Automatic driving resource arrangement and task unloading method based on model segmentation

Abstract

An automatic driving resource arrangement and task unloading method based on model segmentation comprises the steps that in an automatic driving system, each vehicle continuously generates a vehicle-mounted task, and a vehicle task calculation model makes a real-time decision according to collected sensing data; each base station carries out task unloading scheduling on a plurality of vehicles in the coverage range; the MEC server communicates with the base station, and carries out computing resource arrangement and unloading task calculation on the edge side; according to the model segmentation method, a lightweight shallow network is deployed for a vehicle to serve as a vehicle model, and a deep network is deployed for an MEC server to serve as an edge model; adopting a residual network Resnet as a backbone network to realize model segmentation; residual blocks in the Resnet are dynamically segmented by segmentation points; according to resource arrangement and task unloading, collaborative decision-making of resource arrangement and task unloading is modeled as a constraint optimization problem for minimizing long-term task completion time delay, then the constraint optimization problem is decoupled as a resource allocation and task unloading sub-problem, and a double-layer DDPG network is constructed to solve the sub-problem.

Description

Technical Field

[0001] The present invention relates to autonomous driving technology, and in particular to a resource scheduling and task offloading method based on (vehicle task computing) model segmentation in an autonomous driving network. Background Art

[0002] In recent years, the rapid development of autonomous driving is redefining people’s travel methods and providing new impetus for optimizing the existing transportation system. Autonomous vehicles perceive the surrounding traffic environment through advanced sensors, and the vehicle task computing model (generally deployed on the vehicle, usually referred to as the “on-board model” or “vehicle model”) makes real-time decisions based on the collected perception data to control the vehicle’s steering, acceleration, braking and other behaviors [1].

[0003] Due to the space and power consumption limitations of on-board computing devices, vehicles are often unable to respond to high-computing load tasks in a timely manner. A feasible solution is the multi-access edge computing (MEC) paradigm [2], which fully or partially delegates on-board tasks to MEC servers for collaborative computing. Different offload tasks have different resource requirements, so MEC servers use RAN slicing technology to build isolated logical resource slices to provide customized resource services for different offload tasks [3][4]. However, the dynamic environmental conditions and vehicle density in autonomous driving scenarios still make edge collaborative resource orchestration and task offloading face many challenges:

[0004] (1) Demand-driven model segmentation scheme design. Resource-constrained vehicles find it difficult to continuously perform high-computational load tasks for a long time, such as target classification or recognition applications based on deep neural networks (DNNs). Vehicles may choose to fully or partially offload tasks to MEC servers to solve the problem of insufficient on-board computing power. However, during peak traffic hours in the morning and evening, offloading a large number of vehicle tasks to MEC servers will cause resource competition and shortages, significantly increasing task response latency.

[0005] To this end, M.Amiri et al. [5] proposed a model partitioning mechanism that divides the deep model on the device side and the cloud side into a series of blocks. The user selects a combination of blocks on both ends to jointly perform task calculations based on the inference time constraint. Although this reduces the amount of task calculation, it will inevitably lose inference accuracy when fewer block combinations are selected. P.Consul et al. [6] used split learning (SL) to divide the deep learning network into client and server networks, and deployed them on drones and central servers respectively. The drone sends the activation variables from the cutting layer to the central server, and the central server uses the received activation variables to calculate the remaining layers to complete the model inference. Compared with centralized learning, experimental results show that the proposed scheme has an accuracy improvement of 29.31% and a latency reduction of 7.37%. Luo et al. [7] deployed a multi-export split network on smart cars, and achieved distributed dynamic collaborative reasoning between smart cars and the cloud by setting different split points.

[0006] Although the above method designs a model segmentation scheme, it simply divides the deep network into several fixed blocks, and does not discuss in detail how to set the segmentation points driven by task requirements and how to jointly train the client and server networks.

[0007] (2) Slice resource orchestration based on model segmentation. RAN slicing uses network function virtualization (NFV) and software defined networking (SDN) to divide the physical network into multiple logical networks to complete various in-vehicle applications with heterogeneous resource requirements.

