An Edge Offloading Method for Optimizing Vehicular Network Tasks Based on Deep Learning

Abstract

The present invention discloses an edge offloading method for optimizing tasks in the Internet of Vehicles based on deep learning, which is applied to the edge computing system of the Internet of Vehicles. The edge computing system of the Internet of Vehicles includes a roadside unit and an on-vehicle unit. The roadside unit includes an edge computing server. The edge computing system of the Internet of Vehicles further includes a cloud computing server. The edge computing server includes at least two offloading units, as well as a scheduling unit and a penalty determination unit corresponding to each offloading unit. A deep neural network for offloading decision-making is installed in the offloading unit, and tasks exceeding the available computing power of the edge computing server are scheduled from the edge computing server to the cloud computing server. The present invention effectively reduces the total delay and total energy consumption of the edge computing system of the Internet of Vehicles.

Description

Technical Field

[0001] This application belongs to the technical field of vehicular edge computing and deep learning, and particularly relates to a method for optimizing task edge offloading in the Internet of Vehicles based on deep learning. Background Art

[0002] In recent years, with 5G reaching a new peak, the Internet of Things and wireless communication technologies have been continuously developed and improved. Various new mobile applications such as virtual reality and ultra-high-definition video have emerged in large numbers, leading to an explosive growth in the number of mobile terminals and a large consumption of mobile resources. As a key technology of 5G, Mobile Edge Computing (MEC) is considered a powerful paradigm to solve the growing demand for computing tasks and the limited computing power and resources of users themselves. Vehicular Edge Computing (VEC) is an innovative technology that combines MEC with the Internet of Vehicles. As its core technology, edge offloading can sink computationally intensive and latency-sensitive in-vehicle applications / in-vehicle computing tasks to edge servers that are closer to vehicle terminals than remote cloud servers, thus greatly reducing the latency and energy consumption of executing terminal tasks, and can also effectively improve QoS (Quality of Service) and user QoE (Quality of Experience).

[0003] In the prior art, it is difficult to issue offloading decisions for different task requirements of vehicle users. For different types of in-vehicle tasks, issuing targeted offloading decisions will greatly reduce the latency and energy consumption of the VEC system. At the same time, the computing resources of each server are not fully utilized, and it is also necessary to study the resource scheduling of different computing platforms and consider the losses of the servers. Summary of the Invention

[0004] The purpose of this application is to provide a method for optimizing task edge offloading in the Internet of Vehicles based on deep learning to overcome the latency and energy consumption problems caused by offloading decisions in the prior art.

[0005] To achieve the above purpose, the technical solution of this application is as follows:

[0006] A method for optimizing task edge offloading in the Internet of Vehicles based on deep learning, which is applied to the edge computing system of the Internet of Vehicles. The edge computing system of the Internet of Vehicles includes a roadside unit and an on-vehicle unit. The roadside unit includes an edge computing server. The edge computing system of the Internet of Vehicles further includes a cloud computing server. The edge computing server includes at least two offloading units, as well as a scheduling unit and a penalty determination unit corresponding to each offloading unit. A deep neural network for offloading decision is installed in the offloading unit. The method for optimizing task edge offloading in the Internet of Vehicles based on deep learning includes:

[0007] Taking the weighted sum of the total energy consumption and total delay of the edge computing system of the Internet of Vehicles as the system overhead, jointly modeling offloading decision and computing resource allocation to determine the objective function;

[0008] Training the deep neural network. For a roadside unit, taking the received vehicle tasks and their corresponding attributes as the input of the deep neural network in the offloading unit, and obtaining candidate offloading decisions through the deep neural network;

[0009] For the tasks determined to be executed by the edge computing server in the candidate offloading decisions, the scheduling unit compares the computing resources required by the tasks with the idle computing resources of the edge computing server. If the idle computing resources of the edge computing server are sufficient, the task is processed in the edge computing server; otherwise, the task is scheduled to the cloud computing server;

[0010] The penalty determination unit determines whether the execution delay of each vehicle task exceeds the maximum tolerable delay, and calculates the value of the fitness function according to the penalty function. The fitness function is the sum of the objective function and the penalty function, and selects the candidate offloading decision with the smallest value of the fitness function as the optimal offloading decision;

[0011] Putting the vehicle tasks and the optimal offloading decisions into the standard library as training criteria, randomly selecting training criteria from the standard library and assigning them to the deep neural network in the offloading unit, and using the gradient descent algorithm to minimize the loss function to update the network parameters of the deep neural network;

[0012] Continuously training until the training of the deep neural network is completed, using the trained deep neural network to update the offloading unit, and offloading the vehicle tasks through the offloading unit and the scheduling unit.

