A vehicle computing task scheduling method based on vehicle-edge-cloud collaboration

Abstract

The application discloses a vehicle computing task scheduling method based on vehicle-edge-cloud cooperation and belongs to the technical field of intelligent transportation systems and Internet of Vehicles. The method is aimed at the problems of limited computing resources of intelligent networked vehicles, high task delay and large energy consumption. An improved greedy algorithm is used to optimize the task execution order, and local greedy selection is performed according to the data volume, computation volume and maximum tolerable delay of the task to output an optimal platform allocation label. Then, a feedforward neural network is trained using the labels. The trained neural network can quickly predict an optimal platform vehicle-edge-cloud cooperative allocation scheme according to the input task characteristics in a real-time scenario. Through the fusion of the algorithm and the neural network, the application realizes efficient online real-time scheduling, significantly improves the task scheduling efficiency and system reliability, has excellent real-time performance and scalability, effectively reduces the task execution delay and energy consumption, and is suitable for efficient scheduling and resource allocation of large-scale computing tasks in the Internet of Vehicles.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems and vehicle networking technology, specifically to a vehicle computing task scheduling method based on vehicle-edge-cloud collaboration. Background Technology

[0002] With the widespread deployment of applications such as autonomous driving, image recognition, and intelligent navigation in intelligent connected vehicles, the number and complexity of computing tasks generated by in-vehicle systems have increased dramatically. Traditional local computing methods are limited by the vehicle's own computing power and struggle to meet the processing demands of latency-sensitive tasks. While cloud computing solutions offer powerful processing capabilities, they often face problems such as high communication latency, limited bandwidth resources, and high energy consumption in variable network environments. In recent years, with the development of edge computing technology, deploying multi-access edge computing (MEC) servers along roadsides to provide near-field computing support has gradually become an effective intermediate layer supplementing local and cloud computing. This has led to the formation of an in-vehicle-edge-cloud collaborative computing architecture, which has become a crucial foundation for task scheduling in intelligent transportation systems.

[0003] Most task scheduling methods in related technologies suffer from the following problems: Traditional task scheduling methods, such as full permutation search and heuristic algorithms, experience a sharp increase in computational complexity when faced with a large number of tasks, failing to meet real-time requirements; they cannot balance global optimization and real-time performance, as existing task scheduling methods often rely on greedy algorithms, which cannot simultaneously guarantee global optimization and real-time scheduling, easily leading to performance instability; traditional algorithms fail to fully consider the collaborative working mechanisms of in-vehicle, edge, and cloud platforms, resulting in low scheduling decision efficiency and failing to fully utilize the advantages of each platform. Therefore, how to ensure real-time performance while simultaneously achieving globally optimal task allocation has become an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a vehicle computing task scheduling method based on vehicle-edge-cloud collaboration. The proposed scheduling method, based on a vehicle-edge-cloud collaborative architecture, constructs a multi-objective optimization model with the goal of minimizing total system latency and energy consumption. It designs a task scheduling strategy that integrates an improved greedy algorithm and a feedforward neural network. Under the premise of meeting latency constraints, it can perform real-time optimal allocation and scheduling of onboard computing tasks among local, edge, and cloud platforms, minimizing overall latency and energy consumption and improving system service quality.

[0005] To achieve the above technical solution, the specific steps are as follows:

[0006] S1. Define the attributes of in-vehicle computing tasks, construct communication and computing models based on the in-vehicle-edge-cloud collaborative system architecture, and perceive the system status in real time;

[0007] The calculation methods for vehicle computing task attributes include: vehicle computing tasks can be calculated locally, calculated by offloading the task to an edge computing platform, and calculated by offloading it to a cloud computing platform;

[0008] Building a vehicle-edge-cloud collaborative system architecture includes: building a communication model for the vehicle-edge-cloud collaborative system and building a computing model for the vehicle-edge-cloud collaborative system;

[0009] The real-time perception system status is to collect dynamic information such as the vehicle's current speed, connection status with the edge server / cloud server, available computing resources of the edge server / cloud server, link bandwidth, and transmission success rate;

[0010] Vehicle-mounted computing tasks can be performed locally, offloaded to an edge computing platform, or offloaded to a cloud computing platform. The task must consider the workload, the number of computing cycles on each platform, and the maximum tolerable time latency.

[0011]

[0012] In the formula, D is the amount of data generated by the task, and C is the number of computing cycles required by the task. This is the maximum tolerable delay for the task.

[0013] At a certain moment, on a certain road segment, assuming each vehicle generates one computing task, then n vehicles generate n computing tasks; each task can only choose one computing platform for computing, a tasks choose the edge platform for computing, b tasks choose the cloud platform for computing, and nab tasks choose the local platform for computing.

[0014] S1.1 Based on the attributes of vehicle computing tasks, construct a communication model for the vehicle-edge-cloud collaborative system;

[0015] Construct a vehicle-edge-cloud collaborative system (VEC communication model), which specifically includes: a three-level heterogeneous computing node Node consisting of a task vehicle's onboard server (Task Vehicle), an edge server (MEC), and a cloud server (CS); only the uplink communication link E from vehicle to edge to cloud is considered; G is the overall model, and each node needs to consider the communication range radius R, computing power F, location coordinates P, data success reception probability S, and data packet transmission rate λ;

[0016] The specific expression for constructing the communication model of the vehicle-edge-cloud collaborative system is as follows:

[0017]

[0018] S1.2 Based on the attributes of vehicle-mounted computing tasks, construct a vehicle-edge-cloud collaborative system computing model;

[0019] Building a vehicle-edge-cloud collaborative system computing model includes: building a local computing model, building an edge computing model, and building a cloud computing model.

[0020] The specific construction includes the following steps:

[0021] S1.2.1 Define the number of vehicle-mounted tasks and their allocation scheme;

[0022] At a certain moment, on a certain road segment, assuming each vehicle generates one computing task, then n vehicles generate n computing tasks; each task can only choose one computing platform for computing, a tasks choose the edge platform for computing, b tasks choose the cloud platform for computing, and nab tasks choose the local platform for computing.

