A vehicle networking model training method and a readable storage medium

Abstract

The application provides a vehicle networking model training method and readable storage medium for realizing privacy protection and resource optimization, and the vehicle networking model training method is used for controlling a vehicle networking model training system; the vehicle networking model training system comprises an intelligent traffic system, at least one intelligent vehicle, a privacy protection module and a resource optimization module; the at least one intelligent vehicle is internally provided with a local computing and environment module and a sensing module, and is connected with the intelligent traffic system; the privacy protection module is composed of a plurality of roadside unit sets; the resource optimization module is provided with a global model aggregation module, and the global model aggregation module is connected with the privacy protection module. The data privacy of the intelligent vehicle is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle technology, and more specifically, to a method for training a vehicle networking model that achieves privacy protection and resource optimization, and a readable storage medium. Background Technology

[0002] The rapid development of the Internet of Things (IoT) and wireless communication technologies is continuously driving the transformation from traditional vehicle-mounted ad hoc networks to vehicle-to-everything (V2X) networks. In V2X, intelligent vehicles use various sensors, cameras, and navigation systems to collect information from their surroundings and share it with servers or nearby vehicles, improving traffic efficiency and reducing accident rates. Simultaneously, the rapid development of V2X is accelerating the construction and practical application of intelligent transportation systems (ITS). ITS collects various sensory information from intelligent vehicles through roadside units, uploads it to servers using wireless communication technology, and analyzes the data to make decisions beneficial to intelligent vehicles, thereby reducing traffic congestion, lowering fuel consumption and accident rates, and improving the energy utilization rate of electric vehicles and the effective use of grid energy. While ITS utilizes shared data to train global models, it achieves more comprehensive, broader, and intelligent real-time and accurate management of V2X.

[0003] However, the information interaction between intelligent vehicles and intelligent transportation systems suffers from serious communication overhead and data privacy issues, severely hindering the future development of intelligent transportation systems. Traditional intelligent transportation systems rely on raw data collected by onboard devices for interaction. In wireless communication modes, malicious users may intercept wireless communication signals to eavesdrop, delete, edit, and replay messages. Therefore, the traditional interaction method's unencrypted transmission of raw data poses serious risks to data security and privacy. With the deepening integration of federated learning and vehicle-to-everything (V2X) communication, many methods have been implemented to protect customer data privacy to some extent. However, these methods only consider the data heterogeneity of intelligent vehicles and focus on addressing model differences caused by data heterogeneity to improve the accuracy of the global model. However, the wireless communication bandwidth, available CPU and GPU computing resources, and power or fuel levels of intelligent vehicles in V2X are limited, significantly restricting the practical deployment of V2X. To address the practical deployment issues of federated learning in V2X, some methods have emerged and been implemented, but these methods primarily consider reducing the number of communication rounds, which does not fundamentally solve the customer's resource constraints. Summary of the Invention

[0004] Therefore, embodiments of the present invention provide a method for training vehicle network models that achieves privacy protection and resource optimization, effectively improving the local data privacy of intelligent vehicles.

[0005] To address the aforementioned problems, this invention provides a vehicle-to-everything (V2X) model training method that achieves privacy protection and resource optimization. The V2X model training method is used to control a V2X model training system. The V2X model training system includes: an intelligent transportation system; at least one intelligent vehicle, which is equipped with a local computing and environment module and a sensing module, and is connected to the intelligent transportation system; a privacy protection module, which is composed of multiple roadside units; and a resource optimization module, which includes a global model aggregation module connected to the privacy protection module. The V2X model training method specifically includes: Step S100: The intelligent transportation system sends a global model to the multiple roadside units. Intelligent vehicle subset and global task T i The multiple roadside units will define the global model. The information is sequentially sent to the corresponding intelligent vehicles within the subset of intelligent vehicles; Step S200: Road information and video information are collected through the sensing module, and the information is processed according to the global task T. i Step S300: The road information, video information, and intelligent vehicle data applicable to the global task are integrated into a local model and sent to the roadside unit; Step S400: Each roadside unit performs intermediate aggregation on the local model sent by the intelligent vehicle to form an intermediate aggregated model; Step S400: Multiple roadside units send the intermediate aggregated model to the global model aggregation module to obtain the next global model. And the next global model As the next global task T i+1 The initial model; where i is the current communication round number.