[0008] Cui et al. [3] proposed a multi-service resource allocation strategy for V2X, decoupling the resource allocation problem of multiple V2X services into the sub-problems of inter-slice and intra-slice resource allocation. A heuristic algorithm was used to solve the NP-hard sub-problems of resource allocation for two slices, and the QoS of the resulting scheduling strategy was improved by 15% compared with the maximum weighted delay priority algorithm. W. Fan et al. [8] analyzed the properties of tasks, channels, vehicles, and roadside units (RSUs) and designed a method based on generalized curved decomposition and reformulation linearization to flexibly select vehicle tasks to be processed locally or offloaded through V2I or V2V, aiming to balance resource cost and processing latency. Considering the dynamic and diverse communication requirements of vehicle networks, B. Hazarika et al. [9] designed a three-layer open radio access network (O-RAN) slicing model, corresponding to service type, vehicle type, and specific vehicle application. Federated deep reinforcement learning was used to implement hierarchical slicing resource allocation to enhance the structured and efficient network resource management.

[0009] The traditional resource orchestration methods mentioned in the above literature all consider the vehicle model as a whole, and rarely involve the segmentation of the vehicle model. Model segmentation forms different resource requirements on the client and server through splitting and combining, which puts higher requirements on the dynamics, adaptability and scalability of the slice resource orchestration strategy. How to solve an efficient resource orchestration strategy for the task offloading paradigm based on model segmentation is an unresolved problem.

[0010] (3) Feasible decoupling of resource scheduling and task offloading. The communication between high-speed vehicles and RSUs is short-lived and affected by vehicle speed, direction, and road conditions. This makes it increasingly difficult for vehicles to use traditional heuristic algorithms to infer task offloading decisions.

[0011] To this end, R. Narmeen et al.

[10] used end-to-end learning to solve the task offloading problem through deep deterministic policy gradient (DDPG) and actor-critic structure, and improved the task completion efficiency with stronger environmental adaptation and learning ability. Sun et al.

[11] proposed a joint task offloading and job scheduling method based on cloud edge computing, and used ant colony algorithm to solve multi-objective optimization to obtain the optimal offloading strategy. Simple additive weighting (SAW) and multi-criteria decision making (MCDM) were used to evaluate the delay, energy consumption and load balancing of the vehicle system. Wu et al.

[12] introduced a transformer-based long sequence prediction network (TLSFN) to predict the current and future queue lengths of the MEC server, and adopted a transformer-based multi-agent actor-critic (TPDMAAC) method based on the available resources of the queue to provide low-latency offloading services for computationally intensive vehicle tasks. Reference

[13] pointed out that the communication and computing resource allocation of vehicles, unmanned aerial vehicles (UAVs) and RSUs will determine the efficiency of task offloading, among which tasks that are sensitive to delay and have high computational requirements need to be offloaded to UAVs or RSUs for processing. It can be seen that there is coupling between resource scheduling and task offloading in the Internet of Vehicles. Reasonable resource scheduling will improve the efficiency of task offloading, and a reasonable task offloading strategy will balance the load between resource slices.

[0012] Researchers usually divide resource slices according to task type or service type, and there are few slice orchestration solutions for model segmentation. This requires us to re-decouple resource orchestration and task offloading problems based on the model segmentation results, and select appropriate deep reinforcement learning methods to solve the two decoupled sub-problems. Summary of the invention

[0013] To address the above problems, the present invention proposes an autonomous driving resource orchestration and task offloading method based on model segmentation, aiming to form an efficient and stable edge collaborative reasoning system to improve the QoS of in-vehicle services with efficient task response.