[0013] Further, the objective function is as follows:

[0014]

[0015] Among them, C(a,b) is the total system cost, N is the total number of vehicles, M is the number of tasks, and y ,

[0015] , , , nm ,

[0014] , nm , ,

[0013] , ,

[0012] , ,

[0011] is the attribute value corresponding to task a nm corresponding, Indicates the energy consumption of vehicle n for locally executing task m, b nm Indicates task a nm The corresponding decision value Indicates the total energy consumption of vehicle n for offloading task m Indicates the total offloading delay of vehicle n Indicates the total delay consumed by vehicle n for locally executing tasks

[0016] Further, randomly selecting a training criterion from the standard library and assigning it to the deep neural network in the offloading unit includes:

[0017] For k offloading units, randomly select a training criterion from the standard library and assign it to the first offloading unit

[0018] Randomly select a training criterion from the standard library again and assign it to other offloading units in turn

[0019] Further, after the standard library is full, the training criterion stored first is automatically discarded

[0020] Further, the penalty function is expressed as:

[0021]

[0022] where h = max{0, g} Indicates the total offloading delay of vehicle n, and θ(h) represents the penalty function

[0023] An edge offloading method for optimizing tasks in the vehicle network based on deep learning proposed in this application has the following beneficial effects: The joint modeling of offloading decision and resource allocation is based on the "cloud-edge-end" service architecture, and tasks that exceed the available computing power of the MEC server are scheduled from the MEC server to the cloud computing server, effectively reducing the total delay and total energy consumption of the VEC system. The initial tasks are divided into three categories: energy consumption sensitivity, delay sensitivity, and non-sensitivity. Targeted offloading decisions are provided for different types of tasks, effectively reducing the average service delay of vehicle network users. The deep neural network inside the offloading algorithm can issue the optimal offloading decision for new task requests in a very short time through effective training Description of the Drawings

[0024] Figure 1 Is the network model diagram of the VEC system of this application

[0025] Figure 2 Is the flowchart of the edge offloading method for optimizing tasks in the vehicle network based on deep learning of this application

[0026] Figure 3 Is the network structure diagram of the edge offloading of the vehicle network optimization task of this application DETAILED DESCRIPTION

[0027] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments do not constitute a limitation of the present invention.

[0028] The overall concept of this application is to first jointly model offloading decisions and resource allocation as a multi-constrained optimization problem based on task type. Subsequently, a scheduling unit and penalty function are constructed to eliminate the constraints of the optimization problem. Finally, a distributed offloading network based on deep learning is proposed to solve the optimization problem and effectively reduce the overhead of the VEC system through learning and training. The application of deep neural networks (DNNs) to the Internet of Vehicles can enhance the system's environmental perception capabilities and overcome the dimensional limitations of handling offloading decisions and resource allocation problems.

[0029] The application scenario of this application is as follows Figure 1 As shown in the figure, the road side unit RSU (Road Side Unit) in the ETC system is installed on the road side and uses DSRC (Dedicated Short Range Communication) technology to communicate with the on-board unit OBU (On Board Unit) to realize vehicle identification and electronic point deduction, etc. Within the radius of the road test unit RSU, the requesting vehicle can perform task calculations locally or offload the tasks to the RSU equipped with the MEC server on that section of road for calculation. Assuming that different vehicles have computing tasks within the RSU range, the vehicle set is defined as N = {1,2,...,N}. Each vehicle has multiple independent tasks that need to be processed on the local CPU or offloaded to the MEC server for processing. The task set of vehicle n is defined as M = {1,2,...,M}. Define a nm The workload of the mth task of vehicle n is the workload of the vehicle terminal MD. The vehicle terminal MD can decide whether to offload task m to the edge server MEC. The offloading decision is represented by a binary variable b nm ∈{0,1} represents b. nm = 0 means vehicle n will process its task m locally; b nm =1 means that vehicle n offloads its task m to the edge server MEC, and the edge server MEC will transmit the calculation result back to vehicle n.

