A vehicle networking edge network service deployment and request offloading method

Abstract

This invention proposes a method for deploying and offloading edge network services in the field of vehicle-to-everything (V2X) computing. The invention constructs a V2X edge computing system. Based on task request information from vehicles, a central base station formulates a contract set, which is then broadcast to potential service-oriented vehicle-mounted drones for contract selection. An optimization model is constructed using the objective function of maximizing the minimum task completion rate, combined with constraints. Solving this optimization model is transformed into a Markov decision process, and deep reinforcement learning is used to train the service deployment and task offloading model. Based on the trained model, the optimal service deployment and task offloading decisions are made according to the current task request. This invention can fully utilize the computing resources of service nodes, reduce task latency, alleviate the burden on roadside units, and improve task completion rate while protecting privacy.

Description

Technical Field

[0001] This invention relates to the technical field of vehicle networking and edge computing, and more specifically, to a method for deploying and requesting unloading edge network services for vehicle networking. Background Technology

[0002] With the rapid development of vehicle-to-everything (V2X) technology, vehicles in modern transportation systems can connect with each other and communicate with roadside infrastructure through wireless communication technology. This interconnected V2X is widely used in intelligent transportation systems, autonomous driving, road safety warnings, and in-vehicle entertainment services.

[0003] In the vehicle-to-everything (V2X) network, vehicles can form a dynamic network structure through wireless communication, enabling information exchange and collaborative control. Existing solutions typically rely on cloud computing centers or fixed roadside units (RSUs) for services. However, cloud computing centers suffer from high transmission latency, and fixed RSUs have limited coverage and high deployment and maintenance costs. As vehicle intelligence increases, private cars are gradually acquiring stronger computing and storage capabilities, enabling them to act as service nodes, forming in-vehicle fog computing or in-vehicle cloud. Furthermore, this reduces the burden on RSUs when they are busy, significantly improving task completion rates. This type of distributed architecture enables resource sharing and task offloading among vehicles, effectively reducing latency and improving resource utilization.

[0004] Most existing research is based on idealized assumptions, namely that when making service deployment and task offloading decisions, the system is assumed to have complete access to the resource and status information of all vehicles and service nodes, thus achieving optimization under globally visible conditions. However, in real-world connected vehicle environments, vehicles using in-vehicle fog computing are typically unwilling to directly expose their computing power, storage, or energy consumption to protect their privacy and security. This need for incomplete information and privacy protection makes traditional decision-making mechanisms that rely on globally known information significantly limited in practical applications, necessitating new methods to support efficient service deployment and task offloading even when information is partially invisible. Summary of the Invention

[0005] To address the problem that existing technologies cannot achieve efficient service deployment and task offloading when vehicle and service node information is not visible, this invention proposes a method for deploying and requesting offloading services in the edge network of the Internet of Vehicles. This method optimizes task offloading decisions, ensuring reasonable allocation of computing resources while protecting the privacy of service vehicle-mounted drones.

[0006] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: A method for deploying and requesting offloading of edge network services in the Internet of Vehicles (IoV) includes the following steps: A vehicle-to-everything (V2X) edge computing system is constructed, comprising a central base station, several roadside units, several service vehicle-mounted drones, several task-requesting vehicles, and various deep neural network inference service models. The task request vehicle generates computing tasks based on its own business needs and sends the task request information to the central base station; the task request information includes the identifier of the requested deep neural network inference service model, the amount of task data, the computing load, and the maximum tolerable latency; Based on the task request information, a contract set is formulated using the central base station, and the contract set is broadcast to potential service vehicle-mounted drones through roadside units. The potential service vehicle-mounted drones select contract items according to their own type and return the contract item identifiers to the central base station. The central base station determines the amount of computing resources committed to be contributed by each service vehicle-mounted drone based on the contract item identifiers. The contract signing results, service model cache status, and communication link status at the current moment are collected by the central base station. Combined with the task request information, the communication transmission delay of the task is calculated and the total task delay is obtained. With the objective function of maximizing the minimum task completion rate of various services, and with constraints such as unique task unloading constraint, service model caching constraint, service vehicle-mounted drone computing resource capacity constraint, task latency constraint, and contract individual rationality and incentive compatibility constraint, an optimization objective model is constructed. The optimization objective model is transformed into a Markov decision process. Based on deep reinforcement learning, a trained service deployment and request offloading model is obtained, and the service model caching decision and task offloading decision are output.

[0007] Preferably, the vehicle-to-everything (V2X) edge computing system includes a central base station and U roadside units, wherein It includes I task requesting vehicles and M service vehicle-mounted drones, among which , It also includes N different deep neural network inference service models, among which The N different deep neural network inference service models are used to provide different edge AI services; the roadside units are connected to the central base station via wired links, and the roadside units are connected to each other via wired links. The task requesting vehicle is connected to the roadside unit, the task requesting vehicle is connected to the service vehicle-mounted drone, and the service vehicle-mounted drone is connected to the roadside unit via wireless links; the service vehicle-mounted drone is pre-configured with multiple deep neural network inference service model virtual machines.

[0008] Preferably, the process of formulating the contract set includes: based on the pre-entered data distribution of the energy consumption model of the service vehicle-mounted drone, using the central base station to calculate the unit cost coefficient of the service vehicle-mounted drone. They are categorized into K types from highest to lowest, denoted as... ,in, Indicates the first The unit computational cost coefficient for each type of service vehicle-mounted drone; setting the computational resource quantity in the contract as... ,in, Representation type k The computing resources required for the vehicle-mounted drone; constructing a set containing K contract items. ,in, , Payment to the central base station for type The reward for vehicle-mounted drones, when season ,when hour, ,in, This represents the energy consumption conversion factor. This indicates the increased cost compensation due to increased resource contribution. Indicates the first Vehicle-mounted drones for similar services are more efficient than Information rental income obtained from vehicle-mounted drone services.

[0009] Preferably, the service vehicle-mounted drone is determined. The process for determining the required amount of computing resources is as follows: make Indicated as a service vehicle-mounted drone The amount of resources required when serving vehicle-mounted drones Upon receiving the contract set broadcast by the central base station via roadside units, the service vehicle-mounted drone... Its own type To determine whether the individual rationality constraint and incentive compatibility constraint are satisfied, the expression for the individual rationality constraint is: The expression for the incentive compatibility constraint is: ,in The type is The utility of vehicle-mounted drones If the conditions are met, then select the contract. It becomes a node capable of unloading tasks and returns the contract item identifier to the central base station. At this point, i.e., vehicle-mounted drone service Committed computing resources Equal to the contract item .

[0010] Preferably, the process of calculating the communication transmission delay of the task satisfies the expression: , ,