[0014] The proposed autonomous driving resource scheduling and task offloading method is applied to the autonomous driving system. The autonomous driving system consists of a base station set, a vehicle set and a MEC server. In the autonomous driving system, each vehicle continuously generates on-board tasks, and the vehicle task calculation model makes real-time decisions based on the collected perception data, and then the vehicle processor controls the vehicle operation. Each base station performs task offloading scheduling for multiple vehicles within the coverage area. The MEC server communicates with the base station and performs edge-side computing resource scheduling and offloading task calculation.

[0015] The model in the model segmentation is a vehicle task calculation model, and the model segmentation method is as follows: the vehicle deploys a lightweight shallow network as a vehicle model, and the MEC server deploys a deep network as an edge model; the residual network Resnet is used as the backbone network to realize model segmentation; the residual block in Resnet is dynamically segmented by the segmentation point;

[0016] Resource scheduling (i.e. resource allocation) and task offloading are solved using a two-layer DDPG model. Specifically, the collaborative decision-making of resource scheduling and task offloading is modeled as a constrained optimization problem P1 that minimizes the long-term task completion delay. The problem P1 is then decoupled into resource allocation and task offloading sub-problems, and a two-layer DDPG network is constructed to solve the sub-problems.

[0017] The main contributions of the present invention include:

[0018] First, to address problem (1), a model-based RAN slicing framework is proposed to support multi-dimensional resource orchestration and collaborative task offloading in each slicing window. According to the M / M / 1 queuing model, the RAN slicing and task offloading decisions for different slicing modules are modeled as the minimization problem of long-term task completion delay under network resource constraints.

[0019] Second, for problem (2), a slice resource scheduling method suitable for model segmentation is proposed. At the beginning of each slice window, the slice resources of the current window are re-partitioned according to the processing load of the unloaded queue in the MEC server in the previous window. For each window, a deep reinforcement method is applied to solve the communication and computing resource scheduling problems.

[0020] Third, to address problem (3), a two-layer DDPG scheduling strategy is designed to determine the RAN slice resource allocation at a large time scale and the task offloading at a small time scale, which takes into account the load balancing of all offloading queues.

[0021] The “model segmentation” refers to the vehicle task calculation model segmentation.

[0022] Simulation results show that the scheme of the present invention is superior to the benchmark method in terms of resource utilization, task completion rate and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Represents the autonomous driving system model based on model segmentation.

[0024] Figure 2 Represents the model segmentation framework.

[0025] Figure 3 Represents the slice resource orchestration framework.

[0026] Figure 4 Represents a two-layer DDPG network architecture that jointly optimizes resource orchestration and task offloading.

[0027] Figure 5 Represents traffic volume during the day.

[0028] Figure 6(a) and Figure 6(b) show the effect of different learning rates on the performance, where:

[0029] Figure 6(a) shows the loss of the segmentation model under different learning rates;

[0030] Figure 6(b) shows the reward value of the DRL algorithm under different learning rates.

[0031] Figure 7(a) and Figure 7(b) show the impact of different split points on task offloading performance, where:

[0032] Figure 7(a) shows the effect of segmentation points on target classification accuracy;

[0033] FIG7( b ) shows the effect of the segmentation point on the amount of intermediate feature data.

[0034] Figure 8 Represents the impact of knowledge distillation on the performance of the segmentation model.

[0035] Fig. 9 Indicates the impact of window duration on performance.

[0036] Figure 10(a) and Figure 10(b) show the effect of the number of available resources on the task failure rate, where:

[0037] Figure 10(a) shows the impact of computing resource variables;

[0038] Figure 10(b) shows the impact of spectrum resource variables.

[0039] Figure 11(a) and Figure 11(b) show the impact of different resource slicing methods on performance, where:

[0040] Figure 11(a) shows the effect of the slicing method on the task failure rate;

[0041] Figure 11(b) shows the impact of the slicing method on the average task processing delay.

