Vehicle-Road-Cloud Collaborative Task Offloading and Task Queue Stability Control System and Method

Abstract

This invention discloses a vehicle-road-cloud collaborative task offloading and task queue stability control system and method, comprising: a vehicle task manager, which is used to acquire task information generated by the task vehicle at the current moment, and update the local task queue status at the current moment according to the task offloading decision at the previous moment; a vehicle task scheduler, which is used to optimize the offloading strategy according to the acquired channel power gain between the task vehicle and the server, the local task queue status, and the server task queue status at the current moment; and a server manager, which has: a resource allocation update unit, which is used to allocate computing resources to the corresponding task vehicle through the server according to the task offloading decision in the optimized offloading strategy, and transmit the updated server task queue status and computing resource allocation of the task vehicle at the current moment, as well as the channel power gain between the vehicle and the server, to the vehicle task scheduler.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a task offloading and task queue stability control system and method based on vehicle-road-cloud collaboration in urban traffic scenarios. Background Technology

[0002] The development of intelligent connected vehicle technology has significantly improved the intelligence level of intelligent transportation systems. Coupled with the rise of mobile edge computing technology, edge servers can provide reliable computing resources for intelligent connected vehicles, effectively improving their computing performance and ensuring the safe and stable operation of these vehicles. However, as the scale of intelligent connected vehicles continues to increase, the limited computing resources of edge servers may not be able to meet the dense service requests, leading to a decline in the service quality of vehicles. This invention provides a method and system for task offloading and task queue stability control based on vehicle-road-cloud collaboration in urban traffic scenarios. This method can fully utilize the idle resources of vehicles in the system, alleviate the computing pressure on edge servers, and the proposed three-stage solution framework can effectively reduce the solution complexity and achieve long and short-term task queue stability control, ensuring stable system operation.

[0003] Currently, there are existing technologies and patents related to edge computing. Patent CN112148380A discloses a resource optimization method and electronic device for mobile edge computing task offloading. It constructs queue stability and network resource overhead indices with the goal of minimizing the task queue length of user terminals and edge servers, and builds an optimization model with the goal of minimizing network resource overhead. It introduces the Lagrange multiplier method and uses the momentum stochastic gradient descent algorithm for online solution to reduce task queue backlog. Patent CN113064665B discloses a multi-server computing offloading method based on Lyapunov optimization, constructing a problem to minimize the average execution cost of mobile devices, and using Lyapunov... The method eliminates energy causal constraints and uses the Lagrange dual method to obtain the optimal solution for unloading decisions and resource allocation strategies. Patent CN112860409A discloses a mobile cloud computing random task sequence scheduling method based on Lyapunov optimization, which constructs a joint optimization model of total user-end energy consumption and total latency, and obtains the optimal execution scheduling strategy based on the Lyapunov optimization method. However, the optimization models for minimizing user overhead (network overhead, latency, and energy consumption, etc.) constructed by the above technologies are mostly non-convex optimization problems that cannot be solved directly. Existing methods solve them by transforming them into convex optimization problems, so their solution efficiency depends on the complexity of the original problem. Summary of the Invention

[0004] The purpose of this invention is to provide a task offloading and task queue stability control system and method based on vehicle-road-cloud collaboration in urban traffic scenarios, which can solve the problem of high solution complexity of existing task offloading scheme optimization models.

[0005] To achieve the above objectives, the present invention provides a vehicle-road-cloud collaborative task offloading and task queue stability control system, characterized in that it includes:

[0006] The vehicle task manager is used to obtain task information generated by task vehicle k at the current time t, and update the local task queue status at the current time according to the task unloading decision of the previous time.

[0007] The vehicle task scheduler is used to determine the channel power gain G between task vehicle k and server m at the current time t. k,m (t) Local task queue status and server task queue status Optimize the uninstallation strategy;

[0008] The server manager has the following features:

[0009] The resource allocation update unit is used to update the task unloading decision x in the optimized unloading strategy. k,m,t The server allocates computing resources to the corresponding task vehicle and transmits the updated server task queue status of task vehicle k at the current time t to the vehicle task scheduler. and computing resource allocation and the channel power gain G between vehicle k and server m k,m (t).

[0010] Furthermore, the server manager also features:

[0011] The trajectory prediction unit is used to predict the future driving trajectory of a vehicle by collecting historical trajectory information of the vehicle.