[0011] ,

[0012] , , ,

[0013] ,

[0014] , in, Indicates the vehicle number that requested the task. This represents the set of vehicles requesting the task, where the number of elements is... , This indicates the serial number of the vehicle-mounted drone used for service. This represents the set of service vehicle-mounted drones, where the number of elements is... , Indicates the number of the roadside unit, and and Representing different roadside units, Represents the set of roadside units, where the number of elements is... , where n represents the number of the deep neural network inference service model. This represents the set of deep neural network inference service models, where the number of elements is... ; Vehicle representing the task request With service vehicle-mounted drones Channel gain between Vehicle representing the task request With roadside units Channel gain between Represents roadside unit With service vehicle-mounted drones Channel gain between This indicates the gain at a reference distance of 1 meter. Request vehicle for mission With service vehicle-mounted drones Euclidean distance, Request vehicle for mission With roadside units Euclidean distance, roadside unit With service vehicle-mounted drones Euclidean distance, This is the path loss index. Indicates the vehicle requesting the task. Do you request a deep neural network inference service model? If requested, then ,otherwise , Request vehicle for mission Requesting a deep neural network inference service model The amount of data for the task; Request vehicle for mission With service vehicle-mounted drones Uplink transmission rate of direct links between them Indicates the vehicle requesting the task. With service vehicle-mounted drones Bandwidth between Indicates the vehicle requesting the task. Transmission power, Indicates service vehicle-mounted drone Environmental noise; Request vehicle for mission With roadside units Uplink transmission rate between Indicates the vehicle requesting the task. With roadside units Bandwidth between Indicates the vehicle requesting the task. Transmission power, Represents roadside unit Environmental noise; roadside unit With service vehicle-mounted drones Downlink transmission rate between Represents roadside unit With service vehicle-mounted drones Bandwidth between Represents roadside unit Transmission power, Indicates service vehicle-mounted drone Environmental noise; roadside unit With roadside units The transmission rate of the wired link between them; Request vehicle for mission With service vehicle-mounted drones Transmission delay when within the coverage area of ​​the same roadside unit; Request vehicle for mission With service vehicle-mounted drones When within the coverage area of ​​different roadside units, the indirect communication transmission delay via roadside unit relay is adopted; Request vehicle for mission With service vehicle-mounted drones The task transmission delay between them can be divided into two cases: when the task requests the vehicle... With service vehicle-mounted drones Within the coverage area of ​​the same roadside unit, the task transmission time is When the task requests the vehicle With service vehicle-mounted drones When within the coverage area of ​​different roadside units, the task transmission time is .

[0015] Preferably, the process of calculating the total delay of task completion satisfies the expression: ,

[0016]

[0017] in, Indicates the vehicle requesting the task. Deep Neural Network Inference Service Model The computational load, Indicates service vehicle-mounted drone Assigned to the vehicle requesting the task The computational resources for the task Indicates service vehicle-mounted drone Is the deep neural network inference service model cached? If it is cached ,otherwise , Indicate whether to request the vehicle for the mission. Task offloading to service vehicle-mounted drone If so, then ,otherwise , To serve vehicle-mounted drones Start the corresponding deep neural network inference service model The virtual machine, and the deep neural network inference service model Latency when deployed on a virtual machine; Request vehicle for mission The task is carried out by a service vehicle-mounted drone. Total delay of completion; To serve vehicle-mounted drones Calculate the vehicle request task The execution time of the task.

[0018] Preferably, the process of constructing the optimization target model is as follows: The objective function is:

[0019] The constraints are: C1: ; C2: ; C3: ; C4: ; C5: ; C6: ； ; in, This represents the set of model cache decision strategies, containing binary decision variables. , This represents the set of task unloading decision strategies, containing binary decision variables. , , Indicates the first Task completion rate of deep neural network inference services Introducing auxiliary variables The minimum value used to characterize the task completion rate of each service type is set as the optimization objective to maximize the minimum value. At the same time, set For any service type All are true; C1 is the unique unloading constraint for the task, meaning that for any task requesting vehicle... The computational tasks it generates can at most be offloaded to a single service vehicle-mounted drone. Process it; C2 is a service model caching constraint, indicating that if a task requests a vehicle... Request No. This involves providing edge network services and offloading tasks to service-equipped drones. Then the drone The corresponding service model must have been cached. ; C3 represents the computing resource capacity constraint allocated to the service vehicle-mounted drone, indicating the capacity of resources allocated to it. The total computing resources required for all computing tasks of the drone must not exceed the total computing resources required for the drone. Maximum computing resource capacity willing to be given up ; This represents the computing resources allocated to a single task; C4 represents the task delay constraint, indicating the task request vehicle. Total latency of computation task completion It must be less than or equal to the maximum tolerable latency threshold for this task. ; C5 represents the individual reasonableness constraint, indicating the remuneration received by the service vehicle-mounted drone participating in the contract. It must be sufficient to cover the computing resources it provides. Energy costs generated ; C6 represents the incentive compatibility constraint, indicating the type as follows: Service vehicle-mounted drones: Select contract items of their type The utility gained must be greater than or equal to the utility gained from choosing other types. Contract terms The utility that can be obtained.

[0020] Preferably, the optimization objective model solution process is converted into a Markov decision process, specifically as follows: First, the continuous operation process of the vehicle-to-everything (V2X) edge computing system is divided into several scheduling time slots, denoted as time slot index t; the state of the V2X edge computing system remains unchanged or approximately unchanged within each time slot, and a decision update is performed at the beginning of the time slot; second, the state space, action space, and reward function are determined, wherein: state space ,in, This represents the overall channel quality between all nodes in the vehicle-to-everything (V2X) edge computing system. , This represents the channel gain matrix between the requesting vehicle and the serving vehicle-mounted drone. Its elements Corresponding task request vehicle With service vehicle-mounted drones Channel gain, This represents the channel gain matrix between the requesting vehicle and the roadside unit. Its elements Corresponding task request vehicle With roadside units Channel gain, This represents the channel gain matrix between the roadside unit and the serving vehicle-mounted drone. Its elements Corresponding roadside unit With service vehicle-mounted drones Channel gain, This represents the vector of remaining available computing resources for all service vehicle-mounted drones at the current moment. ,in This indicates the remaining available computing resources for serving vehicle-mounted drones. , This represents the set of task request information sent by the task requesting vehicle. This includes the amount of task data, computational load, and maximum tolerable latency; Action space ,in, Its elements Indicates service vehicle-mounted drone Should the cached deep neural network inference service model be executed? The decision, Its elements Indicates the vehicle requesting the task. Should the task be unloaded to the service vehicle-mounted drone? Decisions made on the ground This represents the computing resource contribution requirements set by the central base station for K different types of drones. The agent adjusts this vector to determine the resource terms in the contract menu, and then automatically derives the corresponding optimal reward. ; reward function ,in, , The minimum task completion rate, , The latency violation penalty indicates that the total actual latency of all tasks exceeds the maximum tolerable latency. ; This is a resource overflow penalty, representing the portion of the total computing resources allocated to the task that exceeds the capacity of the service vehicle-mounted drone. ; These are the preset positive trade-off coefficients.

[0021] Preferably, based on the state space Action space and reward function The process of designing actor networks and critic networks separately, and solving the optimization objective function based on deep reinforcement learning to obtain the optimal service deployment and task unloading results is as follows: The actor network is initialized based on a preset probability distribution. parameters Critics Network parameters , wherein the parameters and Specifically, this includes the connection weights and bias terms of all neurons in the input, hidden, and output layers of their respective networks; and initializes the bias parameters to positive numbers; and sets the target actor network. parameters and the target critic network parameters The experience replay pool is initialized to a constant capacity and empty. Specifically, the experience replay pool is constructed as a fixed-capacity first-in-first-out circular buffer. For each time slot index t, the following steps are repeated until training converges: Based on the current state in the state space Use an actor network to select actions and add noise to the actions. The expression is: ,in The actor network indicates the current status. The output deterministic strategy This represents exploratory noise that follows a Gaussian distribution. Execution constraint verification and environment interaction steps: Verify the legality of the action: Determine the action's validity. Check if constraints C1 and C2 are met; if not, the current task is deemed a failure, and a preset invalid action penalty value is given as a reward. To keep the system state unchanged, that is Transfer data Store the data in the experience replay pool to train the agent to avoid outputting such actions; if the conditions are met, then execute the action in the environment. Feasibility verification: During execution, determine whether the system state satisfies the constraints C3, C4, C5, and C6. If not, the task execution is deemed to have failed, and a reward including a high penalty is calculated based on the reward function. and obtain the updated status. The data is stored in the experience replay pool. If the conditions are met, a positive reward based on the minimum task completion rate is calculated, and the next state is obtained. The data will be successfully transferred. Store in the experience replay pool; Multiple data points were randomly sampled from the experience replay pool. ; Update the critic network and calculate the target Q-value, expressed as: ,in, Using the discount factor, the commentator network loss is calculated as follows: ,in, To determine the batch size for empirical replay sampling, gradient descent is used to update the parameters of the critic network. The expression is: ,in, Indicates an update operation; Update the actor network and calculate the policy gradient of the actor network, expressed as: The actor network parameters are updated using gradient ascent, expressed as: ; The target network parameters are updated using a soft update method, expressed as follows: ,in For hyperparameters; Continue until training converges, obtaining the trained service deployment and request unloading model.