[0042] Figure 12(a) and Figure 12(b) show the impact of different offloading strategies on performance, where:

[0043] Figure 12(a) shows the impact of offloading strategy on task failure rate;

[0044] Figure 12(b) shows the impact of the offloading strategy on the average task processing delay. DETAILED DESCRIPTION

[0045] The present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0046] 1 Overview

[0047] In the scenario of autonomous driving (IoV) service, this paper proposes a two-layer optimization framework of joint resource orchestration and task offloading based on model segmentation, which provides quality of service (QoS) guarantee for high-speed vehicles of autonomous driving system to efficiently process on-board tasks.

[0048] The present invention designs a model segmentation method based on knowledge distillation, and accordingly deploys a series of unloading queues and processing queues that adapt to different service requirements on vehicles and MEC servers, and constructs multiple logically isolated computing resource slices that adapt to different unloading methods.

[0049] Based on the queuing model, the collaborative decision-making of resource orchestration and task offloading is modeled as a constrained optimization problem that minimizes the long-term task completion delay. The problem is decoupled into two sub-problems: resource allocation and task scheduling, and a two-layer DDPG architecture is constructed to solve it, forming a double decision closed loop with a time window as the cycle. At the beginning of a new time window, the MEC server orchestrates slice resources according to historical task traffic fluctuations, and then performs offloading scheduling on a small time scale.

[0050] Simulation experiments show that compared with the baseline method, the proposed method effectively reduces the task failure rate while showing better adaptability and scalability.

[0051] 2 System Model

[0052] Figure 1 Shows the base station collection Vehicle collection An autonomous driving system (model) consisting of an MEC server. Each vehicle continuously generates on-board tasks (such as environmental perception, positioning, decision planning, motion control, etc.) and temporarily stores them in the local task queue. According to the computing power of the vehicle and the task queuing situation, each base station performs task offloading scheduling for multiple vehicles within the coverage area. The MEC (multi-access edge computing) server with rich computing resources communicates with the base station through a wired network and is responsible for the computing resource orchestration and offload task calculation on the edge side.

[0053] 2.1 Model Segmentation Framework

[0054] like Figure 2 As shown in the figure, the present invention proposes a model segmentation framework for task offloading, in which the vehicle deploys a lightweight shallow network and the MEC server deploys a deep network capable of high-precision reasoning. Resnet is selected as the backbone network to achieve dynamic model segmentation. The reasons are: (1) it consists of multiple modular residual blocks, which makes it easy to set the segmentation points; (2) the number of Conv blocks in each residual block gradually increases, which happens to meet the needs of lightweight vehicle models and accurate edge model reasoning.

[0055] Residual Block F n (x), then the Resnet containing N residual blocks can be expressed as

[0056]

[0057] in represents the residual block set, Indicates function compound operation. Set split points according to dynamic split strategy The vehicle calculates the first l residual blocks of Resnet and obtains the intermediate feature representation

[0058]

[0059] The intermediate feature z is transmitted from the vehicle to the MEC server, which continues to calculate the remaining Nl residual blocks and obtains the final output

[0060]

[0061] In particular, the vehicle sets split point o1 to indicate that all tasks are offloaded to the MEC server, and split point o2 to indicate that the tasks are completed locally.

[0062] After the above dynamic segmentation, the vehicle has a task processing queue Including the shallow model corresponding to the first L segmentation points and 1 local model; at the same time, the MEC server has a task offloading queue It includes deep models corresponding to the first L segmentation points and 1 fully unloaded model.

[0063] 2.2 Resource Orchestration Framework

[0064] The present invention designs a resource arrangement framework based on model segmentation, such as Figure 3 shown.

[0065] The computing resources of the MEC server and the spectrum resources of the base station BS are arranged into L+1 service slices for full or partial offloading. The division of service slices can be periodically adaptively adjusted according to the network situation. The system time domain is divided into a series of slice windows of equal length, each of which contains multiple scheduling slots of equal length. The duration of the slice window w is denoted as f (w) ,window w The set of scheduling slots included is denoted as T (w) Computing resources are allocated in units of VM instances. The set of all VMs held by the MEC server is represented as Its cardinality is O. In the slice window w At the beginning, the computing resources of the MEC server are rescheduled according to the task scheduling decision of window w-1. k represents the number of VM instances assigned to offload queue k by the MEC server (where ). This resource slicing strategy continues until the slicing window w ends. (w) At the beginning of each scheduling time slot, base station i is the set of covered vehicles according to the received task information. Determine the task offloading strategy. The vehicle and MEC server collaborate to complete the computing task with this strategy. At the end of each slice window, the MEC server collects the set of scheduling decisions in window w for the next window.