[0023] S1.2.2 Based on the defined number of vehicle-mounted tasks and allocation scheme, construct a local computing model;

[0024] Since onboard tasks are computed locally without data transmission, local platform computation only includes computation latency, and local platform computation does not generate transmission energy consumption; energy consumption only includes computation energy consumption, as shown in the following expression:

[0025]

[0026] In the formula, Let represent the computation time of task i (the i-th task) on the local platform, where i ∈ nab; This represents the number of computer computing cycles required for task i. Indicates the computing power of the onboard computer; This represents the transmission delay of task i during local platform computation; This indicates the total latency of tasks in local platform computing; This represents the energy consumption of task i during local platform computation; This represents the computational energy factor of the task in local platform computation. This represents the energy transferred during the computation of task i on the local platform. This represents the total energy consumed by the task in local platform computation; n is the number of vehicles; a represents the task selected for edge platform computation; and b represents the task selected for cloud platform computation.

[0027] S1.2.3. Based on the defined number of vehicle-mounted tasks and allocation scheme, construct an edge computing model;

[0028] Offloading vehicle-mounted tasks to the edge platform for computation requires considering the resulting transmission latency, computation latency, and energy loss from edge computing. The objective function for offloading vehicle-mounted tasks to the edge platform is constructed as follows:

[0029]

[0030] In the formula, Indicates task (Assigned to the edge computing platform) Data transmission delay from one task to the roadside unit. ∈a; Indicates task The size of the data; Indicates task Data transfer rate to RSU; Indicates task Computational latency in multi-access edge environments; Indicates task The required number of edge platform computing power cycles, Indicates task -1 The number of edge platform computing power cycles required, C1 <C2<…<Ca; This indicates the computing power of multiple access edge computers; This represents the total latency of all tasks in edge platform computing; Indicates task Energy consumption for transmitting data to roadside units; This represents the energy transfer factor of a task in edge computing; For the task Transmit power of the uplink to the roadside unit; This represents the computational energy factor of the task in multi-access edge computing; This represents the total energy consumption of edge computing.

[0031] S1.2.4. Based on the defined number of in-vehicle tasks and allocation scheme, construct a cloud computing model;

[0032] Offloading vehicle-mounted tasks to a cloud platform for computation requires consideration of transmission latency, computation latency, and cloud computing energy consumption. The objective function for offloading vehicle-mounted tasks to the cloud platform is constructed as follows:

[0033]

[0034] In the formula, Indicates task (The first one allocated to the cloud computing platform) Data transmission latency from one task to the cloud server ∈b; Indicates task Data transfer rate to cloud server =0.9 ; Indicates task The size of the data; Indicates task Computational latency in cloud servers; This indicates the computing power of the cloud server computer; Indicates task The required number of cloud server computing power cycles, C1 <C2<…<Cb; Indicates task -1 is the number of cloud server computing power cycles required; This represents the total latency of all tasks in cloud computing; Indicates task Energy consumption for data transmission to the cloud server; This represents the energy transfer factor of a task during cloud server computation. For the task The transmit power of the uplink to the cloud server; This represents the computational energy factor of the task in cloud server computing; This represents the total energy consumption of all tasks in cloud computing.

[0035] S2. Construct the globally optimal objective function;

[0036] The method for constructing the globally optimal objective function is as follows: based on the vehicle mission attributes and system state, a multi-objective optimization function is constructed to obtain the globally optimal objective function;

[0037] This invention constructs a globally optimal objective function with the overall goal of minimizing task processing time, minimizing task energy consumption, or optimizing the overall service quality of the system. It then constructs a multi-objective optimization function that integrates energy consumption, time, and cost. Under the condition of satisfying task latency constraints, it determines in real time the optimal computing platform to be selected for each task, i.e., the number of local computing, edge computing, and cloud computing tasks, thereby minimizing task execution time and energy consumption, and deriving the globally optimal objective function. The specific expression is:

[0038]

[0039] In the formula, T represents the time required to process all tasks. This represents the maximum tolerable latency when processing task i. This indicates the time required to process task i locally. This indicates the time required to transmit data to the edge platform for processing task i. Let E represent the time required for the cloud platform to process task i, E represent the energy consumption for processing all tasks, and M1 and M2 represent the weights of time and energy, respectively. M1 represents the weight ratio of task execution time in the overall objective function, and M2 represents the weight ratio of task execution energy consumption in the overall objective function.

[0040] S3. Based on the decomposition of the global optimal objective function into local optimal objective functions, an improved greedy algorithm is used to obtain a preliminary global optimal task allocation strategy and construct a training dataset. The feedforward neural network is then trained offline to obtain a trained feedforward neural network model.

[0041] S3.1. Decompose the global optimal objective function into local optimal objective functions; decompose the global optimal objective function into local optimal objective functions for each task to obtain the optimization index function;

[0042] The global optimal objective function is decomposed into local optimal objective functions for each task, and the task quality function (QoS) is established as the optimization index function, with the following expression:

[0043]

[0044] Among them, T i E represents the time required to process task i. i This represents the energy consumption for processing task i.

[0045] S3.2 Based on the optimization index function, improve the greedy algorithm to obtain the globally optimal task allocation strategy;

[0046] First, the task set is enumerated with all permutations, resulting in n! task sequences. Each task order represents a different scheduling execution strategy. Given the task order, the QoS value of each task is calculated on the local platform, edge platform, and cloud platform. A local greedy strategy is used to select the current optimal platform, and the selection result is accumulated into the current scheduling cost. All permutation strategies are traversed, and the corresponding total QoS cost is recorded. Finally, the combination with the minimum total cost is selected as the globally optimal task allocation strategy.

[0047] S3.3 Construct a dataset based on the global optimal task allocation strategy. After performing preprocessing operations on the dataset, divide it into a training dataset and a test dataset. Use the training set to train the feedforward neural network offline to obtain a trained feedforward neural network model.

[0048] The preprocessing operation on the dataset is as follows: the input features [S, R, D, C, T] of the original task are processed... max ]Simplified to [T f ve , C, T maxWhere S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This represents the latency of the task on the platform; each task feature vector in the final input dataset contains 3×m feature data, where m is the number of vehicles in the batch;

[0049] In this invention, the preprocessed dataset is divided into a training dataset and a test dataset in an 8:2 ratio;

[0050] First, the dataset is preprocessed. To reduce the complexity of the task dataset and improve the training speed and efficiency of the feedforward neural network, the original task input features [S, R, D, C, T] are preprocessed. max ]Simplified to [T f ve , C, T max Where S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This represents the latency of the task on the platform; each task feature vector in the final input dataset contains 3×m feature data, where m is the number of vehicles in the batch;

[0051] The preprocessed training dataset is input into the feedforward neural network, passing through the input layer and hidden layer of the feedforward neural network in sequence. After passing through the first fully connected layer, a linearly transformed output is obtained. Then, a ReLU activation function layer is used to introduce nonlinearity, resulting in nonlinear data. Next, the nonlinear data is passed through the second fully connected layer and linearly transformed again. Finally, the output result is obtained after processing through the second ReLU activation function layer and the batch normalization layer.

