Distributed car networking task unloading method and electronic equipment

Abstract

The invention provides a distributed car networking task unloading method and electronic equipment, and belongs to the technical field of car networking. The method comprises the steps that task information of task vehicles and state information of service vehicles in a preset area at the current moment are collected; and inputting the task information of the task vehicle and the state information of the service vehicle into a pre-trained Internet of Vehicles task unloading model, and optimizing an Internet of Vehicles task unloading strategy by using a federal deep learning algorithm to obtain an Internet of Vehicles task unloading scheme in which the average processing time of all tasks is minimized. According to the invention, the federal deep learning algorithm is adopted to optimize the car networking task unloading strategy, so that the service vehicles required by the unloading task are far less than the service vehicles required by the traditional algorithm, the system cost is reduced, the model training time and the computing resources occupied by the state space are reduced, and the system efficiency is improved. And the task unloading success rate and the long-time link stability are obviously improved.

Description

Technical Field

[0001] The present invention relates to a distributed vehicle networking task unloading method and electronic equipment, belonging to the technical field of vehicle networking. Background Art

[0002] The Internet of Vehicles is a specific application of the Internet of Things in the field of transportation. It is an intelligent network system that connects vehicles, road infrastructure, and pedestrians through wireless communication technology. It can realize information exchange between vehicles and the surrounding environment and provide a variety of services.

[0003] Task offloading technology makes resources more efficiently utilized by rationally allocating computing tasks generated by vehicles to other edge nodes or collaborative vehicles. Through task offloading, vehicles can overcome the limitations of local computing power and storage space and quickly complete tasks such as target recognition, path planning, and vehicle control. Therefore, task offloading technology has become an important support for the efficient operation of the Internet of Vehicles system.

[0004] In the prior art, some use reliable edge servers to offload tasks to target service vehicles to reduce the computational burden of a single vehicle. However, the communication between vehicles relies on edge servers for multi-hop connections, which cannot guarantee the efficient delivery of tasks. At the same time, in an environment where a long-term stable link cannot be established with the edge server, this method cannot be used for task offloading. Some use idle computing resources in decentralized self-organizing vehicle networks to model the task offloading problem as a mixed integer nonlinear programming problem, use heuristic algorithms to find the optimal task offloading strategy, divide lengthy computationally intensive application tasks into subtasks and offload them to selected service vehicles, and reduce the packet loss rate during task offloading. However, this method requires a large number of state transfers and evaluations during the solution process, and the success rate drops rapidly when there is a time limit for task completion. Others consider task offloading in a multi-edge environment, use a deep learning algorithm DL to determine the offloading location and offloading ratio, and quickly make a near-optimal decision. However, the deep learning model is deployed on a central server. As the scale of the Internet of Vehicles expands, each new device needs to upload data to the center for processing, which makes the centralized system face scalability problems. At the same time, a large amount of traffic data is redundantly collected and uploaded to the central server, resulting in a small proportion of samples that are beneficial to improving the generalization ability of the model within the limited storage space.

[0005] In summary, current research on task offloading and offloading algorithms based on deep learning focuses on the computational efficiency of the model and finding the optimal solution, which cannot meet the requirements of Internet of Vehicles applications for task completion time. At the same time, in the scenario where multiple vehicles are performing offloading tasks, the existing learning algorithms are centrally scheduled through the edge network, which occupies too much state space and has a low learning algorithm effect. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a distributed Internet of Vehicles task offloading method and electronic device to solve the problem of task offloading of smart vehicles in a scenario without an edge server.

[0007] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0008] In the first aspect, the present invention provides a distributed Internet of Vehicles task offloading method, including: collecting task information of task vehicles and status information of service vehicles in a preset area at the current moment; inputting the task information of the task vehicles and the status information of the service vehicles into a pre-trained Internet of Vehicles task offloading model, and optimizing the Internet of Vehicles task offloading strategy using a federated deep learning algorithm to obtain an Internet of Vehicles task offloading solution that minimizes the average processing time of all tasks.

[0009] Furthermore, the task information of the task vehicle includes: the location of the task vehicle, the speed of the task vehicle and the task amount of the task generated by the task vehicle; the status information of the service vehicle includes: the location of the service vehicle, the speed of the service vehicle and the available computing resources on the service vehicle.

[0010] Furthermore, the Internet of Vehicles task offloading model includes: a local model set on each service vehicle and a global model set in the cloud.