[0012] The vehicle cloud building unit is used to cluster similar trajectories centered on the mission vehicle, build a vehicle cloud platform based on existing virtualization technology, and aggregate idle vehicle computing resources to build a vehicle cloud virtual computing unit to provide computing resources for the mission vehicle.

[0013] Furthermore, the vehicle task manager updates the local task queue state at the current time t using the following formula (1). The methods specifically include:

[0014]

[0015] in:

[0016]

[0017]

[0018] In the formula, Ak (t) represents the amount of tasks to be processed by task vehicle k at time t, and τ is the length of each time step. Computing resources allocated to the vehicle's local computing unit, ε k x is the CPU computation cycles required for task vehicle k to process each byte of task. k,m,t R represents the task unloading decision made by vehicle k at time t. k,m (t) represents the communication rate between the mission vehicle k and the server m at time t.

[0019]

[0020] In the formula, B k,m The channel bandwidth allocated by server m to task vehicle k, p k,m (t) represents the transmission power between the mission vehicle k and the server m at the current time t, σ 2 This represents the background noise power of the receiver.

[0021] Furthermore, if x k,m,t =1, then task vehicle k chooses to unload the task to the m-th server; if x = 1, then task vehicle k chooses to unload the task to the m-th server; k,m,t =0, then task vehicle k did not choose to unload the task to the m-th server.

[0022] Furthermore, the resource allocation update unit specifically includes:

[0023] The resource allocation subunit is used to allocate resources based on the optimized x. k,m,t The server's computing unit allocates computing resources to the corresponding task vehicles.

[0024] The resource update subunit is used by the computing unit of server m to allocate computing resources to task vehicle k. Get the amount of tasks processed at the current time t. Represented as:

[0025] In the formula, Let m be the task queue state of server m at time t, and its dynamic changes can be represented as follows:

[0026] Furthermore, the vehicle task scheduler has:

[0027] The unloading strategy optimization unit is used to optimize the unloading strategy according to G. k,m (t), and Set the task throughput at each moment to The uninstallation strategy is optimized using the pre-set task uninstallation optimization model P1.

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] In the formula, K represents the total number of mission vehicles, and T represents the total time step. This represents the maximum length of the task queue. The maximum energy consumption of mission vehicle k. The maximum CPU computation cycle for task vehicle k. The maximum transmission power of mission vehicle k. The maximum CPU computation cycle for server m.

[0036] 7. The vehicle-road-cloud collaborative task offloading and task queue stability control system as described in claim 6, characterized in that the offloading strategy optimization unit specifically includes:

[0037] The first computational subunit is used to decompose the original problem corresponding to the task unloading optimization model P1 into continuous-time subproblems using the Lyapunov optimization method.

[0038] The second computational subunit is used to decompose the continuous-time subproblem into the following three stages for solution:

[0039] In the first stage, based on the energy consumption constraint C2 and local computing resource constraint C3 of the mission vehicle at each moment, local computing resources are allocated, and it is determined whether the current task queue meets the queue length constraint. If yes, the next moment is entered; otherwise, the second stage is entered, and the mission vehicle needs to unload the task for computation.

[0040] In the second stage, the task vehicle acts as the intelligent agent. Based on the current task queue status and communication status, the vehicle intelligent agent learns the task unloading decision using deep reinforcement learning methods. After the task vehicle makes the unloading decision, it enters the third stage.

[0041] The third stage involves optimizing the task transmission power and server computing resource allocation strategies.

[0042] This invention also provides a method for unloading vehicle-road-cloud collaborative tasks and controlling the stability of the task queue, which includes:

[0043] Step 1: Obtain the task information generated by task vehicle k at the current time t, and update the local task queue status at the current time according to the task unloading decision of the previous time.

[0044] Step 2, based on the obtained channel power gain G between task vehicle k and server m at the current time t. k,m (t) Local task queue status and server task queue status Optimize the uninstallation strategy;

[0045] Step 3, based on the task uninstallation decision x in the optimized uninstallation strategy. k,m,t The server allocates computing resources to the corresponding task vehicle and transmits the updated server task queue status of task vehicle k at the current time t to the vehicle task scheduler. and computing resource allocation and the channel power gain G between vehicle k and server m k,m (t); where the server includes an edge server and a vehicle cloud platform, and the methods for obtaining the vehicle cloud platform include:

[0046] By collecting historical vehicle trajectory information, the future driving trajectory of the vehicle is predicted. Then, similar trajectories are clustered around the mission vehicle. A vehicle cloud platform is built based on existing virtualization technology, and idle vehicle computing resources are aggregated to build a vehicle cloud virtual computing unit to provide computing resources for the mission vehicle.