[0052] The expression for the vector processing process in a fully connected layer is as follows:

[0053]

[0054] In the formula, x is the input vector, W is the weight matrix, c is the bias vector, and y is the output vector;

[0055] The ReLU activation function expression is as follows:

[0056]

[0057] In the formula, It is the input to the activation function. The ReLU function sets all negative values ​​to 0, while positive values ​​remain unchanged.

[0058] Batch normalization layers are used to reduce internal covariate bias, improving model stability and training speed, and effectively reducing the risk of overfitting. The expression for the batch normalization layer is shown below:

[0059]

[0060] In the formula, It is the input data. It is the variance of the mini-batch data, f(x) b ) is the output data. It is the mean of a small batch of data. >0, used to prevent the denominator from being 0;

[0061] The output layer receives the results from the hidden layer and manipulates the output to finally select the result; the fully connected layer obtains the vector of the computing platform category corresponding to each task from the output of the hidden layer; the Softmax layer calculates the probability of the task belonging to each computing platform; and the regression layer determines the final computing platform for each task based on the probability distribution of the Softmax layer.

[0062] The expression for Softmax processing vectors is shown below:

[0063]

[0064] In the formula, Indicates that the layer belonging to the Softmax layer is the 1st generation. Input values ​​for the task; Indicates to Perform exponentiation to ensure all values ​​are positive. The summation of the exponent values ​​for all categories is used for normalization, so that the sum of the probabilities of all categories is 1.

[0065] The FNN model sorts computing platforms as local, edge, and cloud, and then converts them into one-hot encoding format; 0 indicates that the task does not select the corresponding computing platform, and 1 indicates that the task selects the corresponding computing platform. The output layer finally outputs one-hot labels for the three platform categories (local, edge, and cloud), which are the assignment predictions for each task.

[0066] The remaining computing power (CPU / GPU computing power) of the perception platform (local, edge, cloud) will increase the service cost of subsequent tasks if the current task selects edge computing or cloud computing due to limited computing and communication resources.

[0067] In this invention, the test dataset obtained by the improved greedy algorithm is first used as supervised training data to train the feedforward neural network (FNN) model offline. The dataset is divided into training and test datasets in an 8:2 ratio to train the feedforward neural network model. The fully connected layer of the feedforward neural network (FNN) consists of 200 neurons. The key parameters of the FNN model are as follows: maximum number of training rounds = 200, initial learning rate = 0.001, learning rate scheduling = segmented learning rate scheduling, learning rate decay factor = 0.9, learning rate decay period = 10, mini-batch size = 70, and validation frequency = 30.

[0068] S4. Based on the trained feedforward neural network model and real-time task features, perform preprocessing operations on the real-time task features, predict the optimal allocation scheme after online inference through the trained feedforward neural network model, and output task allocation and scheduling instructions.

[0069] This invention is based on a trained feedforward neural network model and real-time task features. After performing preprocessing operations on the real-time task features, the optimal allocation scheme is predicted through online inference by the trained feedforward neural network model, and task allocation and scheduling instructions are output.

[0070] The method for performing preprocessing operations on real-time task features is as follows: [S, R, D, C, T] max ]Simplified to [T f ve , C, T max Where S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This represents the latency of the task on the platform; each task feature vector in the final input dataset contains 3×m feature data, where m is the number of vehicles in the batch;

[0071] The real-time task feature data is sequentially input into the input layer and hidden layer of the feedforward neural network model. After passing through the first fully connected layer, the input real-time task feature data is linearly transformed. Then, a non-linear transformation is introduced through a ReLU activation function layer to enhance the model's expressive power. The non-linear data is linearly transformed again after passing through the second fully connected layer and then output as a linear result after passing through the second ReLU activation function layer. The linear result is transmitted to the output layer for processing. The fully connected operation of the output layer obtains the computing platform category vector corresponding to each task. The Softmax layer predicts the probability that the task belongs to each platform for computing. After regression processing, the final computing platform for each task is determined. The output is a one-hot label of the task allocation and scheduling instructions for three types of platforms (local, edge, and cloud), which is the classification prediction of each task, realizing real-time online inference of the model.

[0072] S5. Based on task allocation and scheduling instructions, tasks are allocated to designated platforms for execution. The allocation priority is dynamically adjusted according to real-time resource utilization and network conditions. This completes the vehicle computing task scheduling method.

[0073] The designated platforms include local platforms, edge platforms, and cloud platforms;

[0074] This invention allocates tasks to designated platforms (local, edge, cloud) for computation based on task allocation and scheduling instructions, and dynamically adjusts allocation priorities according to real-time resource utilization and network conditions to ensure the optimality of the scheduling strategy in real time.

[0075] The beneficial effects of this invention are:

[0076] (1) By establishing a communication and computing model for the vehicle-edge-cloud collaborative system and rationally allocating tasks among local computing, edge computing and cloud computing platforms, this invention can achieve the goals of minimizing task latency, reducing energy consumption and optimizing overall service quality, thereby effectively improving system performance.

[0077] (2) By introducing the strategy of “decomposing the global target into local targets and combining it with an improved greedy algorithm”, this invention can quickly obtain an approximate optimal solution during task scheduling, significantly reducing the computational complexity caused by traditional global enumeration, meeting the real-time requirements of vehicle computing, and improving task scheduling speed.

[0078] (3) By using the globally optimal allocation strategy generated by the greedy algorithm as training data and combining it with the feedforward neural network for offline supervised learning, this invention can directly predict the optimal allocation result of the task in the online stage, thereby reducing the task execution time and improving the system's utilization of computing resources.