[0011] The local model and the global model include: state space , Action Space and the reward function , the expression is:

[0012] S = [ v i , j , d i , j , f i , j , AC j , drt i , j ] ;

[0013] A = [ u i , j ] ;

[0014] ;

[0015] in,

[0016] In the formula, For mission vehicles With service vehicles The relative speed of For mission vehicles To service vehicles The amount of unloaded tasks, For service vehicles Available computing resources on Mission vehicle With service vehicles The relative direction of is the number of mission vehicles at the current moment, Indicates service vehicle For mission vehicles The time to perform the mission, including the mission vehicle To service vehicles Time to uninstall the task and service vehicles Calculate the time of the task ,in, The expression is:

[0017] ;

[0018] For mission vehicles Whether to service vehicles Uninstall task variables, Mission vehicle To service vehicles Uninstall tasks, Mission vehicle No service vehicles Uninstall tasks; is the average processing time of all tasks at the current moment, The average processing time of all tasks at the next moment.

[0019] Furthermore, each service vehicle is only allowed to unload one of the tasks from one task vehicle per unit time;

[0020] In response to the current moment, , then all tasks at the current moment are unloaded to the service vehicle for calculation; in response to the current moment, , then all tasks at the current moment are sorted in descending order according to the task amount, and Tasks with large workloads are preferentially offloaded to the service vehicle for calculation, and the remaining tasks are calculated locally; is the number of vehicles in service at the current moment, For mission vehicles The number of tasks generated per unit time.

[0021] Furthermore, a federated deep learning algorithm is used to optimize the Internet of Vehicles task offloading strategy, including: obtaining the action information of each service vehicle based on the local model parameters and status information on each service vehicle; solving the pre-established task offloading objective function through a federated deep learning algorithm based on the action information and status information of each service vehicle and the task information of the task vehicle, to obtain the Internet of Vehicles task offloading solution at the current moment.

[0022] Furthermore, the task offloading objective function is:

[0023]

[0024] In the formula, Mission vehicle and service vehicles Connection duration, which is calculated as:

[0025]

[0026] in, is the maximum connectable distance between two vehicles.

[0027] Furthermore, the method of optimizing the Internet of Vehicles task offloading strategy by using a federated deep learning algorithm also includes:

[0028] After completing the current Internet of Vehicles task offloading, the local model on each service vehicle calculates and updates the local model parameters according to the current Internet of Vehicles task offloading plan; the global model on the cloud calculates the global reward according to the current Internet of Vehicles task offloading plan, and updates the global model parameters according to the local model parameters on each service vehicle; and synchronizes the updated global model parameters to the local model on each service vehicle.

[0029] Furthermore, the local model on each service vehicle updates the local model parameters according to the Internet of Vehicles task offloading scheme at the current moment, including:

[0030]

[0031]

[0032]

[0033] In the formula, Accumulate rewards for local models, To express the expectation of future rewards, is the maximum cumulative reward of the local model, For service vehicles The local model parameters on .

[0034] Further, the updating of the global model parameters according to the local model parameters on each service vehicle includes: updating the global model parameters using the average value of the local model parameters uploaded by each service vehicle, and the expression is:

[0035]

[0036] In the formula, are global model parameters.

[0037] In a second aspect, the present invention provides an electronic device comprising: a processor and a memory; the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the distributed Internet of Vehicles task offloading method as described in the first aspect is implemented.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) The distributed Internet of Vehicles task offloading method provided by the present invention adopts a federated deep learning algorithm to optimize the Internet of Vehicles task offloading strategy, so that the service vehicles required for offloading tasks are far less than those required by traditional algorithms. This not only reduces the system cost, but also reduces the model training time and the computing resources occupied by the state space, and significantly improves the task offloading success rate and the stability of long-term links.

[0040] (2) The goal of the federated deep learning algorithm used in the distributed Internet of Vehicles task offloading method provided by the present invention is to ensure that the connection between vehicles is as stable as possible and is less affected by the vehicle's driving speed, thereby ensuring the success rate of task offloading even with less computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the distributed task offloading scenario provided in Embodiment 1 of the present invention;

[0042] Figure 2 A comparison chart of the number of service vehicles required in the simulation experiment provided in Example 3 of the present invention;

[0043] Figure 3 A comparison chart of the success rate of task offloading in the simulation experiment provided in Example 3 of the present invention;

[0044] Figure 4 A comparison chart of the timely completion rate of tasks in the simulation experiment provided in Example 3 of the present invention;

[0045] Figure 5 This is a comparison chart of the average vehicle connection time in the simulation experiment provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0046] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0047] Example 1

[0048] This embodiment provides a distributed Internet of Vehicles task offloading method, including the following steps:

[0049] Step 1: Collect the task information of the task vehicles and the status information of the service vehicles in the preset area at the current moment.