[0047] Furthermore, in step 2, the local task queue state is updated at the current time t using the following formula (1). The methods specifically include:

[0048]

[0049] in:

[0050]

[0051]

[0052] In the formula, A k (t) represents the amount of tasks to be processed by task vehicle k at time t, and τ is the length of each time step. The computing resources allocated to the local computing unit of the mission vehicle, ε k x is the CPU computation cycles required for task vehicle k to process each byte of task. k,m,t Let x be the task unloading decision made by task vehicle k at time t. k,m,t =1, then task vehicle k chooses to unload the task to the m-th server; if x = 1, then task vehicle k chooses to unload the task to the m-th server; k,m,t=0, then task vehicle k did not choose to unload the task to the m-th server, R k,m (t) represents the communication rate between the mission vehicle k and the server m at time t.

[0053]

[0054] In the formula, B k,m The channel bandwidth allocated by server m to task vehicle k, p k,m (t) represents the transmission power between the mission vehicle k and the server m at the current time t, σ 2 This represents the background noise power of the receiver.

[0055] Furthermore, in step 2, according to G k,m (t), and Set the task throughput at each moment to The uninstallation strategy is optimized using the pre-set task uninstallation optimization model P1.

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] In the formula, K represents the total number of mission vehicles. This represents the maximum length of the task queue. The maximum energy consumption of mission vehicle k. The maximum CPU computation cycle for task vehicle k. The maximum transmission power of mission vehicle k. The maximum CPU computation cycle for server m.

[0064] The present invention has the following advantages due to the adoption of the above technical solutions:

[0065] (1) Compared to offloading tasks to edge servers or nearby vehicles, this invention uses deep neural networks to predict vehicle trajectories and uses the task-unloading vehicles as cluster centers. Based on existing virtualization technology, it integrates idle vehicles with similar driving trajectories and builds a vehicle cloud platform. It aggregates idle vehicle resources to build a vehicle cloud virtual computing unit, providing vehicles with abundant computing resources. This can alleviate the problem of insufficient edge server computing resources in large-scale traffic scenarios.

[0066] (2) Compared with traditional convex optimization methods, this invention uses Lyapunov optimization method and deep reinforcement learning method to build a lightweight solution framework, which can be solved directly without transforming the original non-convex optimization problem, significantly reducing computational complexity. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the structure of the vehicle-road-cloud collaborative task unloading and task queue stability control system provided in an embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of the three-stage solution framework provided in an embodiment of the present invention. Detailed Implementation

[0069] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0070] like Figure 1 As shown, in an urban traffic scenario, base stations equipped with edge servers are evenly distributed on one side of the road. Vehicles with task offloading needs are designated as task vehicles in this embodiment, and vehicles with idle computing resources and willing to share their resources are designated as resource vehicles in this embodiment.

[0071] The vehicle task scheduler collects current system status information from the vehicle task manager and server manager, and generates an unloading policy. The specific process is as follows. It should be noted that the following steps do not strictly need to be performed in the order described below. Those skilled in the art can achieve the objectives of this invention by appropriately adjusting the order of the following steps, and all such adjustments are within the scope of protection of this invention.

[0072] like Figure 1 As shown, the vehicle-road-cloud collaborative task offloading and task queue stability control system provided in this embodiment of the invention includes a vehicle task manager, a vehicle task scheduler, and a server manager, wherein:

[0073] Both the vehicle task manager and the vehicle task scheduler are located on the vehicles. The set of task vehicles is represented as K = {1, 2, ..., k, ..., K}, meaning there are a total of K task vehicles on the road, where k is the vehicle number and task vehicle k represents the k-th task vehicle. The set of servers is represented as M = {1, 2, ..., m, ..., M}, meaning there are a total of M servers on the road, where m is the server number and server m represents the m-th server. Assuming that at each time t, the task vehicles randomly generate latency-sensitive tasks, and to ensure timely processing of computational tasks, it is assumed that tasks can be partitioned and processed in parallel. Therefore, collaborative computation of tasks can be achieved on vehicles, roadside edge servers, or the vehicle cloud.