[0030] At the same time, RSU receives tasks initiated by different vehicles and queues all tasks in the order of initiating the requests and stores them in the MEC server cache. It is expressed as T = {τ 11 ,τ 12 ,...,τ 1m ,τ 21 ,...,τ nm}, all tasks will simultaneously generate a computing processing requirement expressed as Among them, represents the maximum tolerable delay for vehicle n to execute a task, and y nm represents the task attribute value, that is, different requirements for delay and power consumption: when 0 < y nm < 0.5, the dependence of this task on delay is greater than that on energy consumption; when y nm = 0.5, the dependence of this task on energy consumption is equal to that on delay; when 0.5 < y nm < 1, the dependence of this task on energy consumption is greater than that on delay. Each RSU in this application has a certain coverage radius, and the tasks generated by vehicles within this radius are attributed to the MEC server equipped on this RSU for processing. Therefore, in this embodiment, the optimization of offloading for vehicles and their tasks within the range of one RSU is considered, and details will not be elaborated below.

[0031] The current MEC server has a maximum computing resource Therefore, when allocating the server processing resources corresponding to the tasks to be processed in the cache, it is necessary to consider the resources occupied by the previously unfinished tasks. At the same time, if the computing resources required by the cached tasks exceed the remaining computing resources of the server, it will lead to task processing failure. Therefore, it is considered to forward the exceeding tasks from the MEC server to the remote cloud server to solve the tasks through the resource scheduling of the server.

[0032] In one embodiment, as Figure 2 shown, a vehicle network optimization task edge offloading method based on deep learning is proposed, including:

[0033] Step S1: Taking the weighted sum of the total energy consumption and total delay of the vehicle network edge computing system as the system overhead, jointly modeling the offloading decision and computing resource allocation to determine the objective function.

[0034] The technical solution of this application is applied to a vehicle network edge computing system, which includes a roadside unit and an on-vehicle unit. The roadside unit includes an edge computing server, and the vehicle network edge computing system also includes a cloud computing server. As Figure 3 shown, the edge computing server includes at least two offloading units, as well as a scheduling unit and a penalty determination unit corresponding to each offloading unit. A deep neural network for offloading decision is installed in the offloading unit.

[0035] This application defines the weighted sum of the total energy consumption and total delay in the vehicle network edge computing system VEC as the system overhead, and jointly models the offloading decision and computing resource allocation to achieve the purpose of reducing the overhead.

[0036] The objective function of this embodiment is expressed as follows:

[0037]

[0038] Among them, C(a, b) is the total cost of the system, N is the total number of vehicles, M is the number of tasks, and y nm is the attribute value corresponding to task a nm ; represents the energy consumption of vehicle n for locally executing task m, and b nm represents the decision value corresponding to task a nm ; represents the total energy consumption of vehicle n for offloading task m, represents the total offloading delay of vehicle n, represents the total delay consumed by vehicle n for locally executing tasks.

[0039] The derivation process of the objective function in this embodiment is as follows:

[0040] Since the OFDMA system is considered, the transmission interference between independent tasks of different vehicles is ignored. At the same time, the bandwidth is evenly allocated to the connected vehicles, so the uplink transmission rate of the mth task of vehicle n can be obtained:

[0041]

[0042] Among them, represents the sub-channel bandwidth of the uplink of vehicle n; L represents the total uplink bandwidth; H n represents the channel gain between vehicle n and the RSU; σ 2 represents the background noise power; represents the transmission power of vehicle n for task m.

[0043] For the case where vehicle n decides to use its local CPU to execute tasks, the computing power of the vehicle's own CPU is set to Since the task is only calculated locally and the transmission delay of the task does not need to be considered, the delay consumed by vehicle n for locally executing task m can be obtained from the following formula:

[0044]

[0045] After the given specific offloading decision {b nm}, the total delay consumed by vehicle n for locally executing tasks can be obtained:

[0046]

[0047] Define as the local energy consumption of vehicle n for each 1 bit of data, so the energy consumption of vehicle n for locally executing task m is:

[0048]

[0049] When the task is offloaded to the edge computing server, the vehicle transmits the task load a through the wireless channel nm to the RSU and further forwards it to the edge computing server MEC for processing. Considering the situation where the MEC server is directly configured on the RSU, the energy consumption and delay when the RSU forwards the task to the MEC server are ignored. Generally, the edge computing server has sufficient power and computing capabilities. Therefore, the transmission delay for vehicle n to offload task m to the edge server is:

[0050]

[0051] Meanwhile, the computing delay for the task to be processed on the edge computing server is:

[0052]