[0079] (4) Through the synergistic effect of the greedy algorithm and the neural network model, this invention can improve the accuracy of task allocation: on the one hand, the greedy algorithm can generate high-quality optimized solutions as training labels under small-scale tasks; on the other hand, the neural network can accurately classify task features and make rapid inferences. The combination of the two ensures the high accuracy and stability of the task allocation results. Attached Figure Description

[0080] Figure 1 This is a flowchart of the steps of the present invention;

[0081] Figure 2 This is a communication model diagram of the present invention;

[0082] Figure 3 This is an example diagram showing the output results of the feedforward neural network (FNN) model of the present invention;

[0083] Figure 4 This is a comparison chart showing the solution time of the FNN model trained according to this invention and the greedy algorithm;

[0084] Figure 5 This is a comparison chart of results in embodiments of the present invention with the goal of minimizing task energy consumption;

[0085] Figure 6 This is a comparison chart showing the results of minimizing task processing time as the optimal objective in an embodiment of the present invention. Detailed Implementation

[0086] The present invention will be further described in detail below with reference to specific embodiments. The preferred embodiments described below are only examples, and those skilled in the art can conceive of other obvious variations.

[0087] Reference Figure 1 As shown, a vehicle-edge-cloud computing task scheduling method based on the fusion of an improved greedy algorithm and a feedforward neural network includes the following steps:

[0088] S1. Define the attributes of in-vehicle computing tasks, construct communication and computing models based on the in-vehicle-edge-cloud collaborative system architecture, and perceive the system status in real time;

[0089] The calculation methods for vehicle computing task attributes include: vehicle computing tasks can be calculated locally, calculated by offloading the task to an edge computing platform, and calculated by offloading it to a cloud computing platform;

[0090] Building a vehicle-edge-cloud collaborative system architecture includes: building a communication model for the vehicle-edge-cloud collaborative system and building a computing model for the vehicle-edge-cloud collaborative system;

[0091] The real-time sensing system status is achieved by collecting dynamic information such as the vehicle's current speed, connection status with the edge server / cloud server, available computing resources of the edge server / cloud server, link bandwidth, and transmission success rate, which is used to update the system status.

[0092] Vehicle-mounted computing tasks can be performed locally, offloaded to an edge computing platform, or offloaded to a cloud computing platform. The task must consider the workload, the number of computing cycles on each platform, and the maximum tolerable time latency.

[0093]

[0094] In the formula, D is the amount of data generated by the task, and C is the number of computing cycles required by the task. This is the maximum tolerable delay for the task.

[0095] At a certain moment, on a certain road segment, assuming each vehicle generates one computing task, then n vehicles generate n computing tasks; each task can only choose one computing platform for computing, a tasks choose the edge platform for computing, b tasks choose the cloud platform for computing, and nab tasks choose the local platform for computing.

[0096] S1.1 Construct a communication model for the vehicle-edge-cloud collaborative system;

[0097] The vehicle-edge-cloud collaborative system (VEC communication model) specifically includes: a three-tiered heterogeneous computing node system consisting of the vehicle's onboard server (TaskVehicle), edge server (MEC), and cloud server (CS); such as Figure 2 As shown, only the uplink communication link E from vehicle to edge to cloud is considered; G is the overall model, and each node needs to consider the communication range radius R, computing power F, location coordinates P, data successful reception probability S, and data packet transmission rate λ; the specific expression is:

[0098]

[0099] S1.2 Construct a computing model for the vehicle-edge-cloud collaborative system;

[0100] Building a vehicle-edge-cloud collaborative system computing model includes: building a local computing model, an edge computing model, and a cloud computing model; the specific construction includes the following steps:

[0101] S1.2.1 Define the number of vehicle-mounted tasks and their allocation scheme;

[0102] At a certain moment, on a certain road segment, assuming each vehicle generates one computing task, then n vehicles generate n computing tasks; each task can only choose one computing platform for computing, a tasks choose the edge platform for computing, b tasks choose the cloud platform for computing, and nab tasks choose the local platform for computing.

[0103] S1.2.2, Construct a local computing model;

[0104] Since onboard tasks are computed locally without data transmission, local platform computation only includes computation latency, and local platform computation does not generate transmission energy consumption; energy consumption only includes computation energy consumption, as shown in the following expression:

[0105]

[0106] In the formula, Let represent the computation time of task i (the i-th task) on the local platform, where i ∈ nab; This represents the number of computer computing cycles required for task i. Indicates the computing power of the onboard computer; This represents the transmission delay of task i during local platform computation; This indicates the total latency of tasks in local platform computing; This represents the energy consumption of task i during local platform computation; This represents the computational energy factor of the task in local platform computation. This represents the energy transferred during the computation of task i on the local platform. This represents the total energy consumed by the task in local platform computation, where n is the number of vehicles, a represents the task selected for edge platform computation, and b represents the task selected for cloud platform computation.

[0107] S1.2.3, Construct an edge computing model;

[0108] Offloading vehicle-mounted tasks to the edge platform for computation requires considering the resulting transmission latency, computation latency, and energy loss from edge computing. The objective function for offloading vehicle-mounted tasks to the edge platform is constructed as follows:

[0109]

[0110] In the formula, Indicates task (Assigned to the edge computing platform) Data transmission delay from one task to the roadside unit. ∈a; Indicates task The size of the data; Indicates task Data transfer rate to RSU; Indicates task Computational latency in multi-access edge environments; Indicates task The required number of edge platform computing power cycles, Indicates task -1 The number of edge platform computing power cycles required, C1 <C2<…<Ca; This indicates the computing power of multiple access edge computers; This represents the total latency of all tasks in edge platform computing; Indicates task Energy consumption for transmitting data to roadside units; This represents the energy transfer factor of a task in edge computing; For the task Transmit power of the uplink to the roadside unit; This represents the computational energy factor of the task in multi-access edge computing; This represents the total energy consumption of edge computing;

[0111] S1.2.4 Constructing an edge computing model;

[0112] Offloading vehicle-mounted tasks to a cloud platform for computation requires consideration of transmission latency, computation latency, and cloud computing energy consumption. The objective function for offloading vehicle-mounted tasks to the cloud platform is constructed as follows:

[0113]

[0114] In the formula, Indicates task (The first one allocated to the cloud computing platform) Data transmission latency from one task to the cloud server ∈b; Indicates task Data transfer rate to cloud server =0.9 ; Indicates task The size of the data; Indicates task Computational latency in cloud servers; This indicates the computing power of the cloud server computer; Indicates task The required number of cloud server computing power cycles, C1 <C2<…<Cb; Indicates task -1 is the number of cloud server computing power cycles required; This represents the total latency of all tasks in cloud computing; Indicates task Energy consumption for data transmission to the cloud server; This represents the energy transfer factor of a task during cloud server computation. For the task The transmit power of the uplink to the cloud server; This represents the computational energy factor of the task in cloud server computing; This represents the total energy consumption of all tasks in cloud computing;