[0074] The vehicle task manager is used to obtain task information generated by task vehicle k at the current time t, and to update the local task queue status at the current time based on the task unloading decision of the previous time.

[0075] The vehicle task scheduler is used to obtain the local task queue status of task vehicle k at the current time t through the vehicle task manager. And the constraint information of the mission vehicle, and based on the channel power gain G between mission vehicle k and server m at the current time t. k,m (t) Local task queue status and server task queue status The unloading strategy is optimized. Here, l represents the marker symbol indicating vehicle parameter information, and s represents the marker symbol indicating server parameter information. The unloading strategy includes task unloading decisions x. k,m,t Local resource allocation Transmission power p k,m (t). Vehicle constraint information includes the vehicle's maximum CPU computation cycle. Task queue length constraint Energy Consumption Constraints and maximum transmission power

[0076] Preferably, the vehicle task scheduler is used to obtain the local task queue status of task vehicle k at the current time t through the vehicle task manager. Methods for obtaining constraint information for the mission vehicle include:

[0077] First, the vehicle task manager obtains the task information generated by task vehicle k at the current time t.

[0078] In this embodiment, the task information generated by task vehicle k at the current time t is represented by a binary array (A). k (t),ε k Describe A k (t) represents the amount of tasks to be processed by task vehicle k at time t, in bits. ε kThe CPU computation cycles required for task k to process each byte of task are expressed in cycles / bit.

[0079] Then, the vehicle task manager obtains the local task queue status of task vehicle k at the current time t based on the computation task. And the constraint information of the mission vehicle.

[0080] The server manager includes a resource allocation update unit, which is used to update task uninstallation decisions based on the optimized uninstallation strategy. k,m,t The server allocates computing resources to the corresponding task vehicle and transmits the updated server task queue status of task vehicle k at the current time t to the vehicle task scheduler. and computing resource allocation and the channel power gain G between vehicle k and server m k,m (t).

[0081] In one embodiment, the server manager also includes a trajectory prediction unit and a vehicle cloud construction unit, wherein:

[0082] The trajectory prediction unit is used to predict the future driving trajectory of a vehicle by collecting historical trajectory information of the vehicle.

[0083] The vehicle cloud building unit is used to cluster similar trajectories centered on the mission vehicle, build a vehicle cloud platform based on existing virtualization technology, and aggregate idle vehicle computing resources to build a vehicle cloud virtual computing unit to provide computing resources for the mission vehicle.

[0084] This embodiment leverages the inherent advantages of edge servers, such as their geographically dispersed nature and sustainable energy supply. For example, it utilizes time-series vehicle historical trajectory information within their coverage area. Vehicle trajectories are predicted using deep learning models like Long Short-Term Memory (LSTM) networks or Transformer methods pre-deployed within the edge servers. Then, existing clustering algorithms, such as K-means or DBSCAN (Density-Based Spatial Clustering of Applications with Noise), are used to cluster vehicles with similar driving trajectories. This allows for the aggregation of idle computing resources from multiple individual vehicles based on existing virtualization technology, thus constructing a vehicle cloud platform.

[0085] Both the vehicle cloud platform and the edge server are used to provide cloud computing services for the mission vehicle. Therefore, for ease of description, they will be referred to as servers in other parts of this article.

[0086] In one embodiment, the vehicle task manager updates the local task queue state at the current time t using the following formula (1). The methods specifically include:

[0087]

[0088] in:

[0089]

[0090]

[0091] In the formula, A k (t) represents the amount of tasks to be processed by task vehicle k at time t, and τ is the length of each time step. To calculate the workload locally, The upper bound of the computing resources allocated to the vehicle's local computing unit is: ε k x is the CPU computation cycles required for task vehicle k to process each byte of task. k,m,t The task unloading decision made by task vehicle k at time t. R represents the amount of task unloaded by the task transfer unit of task vehicle k at time t. k,m (t) represents the communication rate between the mission vehicle k and the server m at time t.