[0066] 2.3 Communication model

[0067] The present invention adopts the Orthogonal Frequency Division Multiplexing (OFDM) communication protocol, and the spectrum is divided into mutually orthogonal sub-channels. At the base station The average channel gain within the coverage area is gi,j The transmission power of base station i and vehicle j are p i and p j When vehicles and base stations communicate, they are affected by path loss and Rayleigh fading, as well as interference from other base stations. In units of subchannels, base station i allocates spectrum resources to the set of covered vehicles. Its cardinality is J i Let the bandwidth of the subchannel be B, and the distance between base station i and vehicle j be d i,j ,u m Indicates the processing queue Upload intermediate feature z m The number of sub-channels required, then vehicle j delivers z m The uplink transmission rate to base station i is

[0068]

[0069] Where α represents the path loss exponent and σ represents the average background noise. Let v k Indicates that unloading queue k returns execution result y k The number of subchannels required, base station i delivers y k The downlink transmission rate to vehicle j is

[0070]

[0071] 2.4 Task Offloading Framework

[0072] The task offloading framework requires dual optimization of the vehicle's computing power and accuracy. The offloading and processing of tasks no longer rely solely on the vehicle or MEC server alone, but allow the vehicle and MEC server to collaborate to complete them. According to the model segmentation results, each vehicle deploys a series of processing queues to cache two types of tasks: local processing and partial offloading. The MEC server correspondingly deploys a series of offloading queues to cache two types of tasks: partial offloading and full offloading. The MEC server allocates resources for offloading tasks according to the first-come, first-served (FCFS) rule, and transmits the final execution results back to the vehicle via the base station. Depending on the current resource usage, the base station will select different offloading strategies for the vehicle. Two examples are described below.

[0073] (1) Peak traffic period. When there is traffic congestion during the morning and evening rush hours and the MEC server unloading queue is serious, the base station tends to process tasks locally in the vehicle or unload tasks from the previous split point to reduce task processing delay and reduce the load pressure on the MEC server.

[0074] (2) Low-peak traffic period. When there are few vehicles at noon or at night and the MEC server unloading queue is idle, the base station tends to choose to unload all vehicles or unload tasks from the subsequent segmentation points to relieve the vehicle's computing pressure and obtain more accurate inference results.

[0075] From the above example, it can be seen that the base station should consider the traffic flow, vehicle and MEC server workload when scheduling tasks for vehicles. Therefore, the task completion delay is derived according to the M / M / 1 queuing theory, including task unloading, processing and result return delay. The vehicle task h is composed of a four-tuple {τ h ,z h ,y h ,ε h} represents, where τ h ,z h,y h ,ε h They respectively represent the amount of task calculation data, the size of intermediate features, the size of final processing results, and the delay constraint of the task.

[0076] Unloading delay

[0077] Unloading delay refers to the time it takes for a task to be unloaded from a vehicle to the MEC server. The base station and the MEC server are connected via an optical fiber link, so the transmission delay can be ignored. Therefore, only the transmission delay between the vehicle and the base station needs to be calculated.

