is assigned to each fog node to receive the data information of the assigned vehicle node; Each fog node trains an intermediate model based on the received data to obtain the corresponding training gradient, and then uploads the training gradient to the cloud. The cloud platform trains a global model based on each of the fog nodes and sends the global model to each of the vehicle nodes; Each of the vehicle nodes acquires real-time data and inputs the real-time data into the global model to obtain the target decision; The step of allocating at least one vehicle node to each fog node based on the processing score, the distance score, the processing density, and the data volume, and receiving data information on the allocated vehicle nodes, includes: Calculate the maximum end-to-end delay based on the distance score; Each of the vehicle nodes is pre-assigned to each of the fog nodes according to a preset scheme; According to the formula Calculate the resource requirements for each vehicle node; among which, , , This represents the resource requirement of the i-th vehicle node. This represents the amount of data for the i-th vehicle node. This represents the processing density of the j-th fog node. This indicates the preset maximum allowable delay. This represents the maximum end-to-end delay of the transmission value from the i-th vehicle node to the j-th fog node. This indicates the preset error parameter. This represents the processing score of the j-th fog node. This represents the total number of vehicle nodes; According to the formula The preset scheme is optimized to obtain an optimized allocation scheme; wherein, , Let represent the revenue of the i-th vehicle node. This means maximizing total revenue. Indicates the number of fog nodes; Based on the optimized allocation scheme, a vehicle node is allocated to each fog node so that the fog node receives the data information of the allocated vehicle node.

2. The vehicle dynamic cooperative control method based on opportunistic federated learning as described in claim 1, characterized in that, After the step of allocating vehicle nodes to each fog node based on the optimized allocation scheme, so that the fog node receives the data information of the allocated vehicle nodes, the method further includes: Update the second location information and data volume of each vehicle node, and recalculate the maximum end-to-end delay between each vehicle node and each fog node; The allocation scheme is re-optimized based on the recalculated maximum end-to-end latency between each vehicle node and each fog node, and the data volume of each vehicle node.

3. The vehicle dynamic cooperative control method based on opportunistic federated learning as described in claim 1, characterized in that, According to the formula After optimizing the preset scheme to obtain the optimized allocation scheme, the method further includes: Determine whether the maximized total revenue is less than a preset revenue threshold; If the total revenue is less than the preset revenue threshold, then multiple vehicle nodes are deleted sequentially according to a preset number until the maximum total revenue is greater than or equal to the preset revenue threshold.

4. The vehicle dynamic cooperative control method based on opportunistic federated learning as described in claim 1, characterized in that, After the step of training a global model based on each of the fog nodes in the cloud and sending the global model to each of the vehicle nodes, the method further includes: The vehicle node extracts global model parameters from the global model; The global model parameters are input into the local initial model to obtain the local intermediate model; The local intermediate model is trained based on the local data stored in the vehicle node to obtain the updated global model.

5. A vehicle dynamic cooperative control system based on opportunistic federated learning, characterized in that, This is implemented using an opportunistic federated learning architecture, which includes a cloud, multiple fog nodes, and individual vehicle nodes. Each fog node is connected to at least one vehicle node, and each fog node is connected to the cloud. The fog nodes receive data collected by the vehicle nodes, train intermediate models to obtain training gradients, and upload these gradients to the cloud. The cloud trains a global model based on the training gradients uploaded by each fog node and sends the global model to each vehicle node. The system includes: The acquisition module is used to acquire in real time the available resource information, processing density, and first location information of each fog node, as well as the data information collected by each vehicle node and the second location information of each vehicle node; The determination module is used to determine the processing score corresponding to the fog node based on the available resource information; The calculation module is used to calculate the distance score from each vehicle node to each fog node based on the first location information and the second location information, and to calculate the data volume based on the data information; The allocation module is used to allocate at least one vehicle node to each fog node based on the processing score, the distance score, the processing density, and the data volume, and to receive data information on the allocation of vehicle nodes. The training module is used by each fog node to train the intermediate model based on the received data information to obtain the corresponding training gradient, and then upload the training gradient to the cloud. The generation module is used to train a global model based on each of the fog nodes in the cloud and send the global model to each of the vehicle nodes; The input module is used for each of the vehicle nodes to acquire real-time data and input the real-time data into the global model to obtain the target decision; The allocation module includes: The delay calculation submodule is used to calculate the maximum end-to-end delay based on the distance score; The pre-allocation submodule is used to pre-allocate each of the vehicle nodes to each of the fog nodes according to a preset scheme; The resource requirement calculation submodule is used to calculate the requirements based on the formula. Calculate the resource requirements for each vehicle node; among which, , , This represents the resource requirement of the i-th vehicle node. This represents the amount of data for the i-th vehicle node. This represents the processing density of the j-th fog node. This indicates the preset maximum allowable delay. This represents the maximum end-to-end delay of the transmission value from the i-th vehicle node to the j-th fog node. This indicates the preset error parameter. This represents the processing score of the j-th fog node. This represents the total number of vehicle nodes; The optimization submodule is used to optimize the formula. The preset scheme is optimized to obtain an optimized allocation scheme; wherein, , Let represent the revenue of the i-th vehicle node. This means maximizing total revenue. Indicates the number of fog nodes; The allocation submodule is used to allocate vehicle nodes to each fog node based on the optimized allocation scheme, so that the fog node receives the data information of the allocated vehicle nodes.

6. The vehicle dynamic cooperative control system based on opportunistic federated learning as described in claim 5, characterized in that, The allocation module further includes: The location update submodule is used to update the second location information and data volume of each vehicle node, and recalculate the maximum end-to-end delay between each vehicle node and each fog node. The re-optimization submodule is used to re-optimize the allocation scheme based on the recalculated maximum end-to-end latency between each vehicle node and each fog node, as well as the data volume of each vehicle node.

7. The vehicle dynamic cooperative control system based on opportunistic federated learning as described in claim 5, characterized in that, The allocation module further includes: The maximum total revenue judgment submodule is used to determine whether the maximum total revenue is less than a preset revenue threshold; The vehicle node deletion submodule is used to delete multiple vehicle nodes sequentially according to a preset number if the total revenue is less than the preset revenue threshold, until the maximum total revenue is greater than or equal to the preset revenue threshold.

8. The vehicle dynamic cooperative control system based on opportunistic federated learning as described in claim 5, characterized in that, The vehicle dynamic cooperative control system based on opportunistic federated learning also includes: A global model parameter extraction module is used to extract global model parameters from the global model of the vehicle node; A global model parameter input module is used to input the global model parameters into the local initial model to obtain a local intermediate model; The local intermediate model training module is used to train the local intermediate model based on the local data stored in the vehicle node to obtain the updated global model.

Images