[0092] According to the channel power gain G k,m (t) and Shannon's theorem, we can obtain the following equation (2):

[0093]

[0094] In the formula, B k,m The channel bandwidth allocated by server m to task vehicle k, p k,m (t) represents the transmission power between mission vehicle k and server m at the current time t, and its upper bound is... σ 2 This represents the background noise power of the receiver.

[0095] Equation (1) in this embodiment describes the dynamic change process of the vehicle's local task queue at time t, from the arrival of the task at A. k (t) and the task departs Jointly maintained, its upper limit is Therefore, unlike existing task offloading schemes that ensure task execution reliability through task latency constraints, this embodiment of the invention can ensure task execution reliability by dynamically adjusting the task queue length constraint. For example, if This indicates that the task needs to be processed in real time, and a backlog in the vehicle task queue is not allowed; if This indicates that the vehicle task queue is allowed to back up, and the tasks do not need to be processed in real time.

[0096] In one embodiment, if x k,m,t =1, then task vehicle k chooses to unload the task to the m-th server; if x = 1, then task vehicle k chooses to unload the task to the m-th server; k,m,t =0, then task vehicle k did not choose to unload the task to the m-th server.

[0097] In one embodiment, the resource allocation update unit specifically includes a resource allocation subunit and a resource update subunit, wherein:

[0098] The resource allocation subunit is used to allocate resources based on the optimized x. k,m,t The server's computing unit allocates computing resources to the corresponding task vehicles.

[0099] The resource update subunit is used by the computing unit of server m to allocate computing resources to task vehicle k. Get the amount of tasks processed at the current time t. Represented as: in, The computing resources allocated to vehicle k by the computing unit of the m-th server are as follows: Since each server can provide computing resources for multiple task vehicles, to ensure fairness in the allocation of computing resources, server m allocates resources based on the proportion of tasks unloaded from task vehicle k to the total number of tasks in the server's task queue. The upper bound of this allocation is... Let m be the task queue state of server m at time t, and its dynamic changes can be represented as follows: Similar to the local task queue, this task queue is formed by the arrival of tasks. Leave the mission Joint maintenance. Simultaneously, server m obtains the channel power gain G based on the current time t and the time-varying distance between task vehicle k and it. k,m (t) is transmitted to the vehicle task scheduler, and the channel power gain G k,m (t) The calculation method can be found in the method described in patent CN111741438A.

[0100] In one embodiment, the vehicle task scheduler has an offloading strategy optimization unit, which is used to optimize the offloading strategy according to G. k,m (t), and Set the task throughput at each moment to The uninstallation strategy is optimized using the pre-set task uninstallation optimization model P1.

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] In the formula, K represents the total number of mission vehicles, and T represents the total time step. This represents the maximum length of the task queue. The maximum energy consumption of mission vehicle k, and the energy consumption of mission vehicle k. Set to: ξ represents the energy consumption coefficient related to automotive chip architecture. The maximum CPU computation cycle for task vehicle k. The maximum transmission power of mission vehicle k. The maximum CPU computation cycle for server m.

[0109] Constraint C1 indicates that the current task queue length of the vehicle must meet the queue length constraint. Note that this embodiment achieves real-time task computation by adding constraint C1; that is, the smaller the task queue length constraint value, the lower the task computation latency. Therefore, the same function can also be achieved through the task computation latency constraint used in existing methods. Constraint C2 indicates that the energy consumption of local task processing at time t must meet the energy consumption constraint. Constraint C3 indicates that the local computing resources allocated at time t do not exceed the vehicle's maximum computing resource constraint. Constraint C4 indicates that vehicle k can only select one server for task offloading at time t. Constraint C5 indicates that the vehicle's task transmission power must meet the maximum power constraint. Constraint C6 indicates that the computing resources allocated by server m to all vehicles at time t do not exceed its maximum computing resources.

[0110] As can be seen from the task offloading optimization model P1, the vehicle's task throughput is related to the allocation of the vehicle's local computing resources. and transmission power p k,m Since all values ​​(t) are directly proportional, the optimal solution for resource allocation and transmission power can be directly obtained through constraints. Existing methods can also convert task throughput into task computation rate, i.e., set the ratio of task throughput to time length τ at each moment.