[0078] Consider the slice window w The task of vehicle j is offloaded to base station i. Suppose the task set initiated by vehicle j is The vehicle task set received by base station i is Its base number is The processing queue set of vehicle j is Its base number is make represents the task of base station i to decide vehicle j Enter the local unloading queue m, and the extracted intermediate features are Its size is otherwise, Assume that vehicle j is in the queue m The number of computing units it has is o i,j,m , where the maximum GPU cycle per computational unit is If the amount of computation required to extract intermediate features for task h is The delay it calculates when processing queue m is

[0079]

[0080] Suppose the task set for processing queue m when vehicle j initiates task h is The queuing delay of this task is

[0081]

[0082] In addition, the computation and uplink transmission delay of the vehicle extracting the intermediate features of the task and offloading them to the MEC server is

[0083]

[0084] According to equations (7) and (8), the unloading delay of task h is

[0085]

[0086] Assume that the average queuing delay, feature extraction and upload delay of all vehicles covered by base station i are

[0087]

[0088] and in express Therefore, the average task unloading delay of all vehicles is

[0089]

[0090] The arrival of offload tasks in the MEC server can be modeled as a Poisson process. represents the arrival rate of task h of vehicle j to unloading queue k, and the corresponding task arrival rate of unloading queue k is

[0091]

[0092] Since each queue in the MEC server only processes one offload task at a time, the task offloading process can be modeled as an M / M / 1 process. The service intensity of the offload queue k is defined as

[0093]

[0094] When the task arrival rate is continuously greater than the output rate, the accumulated tasks will cause the queue to overflow. In order to ensure the stability of the queue, equation (13) should satisfy

[0095]

[0096] Processing delay

[0097] Processing latency refers to the time from when the intermediate feature enters the unloading queue of the MEC server to when the processing is completed. The MEC server needs to allocate different numbers of VM instances to them based on the average processing rate of a series of unloading queues in the previous slice window. Assume that the maximum GPU cycle of each VM instance is τ (max) When the intermediate features of task h arrive at the MEC server, the set of tasks to be unloaded from queue k is If the number of VM instances assigned to the unload queue k is and intermediate features The amount of computation required to continue reasoning is calculate The delay is

[0098]

[0099] Assume that the calculation result Y of task h′ in unloading queue k is i,j,h′,k The size of y i,j,h′,k , the queuing delay and computation delay of task h in the queue are

[0100]

[0101] and in Indicates that task h enters the unloading queue k, otherwise Therefore, the corresponding processing delay is

[0102]

[0103] Therefore, the average queuing delay and average computation delay of all tasks in the offload queue k are

[0104]

[0105] and Accordingly, the average processing delay of all tasks in the offload queue k is

[0106]

[0107] Return delay

[0108] After the MEC server completes the offload task, it will calculate the result Y i,j,h,k The appropriate base station will forward it back to the vehicle. (or ) indicates that the calculation result is returned from base station i (or base station i′) to vehicle j. The travel time for vehicle j to leave base station i and reach the coverage area of ​​base station i′ is ν i,i′ According to formula (5), the return delay of the calculation result of task h to the vehicle is

[0109]

[0110] Therefore, the average return delay of the calculation results in the unloading queue k to the vehicle is

[0111]

[0112] 2.5 Problem Modeling

[0113] Unloading too many tasks from vehicles will cause MEC servers to be overloaded, thus affecting the performance and efficiency of the system. Therefore, for the task offloading framework based on model segmentation, a challenging problem is the joint optimization of MEC server resource orchestration and task offloading scheduling.

[0114] In the Slice Window w In the example, the service delay of task h is

[0115]

[0116] Define a binary variable

[0117]

[0118] in Indicates that the calculation result of task h is successfully delivered within the specified time.

[0119] According to equations (11), (19), and (21), the average service delay of all tasks is the sum of the unloading delay, processing delay, and return delay, that is,

[0120]

[0121] The task miss rate during offloading is defined as

[0122]

[0123] The strategy sets for allocating computing resources and spectrum resources in MEC servers are:

[0124]

[0125] and Note that the slice window w It consists of a series of continuous scheduling slots t. In time slot t, let the set of tasks initiated by vehicle j be And the task set received by base station i is The corresponding uninstallation strategy is Therefore, a set of scheduling slots T in this window (w) The set of task offloading strategies is

[0126]

[0127] Given a slice window index set The optimization goal of system task scheduling is to minimize the average task service delay. Under long-term accumulation, this optimization problem is modeled as