[0006] Compared with existing technologies, the technical effects achieved by this solution are as follows: On the one hand, by setting up a privacy protection module, which consists of multiple roadside unit sets, when intelligent vehicles send raw data, the data is transformed into an intermediate aggregation model through the privacy protection module, and then the global model aggregation module aggregates the multiple intermediate aggregation models into the next global model. Through the next global model Instead of transmitting raw data, it effectively protects the local data privacy of intelligent vehicles; on the other hand, by first converting the raw data into an intermediate aggregation model and then into a global model, the models transmitted on the nodes can be different, which strengthens the ability to resist malicious attackers from capturing and cracking the raw model.

[0007] In one embodiment of the present invention, step S200 specifically includes: step S210: the intelligent vehicle activates wireless communication to receive the global task T from the intelligent transportation system.i Step S220: According to the global task T i Select the corresponding road information, video information, and intelligent vehicle data set to form the local resource status. Step S230: Based on the local resource status With the global task T i The minimum loss minf(w) is set as the training objective to obtain the local model. Among them, minimum loss , For the first The amount of data in the dataset established by the intelligent vehicles participating in this communication round, x i Local resource status The processed vector form, where the superscript k represents the k-th intelligent vehicle participating in this global model iteration, x i and y i The label corresponds to the data vector, where index i represents the i-th training sample, w represents the weight vector of the local model, and l represents the optimization method; Step S240: Upload the local resource status. and the local model To the roadside unit.

[0008] Compared with existing technologies, the technical effects achieved by this solution are as follows: by setting the minimum loss minf(w) as the training objective, and minimizing the loss... By repeatedly training, the preset minimum loss value is achieved, thereby improving training accuracy.

[0009] In one embodiment of the present invention, step S300 specifically includes: step S310: the roadside unit receives the local resource status. and the local model Step S320: For the received local model... Intermediate aggregation is performed to form an intermediate aggregation model; wherein, the intermediate aggregation model N IV This refers to the total number of intelligent vehicles participating in this model iteration.

[0010] Compared with existing technologies, the technical effects achieved by adopting this technical solution are: [the following effects are achieved: local resource status] and local model Implement intermediate aggregation and intermediate aggregation model in roadside units. By delegating the aggregation task to intermediate aggregation nodes, the models transmitted by the two intermediate nodes are inconsistent, thereby strengthening the ability to resist malicious attackers from capturing and cracking the original model.

[0011] In one embodiment of the present invention, step S400 specifically includes: step S410: the plurality of roadside units aggregate the intermediate model. Send to the global model aggregation module; Step S420: The global model aggregation module sends multiple intermediate aggregated models The next global model is obtained by aggregation. The next global model N RSU This refers to the total number of intelligent vehicles participating in this model iteration; Step S430: The next global model... This will serve as the initial model for the next global task Ti+1.

[0012] Compared with existing technologies, the technical effects achieved by adopting this technical solution are: through the intermediate aggregation model Send to the global model aggregation module, and combine multiple intermediate aggregated models. The next global model is obtained by aggregation. And the next global model The next global model As the initial model for the next global task Ti+1, the detection accuracy of the global model is continuously improved.

[0013] In one embodiment of the present invention, the method further includes the following steps before step S100: step S10: setting the target accuracy; step S20: when the global model... If the target accuracy is not achieved, then repeat steps S100, S200, S300, and S400.

[0014] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: By setting a target accuracy, when the global model... If the target accuracy is not achieved, the test is repeated until the target accuracy is achieved.

[0015] In one embodiment of the present invention, before step S100, the method further includes: the intelligent transportation system sending an initial global task T1 to the resource optimization module, and the resource optimization module establishing an initial global model based on the initial global task T1. .

[0016] Compared with existing technologies, the technical effects achieved by adopting this technical solution are: establishing an initial global model manually. And the initial global model It can establish a target accuracy that fits the initial global task T1, thereby reducing the number of training iterations and achieving the target accuracy faster.

[0017] In one embodiment of the present invention, the resource optimization module further includes: an intelligent vehicle strategy library, which employs deep reinforcement learning as the intelligent vehicle selection algorithm. By collecting the resource status of the intelligent vehicles and calculating rewards, the mapping function is continuously updated, and the most participating subset of intelligent vehicles is obtained from the mapping function. The intelligent vehicles with the longest long-term benefits are selected to participate in the next global task.