[0111] In one embodiment, analysis of the above task unloading optimization model P1 reveals that, due to coupling between multiple optimization variables and between consecutive time slots, the original problem corresponding to task unloading optimization model P1 is a non-convex optimization problem. To solve this optimization problem, the unloading strategy optimization unit specifically includes a first computational subunit and a second computational subunit, wherein:

[0112] The first computational subunit is used to decompose the original problem corresponding to the task unloading optimization model P1 into continuous-time subproblems using the Lyapunov optimization method. As described in the Lyapunov drift-penalty function construction method in patent CN112860409A, the Lyapunov drift-penalty function for the original problem P1 can be constructed to maximize the average task throughput at time t (penalty) while reducing the difference in task queues between time t and time t+1 (drift). Therefore, the decomposed continuous-time subproblem P2 can be obtained as follows:

[0113]

[0114] st: C1~C6

[0115] Where V>0 is the Lyapunov weighting coefficient, which represents the importance of task throughput relative to Lyapunov drift.

[0116] Since the multiple optimization variables of the continuous-time subproblem P2 are still coupled, the second computational unit is used to decompose the continuous-time subproblem into the following three stages for solution, as follows: Figure 2 As shown:

[0117] Phase 1: The task vehicle allocates local computing resources based on energy consumption constraint C2 and local computing resource constraint C3, and determines the current task queue. Check if the queue length constraint is met. If so, update the local task queue. If not, the task vehicle needs to unload tasks, partially unloading them to satisfy the queue length constraint, and then proceed to the second stage to solve the task unloading strategy. The methods for allocating local computing resources include:

[0118] Solving for the optimal local computing resource allocation strategy At the same time, local computing task volume can be obtained. Get local task queue

[0119] Phase Two: Since the vehicle can only handle a portion of the tasks under energy constraints in Phase One, it needs to offload excess tasks to the server to satisfy queue length constraint C1. Because the urban traffic environment is highly dynamic, traditional optimization methods struggle to capture and solve rapidly changing high-dimensional environmental state information. Therefore, this embodiment uses the vehicle as an agent, learning task offloading decisions based on deep reinforcement learning according to the current task queue and communication states. For example, this embodiment uses deep Q-learning, setting subproblem P2 as the agent's reward value. A Q-prediction network and a Q-target network are constructed. The agent generates and executes offloading actions using ∈-greedy based on the Q-value output by the Q-prediction network. Simultaneously, the offloading actions, reward value, and next-time state information are stored in an experience replay pool. A certain number of batches are randomly selected for training the neural network until convergence. Updated parameters are periodically synchronized to the Q-target network, ultimately outputting the optimal task offloading decision.

[0120] Phase Three: After the task vehicle makes the optimal unloading decision, the task will be transmitted to the server for calculation. To ensure the reliability of task transmission, the transmission power needs to be optimized. From constraints C1 and C5, the optimal transmission power can be obtained as follows:

[0121]

[0122] in, This is the optimal local computing resource allocation strategy obtained from the first stage solution.

[0123] Therefore, the task transmission volume is: After a task is transmitted to the server, it enters the task queue, and the server's computing unit allocates computing resources to it. To be fair, server computing resources The allocation strategy is set as the ratio of the number of tasks unloaded from vehicle k to server m to the total number of tasks unloaded from all vehicles to server m:

[0124]

[0125] From this, we can obtain the amount of work calculated by the server. Further update the server task queue Then proceed to the next time step t+1. This cycle continues until the final time T is reached.

[0126] This invention also provides a method for unloading vehicle-road-cloud collaborative tasks and controlling the stability of the task queue, which includes:

[0127] Step 1: Obtain the task information generated by task vehicle k at the current time t, and update the local task queue status at the current time according to the task unloading decision of the previous time.

[0128] Step 2, based on the obtained channel power gain G between task vehicle k and server m at the current time t. k,m (t) Local task queue status and server task queue status Optimize the uninstallation strategy;

[0129] Step 3, based on the task uninstallation decision x in the optimized uninstallation strategy. k,m,t The server allocates computing resources to the corresponding task vehicle and transmits the updated server task queue status of task vehicle k at the current time t to the vehicle task scheduler. and computing resource allocation and the channel power gain G between vehicle k and server m k,m (t); where the server includes an edge server and a vehicle cloud platform, and the methods for obtaining the vehicle cloud platform include:

[0130] By collecting historical vehicle trajectory information, the future driving trajectory of the vehicle is predicted. Then, similar trajectories are clustered around the mission vehicle. A vehicle cloud platform is built based on existing virtualization technology, and idle vehicle computing resources are aggregated to build a vehicle cloud virtual computing unit to provide computing resources for the mission vehicle.