[0022] Preferably, the central base station, based on the trained service deployment and request offloading model, issues control commands for service model caching decisions and task offloading decisions according to the current status of the vehicle-to-everything (V2X) edge computing system and task requests. The service vehicle-mounted drone responds to the control commands and starts the corresponding deep neural network inference service model virtual machine in advance. The task requesting vehicle responds to the control commands and offloads the computing task to the designated service vehicle-mounted drone for processing.

[0023] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention proposes a method for service deployment and task offloading in the edge network of a vehicle-to-everything (V2X) network. It constructs an V2X edge computing system that integrates mobile and fixed edge resources, improving coverage and service availability. Based on task request information from vehicles, a contract set is established using a central base station and broadcast to potential service-oriented vehicle-mounted drones (V2Xs). These V2Xs then select a contract, effectively inducing them to truthfully disclose their capabilities and contribute computing resources commensurate with their abilities. An optimization model is constructed using the objective function of maximizing the minimum task completion rate, combined with constraints. Solving this model is transformed into a Markov decision process, and deep reinforcement learning is used to train the service deployment and task offloading model. This avoids resource monopolies by a few dominant nodes or resource crowding out by simple tasks in the edge network, effectively ensuring that vehicles requesting services at the network edge or with poor channel conditions can receive necessary service responses, thereby improving the coverage quality and service consistency of the entire V2X edge network service. Based on the trained model, optimal service deployment and task offloading decisions are made according to the current task request. This invention fully utilizes the computing resources of service nodes, reduces task latency, alleviates the burden on roadside units, and improves task completion rate while protecting privacy. In terms of service deployment, it effectively overcomes information asymmetry and achieves precise incentives and allocation of resources, thereby improving the coverage quality and service consistency of the entire vehicle-to-everything (V2X) edge network service. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a method for deploying and requesting unloading edge network services for vehicle-to-everything (V2X) proposed in this embodiment of the invention. Figure 2 This is a schematic diagram of the vehicle-to-everything (V2X) edge computing system proposed in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the deep reinforcement learning training framework proposed in this embodiment of the invention; Figure 4 This is a schematic diagram showing the relationship between reward and training rounds during the deep reinforcement learning training process proposed in this embodiment of the invention. Figure 5 A schematic diagram showing the comparison curves of task completion rate as a function of the number of vehicles requesting tasks in an embodiment of the present invention; Figure 6 This is a bar chart showing the comparison of average task latency as a function of task data volume in embodiments of the present invention. Detailed Implementation

[0025] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions; It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Example 1 This embodiment proposes a method for deploying and requesting unloading edge network services in the vehicle-to-everything (V2X) network. (See also...) Figure 1 This includes the following steps: S1: Construct a vehicle-to-everything (V2X) edge computing system, which includes a central base station, several roadside units, several service vehicle-mounted drones, several task-requesting vehicles, and various deep neural network inference service models. S2: The task request vehicle generates a computing task based on its own business needs and sends the task request information to the central base station; the task request information includes the identifier of the requested deep neural network inference service model, the amount of task data, the computing load, and the maximum tolerable latency; S3: Based on the task request information, the central base station formulates a contract set and broadcasts the contract set to potential service vehicle-mounted drones through roadside units. The potential service vehicle-mounted drones select contract items according to their own type and return the contract item identifiers to the central base station. The central base station determines the amount of computing resources committed to be contributed by each service vehicle-mounted drone based on the contract item identifiers. S4: Use the central base station to collect the contract signing results, service model cache status and communication link status at the current moment, and combine them with the task request information to calculate the communication transmission delay of the task and obtain the total task delay. S5: With the objective function of maximizing the minimum task completion rate of various services, and with constraints such as unique task unloading constraint, service model caching constraint, service vehicle-mounted drone computing resource capacity constraint, task latency constraint, and contract individual rationality and incentive compatibility constraint, an optimization objective model is constructed. S6: The optimization objective model solution is transformed into a Markov decision process. Based on deep reinforcement learning, a trained service deployment and request offloading model is obtained, and the service model caching decision and task offloading decision are output.

[0028] In this embodiment, a vehicle-to-everything (V2X) edge computing system is constructed. This system integrates mobile and fixed edge resources, improving coverage and service availability. The transmission efficiency and latency within the computing system provide a communication basis for task offloading. Based on task requests issued by vehicles, a contract set is formulated using a central base station and broadcast to potential service vehicle-mounted drones. These potential service vehicle-mounted drones possess service capabilities but are not activated by the existing system; they need to be transformed into actual service nodes through contracts. The contract set incentivizes service vehicle-mounted drones to participate in task offloading and resolves the information asymmetry problem. This maximizes the minimum task requirement for various services. The task completion rate is used as the objective function. An optimization target model is constructed using constraints such as unique task offloading, service model caching, service vehicle-mounted drone computing resource capacity, task latency, and individual contract rationality and incentive compatibility. This ensures the fairness of the task completion rate and optimizes the overall system performance. Based on Markov decision processes and deep reinforcement learning, the system's adaptability and robustness are improved. The optimization target model is solved, and the optimal service deployment and task offloading model is trained. This model is deployed to the central base station, and optimal service deployment and task offloading decisions are made based on the current state of the vehicle-to-everything (V2X) edge computing system and task requests. The method proposed in this invention can fully utilize the computing resources of service nodes, reduce task latency, alleviate the burden on roadside units, and improve the task completion rate while protecting privacy.

[0029] Example 2 This embodiment proposes a method for deploying and requesting unloading edge network services in the vehicle-to-everything (V2X) network. (See also...) Figure 2 The vehicle-to-everything (V2X) edge computing system includes a central base station and U roadside units, among which... It includes I task requesting vehicles and M service vehicle-mounted drones, among which , It also includes N different deep neural network inference service models, among which N different deep neural network inference service models are used to provide different edge AI services. Roadside units are connected to the central base station via wired links, and roadside units are connected to each other via wired links. The task requesting vehicle is connected to the roadside unit, the task requesting vehicle is connected to the service vehicle-mounted drone, and the service vehicle-mounted drone is connected to the roadside unit via wireless links. Multiple deep neural network inference service model virtual machines are pre-configured on the service vehicle-mounted drone.

[0030] In this embodiment, the service vehicle-mounted drone can deploy different deep neural network inference service models. The roadside units are connected to the central base station via wired fiber optic cables with a link transmission rate of 10 Gbps. The roadside units are also connected to each other via wired fiber optic cables with a link transmission rate of 10 Gbps. The system indicates that vehicles and roadside units are connected wirelessly with a bandwidth of 20MHz, and task-requesting vehicles and service vehicle-mounted drones are also connected wirelessly with a bandwidth of 10MHz. The maximum tolerable latency for task-requesting vehicles is 0.5 seconds. Roadside units and vehicles are evenly distributed on a two-lane, one-way road, and there are 10 different deep neural network inference service models.

[0031] In one optional embodiment, the process of formulating the contract set includes: based on the pre-recorded distribution of energy consumption model data for the service vehicle-mounted drone, using the central base station to calculate the unit cost coefficient for the service vehicle-mounted drone. They are categorized into K types from highest to lowest, denoted as... ,in, Indicates the first The unit computational cost coefficient for each type of service vehicle-mounted drone; setting the computational resource quantity in the contract as... ,in, Representation type k The computing resources required for the vehicle-mounted drone; constructing a set containing K contract items. ,in, , Payment to the central base station for type The reward for vehicle-mounted drones, when season ,when hour, ,in, This represents the energy consumption conversion factor. This indicates the increased cost compensation due to increased resource contribution. Indicates the first Vehicle-mounted drones for similar services are more efficient than Information rental income obtained from vehicle-mounted drone services.