[0050] Specifically, Figure 1 As shown, in an environment without edge server coverage, a suitable area is selected, and vehicles entering the area are equipped with necessary hardware and all vehicles have the same hardware specifications, where there are M task vehicles and N service vehicles. The task information of the task vehicle includes: the location of the task vehicle, the speed of the task vehicle, and the task volume of the task generated by the task vehicle; the status information of the service vehicle includes: the location of the service vehicle, the speed of the service vehicle, and the available computing resources on the service vehicle.

[0051] Step 2: Input the task information of the task vehicle and the status information of the service vehicle into the pre-trained IoV task offloading model, and use the federated deep learning algorithm to optimize the IoV task offloading strategy to obtain the IoV task offloading solution that minimizes the average processing time of all tasks.

[0052] Specifically, the Internet of Vehicles task offloading model includes: a local model set on each service vehicle and a global model set in the cloud.

[0053] After multiple task vehicles in the area generate unloading tasks at the same time, the service vehicles in the area observe the status information of their own vehicles and the unloading task information, select the appropriate task vehicle to connect with, accept the unloading task and calculate the result. According to the optimization results of the federated deep learning algorithm, each task selects a service vehicle from the response to unload, and each service vehicle is only responsible for unloading the tasks assigned to them.

[0054] In a specific embodiment, the local model and the global model include: state space , Action Space and the reward function , the expression is:

[0055] S = [ v i , j , d i , j , f i , j , AC j , drt i , j ] ;

[0056] A = [ u i , j ] ;

[0057] ;

[0058] in,

[0059] In the formula, For mission vehicles With service vehicles The relative speed of Indicates service vehicle For mission vehicles The time to perform the task, For mission vehicles To service vehicles The amount of unloaded tasks, For service vehicles Available computing resources on Mission vehicle With service vehicles The relative direction of For mission vehicles Whether to service vehicles Uninstall task variables, is the average processing time of all tasks at the current moment, The average processing time of all tasks at the next moment.

[0060] In a more specific embodiment, the service vehicle For mission vehicles Time to perform the task Including mission vehicles To service vehicles Time to uninstall the task and service vehicles Calculate the time of the task ,in, The expression is:

[0061] .

[0062] In a more specific embodiment, Mission vehicle To service vehicles Uninstall tasks, Mission vehicle No service vehicles Uninstall the task.

[0063] In a more specific embodiment, the service vehicle The available computing resources on the service vehicle The ratio of unallocated resources With service vehicles Total resources Multiply to obtain.

[0064] In a specific embodiment, if the number of all service vehicles within the area at the current moment is less than the number of all tasks at the current moment, that is, , then all tasks at the current moment are sorted in descending order according to the task amount, and the previous Tasks with large workloads are offloaded to The computation is performed by a service vehicle, and the remaining tasks are computed locally.

[0065] On the contrary, if the number of all service vehicles within the area at the current moment is not less than the number of all tasks at the current moment, all tasks at the current moment will be unloaded to the service vehicles for calculation.

[0066] Example 2

[0067] Based on Example 1, this example provides specific content of optimizing the Internet of Vehicles task offloading strategy using a federated deep learning algorithm.

[0068] When a service vehicle enters the environment, the local model on the service vehicle will be provided with the latest global model parameters. The service vehicle decides to accept or reject the task within the communication range based on its own status information through the local model, that is, obtains the action information of each service vehicle.

[0069] Based on the action information and status information of each service vehicle and the task information of the task vehicle, the pre-established task offloading objective function is solved through the federated deep learning algorithm to obtain the current Internet of Vehicles task offloading plan. Then, the task vehicle offloads the task to the corresponding service vehicle according to this plan, and the service vehicle performs calculations and then feeds back the calculation results.

[0070] In some specific embodiments, the task offloading objective function is:

[0071]

[0072] In the formula, Mission vehicle and service vehicles Connection duration, which is calculated as:

[0073]

[0074] in, is the maximum connectable distance between two vehicles.

[0075] When performing task offloading according to the Internet of Vehicles task offloading solution, the service vehicle will track and record the processing time of the task until the task processing is completed.

[0076] After completing the task offloading at the current moment, the local model on each service vehicle calculates and updates the local model parameters according to the current Internet of Vehicles task offloading plan; the global model on the cloud calculates the average processing time of all tasks at the current moment according to the current Internet of Vehicles task offloading plan, and calculates the corresponding global reward.