[0128]

[0129] question It is to allocate spectrum and computing resources to each slice, aiming to meet the resource requirements of each unloading queue and minimize the long-term average task completion delay. Constraints (28a) and (28b) limit the sum of the spectrum and computing resources allocated to the unloading queue to not exceed the total amount. Constraint (28c) maintains the stability of each unloading queue. Constraint (28d) indicates that the number of base station subchannels should not exceed its own resources. Constraint (28e) stipulates that each task can only select one base station to return the processing result. Constraint (28f) limits each task to be unloaded through only one base station and enter the designated unloading queue. Finally, constraint (28g) requires that the unloading task incomplete rate is lower than the expected threshold to meet the set performance standard.

[0130] 3. Design

[0131] In order to obtain the exact optimal solution, It is decoupled into three sub-problems: 1) dynamic model segmentation; 2) computing resource allocation (large time scale); 3) task offloading scheduling (small time scale). DDPG networks are deployed on MEC servers and base stations respectively. These sub-problems are solved alternately by the two-layer DDPG network to form a closed loop that runs continuously. Figure 4 As shown in the figure, the short-term scheduling behavior accumulated by each base station in the lower layer provides the long-term resource status of the system for the upper MEC server. When the continuous operation of the two layers reaches a balance, the system will have the ability to handle a large number of vehicle-mounted tasks. The specific operation of each layer is described below. Distillation training of segmentation model: Design distillation training of vehicle model and MEC edge model for dynamic segmentation of the model.

[0132] Training scheme to achieve collaborative reasoning of the two models.

[0133] Resource slicing (upper layer): The MEC server designs computing and spectrum resources for the current window by analyzing historical scheduling data.

[0134] Source allocation strategy The resource allocation strategy of the slice continues until the end of window w, until the next window is reallocated.

[0135] Task offloading scheduling (lower layer): At the beginning of the scheduling time slot t, Send to the lower-level Actor network to confirm

[0136] Determine the task offloading schedule for this time At the end of the last time slot of each slice window, all offloading decisions within that window All will be recorded and used to determine resource allocation for the next window

[0137] The detailed solutions to the three sub-problems in the two-layer DDPG network are discussed below.

[0138] 3.1 Frozen Distillation Training of Segmentation Model

[0139] In order to ensure that the intermediate features of vehicle calculations can be directly fed into the deep network of the edge model, a frozen distillation training method is designed, where the vehicle model and the edge model are regarded as the teacher model and the student model, respectively.

[0140] First, the vehicle model is pre-trained using local data.

[0141] Next, for a given split point l, the vehicle model parameters are copied to the corresponding part of the edge model and the parameters are frozen to ensure that the parameters of the two models are consistent before the split point l. Subsequently, the knowledge distillation technology transfers the knowledge of the vehicle model to the edge model. The temperature factor H can be used to soften the output target of the vehicle model and control the smoothness of the soft target. When the temperature H>1, the calculation result Y of the vehicle model i,j,m Softened by Softmax function and converted into probability distribution

[0142]

[0143] On this basis, KL divergence is used as the cross entropy loss function

[0144]

[0145] where Y i,j,m ,Y i,j,k Denote the logarithmic vectors generated by the vehicle model and the edge model respectively. According to formula (30), multiple rounds of frozen distillation training are performed, and then the edge model corresponding to the split point l can be deployed to the lth unloading queue. The frozen distillation training is summarized as Algorithm 1 below.

[0146]

[0147] 3.2 Resource Orchestration for Model Segmentation

[0148] The number of task requests from actual vehicles is time-varying and uncertain. A fixed resource allocation method will not be able to capture the fluctuations in task arrivals. During peak traffic hours, the MEC server should reduce the computing resources of the full unloading queue and add them to other partial unloading queues to adapt to high traffic volume. During non-peak hours, the number of requested tasks is relatively stable, and appropriately increasing the computing resources of the full unloading queue can reduce the computing burden on vehicles.