[0018] Compared with existing technologies, the technical effects achieved by this solution are as follows: On the one hand, by using data-driven deep reinforcement learning to replace the original static selection algorithms, such as random algorithms and greedy algorithms, the intelligent vehicle selection algorithm of this invention selects intelligent vehicles with longer-term benefits to participate in global model updates through analysis of the environment and intelligent vehicle resources, thereby reducing the uncertainty brought by the original static algorithm, enhancing the robustness of the algorithm of this invention, and ultimately improving the overall resource utilization of the system. On the other hand, deep reinforcement learning establishes an experience replay area through interaction with the environment, and uses the samples in the experience replay area as training samples. In the early stage of training, the required data can be continuously generated through simulation training and updated offline, with low requirements for real-time data.

[0019] In one embodiment of the present invention, the intelligent vehicle policy library adopts DDQN as the main framework, and DNN as the Q network and target Q network.

[0020] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: The intelligent vehicle strategy library uses DDQN as the main framework, and DNN as the Q network and target Q network, which improves the scalability of the intelligent vehicle selection algorithm.

[0021] In one embodiment of the present invention, the intelligent vehicle policy library adopts DDQN as the main framework, with DNN as the Q-network and target Q-network. Specifically, it includes: Step S500: Initialize experience replay, Q-network, and target Q-network with a learning rate of 0.9, a discount factor of 0.5, an initial selection probability ε of 0.1, and a selection probability ε decay coefficient of 0.001; Step S510: If the intelligent vehicle selection algorithm is not in a convergent state, repeat steps S520-S590; Step S520: If the global model... If the accuracy is less than the target or the maximum number of communication rounds, repeat steps S530-S590; Step S530: Randomly select action a with probability ε, and select the action with the maximum reward value a=argmaxQ(s,a;θ) with probability (1-ε), where action a is the subset of intelligent vehicles, probability ε decreases continuously with the increase of communication rounds, and θ is the model weight of the Q network; Step S540: Execute action a, that is, select the corresponding subset of intelligent vehicles to participate in the global model of this communication round. Update; Step S550: After executing action a, obtain the reward value R from environmental feedback, and the resource state s′ for the next communication round, where the reward value is set as R = -α·power consumption (fuel consumption) - β·time delay, s is the resource state for the current communication round, and the quadruple is set.<s,a,R,s′> Store the experience replay, where α and β are coefficients; Step S560: If the experience replay is full, execute steps S570-S590, otherwise skip; Step S570: Randomly select the required number of quadruplets from the experience replay.<s,a,R,s′> The sample is used as the minimum batch processing sample; Step S580: Update the Q network using the extracted samples, the update method is as follows. , where η is a coefficient; This represents the updated Q value after performing action a in state s; This represents the Q-value obtained through the Q-network when action a is performed in state s. This represents the reward obtained by performing action a in state s; This represents the Q value obtained by taking the best action a' under the resource state in the next communication round; a' represents the best action in the next communication round; A represents the set of all actions; Step S590: Periodically update the target Q network by assigning the model parameters of the Q network to the target Q network.

[0022] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: by periodically updating the target Q network, the update method is to assign the model parameters of the Q network to the target Q network, which can cope with the rapid changes in local resources of connected vehicle intelligent vehicles, and can reduce the uncertainty caused by unknown states through neural network mapping, thus having stronger adaptability and reducing the pressure on the backbone network.

[0023] In another aspect, embodiments of the present invention also provide a readable storage medium, the readable storage medium including stored computer-executable instructions, wherein, when the computer-executable instructions are executed by a processor, the device where the storage medium is located controls the execution of the steps of the vehicle network model training method that achieves privacy protection and resource optimization as described in any of the above embodiments.

[0024] The readable storage medium in this embodiment includes stored computer-executable instructions, and the computer-executable instructions are processed by a processor as in any embodiment of the present invention to implement a vehicle network model training method that achieves privacy protection and resource optimization. Therefore, it has all the beneficial effects of the vehicle network model training method that achieves privacy protection and resource optimization as in any embodiment of the present invention, which will not be repeated here.