[0053] Among them, it is defined that is the processing resource allocated by the edge computing server to this task. Since the data size of the computing result fed back by the MEC server to vehicle n is much smaller than the input data, the downlink transmission and the delay and energy consumption of vehicle n receiving the computing result are not considered. After a specific offloading decision {b nm} is given, the total offloading delay of vehicle n can be obtained:

[0054]

[0055] It is defined that is the local energy consumption for vehicle n to transmit each 1-bit data. Therefore, the energy consumption for vehicle n to transmit task m is:

[0056]

[0057] Therefore, the total energy consumption for vehicle n to offload its task m to the edge computing server can be obtained:

[0058]

[0059] Among them, it is defined that P mec is the device power of the MEC server, and μ is the energy consumption ratio of the MEC server to execute the task. When μ = 0, only the energy consumed by vehicle n is considered for the entire offloading process.

[0060] When the computing resources required by the offloading tasks in the queue exceed the idle computing resources of the MEC, the transcendent tasks will be first sent from the MEC server to the base station BS via optical fiber, and then forwarded from the BS to the remote cloud computing server via optical fiber. Since the vehicle has offloaded the tasks to the MEC server and all subsequent offloading paths are via optical fiber, the uplink and downlink transmission delays and energy consumption do not need to be considered. Therefore, the computing delay of task m of vehicle n on the cloud computing server is:

[0061]

[0062] where f cloud is the computing resources provided by the cloud server. Define P cloud as the device power of the cloud server, then the computing energy consumption of task m of vehicle n on the cloud server is:

[0063]

[0064] where λ is the proportion of energy consumption for the cloud server to execute tasks. When λ = 0, only the energy consumed by vehicle n is considered for the entire offloading process. The cloud computing model is based on the edge computing model. After a specific offloading decision {b nm} is given, tasks with b nm = 1 will have two offloading methods. Therefore, the new total offloading delay of vehicle n and the new total energy consumption of task m offloaded by vehicle n can be obtained:

[0065]

[0066] where, when d = 1, this task needs to be offloaded to the cloud computing server; when d = 0, this task does not need to be offloaded to the cloud computing server.

[0067] Based on the total delay T and total energy consumption E generated in the above three scenarios, the system cost C is introduced, which represents the weighted value of energy consumption and delay for executing all tasks:

[0068]

[0069] As described above, the goal is to minimize the total system cost C(a, b) under multiple constraints. Therefore, the task processing of the system is modeled as P1:

[0070]

[0071] C2: b nm ∈ {0, 1};

[0072]

[0073] C4: y nm ∈ (0, 1);

[0074] C5: n ∈ N, m ∈ M;

[0075] where a = {a nm | n ∈ N, m ∈ M}, b = {b nm | n ∈ N, m ∈ M}, is the maximum computing resource of the current MEC server. C1 gives the maximum tolerable delay for vehicle n to execute all tasks. C2 gives a binary offloading decision. C3 limits that the computing resources allocated to the offloading tasks in the queue cannot exceed the maximum computing resource of the MEC server. C4 gives the degree of dependence of each task on delay and energy consumption. C5 gives the number of vehicles and tasks under this system.

[0076] Step S2, train the deep neural network. For a roadside unit, take the received vehicle tasks and their corresponding attributes as the input of the deep neural network in the offloading unit, and obtain the candidate offloading decision through the deep neural network.

[0077] As Figure 3 shown, this application is essentially a parallel VEC network offloading architecture based on DNN. The overall architecture consists of four parts: K independent and parallel offloading units, a scheduling unit, a penalty function, and a memory unit with a finite size. All are in the edge computing server and located in a roadside unit. The parallel distributed offloading units and the scheduling unit are used to implement the decision generation part. Among them, each offloading unit contains a deep neural network DNN with exactly the same structure; the penalty function is used to handle the constraint conditions; the memory unit is used to implement the learning and training part in the algorithm.

[0078] For the tasks of different vehicles that initiate requests, obtain the training input S. The tasks enter the cache queue in the order of vehicle request initiation. In addition to the cache for storing the input tasks, another set of caches is generated to store the attribute values Y = {A, B, C} of each task. Among them, label A corresponds to the demand y nm ∈ (0.5, 1); label B corresponds to y nm = 0.5; label C corresponds to y nm ∈ (0.5, 1).