[0032] Determine the service vehicle-mounted drone The process for determining the required amount of computing resources is as follows: make Indicated as a service vehicle-mounted drone The amount of resources required when serving vehicle-mounted drones Upon receiving the contract set broadcast by the central base station via roadside units, the service vehicle-mounted drone... Its own type To determine whether the individual rationality constraint and incentive compatibility constraint are satisfied, the expression for the individual rationality constraint is: The expression for the incentive compatibility constraint is: ,in The type is The utility of vehicle-mounted drones If the conditions are met, then select the contract. It becomes a node capable of unloading tasks and returns the contract item identifier to the central base station. At this point, i.e., vehicle-mounted drone service Committed computing resources Equal to the contract item .

[0033] In this embodiment, the central base station calculates the unit cost coefficient for all service vehicle-mounted drones based on the pre-recorded distribution data of the service vehicle-mounted drone energy consumption model. They are categorized into K types from highest to lowest, denoted as... ,in This represents the energy consumed per unit of computing resources, measured in joules per CPU cycle (J / cycle); correspondingly, the amount of computing resources in the contract is set to... ,in Representation type k The computing resources required for the vehicle-mounted drone are measured in CPU cycles; construct a set containing K contract items. ,in Pay the type of central base station The compensation for vehicle-mounted drones is in yuan.

[0034] Based on individual rationality constraints and incentive compatibility constraints, an optimal reward solution model is constructed, and the specific steps include: Step 1: Establish boundary conditions for the highest cost type. By setting its individual rationality constraint as a tight constraint and setting its utility to zero to minimize base station expenditure, we obtain the initial reward formula: ; Step 2: Simplify constraints, for type Using cost coefficients The monotonically decreasing property simplifies global excitation compatibility constraints to local adjacent excitation compatibility constraints, requiring type Select the corresponding contract for the service vehicle-mounted drone. Its utility is no less than that of its imitation of adjacent inefficient types. The utility of is expressed as: ; Step 3: Derive the recursive formula. By setting the local adjacent incentive compatibility constraint as an equation, the recursive relationship of rewards between adjacent types is derived. The calculation formula is as follows: ; Among them, energy consumption conversion factor The unit is yuan per joule (yuan / J). This indicates the increased cost compensation due to increased resource contribution. Indicates the first Vehicle-mounted drones for similar services are more efficient than The information rent obtained from vehicle-mounted drones serving different types of services is an additional incentive revenue that the central base station needs to provide to high-efficiency types of vehicle-mounted drones under conditions of information asymmetry, in order to ensure that different types of vehicle-mounted drones truthfully select contract items that match their type.

[0035] In this embodiment, K=3. Contract resource terms Energy consumption conversion factor ,type The utility function expression is: ; Let the lowest utility type Individual rationality constraints take tight constraints ,but: ; for The recursive formula for the reward is: ; when Sometimes, ; in, , ; Therefore, there is ; when Sometimes, ; in, , ; therefore, .

[0036] The result obtained by the above recursive formula It can guarantee that the type is Vehicle-mounted drones can only be used when selecting the corresponding Only when the time is right can the vehicle-mounted drone achieve its maximum utility, thus enabling it to automatically and accurately contribute resources.

[0037] In the aforementioned contract mechanism, the service vehicle-mounted drone does not need to explicitly upload its actual type parameters to the central base station. Instead, it calculates the utility of each contract item locally based on its own private parameters. The system selects the contract that yields the greatest benefit and signs it accordingly. During operation, the central base station only needs to know the selected contract number and does not need to know the... The specific values ​​enable the central base station to avoid directly collecting and storing sensitive parameters of the service vehicle-mounted drone while making service deployment and task offloading decisions.

[0038] Due to the heterogeneity of service vehicle-mounted drone resources, and to incentivize their participation in task offloading, all service vehicle-mounted drones are categorized into K types. This provides a basis for contract formulation even under information asymmetry. When a vehicle requests a task, the central base station creates a set of contracts based on the different tasks. These contracts are then broadcast to potential service vehicle-mounted drones via roadside units. Each drone, based on its type, assesses individual rationality and incentive compatibility, selects the most suitable contract, and joins the task offloading node set.

[0039] In an optional embodiment, the process of calculating the communication transmission delay of the task satisfies the expression: , ,

[0040] ,

[0041] , , ,

[0042] ,

[0043] , in, Indicates the vehicle number that requested the task. This represents the set of vehicles requesting the task, where the number of elements is... , This indicates the serial number of the vehicle-mounted drone used for service. This represents the set of service vehicle-mounted drones, where the number of elements is... , Indicates the number of the roadside unit, and and Representing different roadside units, Represents the set of roadside units, where the number of elements is... , where n represents the number of the deep neural network inference service model. This represents the set of deep neural network inference service models, where the number of elements is... ; Vehicle representing the task request With service vehicle-mounted drones Channel gain between Vehicle representing the task request With roadside units Channel gain between Represents roadside unit With service vehicle-mounted drones Channel gain between This indicates the gain at a reference distance of 1 meter. Request vehicle for mission With service vehicle-mounted drones Euclidean distance, Request vehicle for mission With roadside units Euclidean distance, roadside unit With service vehicle-mounted drones Euclidean distance, This is the path loss index. Indicates the vehicle requesting the task. Do you request a deep neural network inference service model? If requested, then ,otherwise , Request vehicle for mission Requesting a deep neural network inference service model The amount of data for the task; Request vehicle for mission With service vehicle-mounted drones Uplink transmission rate of direct links between them Indicates the vehicle requesting the task. With service vehicle-mounted drones Bandwidth between Indicates the vehicle requesting the task. Transmission power, Indicates service vehicle-mounted drone Environmental noise; Request vehicle for mission With roadside units Uplink transmission rate between Indicates the vehicle requesting the task. With roadside units Bandwidth between Indicates the vehicle requesting the task. Transmission power, Represents roadside unit Environmental noise; roadside unit With service vehicle-mounted drones Downlink transmission rate between Represents roadside unit With service vehicle-mounted drones Bandwidth between Represents roadside unit Transmission power, Indicates service vehicle-mounted drone Environmental noise; roadside unit With roadside units The transmission rate of the wired link between them; Request vehicle for mission With service vehicle-mounted drones Transmission delay when within the coverage area of ​​the same roadside unit; Request vehicle for mission With service vehicle-mounted drones When within the coverage area of ​​different roadside units, the indirect communication transmission delay via roadside unit relay is adopted; Request vehicle for mission With service vehicle-mounted drones The task transmission delay between them can be divided into two cases: when the task requests the vehicle... With service vehicle-mounted drones Within the coverage area of ​​the same roadside unit, the task transmission time is When the task requests the vehicle With service vehicle-mounted drones When within the coverage area of ​​different roadside units, the task transmission time is .

[0044] In this embodiment, This represents the gain at a reference distance of 1 meter, taken as -38.5 dB. This is the path loss index, with a value of 3.

[0045] Indicates the vehicle requesting the task. With service vehicle-mounted drones The bandwidth between them is 10 MHz. Indicates the vehicle requesting the task. The transmission power is set to 0.31W. The ambient noise level for the vehicle-mounted drone is represented by a value of 3.9 x 10. -21 W.

[0046] Indicates the vehicle requesting the task. With roadside units The bandwidth between them is 10 MHz. Indicates the vehicle requesting the task. The transmission power is set to 0.31W. Represents roadside unit The ambient noise level is 3.9 x 10⁻⁶. -21 W.