[0025] By adopting the technical solution of the present invention, the following technical effects can be achieved: (1) By setting up a privacy protection module, which consists of multiple roadside unit sets, when the intelligent vehicle sends raw data, it will be converted into an intermediate aggregation model through the privacy protection module, and then the multiple intermediate aggregation models will be aggregated into the next global model through the global model aggregation module. Through the next global model Instead of transmitting raw data, it effectively protects the local data privacy of intelligent vehicles; (2) First, the original data is transformed into an intermediate aggregation model and then the intermediate aggregation model is transformed into a global model. This can make the models transmitted on the nodes different, thus strengthening the ability to resist malicious attackers from capturing and cracking the original model. (3) By periodically updating the target Q network, the update method is to assign the model parameters of the Q network to the target Q network. This can cope with the rapid changes in local resources of intelligent vehicles connected to the Internet of Vehicles, and can reduce the uncertainty caused by unknown states through the mapping of neural networks. It has a stronger adaptability and reduces the pressure on the backbone network. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings to be used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is one of the architecture block diagrams for controlling the vehicle network model training system according to the first embodiment of the present invention; Figure 2 This is the second architectural block diagram of the vehicle network model training method described in the first embodiment of the present invention for controlling the vehicle network model training system; Figure 3 This is a flowchart of the vehicle network model training method for achieving privacy protection and resource optimization as described in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the transmission between the Q network and the target Q network in the first embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the readable storage medium according to the second embodiment of the present invention.

[0028] Explanation of reference numerals in the attached figures: 200 represents readable storage media; 210 represents computer-executable instructions. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] [First Embodiment] Combination Figure 1-3 The first embodiment of the present invention provides a vehicle-to-everything (V2X) model training method for achieving privacy protection and resource optimization. The V2X model training method is used to control a V2X model training system. The V2X model training system includes: an intelligent transportation system; at least one intelligent vehicle, wherein the at least one intelligent vehicle is equipped with a local computing and environment module and a sensing module, and the at least one intelligent vehicle is connected to the intelligent transportation system; a privacy protection module, which is composed of multiple roadside unit sets; and a resource optimization module, wherein the resource optimization module is equipped with a global model aggregation module, and the global model aggregation module is connected to the privacy protection module. The vehicle-to-everything (V2X) model training method specifically includes: Step S100: The intelligent transportation system sends the global model to the multiple roadside units. The multiple roadside units will integrate the global model with a subset of intelligent vehicles and a global task Ti. The data is sequentially sent to the corresponding intelligent vehicles within the subset of intelligent vehicles. Step S200: Collect road information and video information through the sensing module, and according to the global task T i The road information, video information, and intelligent vehicle data applicable to the global task are integrated into a local model and sent to the roadside unit. Step S300: Each roadside unit performs intermediate aggregation on the local model sent by the intelligent vehicle to form an intermediate aggregated model; Step S400: The multiple roadside units send the intermediate aggregated model to the global model aggregation module to obtain the next global model. And the next global model As the initial model for the next global task Ti+1; It should be noted that i represents the current communication round number.

[0031] For example, on the one hand, by setting up a privacy protection module, which consists of multiple roadside unit sets, when the intelligent vehicle sends raw data, it will be transformed into an intermediate aggregation model through the privacy protection module, and then the global model aggregation module will aggregate the multiple intermediate aggregation models into the next global model. Through the next global model Instead of transmitting raw data, it effectively protects the local data privacy of intelligent vehicles; on the other hand, by first converting the raw data into an intermediate aggregation model and then into a global model, the models transmitted on the nodes can be different, which strengthens the ability to resist malicious attackers from capturing and cracking the raw model.

[0032] Specifically, in step S200, road information and video information are collected through the sensing module, and based on the global task T... i The road information, video information, and intelligent vehicle data applicable to the global task are integrated into a local model and sent to the roadside unit; specifically, step S210: the intelligent vehicle activates wireless communication to receive the global task T from the intelligent transportation system. i Step S220: According to the global task T i Select the corresponding road information, video information, and intelligent vehicle data set to form the local resource status. Step S230: Based on the local resource status With the global task T i The minimum loss minf(w) is set as the training objective to obtain the local model. Among them, minimum loss , x is the amount of data in the dataset established by the kth intelligent vehicle participating in this communication round. i Local resource status The processed vector form, y i Let w be the label corresponding to the data vector, w be the weight vector of the local model, and l be the optimization method; Step S240: Upload the status of the local resources. and the local model To the roadside unit.