[0115] This embodiment will illustrate an exemplary implementation scenario of the vehicle-edge-cloud computing task scheduling strategy that integrates an improved greedy algorithm with a feedforward neural network; in this embodiment, each task has three corresponding computing platform categories;

[0116] To verify the superiority of the method of this invention, joint simulation was performed using Matlab and SUMO. A real-world scenario was constructed, consisting of a 1000-meter-long, 8-meter-wide one-way, two-lane highway. A Roadside Unit (RSU) was installed at each end of the road, and the cloud server CS was connected to the RSU via wired communication. The traffic flow was 2000 vehicles per hour. The simulation time slot was 0.1 seconds, and the simulation time was set to 70 seconds. To achieve the overall goal of minimizing latency, reducing energy consumption, or optimizing the overall service quality of the system, we adopted the task scheduling method proposed in this strategy to rationally allocate computing platforms.

[0117] S2. Construct the globally optimal objective function;

[0118] The method for constructing the globally optimal objective function is as follows: based on the vehicle mission attributes and system state, a multi-objective optimization function is constructed to obtain the globally optimal objective function;

[0119] This invention constructs a globally optimal objective function with the overall goal of minimizing task processing time, minimizing task energy consumption, or optimizing the overall service quality of the system. It then constructs a multi-objective optimization function that integrates energy consumption, time, and cost. Under the condition of satisfying task latency constraints, it determines in real time the optimal computing platform to be selected for each task, i.e., the number of local computing, edge computing, and cloud computing tasks, thereby minimizing task execution time and energy consumption, and deriving the globally optimal objective function. The specific expression is:

[0120]

[0121] In the formula, T represents the time required to process all tasks. This represents the maximum tolerable latency when processing task i. This indicates the time required to process task i locally. This indicates the time required to transmit data to the edge platform for processing task i. E represents the time required for the cloud platform to process task i, E represents the energy consumption for processing all tasks, and M1 and M2 are the weights of time and energy, which are determined by different switching strategies.

[0122] In the simulation verification, two control parameters, M1 and M2, were set to represent the weight ratios of task execution time and energy consumption in the overall objective function, respectively. A joint simulation was performed using Matlab and SUMO. A real-world scenario was constructed: a one-way, two-lane highway with a length of 1000 meters and a width of 8 meters. By setting M1=1 and M2=0, the goal was to minimize task delay. Figure 6 As shown, or by setting M1=0 and M2=1 with the goal of minimizing energy consumption, such as... Figure 5 As shown, the scheduling results indicate that the method of this invention can flexibly adapt to different target orientations and consistently demonstrates better overall performance than the single-platform strategy in simulations, verifying the effectiveness of the proposed global optimization model.

[0123] S3. Based on the decomposition of the global optimal objective function into local optimal objective functions, an improved greedy algorithm is used to obtain a preliminary global optimal task allocation strategy and construct a training dataset. The feedforward neural network is then trained offline to obtain a trained feedforward neural network model.

[0124] S3.1. Decompose the global optimal objective function into local optimal objective functions; decompose the global optimal objective function into local optimal objective functions for each task to obtain the optimization index function;

[0125] The global optimal objective function is decomposed into local optimal objective functions for each task, and the task quality function (QoS) is established as the optimization index function, with the following expression:

[0126]

[0127] Among them, T i E represents the time required to process task i. i This represents the energy consumption for processing task i;

[0128] S3.2 Based on the optimization index function, improve the greedy algorithm to obtain the globally optimal task allocation strategy;

[0129] First, the task set is enumerated by full permutation, with n! task sequences. Each task order represents a different scheduling execution strategy. Given the task order, the QoS value of each task is calculated on the local platform, edge platform, and cloud platform. A local greedy strategy is used to select the current optimal platform, and the selection result is accumulated into the current scheduling cost. All permutation strategies are traversed, and the corresponding total QoS cost is recorded. Finally, the combination with the minimum total cost is selected as the globally optimal task allocation strategy.

[0130] S3.3 Construct a dataset based on the global optimal task allocation strategy. After performing preprocessing operations on the dataset, divide it into a training dataset and a test dataset. Use the training set to train the feedforward neural network offline to obtain a trained feedforward neural network model.

[0131] The preprocessing operation on the dataset is as follows: the input features [S, R, D, C, T] of the original task are processed... max ]Simplified to [T f ve , C, T max Where S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This represents the latency of the task on the platform; each task feature vector in the final input dataset contains 3×m feature data, where m is the number of vehicles in the batch;

[0132] In this invention, the preprocessed dataset is divided into a training dataset and a test dataset in an 8:2 ratio;

[0133] First, the dataset is preprocessed. To reduce the complexity of the task dataset and improve the training speed and efficiency of the feedforward neural network, the original task input features [S, R, D, C, T] are preprocessed. max ]Simplified to [T f ve , C, T max Where S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This represents the latency of the task on the platform; each task feature vector in the final input dataset contains 3×m feature data.

[0134] The preprocessed training dataset is input into the feedforward neural network, passing through the input layer and hidden layer of the feedforward neural network in sequence. After passing through the first fully connected layer, a linearly transformed output is obtained. Then, a ReLU activation function layer is used to introduce nonlinearity, resulting in nonlinear data. Next, the nonlinear data is passed through the second fully connected layer and linearly transformed again. Finally, the output result is obtained after processing through the second ReLU activation function layer and the batch normalization layer.

[0135] The expression for the vector processing process in a fully connected layer is as follows:

[0136]

[0137] In the formula, x is the input vector, W is the weight matrix, c is the bias vector, and y is the output vector;

[0138] The ReLU activation function expression is as follows:

[0139]

[0140] In the formula, It is the input to the activation function; the ReLU function sets all negative values ​​to 0, while positive values ​​remain unchanged;

[0141] Batch normalization layers are used to reduce internal covariate bias, improving model stability and training speed, and effectively reducing the risk of overfitting. The expression for the batch normalization layer is shown below:

[0142]

[0143] In the formula, It is the input data. It is the variance of the mini-batch data, f(x) b ) is the output data. It is the mean of a small batch of data. >0, used to prevent the denominator from being 0;

[0144] The output layer receives the results from the hidden layer and manipulates the output to finally select the result; the fully connected layer obtains the vector of the computing platform category corresponding to each task from the output of the hidden layer; the Softmax layer calculates the probability of the task belonging to each computing platform; and the regression layer determines the final computing platform for each task based on the probability distribution of the Softmax layer.