[0047] Represents roadside unit With service vehicle-mounted drones The bandwidth between them is 10 MHz. Represents roadside unit The transmission power is set to 1W. Indicates service vehicle-mounted drone The ambient noise level is 3.9 x 10⁻⁶. -21 W.

[0048] In an optional embodiment, the process of calculating the total delay of task completion satisfies the expression: ,

[0049]

[0050] in, Indicates the vehicle requesting the task. Deep Neural Network Inference Service Model The computational load, Indicates service vehicle-mounted drone Assigned to the vehicle requesting the task The computational resources for the task Indicates service vehicle-mounted drone Is the deep neural network inference service model cached? If it is cached ,otherwise , Indicate whether to request the vehicle for the mission. Task offloading to service vehicle-mounted drone If so, then ,otherwise , To serve vehicle-mounted drones Start the corresponding deep neural network inference service model The virtual machine, and the deep neural network inference service model Latency when deployed on a virtual machine; Request vehicle for mission The task is carried out by a service vehicle-mounted drone. Total delay of completion; To serve vehicle-mounted drones Calculate the vehicle request task The execution time of the task.

[0051] In this embodiment, when the vehicle and the service vehicle-mounted drone are within the same roadside unit coverage area, the vehicle and the service vehicle-mounted drone are close, the transmission rate is high, and the latency is low, so direct unloading is preferred; when they are within the coverage area of ​​adjacent roadside units, they need to be forwarded through intermediate roadside units, the transmission rate is low, and the latency is high, so the transmission latency and the service vehicle-mounted drone resources need to be weighed.

[0052] Each task has an actual completion time. and the maximum tolerable time threshold for execution on service vehicle-mounted drones ,when In other words, if the actual completion time does not exceed this threshold, the task will be offloaded to the service vehicle-mounted drone; otherwise, the task will be offloaded to the roadside unit. Only tasks that cannot be completed within the maximum tolerable delay time will be offloaded to the roadside unit and the service vehicle-mounted drone. The maximum limit on the amount of computing resources that can be transferred is the limit at the time the contract is signed. Assigned to different task request vehicles The total computing resources cannot exceed .

[0053] Since the amount of data returned after task processing is very small, it can be ignored. The total delay for the task requesting vehicle to complete task processing is... , Mission request vehicle Should the task be unloaded to the service vehicle drone? It depends on whether the total latency meets the constraints. Including the latency of virtual machine startup and model deployment in the total latency is more in line with actual edge computing scenarios.

[0054] Example 3 This embodiment further improves upon Embodiment 2 by providing a method for deploying and requesting offloading of vehicle-to-everything (V2X) edge network services. The process of constructing and optimizing the target model is as follows: The objective function is:

[0055] The constraints are: C1: ; C2: ; C3: ; C4: ; C5: ; C6: ； ; in, This represents the set of model cache decision strategies, containing binary decision variables. , This represents the set of task unloading decision strategies, containing binary decision variables. , , Indicates the first Task completion rate of deep neural network inference services Introducing auxiliary variables The minimum value used to characterize the task completion rate of each service type is set as the optimization objective to maximize the minimum value. At the same time, set For any service type All are true; C1 is the unique unloading constraint for the task, meaning that for any task requesting vehicle... The computational tasks it generates can at most be offloaded to a single service vehicle-mounted drone. Process it; C2 is a service model caching constraint, indicating that if a task requests a vehicle... Request No. This involves providing edge network services and offloading tasks to service-equipped drones. Then the drone The corresponding service model must have been cached. ; C3 represents the computing resource capacity constraint allocated to the service vehicle-mounted drone, indicating the capacity of resources allocated to it. The total computing resources required for all computing tasks of the drone must not exceed the total computing resources required for the drone. Maximum computing resource capacity willing to be given up ; This represents the computing resources allocated to a single task; C4 represents the task delay constraint, indicating the task request vehicle. Total latency of computation task completion It must be less than or equal to the maximum tolerable latency threshold for this task. ; C5 represents the individual reasonableness constraint, indicating the remuneration received by the service vehicle-mounted drone participating in the contract. It must be sufficient to cover the computing resources it provides. Energy costs generated ; C6 represents the incentive compatibility constraint, indicating the type as follows: Service vehicle-mounted drones: Select contract items of their type The utility gained must be greater than or equal to the utility gained from choosing other types. Contract terms The utility that can be obtained.

[0056] In this embodiment, As an auxiliary variable, represents the minimum computational task completion rate across all service types. The optimization objective aims to maximize this minimum value to ensure service fairness. C1 indicates that for any task requesting vehicle i, the computational task it generates can be offloaded to at most one service vehicle-mounted drone. Processing is performed to ensure the uniqueness of task processing. It is a binary variable; if the task requests a vehicle... Offload its computing tasks to a service vehicle-mounted drone C2 is 1 if the task requests a vehicle, otherwise it is 0; Request No. Edge network services ( And the task will be offloaded to the service vehicle-mounted drone. ( Then the drone The corresponding service model must have been cached. （ ） The constraints ensure the matching of tasks, models, and resources, avoiding invalid unloading; C3 indicates the allocation of the service vehicle-mounted drone. The total computing resources required for all computing tasks of the drone must not exceed the total computing resources required for the drone. Maximum computing resource capacity willing to be given up ; C4 represents the computing resources allocated to a single task; C4 represents the task request vehicle. Total latency of computation task completion It must be less than or equal to the maximum tolerable latency threshold for this task. C5 indicates the compensation received by the service provider for vehicle-mounted drones participating in the contract. It must be sufficient to cover the computing resources it provides. Energy costs generated To ensure their willingness to participate; C6 indicates type. Service vehicle-mounted drones: Select contract items of their type The utility gained must be greater than or equal to the utility gained from choosing other types. Contract terms The utility that can be obtained will incentivize drones to truthfully report their type.

[0057] In an optional embodiment, the optimization objective model solution process is transformed into a Markov decision process, specifically: First, the continuous operation process of the vehicle-to-everything (V2X) edge computing system is divided into several scheduling time slots, denoted as time slot index t; the state of the V2X edge computing system remains unchanged or approximately unchanged within each time slot, and a decision update is performed at the beginning of the time slot; second, the state space, action space, and reward function are determined, wherein: state space ,in, This represents the overall channel quality between all nodes in the vehicle-to-everything (V2X) edge computing system. , This represents the channel gain matrix between the requesting vehicle and the serving vehicle-mounted drone. Its elements Corresponding task request vehicle With service vehicle-mounted drones Channel gain, This represents the channel gain matrix between the requesting vehicle and the roadside unit. Its elements Corresponding task request vehicle With roadside units Channel gain, This represents the channel gain matrix between the roadside unit and the serving vehicle-mounted drone. Its elements Corresponding roadside unit With service vehicle-mounted drones Channel gain, This represents the vector of remaining available computing resources for all service vehicle-mounted drones at the current moment. ,in This indicates the remaining available computing resources for serving vehicle-mounted drones. , This represents the set of task request information sent by the task requesting vehicle. This includes the amount of task data, computational load, and maximum tolerable latency; Action space ,in, Its elements Indicates service vehicle-mounted drone Should the cached deep neural network inference service model be executed? The decision, Its elements Indicates the vehicle requesting the task. Should the task be unloaded to the service vehicle-mounted drone? Decisions made on the ground This represents the computing resource contribution requirements set by the central base station for K different types of drones. The agent adjusts this vector to determine the resource terms in the contract menu, and then automatically derives the corresponding optimal reward. ; reward function ,in, , To achieve the minimum task completion rate, , The latency violation penalty indicates that the total actual latency of all tasks exceeds the maximum tolerable latency. ; This is a resource overflow penalty, representing the portion of the total computing resources allocated to the task that exceeds the capacity of the service vehicle-mounted drone. ; These are the preset positive trade-off coefficients.