[0033] For example, by setting the minimum loss minf(w) as the training objective, and minimizing the loss... By repeatedly training, the preset minimum loss value is achieved, thereby improving training accuracy.

[0034] Further, in step S300, each roadside unit performs intermediate aggregation on the local model sent by the intelligent vehicle to form an intermediate aggregated model, specifically including: Step S310: The roadside unit receives the local resource status. and the local model Step S320: For the received local model... Intermediate aggregation is performed to form an intermediate aggregation model; wherein, the intermediate aggregation model .

[0035] For example, local resource status and local model Implement intermediate aggregation and intermediate aggregation model in roadside units. By delegating the aggregation task to intermediate aggregation nodes, the models transmitted by the two intermediate nodes are inconsistent, thereby strengthening the ability to resist malicious attackers from capturing and cracking the original model.

[0036] Furthermore, in step S400, the multiple roadside units send the intermediate aggregated model to the global model aggregation module to obtain the next global model. And the next global model As the next global task T i+1 The initial model specifically includes: Step S410: The multiple roadside units aggregate the intermediate model. Send to the global model aggregation module; Step S420: The global model aggregation module sends multiple intermediate aggregated models The next global model is obtained by aggregation. The next global model Step S430: The next global model... This will serve as the initial model for the next global task Ti+1.

[0037] For example, by using an intermediate aggregation model Send to the global model aggregation module, and combine multiple intermediate aggregated models. The next global model is obtained by aggregation. And the next global model The next global model As the initial model for the next global task Ti+1, the detection accuracy of the global model is continuously improved.

[0038] Preferably, the method further includes the following steps before step S100: Step S10: Set the target precision; Step S20: When the global model If the target accuracy is not achieved, then repeat steps S100, S200, S300, and S400.

[0039] For example, by setting a target accuracy, when the global model... If the target accuracy is not achieved, the test is repeated until the target accuracy is achieved.

[0040] Preferably, before step S100, the method further includes: the intelligent transportation system sending an initial global task T1 to the resource optimization module, and the resource optimization module establishing an initial global model based on the initial global task T1. .

[0041] For example, by manually creating an initial global model And the initial global model It can establish a target accuracy that fits the initial global task T1, thereby reducing the number of training iterations and achieving the target accuracy faster.

[0042] Preferably, the resource optimization module further includes: an intelligent vehicle strategy library, which uses deep reinforcement learning as the intelligent vehicle selection algorithm. By collecting the resource status of the intelligent vehicles and calculating rewards, the mapping function is continuously updated, and the most participating subset of intelligent vehicles is obtained from the mapping function. The intelligent vehicles with the longest long-term benefits are selected to participate in the next global task.

[0043] For example, on the one hand, by using data-driven deep reinforcement learning to replace the original static selection algorithms, such as random algorithms and greedy algorithms, the intelligent vehicle selection algorithm of this invention selects intelligent vehicles with longer-term benefits to participate in global model updates through analysis of the environment and intelligent vehicle resources, thereby reducing the uncertainty brought about by the original static algorithm, enhancing the robustness of the algorithm of this invention, and ultimately improving the overall resource utilization of the system. On the other hand, deep reinforcement learning establishes an experience replay area through interaction with the environment, and uses the samples in the experience replay area as training samples. In the early stage of training, the required data can be continuously generated through simulation training and updated offline, with low requirements for real-time data.

[0044] Preferred, see Figure 4 The intelligent vehicle policy library uses DDQN as its main framework, with DNN serving as both the Q-network and the target Q-network. For example, using DDQN as the main framework and DNN as both the Q-network and the target Q-network improves the scalability of the intelligent vehicle selection algorithm.

[0045] Preferably, the intelligent vehicle selection algorithm includes, but is not limited to, DQN, DDQN, DDPG, and A3C; the intelligent vehicle strategy mapping model includes, but is not limited to, DNN, CNN, seq2seq, and RNN.