[0131] In one embodiment, in step 2, the local task queue state is updated at the current time t using the following formula (1).

[0132] This invention considers the real-time mobility and communication status changes of vehicles in urban traffic scenarios. It establishes a task offloading optimization model with the goal of maximizing the task throughput of local vehicle computation and offloading. It comprehensively considers vehicle task queue length constraints, energy consumption constraints, and transmission power constraints. It combines Lyapunov optimization method and deep reinforcement learning method to construct a lightweight solution framework, thereby achieving long and short task queue stability and reducing solution complexity.

[0133] 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. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; 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 vehicle-road-cloud cooperative task offloading and task queue stability control system, characterized in that, Comprise: A vehicle task manager is configured to obtain task information generated by a task vehicle k at a current time t, and update a local task queue state at the current time according to a task offloading decision at a previous time; a vehicle task scheduler configured to determine, according to the obtained channel power gain G between the vehicle k and the server m at the current time t, a task scheduling strategy for the vehicle k k,m (t), a local task queue state and a server task queue state optimizing the offloading strategy; wherein the offloading strategy optimization unit of the vehicle task scheduler comprises: A first calculation subunit is configured to decompose an original problem corresponding to a task offloading optimization model P1 into continuous time subproblems by using a Lyapunov optimization method; A second calculation subunit is configured to decompose the continuous time subproblems into three stages for solving as follows: In a first stage, local computing resource allocation is performed according to energy consumption constraints C2 and local computing resource constraints C3 of the task vehicle at each time, and it is determined whether the current task queue satisfies a queue length constraint, if yes, the next time is entered; if not, the task vehicle needs to perform offloading calculation in a second stage; In the second stage, the task vehicle is taken as an agent, a task offloading decision is learned by a vehicle agent based on a deep reinforcement learning method according to a current task queue state and a communication state, and the task vehicle enters a third stage after making the offloading decision; In the third stage, a task transmission power and a server computing resource allocation strategy are optimized; A server manager has: a resource allocation updating unit configured to allocate, according to the task offloading decision x in the optimized offloading strategy, a computing resource for a corresponding task vehicle to a vehicle task scheduler, and transmit, to the vehicle task scheduler, an updated server task queue state of the task vehicle k at a current time t k,m,t , a server task queue state of the task vehicle k at the current time t , a computing resource allocation , and a channel power gain G k,m (t) between the vehicle k and the server m; A trajectory prediction unit is configured to predict a future driving trajectory of a vehicle by collecting vehicle historical trajectory information; A vehicle cloud construction unit is configured to cluster similar trajectories with the task vehicle as the center, construct a vehicle cloud platform based on existing virtualization technology, and aggregate idle computing resources of the vehicle to construct a vehicle cloud virtual computing unit for providing computing resources for the task vehicle. 2.The V2X cooperative task offloading and task queue stability control system of claim 1, wherein, The vehicle task manager updates the local task queue state at a current time t using the following equation (1) The method specifically comprises: Wherein: where A k (t) is the amount of tasks to be processed by task vehicle k at time t, τ is the length of each time interval, is the computing resource allocated to the vehicle local computing unit, ε k is the CPU computing cycles required by task vehicle k to process each byte of task, x k,m,t is the task offloading decision made by task vehicle k at time t, R k,m (t) is the communication rate between task vehicle k and server m at time t; In the formula, B k,m is the channel bandwidth allocated to the task vehicle k by the server m, p k,m (t) is the transmission power of the task vehicle k at the current time t to the server m, σ 2 is the background noise power of the receiver. 3.The V2X task offloading and task queue stability control system of claim 2, wherein, If x k,m,t = 1, the task vehicle k selects to unload the task to the mth server; if x k,m,t = 0, the task vehicle k does not select to unload the task to the mth server. 4.The V2X cooperative task offloading and task queue stability control system of claim 2, wherein, The resource allocation updating unit specifically comprises: a resource allocation subunit for allocating computing resources to the respective task vehicles according to the optimized x k,m,t by the computing unit of the server for the respective task vehicle; a resource update subunit for the computing unit of the server m to allocate the computing resource to the task vehicle k obtaining the amount of tasks processed at the current time t is expressed as: wherein is the task queue state of server m at time t, which changes dynamically and can be represented as 5.The V2X task offloading and task queue stability control system of claim 4, wherein, The offloading strategy optimization unit is configured to optimize the offloading strategy according to the G k,m (t), and Set the task throughput at each moment as Optimize the offloading strategy by using the pre-set task offloading optimization model P1. where K is the total number of task vehicles, T is the total time step, is the maximum value of the task queue length, is the maximum energy consumption of task vehicle k, is the maximum CPU computation period of task vehicle k, is the maximum transmission power of task vehicle k, is the maximum CPU computation period of server m.