[0145] The expression for Softmax processing vectors is shown below:

[0146]

[0147] In the formula, Indicates that the layer belonging to the Softmax layer is the 1st generation. Input values ​​for the task; Indicates to Perform exponentiation to ensure all values ​​are positive. The summation of the exponent values ​​for all categories is used for normalization, so that the sum of the probabilities of all categories is 1.

[0148] The FNN model sorts computing platforms as local, edge, and cloud, and then converts them into one-hot encoding format; 0 indicates that the task does not select the corresponding computing platform, and 1 indicates that the task selects the corresponding computing platform. The output layer finally outputs one-hot labels for the three platform categories (local, edge, and cloud), which are the assignment predictions for each task.

[0149] The remaining computing power (CPU / GPU computing power) of the perception platform (local, edge, cloud) will increase the service cost of subsequent tasks if the current task selects edge computing or cloud computing due to limited computing and communication resources.

[0150] In this embodiment, the test dataset obtained by the improved greedy algorithm is first used as supervised training data to train the feedforward neural network (FNN) model offline. The greedy algorithm minimizes the function in different task sequences. To optimize the target, local, edge, or cloud platforms were progressively selected as task allocation platforms, resulting in multiple optimal task allocation results. This embodiment obtained 14,000 sets of vehicle-edge-cloud task scheduling data through a globally optimal task allocation strategy. These data were divided into training and testing datasets in an 8:2 ratio for training a feedforward neural network model. The fully connected layer of the feedforward neural network (FNN) consists of 200 neurons, with the following key parameters: maximum training epochs = 200, initial learning rate = 0.001, learning rate scheduling = segmented learning rate scheduling, learning rate decay factor = 0.9, learning rate decay period = 10, mini-batch size = 70, and validation frequency = 30. Figure 3 As shown in the figure, the FNN model output is an example. After training, the FNN model can quickly predict any given task feature vector. Without having to traverse all task combinations, it can directly output the optimal allocation strategy of the current task in three computing platforms (local, edge, and cloud), thus improving the real-time performance of scheduling decisions.

[0151] S4. Based on the trained feedforward neural network model and real-time task features, perform preprocessing operations on the real-time task features, predict the optimal allocation scheme after online inference through the trained feedforward neural network model, and output task allocation and scheduling instructions.

[0152] This invention is based on a trained feedforward neural network model and real-time task features. After performing preprocessing operations on the real-time task features, the optimal allocation scheme is predicted through online inference by the trained feedforward neural network model, and task allocation and scheduling instructions are output.

[0153] The method for performing preprocessing operations on real-time task features is as follows: [S, R, D, C, T] max ]Simplified to [T f ve , C, T max Where S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This represents the latency of the task on the platform; each task feature vector in the final input dataset contains 3×m feature data, where m is the number of vehicles in the batch;

[0154] The real-time task feature data is sequentially input into the input layer and hidden layer of the feedforward neural network model. After passing through the first fully connected layer, the input real-time task feature data is linearly transformed. Then, a non-linear transformation is introduced through a ReLU activation function layer to enhance the model's expressive power. The non-linear data is linearly transformed again after passing through the second fully connected layer and then output as a linear result after passing through the second ReLU activation function layer. The linear result is transmitted to the output layer for processing. The fully connected operation of the output layer obtains the computing platform category vector corresponding to each task. The Softmax layer predicts the probability that a task belongs to each platform for computing. After regression processing, the final computing platform for each task is determined. The output layer outputs one-hot labels for task allocation and scheduling instructions for three types of platforms (local, edge, and cloud), which is the classification prediction for each task, realizing real-time online inference of the model.

[0155] S5. Based on task allocation and scheduling instructions, tasks are allocated to designated platforms for execution. The allocation priority is dynamically adjusted according to real-time resource utilization and network conditions. This completes the vehicle computing task scheduling method.

[0156] The designated platforms include local platforms, edge platforms, and cloud platforms;

[0157] This invention allocates tasks to designated platforms (local, edge, cloud) for computation based on task allocation and scheduling instructions, and dynamically adjusts allocation priorities according to real-time resource utilization and network conditions to ensure the optimality of the scheduling strategy in real time.

[0158] To verify the real-time performance and effectiveness of the aforementioned FNN model, this embodiment further compares the trained FNN model with the improved greedy algorithm during task scheduling; for example... Figure 4 As shown, within the simulation time range of 0 to 70 seconds, the computation time required for the FNN model and the greedy algorithm to output the optimal task allocation strategy during task scheduling was recorded respectively. In the figure, the black curve represents the solution time of the greedy algorithm, the red curve represents the prediction time of the FNN model, the horizontal axis is the simulation time, and the vertical axis is the time required for the two algorithms to calculate an allocation scheme (unit: milliseconds).

[0159] The experimental results show that the inference time of the FNN model remains relatively constant, while the solution time of the greedy algorithm increases significantly with the number of tasks. The results can be divided into five states: State 1 (0–4 seconds), when there are 1–2 tasks, the average solution time of the greedy algorithm is only a few milliseconds, slightly better than the FNN; State 2 (61–64 seconds), when there are 3 tasks, the average solution time of the greedy algorithm is similar to or slightly longer than that of the FNN; State 3 (5–12 seconds), when there are 4 tasks, the average solution time of the greedy algorithm exceeds 10 milliseconds, significantly longer than the FNN model; State 4 (25–29 seconds), when there are 5 tasks, the average solution time of the greedy algorithm exceeds 70 milliseconds; State 5 (16–19 seconds), when there are 6 tasks, the solution time of the greedy algorithm exceeds 400 milliseconds, far exceeding the inference time of the FNN model.

[0160] In summary, the results show that the FNN model has a significant advantage in inference speed when dealing with large-scale multi-task scheduling problems. Its performance is not affected by the increase in the number of tasks, and it can stably and quickly output task allocation strategies, thereby meeting the real-time requirements of heterogeneous computing environments of vehicle-edge-cloud, verifying the effectiveness and feasibility of the method proposed in this invention.