[0079] First, initialize the network parameters of the DNNs with the same structure placed in the offloading units One DNN for one offloading unit. Among them, the linear relationship coefficient w k is initialized with the initial value defined by the normal distribution tf.random_normal_initializer, and the bias coefficient j kInitialize the initial value with the constant initializer tf.constant_initializer, and then input the task with attribute values into the DNN network. The DNN network outputs the candidate offloading decision b k The offloading policy function of the task to obtain k binary decisions through the deep neural network can be expressed in parametric form:

[0080]

[0081] Once k binary offloading decisions are obtained, the constraints C2, C4, and C5 in the original problem P1 can be eliminated, and the problem P1 will be transformed into a resource allocation and tolerable delay constraint problem (P2):

[0082]

[0083] Step S3: For the tasks determined to be executed by the edge computing server in the candidate offloading decisions, the scheduling unit compares the computing resources required by the tasks with the idle computing resources of the edge computing server. If the idle computing resources of the edge computing server are sufficient, the task is processed in the edge computing server; otherwise, the task is scheduled to the cloud computing server.

[0084] For the candidate offloading decisions obtained in the previous steps, the scheduling unit determines the processing method of the offloading tasks and updates the cost value corresponding to the offloading decisions.

[0085] As Figure 3 shown, pre-allocate the server computing resources for the size of the tasks with b nm = 1 in the cache. Before processing each cached task with b nm = 1, compare the computing resources it requires with the idle computing resources of the current MEC. Once the result is greater than the idle resources, it will be regarded as an over-item task and then scheduled from the MEC server to the remote cloud computing server to assist in processing the computing task with the computing resources of the cloud computing server. If the idle computing resources of the MEC are sufficient, the task is executed by the MEC server.

[0086] After reasonably allocating the tasks offloaded to the edge computing server and the cloud computing server, the cost value C * (S, b k ) output from the k offloading units can be updated. After solving the resource allocation problem, only the last constraint C1 remains. Therefore, the problem P2 will be transformed into a maximum tolerable delay constraint problem (P3):

[0087]

[0088] Step S4: The penalty determination unit determines whether the execution delay of each vehicle task exceeds the maximum tolerable delay, and calculates the value of the fitness function according to the penalty function. The fitness function is the sum of the objective function and the penalty function. The candidate offloading decision with the smallest value of the fitness function is selected as the optimal offloading decision.

[0089] For the only inequality constraint condition C1 in the P3 problem, a penalty function is constructed to completely eliminate the constraint on the objective function, so as to solve this nonlinear programming problem.

[0090] Transpose the constraint condition C1:

[0091]

[0092] Then construct a penalty function to determine whether the execution delay of vehicle n task exceeds the maximum tolerable delay:

[0093] h = max{0, g};

[0094] Among them, when h ≤ 0, there is no need to penalize the objective function; when h > 0, the objective function needs to be penalized.

[0095] Considering that the RSU coverage range is limited, relatively long delays will not only reduce the user experience, but also cause task execution failures. Therefore, it is set that exceeding the maximum tolerable delay will have an exponential impact on the system cost, and the penalty function is obtained:

[0096]

[0097] Among them, h = max{0, g}, represents the total delay consumed by vehicle n to execute the task locally, represents the total offloading delay of vehicle n, and θ(h) represents the penalty function. In the formula respectively represent the preset time thresholds, for example, equal to 1, 3, and 10 seconds respectively. Based on the above objectives and the establishment of the penalty function, the fitness function can calculate the total cost of the entire system under the constraint of the maximum tolerable delay. The larger the obtained fitness function value, the greater the cost consumed by the system, and vice versa. Adding the objective function and the penalty function can obtain the fitness function:

[0098] Fitnesss(h) = Fitness(S, b k ) = C * (S, b k ) + θ(h).

[0099] After solving the maximum tolerable latency problem (P3), among the k groups of candidate objects that have passed through the offloading unit, scheduling unit, and penalty determination, the offloading decision with the minimum system cost is selected as the optimal offloading decision for output. This set of optimal offloading decisions will be merged with the corresponding input S as the training criterion for the neural network. The process of obtaining the optimal offloading decision can be expressed as:

[0100] B = argmin Fitness(S, b k ).

[0101] Step S5: Put the vehicle task and the optimal offloading decision into the standard library as the training criterion. Randomly select a training criterion from the standard library and assign it to the deep neural network in the offloading unit. Use the gradient descent algorithm to minimize the loss function and thus update the network parameters of the deep neural network.

[0102] For the optimal offloading decision obtained from a set of input S through the offloading unit, scheduling unit, and penalty function, it is stored in the standard library Standard as the training criterion for the DNN. After completing one content storage, randomly sample k groups of training criteria and assign them to all DNNs. Use the gradient descent algorithm to minimize the loss function and thus update the DNN network parameters. When the memory in the standard library Standard is full, the most original data is automatically discarded and new data is added as a supplement.