6. A vehicle-road-cloud cooperative task offloading and task queue stability control method, characterized in that, Comprise: Step 1, obtaining task information generated by a task vehicle k at a current time t, and updating a local task queue state at the current time according to a task offloading decision at a previous time; Step 2, according to the obtained channel power gain G between the task vehicle k and the server m at the current time t k,m (t), local task queue state and server task queue state optimizing the offloading strategy; wherein the offloading strategy optimization specifically comprises: Decomposing an original problem corresponding to a task offloading optimization model P1 into continuous time subproblems by using a Lyapunov optimization method; Decomposing the continuous time subproblems into three stages for solving as follows: In a first stage, local computing resource allocation is performed according to energy consumption constraints C2 and local computing resource constraints C3 of the task vehicle at each time, and it is determined whether the current task queue satisfies a queue length constraint, if yes, the next time is entered; if not, the task vehicle needs to perform offloading calculation in a second stage; In the second stage, the task vehicle is taken as an agent, a task offloading decision is learned by a vehicle agent based on a deep reinforcement learning method according to a current task queue state and a communication state, and the task vehicle enters a third stage after making the offloading decision; In the third stage, a task transmission power and a server computing resource allocation strategy are optimized; Step 3, according to the optimized offloading strategy, task offloading decision x k,m,t , by the server for the corresponding task vehicle to allocate computing resources, to the vehicle task scheduler Transmission of the updated server task queue state of the task vehicle k at the current time t and computing resource allocation and the channel power gain G k,m (t) between vehicle k and server m; wherein the server includes an edge server and a vehicle cloud platform, and the acquisition method of the vehicle cloud platform comprises: Predicting a future driving trajectory of a vehicle by collecting vehicle historical trajectory information, clustering similar trajectories with the task vehicle as the center, constructing a vehicle cloud platform based on existing virtualization technology, and aggregating idle computing resources of the vehicle to construct a vehicle cloud virtual computing unit for providing computing resources for the task vehicle; Predicting a future driving trajectory of a vehicle by collecting vehicle historical trajectory information; Clustering similar trajectories with the task vehicle as the center, constructing a vehicle cloud platform based on existing virtualization technology, and aggregating idle computing resources of the vehicle to construct a vehicle cloud virtual computing unit for providing computing resources for the task vehicle.

7. The V2X task offloading and task queue stability control method of claim 6, wherein, In step 2, the local task queue state is updated at the current time t using the following formula (1) The method specifically comprises: Wherein: wherein A k (t) is the amount of tasks to be processed by task vehicle k at time t, τ is the length of each time interval, is the computing resource allocated to the local computing unit of task vehicle, ε k is the CPU computing cycles required by task vehicle k to process each byte of task, x k,m,t is the task offloading decision made by task vehicle k at time t, if x k,m,t = 1, task vehicle k chooses to offload tasks to the mth server; if x k,m,t = 0, task vehicle k does not choose to offload tasks to the mth server, R k,m (t) is the communication rate between task vehicle k and server m at time t; In the formula, B k,m is the channel bandwidth allocated to the task vehicle k by the server m, p k,m (t) is the transmission power of the task vehicle k at the current time t to the server m, σ 2 is the background noise power of the receiver. 8.The method according to claim 6 or 7, wherein, In step 2, according to G k,m (t), and The task throughput at each time is set as The task offloading strategy is optimized by using the pre-set task offloading optimization model P1. where K is the total number of task vehicles, is the maximum length of the task queue, is the maximum energy consumption of task vehicle k, is the maximum CPU computation period of task vehicle k, is the maximum transmission power of task vehicle k, is the maximum CPU computation period of server m.

Images