[0161] This invention's scheduling system supports dynamically allocating each task to one of three types of computing platforms based on its current state: local platform, edge platform, or cloud platform. During inference, the trained feedforward neural network (FNN) model receives the task feature vector as input and, combined with the current system resource state, makes the optimal platform allocation decision for the task. Experimental results show that the FNN model maintains a stable and fast inference speed even with an increasing number of tasks. Compared to traditional greedy algorithms, its decision latency is significantly lower, making it suitable for in-vehicle computing scenarios requiring rapid response. Furthermore, the task scheduling model incorporates a system state awareness mechanism, enabling it to generate new allocation strategies and update execution priorities in real time when resource utilization, task load, and network conditions change, thereby achieving dynamic adaptive task scheduling and target optimization.

[0162] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the design concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A vehicle computing task scheduling method based on vehicle-edge-cloud collaboration, characterized in that, Includes the following steps: S1. Define the attributes of in-vehicle computing tasks, construct communication and computing models based on the in-vehicle-edge-cloud collaborative system architecture, and perceive the system status in real time; The calculation methods for the vehicle computing task attributes include: the vehicle computing task can be calculated locally, calculated by offloading the task to an edge computing platform, and calculated by offloading it to a cloud computing platform. The construction of the vehicle-edge-cloud collaborative system architecture includes: constructing a communication model for the vehicle-edge-cloud collaborative system and constructing a computing model for the vehicle-edge-cloud collaborative system; The real-time perception system status includes collecting the vehicle's current speed and connection status with the edge server / cloud server; collecting dynamic information on the vehicle's current speed, connection status with the edge server / cloud server, available computing resources of the edge server / cloud server, link bandwidth, and transmission success rate. S2. Construct the globally optimal objective function; The method for constructing the globally optimal objective function is as follows: based on the vehicle mission attributes and system state, a multi-objective optimization function is constructed to obtain the globally optimal objective function; S3. Based on the decomposition of the global optimal objective function into local optimal objective functions, an improved greedy algorithm is used to obtain a preliminary global optimal task allocation strategy and construct a training dataset. The feedforward neural network is then trained offline to obtain a trained feedforward neural network model. S3.

1. Decompose the global optimal objective function into local optimal objective functions; decompose the global optimal objective function into local optimal objective functions for each task to obtain the optimization index function; The global optimal objective function is decomposed into local optimal objective functions for each task, and the task quality function (QoS) is established as the optimization index function, with the following expression: ； Among them, T i E represents the time required to process task i. i M1 represents the energy consumption for processing task i, M2 represents the weight ratio of task execution time in the overall objective function, and M3 represents the weight ratio of task execution energy consumption in the overall objective function. S3.2 Based on the optimization index function, improve the greedy algorithm to obtain the globally optimal task allocation strategy; The task set is enumerated by performing full permutations, resulting in n! task sequences. Given the task order, the QoS value of each task is calculated on the local platform, edge platform, and cloud platform. A local greedy strategy is used to select the current optimal platform, and the selection result is accumulated into the current scheduling cost. All permutation strategies are iterated over, and the corresponding total QoS values ​​are recorded. Finally, the combination with the minimum total cost is selected as the globally optimal task allocation strategy. S3.3 Construct a dataset based on the global optimal task allocation strategy. After performing preprocessing operations on the dataset, divide it into a training dataset and a test dataset. Use the training set to train the feedforward neural network model offline to obtain a trained feedforward neural network model. S4. Based on the trained feedforward neural network model and real-time task features, perform preprocessing operations on the real-time task features, predict the optimal allocation scheme after online inference through the trained feedforward neural network model, and output task allocation and scheduling instructions. S5. Based on the task allocation and scheduling instructions, the task is allocated to the designated platform for execution. The allocation priority is dynamically adjusted according to the real-time resource utilization and network conditions. The task is then allocated to the designated platform for execution, thus completing the vehicle computing task scheduling method. The designated platforms include local platforms, edge platforms, and cloud platforms.

2. The vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 1, characterized in that, The specific expression for the vehicle computing task attribute Task in step S1 is as follows: ； In the formula, D is the amount of data generated by the task, and C is the number of computing cycles required by the task. This is the maximum tolerable delay for the task.

3. The vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 1, characterized in that, The construction of the vehicle-edge-cloud collaborative system architecture in step S1 specifically includes the following steps: S1.1 Based on the attributes of vehicle computing tasks, construct a communication model for the vehicle-edge-cloud collaborative system; The communication model for the vehicle-edge-cloud collaborative system includes: a three-level heterogeneous computing node consisting of the vehicle's onboard server, edge server, and cloud server, and an uplink communication link from the vehicle to the edge to the cloud. S1.2 Based on the attributes of vehicle-mounted computing tasks, construct a vehicle-edge-cloud collaborative system computing model; The construction of the vehicle-edge-cloud collaborative system computing model includes: constructing a local computing model, constructing an edge computing model, and constructing a cloud computing model.

4. The vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 3, characterized in that, In step S1.1, constructing the vehicle-edge-cloud collaborative system communication model specifically includes: a three-tiered heterogeneous computing node Node consisting of the vehicle's onboard server, edge server, and cloud server, and an uplink communication link E from the vehicle to the edge to the cloud; the specific expression of the vehicle-edge-cloud collaborative system communication model is as follows: ； In the formula, G represents the overall model, R represents the communication range radius, F represents computing power, P represents location coordinates, S represents the probability of successful data reception, λ represents the data packet transmission rate, Task represents the vehicle-mounted computing task attribute, D represents the amount of data generated by the task, and C represents the number of computing cycles required by the task. This represents the maximum tolerable delay for the task.

5. A vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 3, characterized in that, In step S1.2, constructing the vehicle-edge-cloud collaborative system computing model specifically includes the following steps: S1.2.1 Define the number of vehicle-mounted tasks and their allocation scheme; Each vehicle generates one computing task, so n vehicles generate n computing tasks; each task can only choose one computing platform for computing, a tasks choose the edge platform for computing, b tasks choose the cloud platform for computing, and nab tasks choose the local platform for computing, where n is the number of vehicles, a is the task that chooses the edge platform for computing, and b is the task that chooses the cloud platform for computing. S1.2.2 Based on the defined number of vehicle-mounted tasks and allocation scheme, construct a local computing model; Since onboard tasks are computed locally without data transmission, local platform computation only includes computation latency, and local platform computation does not generate transmission energy consumption; energy consumption only includes computation energy consumption. The local computation model expression is as follows: ； In the formula, This represents the time taken for task i to be computed on the local platform, where i is the local compute task, and i ∈ nab; This represents the number of computer computing cycles required for task i. Indicates the computing power of the onboard computer; This represents the transmission delay of task i during local platform computation; This indicates the total latency of tasks in local platform computing; This represents the energy consumption of task i during local platform computation; This represents the computational energy factor of the task in local platform computation. This represents the energy transferred during the computation of task i on the local platform. This represents the total energy consumed by the task in local platform computation, where n is the number of vehicles, a represents the task selected for edge platform computation, and b represents the task selected for cloud platform computation. S1.2.