[0058] In this embodiment, in order to balance system utility and constraint penalties, the tradeoff coefficients in the reward function are set as follows: Set utility coefficient =100, setting the latency penalty coefficient. Set resource penalty coefficient Due to the task completion rate It is a decimal between [0,1], and the delay penalty and resource penalties The numerical magnitude may be large; to prevent the model from ignoring smaller values... Gain, by increasing This amplifies the utility to the same order of magnitude as the penalty. Presupposed positive trade-off coefficient. This is used to eliminate differences in the numerical magnitude of task completion rate, latency, and resource consumption, and to balance the optimization objective and constraints. The penalty coefficient is used in this context. The value of needs to be such that the penalty value generated when the constraint is violated is greater than the gain brought by the task completion rate, so as to guide the model to learn a strategy that satisfies the constraint.

[0059] The state space defines the current state of the system as perceived by the agent, and must cover all key factors affecting decision-making. The quality of the communication link between roadside units directly affects the efficiency of task transmission. This limits the drone's mission processing capabilities; The specific task requirements are clearly defined, reflecting the feasibility of the allocation. The action space defines the set of policies that the agent can adjust, and must cover the key dimensions of the optimization objective. It affected service startup latency and model availability. It affected the overall task latency and roadside unit load. The reward function determines the economic cost of the system and is a key decision variable for balancing task completion rate and incentive costs. It is a core indicator for measuring the quality of actions and is strongly correlated with optimization goals.

[0060] In an alternative embodiment, based on state space Action space and reward function The process of designing actor networks and critic networks separately, and using deep reinforcement learning to solve for and optimize the objective function to obtain the optimal service deployment and task unloading results is as follows: The actor network is initialized based on a preset probability distribution. parameters Critics Network parameters , where the parameters and Specifically, this includes the connection weights and bias terms of all neurons in the input, hidden, and output layers of their respective networks; and initializes the bias parameters to positive numbers; and sets the target actor network. parameters and the target critic network parameters The experience replay pool is initialized to a constant capacity and empty. Specifically, the experience replay pool is constructed as a fixed-capacity first-in-first-out circular buffer. For each time slot index t, the following steps are repeated until training converges: Based on the current state in the state space Use an actor network to select actions and add noise to the actions. The expression is: ,in The actor network indicates the current status. The output deterministic strategy This represents exploratory noise that follows a Gaussian distribution. Execution constraint verification and environment interaction steps: Verify the legality of the action: Determine the action's validity. Check if constraints C1 and C2 are met; if not, the current task is deemed a failure, and a preset invalid action penalty value is given as a reward. To keep the system state unchanged, that is Transfer data Store the data in the experience replay pool to train the agent to avoid outputting such actions; if the conditions are met, then execute the action in the environment. Feasibility verification: During execution, determine whether the system state satisfies constraints C3, C4, C5, and C6. If not, the task execution is deemed a failure, and a reward including a high penalty is calculated based on the reward function. and obtain the updated status. The data is stored in the experience replay pool. If the conditions are met, a positive reward based on the minimum task completion rate is calculated, and the next state is obtained. The data will be successfully transferred. Store in the experience replay pool; Multiple data points were randomly sampled from the experience replay pool. ; Update the critic network and calculate the target Q-value, expressed as: ,in, Using the discount factor, the commentator network loss is calculated as follows: ,in, To determine the batch size for empirical replay sampling, gradient descent is used to update the parameters of the critic network. The expression is: ,in, Indicates an update operation; Update the actor network and calculate the policy gradient of the actor network, expressed as: The actor network parameters are updated using gradient ascent, expressed as: ; The target network parameters are updated using a soft update method, expressed as follows: ,in For hyperparameters; Continue until training converges, obtaining the trained service deployment and request unloading model.

[0061] Based on the trained service deployment and request offloading model, the central base station issues control commands for service model caching and task offloading decisions according to the current status of the vehicle-to-everything (V2X) edge computing system and task requests. The service vehicle-mounted drone responds to the control commands and starts the corresponding deep neural network inference service model virtual machine in advance. The task requesting vehicle responds to the control commands and offloads the computing task to the designated service vehicle-mounted drone for processing.

[0062] In this embodiment, based on state space Action space and reward function Design separate actor networks and critic networks, see [link / reference] Figure 3The process of solving and optimizing the target model based on deep reinforcement learning to obtain the optimal service deployment and task unloading results is as follows: The neural network's structural parameters are set as follows: learning rate is 0.0003, discount factor is 0.99, experience replay pool size is 10000, soft update coefficient is 0.01, and batch size is 64.

[0063] The network is configured with I = 10 task requesting vehicles, M = 5 service vehicle-mounted drones, and U = 5 roadside units. After flattening and concatenating the matrices in the state space, the input layer dimension is set to 84. To ensure sufficient feature extraction capability and avoid overfitting, both the actor network and the critic network contain two hidden layers, with the number of neurons in each layer being... All values ​​are set to 256. The activation function used is ReLU, and the output layer dimension of the actor network is 3.

[0064] Initialize actor network parameters Critics Network parameters For the weight parameter W of each layer, its initial value starts from a mean of 0 and a variance of . Sampling is performed in a Gaussian distribution, where Set the number of input neurons for this layer. Initialize the bias parameters of all hidden and output layers to a small positive number (0.01) to ensure that neurons are active during the initial training phase. Configure the target actor network. parameters and the target critic network parameters Initialize the experience replay pool; for each time step t, where time step refers to a discrete decision time slot, repeat the following steps: Based on the current state in the state space Use an actor network to select actions and add noise to the actions. The expression is: ; Execution constraint verification and environment interaction steps: (1) Verify the legality of the action: judge the action Check if constraints C1 and C2 are met; if not, the current task is deemed a failure, and a preset invalid action penalty value is given as a reward. To keep the system state unchanged, that is Transfer data Store the data in the experience replay pool to train the agent to avoid outputting such actions; if the conditions are met, then execute the action in the environment. Get the updated status (2) Verify the feasibility of the state: During the execution process, determine whether the system state meets the constraints C3, C4, C5 and C6. If not, the task execution is deemed to have failed. Calculate the reward containing a high penalty item according to the reward function of Reward. and obtain the updated status. The data is stored in the experience replay pool. If the conditions are met, a positive reward based on the minimum task completion rate is calculated, and the next state is obtained. The data will be successfully transferred. Store in the experience replay pool; Multiple data points were randomly sampled from the experience replay pool. ; Update the critic network and calculate the target Q-value, expressed as: , The network loss of commenters is calculated using the following expression: Update the parameters of the critic network using gradient descent. The expression is: ; Update the actor network and calculate the policy gradient of the actor network, expressed as: The actor network parameters are updated using gradient ascent, expressed as: ; The target network parameters are updated using a soft update method, expressed as follows: ; The training continues until convergence, resulting in a trained service deployment and request offloading model. This model is then deployed at the central base station, and during the runtime phase, it is invoked in each scheduling time slot to generate service model caching decisions and task offloading decisions.

[0065] To verify the effectiveness and convergence of the method proposed in this invention, deep reinforcement learning training was performed based on the above simulation parameter settings. Experimental results are as follows: Figure 4 As shown, Figure 4 The graph shows the change in the system's average reward value as a function of the number of training episodes. As can be seen from the graph, in the early stages of training (approximately the first 10 episodes), the reward value fluctuates significantly and is relatively low because the agent is in the exploration phase, indicating that the agent is trying different resource allocation and unloading strategies through trial and error. As training progresses (approximately 10-20 episodes), the reward value shows a rapid upward trend, indicating that the actor network gradually learns to satisfy the constraints (C1-C6) and improve the objective function. An effective strategy for (minimum task completion rate) was found. After approximately 30 rounds of training, the reward curve stabilized and converged to a relatively high value (around 2000), without significant oscillations or declines. This fully demonstrates that the method proposed in this invention has good convergence speed and stability in complex vehicle-to-everything (V2X) edge computing environments, and can successfully find near-optimal service deployment and task offloading strategies while satisfying latency and resource constraints, thus verifying the effectiveness of the technical solution proposed in this embodiment.