[0077] Each service vehicle stores the status information, action information and reward of the task being executed in the experience convolution and updates the local model parameters. When a round of task offloading is completed, the service vehicle uploads the latest parameters to the cloud to update the global model parameters, thereby continuously improving the performance of the Internet of Vehicles task offloading model. Among them, the service vehicle updates the local model parameters, including:

[0078]

[0079]

[0080]

[0081] The global model parameters are updated using the average value of the local model parameters uploaded by each service vehicle. The expression is:

[0082]

[0083] In the formula, Accumulate rewards for local models, To express the expectation of future rewards, is the maximum cumulative reward of the local model, For service vehicles Local model parameters on , are global model parameters.

[0084] Example 3

[0085] This embodiment provides simulation experiments and simulation results for evaluating federated deep learning algorithms.

[0086] To ensure fairness, the proposed algorithm and the comparison algorithm are implemented in Python 3.8 environment. The experiments are conducted on a system with Windows 10, 2.4GHz Intel Core i5-9300H CPU and 16GB RAM. In addition, the deep reinforcement learning environment of this article is based on TensorFlow 2.8.0 and NumPy 1.22.3.

[0087] In the simulation experiment, this paper assumes that the vehicles in the experimental area are driving on roads with speeds randomly generated according to Gaussian distribution in the range of [60km / h ~ 120km / h]. In order to explore the impact of task vehicles on the completion of all tasks when only service vehicles can assist in task calculation, the experiment is set on an obstacle-free highway. The vehicle travels at a constant speed along a straight line without considering turns. The experimental area is set to 2km * 10km, and the vehicle wireless distance is 100 meters. The computing power of the vehicle is cycles / s. In order to analyze the effect of processing computationally intensive and delay-sensitive tasks, 1 to 5 tasks of 5 to 10 mb in size are generated on randomly selected vehicles and unloaded in a specific time period after a specified time interval. The relevant simulation experiment parameters are shown in Table 1:

[0088] surface Simulation experiment parameter table

[0089] Parameters and description Numeric Simulation area 2km*10km Simulation time 60min Number of vehicles 18 / min Vehicle communication range 100m Vehicle speed range 60km / h~120km / h Local computing power in the vehicle 3.2*10^5 cycles / s Task size 5~10mb Number of tasks 1~5 V2V Bandwidth 10MHz Connection Type UDP

[0090] This experiment verifies the effectiveness of the federated deep learning algorithm by comparing the system response time and algorithm execution time in different scenarios with the three existing algorithms. The three existing algorithms are:

[0091] (1) Task replication offloading algorithm: A learning algorithm that replicates tasks and offloads them to all available service vehicles. The algorithm gradually learns the average offloading delay of each service vehicle before converging.

[0092] (2) Fuzzy logic algorithm: Due to the dynamic and high mobility of service vehicles, the selection of service vehicles is affected by many factors, including communication and computing. The fuzzy inference system (FIS) combines all input data and determines the optimal unloading strategy through relevant fuzzy parameters.

[0093] (3) Particle Swarm Optimization Algorithm: The PSO-BAS algorithm controls the balance between local search and global search by adjusting frequency and loudness. In addition, it has an automatic scaling function to ensure fast convergence in early iterations.

[0094] The simulation results are as follows Figures 2 to 5As shown in the figure, TROS represents the task replication offloading algorithm, FIS represents the fuzzy logic algorithm, BAS represents the particle swarm optimization algorithm, and FDL represents the federated deep learning algorithm.

[0095] Depend on Figure 2 It can be seen that the service vehicles required by the FDL task offloading algorithm are far less than those required by other task offloading algorithms. This is because traditional task offloading algorithms will offload tasks to multiple relatively optimal service vehicles to ensure the completion of tasks, and in order to save limited computing resources, task vehicles usually only accept the fastest results, and redundant task offloading will cause a waste of computing resources.

[0096] Depend on Figure 3 It can be seen that the FDL task offloading algorithm also ensures the success rate of task offloading when the required computing resources are small, and is less affected by the vehicle's driving speed. This is because the goal of the FDL task offloading algorithm is to ensure that the connection between vehicles is as stable as possible, so the success rate of task offloading has basically not changed.

[0097] When considering offloading problems in practice, there are usually two issues to consider: computationally intensive issues and latency-sensitive issues. Figure 4 As shown in the figure, considering the delay-sensitive Internet of Vehicles task offloading problem, the shorter the acceptable task processing time, the lower the task completion rate of the traditional algorithm. However, the FDL task offloading algorithm will offload tasks to the service vehicle with the most computing resources, so it can still maintain a high task completion rate.

[0098] Figure 5 The average vehicle connection time in different speed ranges is given. It can be seen from the figure that the FDL task offloading algorithm maintains the stability of the connection during each round of task offloading, which means that the FDL task offloading algorithm can make the connection between vehicles as stable as possible while saving computing resources, thereby improving the success rate of task offloading.