3. Based on the defined number of vehicle-mounted tasks and allocation scheme, construct an edge computing model; Offloading vehicle-mounted tasks to the edge platform for computation requires consideration of transmission latency, computation latency, and energy loss from edge computing. The objective function for offloading vehicle-mounted tasks to the edge platform is constructed, and the edge computing model expression is as follows: ； In the formula, Indicates task Data transmission delay to roadside units, For tasks assigned to the edge computing platform, ∈a; Indicates task The size of the data; Indicates task Data transmission rate to the Road Test Unit (RSU); Indicates task Computational latency in multi-access edge environments; Indicates task The required number of edge platform computing power cycles, Indicates task -1 The number of edge platform computing power cycles required, C1 <C2<…<Ca; This indicates the computing power of multiple access edge computers; This represents the total latency of all tasks in edge platform computing; Indicates task Energy consumption for transmitting data to roadside units; This represents the energy transfer factor of a task in edge computing; For the task Transmit power of the uplink to the roadside unit; This represents the computational energy factor of the task in multi-access edge computing; This represents the total energy consumption of edge computing. This indicates the computing power of edge computing. S1.2.

4. Based on the defined number of in-vehicle tasks and allocation scheme, construct a cloud computing model; Offloading vehicle-mounted tasks to a cloud platform for computation requires consideration of transmission latency, computation latency, and cloud computing energy consumption. The objective function for offloading vehicle-mounted tasks to the cloud platform is constructed, and the cloud computing model expression is as follows: ； In the formula, Indicates task Data transmission latency to the cloud server, For tasks assigned to the cloud computing platform, ∈b; Indicates task Data transfer rate to cloud server =0.9 ; Indicates task The size of the data; Indicates task Computational latency in cloud servers; This indicates the computing power of the cloud server computer; Indicates task The required number of cloud server computing power cycles, C1 <C2<…<Cb; Indicates task -1 is the number of cloud server computing power cycles required; This represents the total latency of all tasks in cloud computing; Indicates task Energy consumption for data transmission to the cloud server; This represents the energy transfer factor of a task during cloud server computation. For the task The transmit power of the uplink to the cloud server; This represents the computational energy factor of the task in cloud server computing; This represents the total energy consumption of all tasks in cloud computing.

6. The vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 1, characterized in that, The global optimal objective function in step S2 The specific expression is: ； In the formula, T represents the time required to process all tasks. This represents the maximum tolerable latency when processing task i. This indicates the time required to process task i locally. This indicates the time required to transmit data to the edge platform for processing task i. Let E represent the time required for the cloud platform to process task i, E represent the energy consumption for processing all tasks, M1 represent the weight ratio of task execution time in the overall objective function, and M2 represent the weight ratio of task execution energy consumption in the overall objective function.

7. The vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 1, characterized in that, Step S3 includes the following steps: The preprocessing operation on the dataset is performed as follows: the dataset is input with features [S, R, D, C, T]. max ]Simplified to [T f ve , C, T max Where S represents the probability of successfully receiving the task data packet; R represents the communication radius of the node; D is the amount of data generated by the task; C represents the number of task calculation cycles; T max T represents the maximum tolerable delay for the task. f ve This indicates the delay of the task on the platform; The preprocessed dataset was divided into a training dataset and a test dataset in an 8:2 ratio; The training dataset is input into the feedforward neural network, passing through the input layer and hidden layer of the feedforward neural network in sequence. After passing through the first fully connected layer, a linearly transformed output is obtained. Then, a ReLU activation function layer is used to introduce non-linearity, resulting in non-linear data. Next, the non-linear data is passed through the second fully connected layer and linearly transformed again. Finally, the output result is obtained after processing through the second ReLU activation function layer and the batch normalization layer. The expression for the vector processing process in a fully connected layer is as follows: ； In the formula, x is the input vector, W is the weight matrix, c is the bias vector, and y is the output vector; The ReLU activation function expression is as follows: ； In the formula, It is the input to the activation function; Batch normalization layers are used to reduce internal covariate bias, improving model stability and training speed, and effectively reducing the risk of overfitting. The expression for the batch normalization layer is shown below: ； In the formula, It is the input data. It is the variance of the mini-batch data, f(x) b ) is the output data. It is the mean of a small batch of data. >0, used to prevent the denominator from being 0; The output layer receives the results from the hidden layer and manipulates the output to finally select the result; the fully connected layer obtains the vector of the computing platform category corresponding to each task from the output of the hidden layer; the Softmax layer calculates the probability that the task belongs to each computing platform; and the regression layer determines the final computing platform for each task based on the probability distribution of the Softmax layer.

8. A vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 7, characterized in that, In the offline training of the feedforward neural network model using the training set, the training parameters of the feedforward neural network model are as follows: the fully connected layer of the feedforward neural network model consists of 200 neurons; the training parameters of the feedforward neural network model include: maximum number of training rounds = 200, initial learning rate = 0.001, learning rate scheduling = segmented learning rate scheduling, learning rate decay factor = 0.9, learning rate decay period = 10, mini-batch size = 70, and validation frequency = 30.

9. A vehicle computing task scheduling method based on vehicle-edge-cloud collaboration according to claim 1, characterized in that... The S4 outputs task allocation and scheduling instructions as follows: Real-time task feature data is sequentially input into the input layer and hidden layer of the feedforward neural network model; the input real-time task feature data is linearly transformed after passing through the first fully connected layer, and then nonlinear transformation is introduced through the ReLU activation function layer to enhance the model's expressive power. The nonlinear data is linearly transformed again after passing through the second fully connected layer, and then a linear result is output after passing through the second ReLU activation function layer; the linear result is transmitted to the output layer for processing. The fully connected operation of the output layer obtains the computing platform category vector corresponding to each task. The Softmax layer predicts the probability that a task belongs to the computing of each platform, and after regression processing, the final computing platform for each task is determined; one-hot labels for task allocation and scheduling instructions for local, edge, and cloud platforms are output.

Images