[0066] Furthermore, to verify the performance advantages of this method under different system loads, multiple sets of comparative experiments were conducted. For example... Figure 5 As shown, Figure 5 This paper demonstrates how the minimum task completion rate changes as the number of vehicles requesting tasks increases. It shows that as the number of vehicles increases from 10 to 60, the task completion rate of all methods decreases due to increased channel congestion and competition for computing resources. However, our proposed method consistently outperforms greedy and random offloading strategies. Especially under high-load scenarios (e.g., with 60 vehicles), this invention maintains a completion rate of over 80%, while the comparative algorithms show a significant decline. This is attributed to the fact that this invention employs an objective function that maximizes the minimum task completion rate. This effectively ensures the task processing of weak nodes and improves the overall robustness of the system.

[0067] like Figure 6 As shown, Figure 6 The average latency comparison is shown for different task data sizes. Compared to fully local execution and greedy offloading, this method exhibits a significant low-latency advantage when handling large data volume tasks (such as 5MB). This is because the present invention combines contract theory and deep reinforcement learning. On the one hand, it uses contracts to select high-performance, computationally efficient services for vehicle-mounted drones; on the other hand, it uses a deep reinforcement learning agent to jointly optimize communication links and computing resources, thereby avoiding the allocation of heavy-load tasks to inefficient nodes or congested links.

[0068] The embodiments described are merely examples to clearly illustrate the present invention and are not intended to limit the implementation of the invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all possible implementations. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for deploying and requesting unloading edge network services in a vehicle-to-everything (V2X) network, characterized in that, Includes the following steps: A vehicle-to-everything (V2X) edge computing system is constructed, comprising a central base station, several roadside units, several service vehicle-mounted drones, several task-requesting vehicles, and various deep neural network inference service models. The task request vehicle generates computing tasks based on its own business needs and sends the task request information to the central base station; the task request information includes the identifier of the requested deep neural network inference service model, the amount of task data, the computing load, and the maximum tolerable latency; Based on the task request information, a contract set is formulated using the central base station, and the contract set is broadcast to potential service vehicle-mounted drones through roadside units. The potential service vehicle-mounted drones select contract items according to their own type and return the contract item identifiers to the central base station. The central base station determines the amount of computing resources committed to be contributed by each service vehicle-mounted drone based on the contract item identifiers. The contract signing results, service model cache status, and communication link status at the current moment are collected by the central base station. Combined with the task request information, the communication transmission delay of the task is calculated and the total task delay is obtained. With the objective function of maximizing the minimum task completion rate of various services, and with constraints such as unique task unloading constraint, service model caching constraint, service vehicle-mounted drone computing resource capacity constraint, task latency constraint, and contract individual rationality and incentive compatibility constraint, an optimization objective model is constructed. The optimization objective model is transformed into a Markov decision process. Based on deep reinforcement learning, a trained service deployment and request offloading model is obtained, and the service model caching decision and task offloading decision are output.

2. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 1, characterized in that, The vehicle-to-everything (V2X) edge computing system includes a central base station and U roadside units, wherein... It includes I task requesting vehicles and M service vehicle-mounted drones, among which , It also includes N different deep neural network inference service models, among which The N different deep neural network inference service models are used to provide different edge AI services; The roadside units are connected to the central base station via wired links, and the roadside units are connected to each other via wired links. The task requesting vehicle is connected to the roadside unit, the task requesting vehicle is connected to the service vehicle-mounted drone, and the service vehicle-mounted drone is connected to the roadside unit via wireless links. The service vehicle-mounted drone is pre-configured with multiple deep neural network inference service model virtual machines.

3. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 1, characterized in that, The process of developing a contract set includes: based on the pre-entered data distribution of energy consumption models for service vehicle-mounted drones, using the central base station to calculate the unit cost coefficient for the service vehicle-mounted drones. They are categorized into K types from highest to lowest, denoted as... ,in, Indicates the first The unit computational cost coefficient for each type of service vehicle-mounted drone; setting the computational resource quantity in the contract as... ,in, Representation type k The computing resources required for the vehicle-mounted drone; constructing a set containing K contract items. ,in, , Payment to the central base station for type The reward for vehicle-mounted drones, when season ,when hour, ,in, This represents the energy consumption conversion factor. This indicates the increased cost compensation due to increased resource contribution. Indicates the first Vehicle-mounted drones for similar services are more efficient than Information rental income obtained from vehicle-mounted drone services.

4. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 3, characterized in that, Determine the service vehicle-mounted drone The process for determining the required amount of computing resources is as follows: make Indicated as a service vehicle-mounted drone The amount of resources required when serving vehicle-mounted drones Upon receiving the contract set broadcast by the central base station via roadside units, the service vehicle-mounted drone... Its own type To determine whether the individual rationality constraint and incentive compatibility constraint are satisfied, the expression for the individual rationality constraint is: The expression for the incentive compatibility constraint is: ,in The type is The utility of vehicle-mounted drones If the conditions are met, then select the contract. It becomes a node capable of unloading tasks and returns the contract item identifier to the central base station. At this point, i.e., vehicle-mounted drone service Committed computing resources Equal to the contract item .

5. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 1, characterized in that, The process of calculating the communication transmission delay of a computation task satisfies the expression: , , , , , , , , in, Indicates the vehicle number that requested the task. This represents the set of vehicles requesting the task, where the number of elements is... , This indicates the serial number of the vehicle-mounted drone used for service. This represents the set of service vehicle-mounted drones, where the number of elements is... , Indicates the number of the roadside unit, and and Representing different roadside units, Represents the set of roadside units, where the number of elements is... , where n represents the number of the deep neural network inference service model. This represents the set of deep neural network inference service models, where the number of elements is... ; Vehicle representing the task request With service vehicle-mounted drones Channel gain between Vehicle representing the task request With roadside units Channel gain between Represents roadside unit With service vehicle-mounted drones Channel gain between This indicates the gain at a reference distance of 1 meter. Request vehicle for mission With service vehicle-mounted drones Euclidean distance, Request vehicle for mission With roadside units Euclidean distance, roadside unit With service vehicle-mounted drones Euclidean distance, This is the path loss index. Indicates the vehicle requesting the task. Do you request a deep neural network inference service model? If requested, then ,otherwise , Request vehicle for mission Requesting a deep neural network inference service model The amount of data for the task; Request vehicle for mission With service vehicle-mounted drones Uplink transmission rate of the direct link between them Indicates the vehicle requesting the task. With service vehicle-mounted drones Bandwidth between Indicates the vehicle requesting the task. Transmission power, Indicates service vehicle-mounted drone Environmental noise; Request vehicle for mission With roadside units Uplink transmission rate between Indicates the vehicle requesting the task. With roadside units Bandwidth between Indicates the vehicle requesting the task. Transmission power, Represents roadside unit Environmental noise; roadside unit With service vehicle-mounted drones Downlink transmission rate between Represents roadside unit With service vehicle-mounted drones Bandwidth between Represents roadside unit Transmission power, Indicates service vehicle-mounted drone Environmental noise; roadside unit With roadside units The transmission rate of the wired link between them; Request vehicle for mission With service vehicle-mounted drones Transmission delay when within the coverage area of ​​the same roadside unit; Request vehicle for mission With service vehicle-mounted drones When within the coverage area of ​​different roadside units, the indirect communication transmission delay via roadside unit relay is used; Request vehicle for mission With service vehicle-mounted drones The task transmission delay between them can be divided into two cases: when the task requests the vehicle... With service vehicle-mounted drones Within the coverage area of ​​the same roadside unit, the task transmission time is When the task requests the vehicle With service vehicle-mounted drones When within the coverage area of ​​different roadside units, the task transmission time is .

6. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 5, characterized in that, The process of calculating the total delay of task completion satisfies the expression: , in, Indicates the vehicle requesting the task. Deep Neural Network Inference Service Model The computational load, Indicates service vehicle-mounted drone Assigned to the vehicle requesting the task The computational resources for the task Indicates service vehicle-mounted drone Is the deep neural network inference service model cached? If it is cached ,otherwise , Indicate whether to request the vehicle for the mission. Task offloading to service vehicle-mounted drone If so, then ,otherwise , To serve vehicle-mounted drones Start the corresponding deep neural network inference service model The virtual machine, and the deep neural network inference service model Latency when deployed on a virtual machine; Request vehicle for mission The task is carried out by a service vehicle-mounted drone. Total delay of completion; To serve vehicle-mounted drones Calculate the vehicle request task The execution time of the task.

7. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 6, characterized in that, The process of constructing the optimization target model is as follows: The objective function is: The constraints are: C1: ; C2: ； C3: ； C4: ； C5: ； C6: ； ； in, This represents the set of model cache decision strategies, containing binary decision variables. , This represents the set of task unloading decision strategies, containing binary decision variables. , , Indicates the first Task completion rate of deep neural network inference services Introducing auxiliary variables The minimum value used to characterize the task completion rate of each service type is set as the optimization objective to maximize the minimum value. At the same time, set For any service type All are true; C1 is the unique unloading constraint for the task, meaning that for any task requesting vehicle... The computational tasks it generates can at most be offloaded to a single service vehicle-mounted drone. Process it; C2 is a service model caching constraint, indicating that if a task requests a vehicle... Request No. This involves providing edge network services and offloading tasks to service-equipped drones. Then the drone The corresponding service model must have been cached. ; C3 represents the computing resource capacity constraint allocated to the service vehicle-mounted drone, indicating the capacity of resources allocated to it. The total computing resources required for all computing tasks of the drone must not exceed the total computing resources required for the drone. Maximum computing resource capacity willing to be given up ; This represents the computing resources allocated to a single task; C4 represents the task delay constraint, indicating the task request vehicle. Total latency of computation task completion It must be less than or equal to the maximum tolerable latency threshold for this task. ; C5 represents the individual reasonableness constraint, indicating the remuneration received by the service vehicle-mounted drone participating in the contract. It must be sufficient to cover the computing resources it provides. Energy costs generated ; C6 represents the incentive compatibility constraint, indicating the type as follows: Service vehicle-mounted drones: Select contract items of their type The utility gained must be greater than or equal to the utility gained from choosing other types. Contract terms The utility that can be obtained.

8. The method for deploying and requesting unloading edge network services for vehicle networking according to claim 7, characterized in that, The optimization objective model solution process is transformed into a Markov decision process, specifically: First, the continuous operation of the vehicle-to-everything (V2X) edge computing system is divided into several scheduling time slots, denoted as time slot index t; the state of the V2X edge computing system remains unchanged or approximately unchanged within each time slot, and a decision update is performed at the beginning of the time slot; second, the state space, action space, and reward function are determined, wherein: state space ,in, This represents the overall channel quality between all nodes in the vehicle-to-everything (V2X) edge computing system. , This represents the channel gain matrix between the requesting vehicle and the serving vehicle-mounted drone. Its elements Corresponding task request vehicle With service vehicle-mounted drones Channel gain, This represents the channel gain matrix between the requesting vehicle and the roadside unit. Its elements Corresponding task request vehicle With roadside units Channel gain, This represents the channel gain matrix between the roadside unit and the serving vehicle-mounted drone. Its elements Corresponding roadside unit With service vehicle-mounted drones Channel gain, This represents the vector of remaining available computing resources for all service vehicle-mounted drones at the current moment. ,in This indicates the remaining available computing resources for serving vehicle-mounted drones. , This represents the set of task request information sent by the task requesting vehicle. This includes the amount of task data, computational load, and maximum tolerable latency; Action space ,in, Its elements Indicates service vehicle-mounted drone Should the cached deep neural network inference service model be executed? The decision, Its elements Indicates the vehicle requesting the task. Should the task be unloaded to the service vehicle-mounted drone? Decisions made on the ground This represents the computing resource contribution requirements set by the central base station for K different types of drones. The agent adjusts this vector to determine the resource terms in the contract menu, and then automatically derives the corresponding optimal reward. ; reward function ,in, , The minimum task completion rate, , The latency violation penalty indicates that the total actual latency of all tasks exceeds the maximum tolerable latency. ; This is a resource overflow penalty, representing the portion of the total computing resources allocated to the task that exceeds the capacity of the service vehicle-mounted drone. ; These are the preset positive trade-off coefficients.

9. A method for deploying and requesting unloading edge network services for vehicle networking according to claim 8, characterized in that, Based on the state space Action space and reward function The process of designing actor networks and critic networks separately, and solving the optimization objective function based on deep reinforcement learning to obtain the optimal service deployment and task unloading results is as follows: The actor network is initialized based on a preset probability distribution. parameters Critics Network parameters , wherein the parameters and Specifically, this includes the connection weights and bias terms of all neurons in the input, hidden, and output layers of their respective networks; and initializes the bias parameters to positive numbers; and sets the target actor network. parameters and the target critic network parameters The experience replay pool is initialized to a constant capacity and empty. Specifically, the experience replay pool is constructed as a fixed-capacity first-in-first-out circular buffer. For each time slot index t, the following steps are repeated until training converges: Based on the current state in the state space Use an actor network to select actions and add noise to the actions. The expression is: ,in The actor network indicates the current status. The output deterministic strategy This represents exploratory noise that follows a Gaussian distribution. Execution constraint verification and environment interaction steps: Verify the legality of the action: Determine the action's validity. Check if constraints C1 and C2 are met; if not, the current task is deemed a failure, and a preset invalid action penalty value is given as a reward. To keep the system state unchanged, that is Transfer data Store the data in the experience replay pool to train the agent to avoid outputting such actions; if the conditions are met, then execute the action in the environment. Feasibility verification: During execution, determine whether the system state satisfies the constraints C3, C4, C5, and C6. If not, the task execution is deemed to have failed, and a reward including a high penalty is calculated based on the reward function. and obtain the updated status. The data is stored in the experience replay pool. If the conditions are met, a positive reward based on the minimum task completion rate is calculated, and the next state is obtained. The data will be successfully transferred. Store in the experience replay pool; Multiple data points were randomly sampled from the experience replay pool. ; Update the critic network and calculate the target Q-value, expressed as: ,in, Using the discount factor, the commentator network loss is calculated as follows: ,in, To determine the batch size for empirical replay sampling, gradient descent is used to update the parameters of the critic network. The expression is: ,in, Indicates an update operation; Update the actor network and calculate the policy gradient of the actor network, expressed as: The actor network parameters are updated using gradient ascent, expressed as: ; The target network parameters are updated using a soft update method, expressed as follows: ,in For hyperparameters; Continue training until convergence, obtaining the trained service deployment and request offloading model.

10. A method for deploying and requesting unloading edge network services for vehicle networking according to claim 9, characterized in that, Based on the trained service deployment and request offloading model, the central base station issues control commands for service model caching and task offloading decisions according to the current status of the vehicle-to-everything (V2X) edge computing system and task requests. The service vehicle-mounted drone responds to the control commands and starts the corresponding deep neural network inference service model virtual machine in advance. The task requesting vehicle responds to the control commands and offloads the computing task to the designated service vehicle-mounted drone for processing.

Images