[0103] Step S6: Continuously train until the training of the deep neural network is completed. Update the offloading unit with the trained deep neural network, and offload the vehicle task through the offloading unit and the scheduling unit.

[0104] Continue to train. The network of each DNN in the offloading unit will gradually converge after multiple trainings until the offloading decisions output by each DNN are almost the same in the end. Therefore, the more DNNs there are, the faster the convergence. Then, what is actually used in practical applications should be the trained DNN. So every time a new task set is input, the trained DNN can output the optimal offloading decision in a very short time.

[0105] For the vehicle network edge computing system proposed in this application, the network parameters of the DNN can be effectively trained through tens of thousands of randomly generated inputs. Different numbers of DNNs or learning rates will change the convergence speed of this offloading algorithm. After completing the training, for the tasks received by the roadside unit, they can be input into the offloading unit to output offloading decisions. For the tasks determined to be executed by the edge computing server in the offloading decision, the scheduling unit compares the computing resources required for the tasks with the idle computing resources of the edge computing server. If the idle computing resources of the edge computing server are sufficient, the task is processed in the edge computing server; otherwise, the task is scheduled to the cloud computing server.

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and deformations according to the present invention, but these corresponding changes and deformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for edge offloading of vehicle network optimization tasks based on deep learning, which is applied to a vehicle network edge computing system. The vehicle network edge computing system includes a roadside unit and an on-vehicle unit, and the roadside unit includes an edge computing server, characterized in that, The vehicle networking edge computing system further includes a cloud computing server. The edge computing server includes at least two offloading units, as well as a scheduling unit and a penalty determination unit corresponding to each offloading unit. A deep neural network for offloading decision-making is installed in the offloading unit. The edge offloading method for vehicle networking optimization tasks based on deep learning includes: Taking the weighted sum of the total energy consumption and total time delay of the vehicle networking edge computing system as the system overhead, jointly modeling the offloading decision and computing resource allocation to determine the objective function; Training the deep neural network. For a roadside unit, taking the received vehicle tasks and their corresponding attributes as the input of the deep neural network in the offloading unit, and obtaining candidate offloading decisions through the deep neural network; For the tasks determined to be executed by the edge computing server in the candidate offloading decisions, the scheduling unit compares the computing resources required by the tasks with the idle computing resources of the edge computing server. If the idle computing resources of the edge computing server are sufficient, the tasks are processed in the edge computing server; otherwise, the tasks are scheduled to the cloud computing server; The penalty determination unit determines whether the execution time delay of each vehicle task exceeds the maximum tolerable time delay, and calculates the value of the fitness function according to the penalty function. The fitness function is the sum of the objective function and the penalty function, and selects the candidate offloading decision with the smallest value of the fitness function as the optimal offloading decision; Putting the vehicle tasks and the optimal offloading decisions into the standard library as training criteria, randomly selecting training criteria from the standard library and assigning them to the deep neural network in the offloading unit, and using the gradient descent algorithm to minimize the loss function to update the network parameters of the deep neural network; Continuously training until the training of the deep neural network is completed, using the trained deep neural network to update the offloading unit, and offloading the vehicle tasks through the offloading unit and the scheduling unit; Among them, the objective function is as follows: Among them, C(a, b) is the total cost of the system, N is the total number of vehicles, M is the number of tasks, and y nm is the attribute value corresponding to task a nm , represents the energy consumption of vehicle n for locally executing task m, and b nm represents the decision value corresponding to task a nm , represents the total energy consumption of vehicle n for offloading task m, represents the total offloading delay of vehicle n, represents the total delay consumed by vehicle n for locally executing tasks; The penalty function is expressed as: where h = max{0, g}, represents the total unloading delay of vehicle n, and θ(h) represents the penalty function.

2. The edge offloading method for optimizing tasks in the vehicle network based on deep learning according to claim 1, wherein The randomly selecting training criteria from the standard library and assigning them to the deep neural network in the offloading unit includes: For k offloading units, randomly select a training criterion from the standard library and assign it to the first offloading unit; Randomly select a training criterion from the standard library again and assign it to the other offloading units in turn.

3. The edge offloading method for optimizing vehicle networking tasks based on deep learning according to claim 1, wherein After the standard library is full, the first stored training criterion is automatically discarded.

Images