[0046] Specifically, the intelligent vehicle policy library adopts DDQN as the main framework, and DNN as the Q-network and target Q-network specifically includes: Step S500: Initialize the experience replay, Q network, target Q network, learning rate 0.9, discount factor 0.5, initial selection probability ε 0.1, and selection probability ε decay coefficient 0.001; Step S510: If the intelligent vehicle selection algorithm is not in a convergent state, repeat steps S520-S590: Step S520: If the global model If the accuracy is less than the target accuracy or less than the maximum number of communication rounds, repeat steps S530-S590; Step S530: Randomly select action a with probability ε, and select the action with the maximum reward value a=argmaxQ(s,a;θ) with probability (1-ε), where action a is the subset of intelligent vehicles, probability ε decreases continuously with the increase of communication rounds, and θ is the model weight of Q network; Step S540: Execute action a, that is, select the corresponding subset of intelligent vehicles to participate in the global model of this communication round. renew; Step S550: After executing action a, obtain the reward value R from environmental feedback, and the resource state s′ for the next communication round, where the reward value is set as R = -α·power consumption (fuel consumption) - β·time delay, s is the resource state for the current communication round, and the quadruple is set.<s,a,R,s′> Store the experience replay, where α and β are coefficients; Step S560: If the experience replay is full, proceed to steps S570-S590; otherwise, skip. Step S570: Randomly select the required number of quadruplets from the experience replay.<s,a,R,s′> , as the smallest batch sample; Step S580: Update the Q-network using the extracted samples, the update method is as follows: , where η is a coefficient; This represents the updated Q value after performing action a in state s; This represents the Q-value obtained through the Q-network when action a is performed in state s. This represents the reward obtained by performing action a in state s; Let Q represent the Q value obtained by taking the best action a' in the next communication round under the resource conditions; a' represents the best action in the next communication round; A represents the set of all actions. Step S590: Periodically update the target Q network by assigning the model parameters of the Q network to the target Q network.

[0047] For example, by periodically updating the target Q network, the update method is to assign the model parameters of the Q network to the target Q network. This can cope with the rapid changes in local resources of connected vehicle intelligent vehicles, and can reduce the uncertainty caused by unknown states through neural network mapping, thus having stronger adaptability and reducing the pressure on the backbone network.

[0048] In one specific embodiment, the convolutional neural network model size of the MNIST digit recognition task dataset is about 1MB, while the size of a single digit image is about 700kB. If each intelligent vehicle transmits 600 images, the amount of data transmitted is about 410MB. If the original data is replaced by transmitting the local model, the amount of data transmitted is reduced by more than 99%.

[0049] [Second Embodiment] See Figure 5 This embodiment also provides a readable storage medium 200, which stores computer-executable instructions 210. When the computer-executable instructions 210 are read and executed by the processor, they control the air conditioner where the readable storage medium 200 is located to implement the vehicle network model training method for achieving privacy protection and resource optimization as described in the first embodiment.