[0099] Example 4

[0100] This embodiment provides an electronic device, including: a processor and a memory; the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the distributed Internet of Vehicles task unloading method as described in any one of Embodiments 1-3 is implemented.

[0101] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A distributed Internet of Vehicles task offloading method, characterized in that: include: Collect the task information of the task vehicles and the status information of the service vehicles in the preset area at the current moment; The task information of the task vehicle and the status information of the service vehicle are input into a pre-trained Internet of Vehicles task offloading model, and the Internet of Vehicles task offloading strategy is optimized using a federated deep learning algorithm to obtain an Internet of Vehicles task offloading solution that minimizes the average processing time of all tasks.

2. The distributed Internet of Vehicles task offloading method according to claim 1 is characterized in that: The task information of the task vehicle includes: the position of the task vehicle, the speed of the task vehicle and the task amount of the task generated by the task vehicle; The status information of the service vehicle includes: the location of the service vehicle, the speed of the service vehicle and the available computing resources on the service vehicle.

3. The distributed vehicle networking task offloading method according to claim 2 is characterized in that: The Internet of Vehicles task offloading model includes: a local model set on each service vehicle and a global model set on the cloud; The local model and the global model include: state space , Action Space And the reward function , the expression is: ； ； ； in, ; In the formula, For mission vehicles With service vehicles The relative speed of For mission vehicles To service vehicle The amount of unloaded tasks, For service vehicles Available computing resources on Mission vehicle With service vehicles The relative direction of is the number of mission vehicles at the current moment, Indicates service vehicle For mission vehicles The time to perform the mission, including the mission vehicle To service vehicle Time to uninstall the task and service vehicles Calculate the time of the task ,in, The expression is: ； For mission vehicles Whether to service vehicles Uninstall task variables, Mission vehicle To service vehicle Uninstall tasks, Mission vehicle No service vehicles Uninstall tasks; is the average processing time of all tasks at the current moment, The average processing time of all tasks at the next moment.

4. The distributed vehicle networking task offloading method according to claim 3 is characterized in that: Each service vehicle is only allowed to unload one of the tasks from one task vehicle per unit time; In response to the current moment, , then all tasks at the current moment are offloaded to the service vehicle for calculation; In response to the current moment, , then all tasks at the current moment are sorted in descending order according to the task amount, and Tasks with large workloads are offloaded to the service vehicle for calculation first, and the remaining tasks are calculated locally. in, is the number of vehicles in service at the current moment, For mission vehicles The number of tasks generated per unit time.

5. The distributed Internet of Vehicles task offloading method according to claim 4 is characterized in that: The method of optimizing the Internet of Vehicles task offloading strategy by using a federated deep learning algorithm includes: Obtaining action information of each service vehicle according to local model parameters and state information on each service vehicle; According to the action information and status information of each service vehicle and the task information of the task vehicle, the pre-established task offloading objective function is solved through the federated deep learning algorithm to obtain the Internet of Vehicles task offloading solution at the current moment.

6. The distributed vehicle networking task offloading method according to claim 5 is characterized in that: The task offloading objective function is: ； In the formula, Mission vehicle and service vehicles Connection duration, which is calculated as: ； in, is the maximum connectable distance between two vehicles.

7. The distributed vehicle networking task offloading method according to claim 5 is characterized in that: The method of optimizing the Internet of Vehicles task offloading strategy by using a federated deep learning algorithm also includes: After completing the current Internet of Vehicles task offloading, the local model on each service vehicle calculates and updates the local model parameters according to the current Internet of Vehicles task offloading solution; The global model in the cloud calculates the global reward based on the current Internet of Vehicles task offloading plan, and updates the global model parameters based on the local model parameters on each service vehicle; and synchronizes the updated global model parameters to the local model on each service vehicle.

8. The distributed vehicle networking task offloading method according to claim 7 is characterized in that: The local model on each service vehicle updates the local model parameters according to the Internet of Vehicles task offloading scheme at the current moment, including: ； ； ； In the formula, Accumulate rewards for local models, To express the expectation of future rewards, is the maximum cumulative reward of the local model, For service vehicles The local model parameters on .

9. The distributed vehicle networking task offloading method according to claim 8 is characterized in that: The updating of the global model parameters according to the local model parameters on each service vehicle includes: updating the global model parameters using the average value of the local model parameters uploaded by each service vehicle, and the expression is: ； In the formula, are global model parameters.

10. An electronic device, characterized in that: include: Processor and memory; The memory stores computer-readable instructions, which, when executed by the processor, implement the distributed Internet of Vehicles task offloading method according to any one of claims 1-9.

Images