[0050] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0051] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a vehicle-to-everything (V2X) model that achieves privacy protection and resource optimization, characterized in that, The vehicle-to-everything (V2X) model training method is used to control the V2X model training system; The vehicle-to-everything (V2X) model training system includes: Intelligent Transportation Systems; At least one intelligent vehicle, wherein the at least one intelligent vehicle is equipped with a local computing and environment module and a sensing module, and the at least one intelligent vehicle is connected to the intelligent transportation system; The privacy protection module is composed of multiple roadside unit sets; The resource optimization module includes a global model aggregation module, which is connected to the privacy protection module. The vehicle-to-everything (V2X) model training method specifically includes: Step S100: The intelligent transportation system sends the global model to the multiple roadside units. Intelligent vehicle subset and global task T i The multiple roadside units will define the global model. The data is sequentially sent to the corresponding intelligent vehicles within the subset of intelligent vehicles. Step S200: Collect road information and video information through the sensing module, and according to the global task T i The road information, video information, and intelligent vehicle data applicable to the global task are integrated into a local model and sent to the roadside unit. Step S300: Each roadside unit performs intermediate aggregation on the local model sent by the intelligent vehicle to form an intermediate aggregated model; Step S400: The multiple roadside units send the intermediate aggregated model to the global model aggregation module to obtain the next global model. And the next global model As the initial model for the next global task Ti+1; Where i is the current communication round number; Step S200 specifically includes: Step S210: The intelligent vehicle activates wireless communication to receive and connect to the intelligent transportation system to receive the global task T. i Step S220: According to the global task T i Select the corresponding road information, video information, and intelligent vehicle data set as the local resource state. Step S230: Based on the local resource status , with the global task T i The minimum loss minf(w) is set as the training objective to obtain the local model. Among them, minimum loss , x is the amount of data in the dataset established by the kth intelligent vehicle participating in this communication round. i Local resource status The processed vector form, where the superscript k represents the k-th intelligent vehicle participating in this global model iteration, x i and y i The label corresponds to the data vector, where index i represents the i-th training sample, w represents the weight vector of the local model, and l represents the optimization method; Step S240: Upload the status of the local resources. and the local model To the roadside unit; The resource optimization module also includes: The intelligent vehicle strategy library uses deep reinforcement learning as the selection algorithm for the intelligent vehicles. By collecting the resource status of the intelligent vehicles and calculating the rewards, the mapping function is continuously updated and the most participating subset of intelligent vehicles is obtained from the mapping function. The intelligent vehicles with the longest long-term benefits are selected to participate in the next global task. The intelligent vehicle strategy library adopts DDQN as the main framework, and DNN as the Q network and target Q network. The intelligent vehicle policy library adopts DDQN as its main framework, with DNN serving as both the Q-network and the target Q-network, specifically including: Step S500: Initialize the experience replay, Q network, target Q network, learning rate 0.9, discount factor 0.5, initial selection probability ε 0.1, and selection probability ε decay coefficient 0.001; Step S510: If the intelligent vehicle selection algorithm is not in a convergent state, repeat steps S520-S590: Step S520: If the global model If the accuracy is less than the target accuracy or less than the maximum number of communication rounds, repeat steps S530-S590; Step S530: Randomly select action a with probability ε, and select the action with the maximum reward value a=argmaxQ(s,a;θ) with probability (1-ε), where action a is the subset of intelligent vehicles, probability ε decreases continuously with the increase of communication rounds, and θ is the model weight of Q network; Step S540: Execute action a, that is, select the corresponding subset of intelligent vehicles to participate in the global model of this communication round. renew; Step S550: After executing action a, obtain the reward value R from the environmental feedback, and the resource state s′ of the next communication round, where the reward value is set to R = -α·power consumption β·time delay, s is the resource state of the current communication round, and set the quadruple...<s,a,R,s′> Store the experience replay, where α and β are coefficients; Step S560: If the experience replay is full, proceed to steps S570-S590; otherwise, skip. Step S570: Randomly select the required number of quadruplets from the experience replay.<s,a,R,s′> , as the smallest batch sample; Step S580: Update the Q-network using the extracted samples, the update method is as follows: , where η is a coefficient; This represents the updated Q value after performing action a in state s; This represents the Q-value obtained through the Q-network when action a is performed in state s. This represents the reward obtained by performing action a in state s; Let Q represent the Q value obtained by taking the best action a' in the next communication round under the resource conditions; a' represents the best action in the next communication round; A represents the set of all actions. Step S590: Periodically update the target Q network by assigning the model parameters of the Q network to the target Q network.

2. The vehicle-to-everything (V2X) model training method according to claim 1, characterized in that, Step S300 specifically includes: Step S310: The roadside unit receives the local resource status. and the local model ; Step S320: For the received local model Intermediate aggregation is performed to form an intermediate aggregation model; wherein, the intermediate aggregation model N IV This refers to the total number of intelligent vehicles participating in this model iteration.

3. The vehicle-to-everything (V2X) model training method according to claim 2, characterized in that, Step S400 specifically includes: Step S410: The multiple roadside units will aggregate the intermediate model. Send to the global model aggregation module; Step S420: The global model aggregation module aggregates multiple intermediate models. The next global model is obtained by aggregation. The next global model N RSU This refers to the total number of intelligent vehicles participating in this model iteration; Step S430: The next global model As the next global task T i+1 The initial model.

4. The vehicle-to-everything (V2X) model training method according to claim 3, characterized in that, The procedure preceding step S100 also includes: Step S10: Set the target precision; Step S20: When the global model If the target accuracy is not achieved, then repeat steps S100, S200, S300, and S400.

5. The vehicle-to-everything (V2X) model training method according to claim 4, characterized in that, The steps preceding step S100 also include: The intelligent transportation system sends an initial global task T1 to the resource optimization module, and the resource optimization module establishes an initial global model based on the initial global task T1. .

6. A readable storage medium, characterized in that, The readable storage medium includes stored computer-executable instructions, wherein when the computer-executable instructions are executed by a processor, they control the device where the storage medium is located to perform the steps of the vehicle network model training method that achieves privacy protection and resource optimization as described in any one of claims 1-5.

Images