[0149] The resource allocation sub-problem is to minimize the long-term average task completion delay of the system by allocating resources across MEC slices.

[0150]

[0151] Since the decision of each slice window is independent, the scheduling of computing resources within the window is also independent. In the real world, the number of tasks in adjacent slice windows usually does not fluctuate greatly. According to the task scheduling strategy in window w-1 and the vehicle density in window w, the MEC server can predict the computing resources required for each offloading queue in window w and the spectrum resources required by the base station.

[0152] Considering the dynamics of task offloading, the present invention defines the resource allocation problem in a single window as a Markov decision process and solves it using the deep deterministic policy gradient (DDPG) algorithm. Therefore, the long-term optimization problem of P1.1 is decomposed into a short-term optimization problem for a single slice window, and the MEC server is set as an agent. By observing the current environment state s (w) , the agent uses the Actor network to infer the optimal allocation strategy a (w) , and get feedback from the environment (w) Agent is based on the transition probability distribution Pr(s (w+1) |s (w) ,a (w) ) makes the environment state transfer to the next state s (w+1) The action space, state space, and reward function of DDPG considering resource allocation are described as follows:

[0153] State space: The computational resource allocation strategy of the slice window w depends on the vehicle density ρ of the current window (w) and the previous

[0154] The window's uninstall decision set The spectrum resource allocation strategy depends on the available subchannels q in the current window. (w) .

[0155] Therefore, the environmental state is

[0156]

[0157] Action space: For the optimization variables in subproblem P1.1, the actions are the allocation strategies of computation and spectrum resources.

[0158]

[0159] Each action element uses a positive integer to represent the number of VM instances and subchannels assigned to the task queue.

[0160] Reward function: The reward is the system's evaluation of the quality of the action performed. The system goal is transformed from minimizing the average task service delay to maximizing the reward. The reward function is

[0161] r (w) = -logD (w) +Λ (33)

[0162] The first term in (33) is the incentive given by the system when task offloading scheduling is feasible; the second term is the vehicle density exceeding the system capacity (i.e., ρ (w) >ρ (max) ) causes the system to impose penalties when a task fails to prevent uneven resource allocation.

[0163] 3.3 Task offloading for model segmentation

[0164] Given the uncertainty of vehicle environment and communication conditions, the task offloading subproblem is to enable the vehicle to select the best offloading strategy to minimize the total delay of the system task.

[0165]

[0166] As described in Section 3.2, resource allocation for each slice window is independent. Once resource scheduling is determined, the scheduling of small-time-scale tasks in each slice window is also independent. The long-term optimization problem in P1.2 can be decomposed into scheduling slots t∈T (w) The MDP process can also be solved by DDPG.

[0167] The base station acts as an agent to allocate resources and vehicle status s t Input the DDPG network to obtain the task offloading strategy. The action space, state space and reward function of the DDPG are described as follows:

[0168] Action space: The action goal of the task offloading subproblem should be the task offloading decision

[0169]

[0170] State space: Task offloading needs to consider real-time information related to tasks, vehicles, and MEC servers. Indicates vehicle speed and d i,j represents the distance between vehicle i and the currently communicating base station j, and the system state is

[0171]

[0172] Reward function: The effect of the unloading decision can be reflected by the total system delay. The reward function is

[0173]

[0174] In order to better solve the two related problems of resource allocation and task offloading, a two-layer DDPG network is proposed to jointly solve problem P1. The upper-layer resource allocation network is based on the current state s (w) Generate a strategy for allocating computing and spectrum resources (w) Then, at each scheduling slot t, the lower-level task offloading network and a (w) Generating task offloading decisions Joint decision making Applied to edge-cloud collaborative computing, the obtained reward value is used to guide the parameter update of the double-layer DDPG network. Additionally, considering that dynamic tasks cause inconsistent action space dimensions, an additional policy network and linear layer

[14] can be used to adapt to different action space dimensions. The policy network determines the relevance by calculating the attention scores of the input and output of the network.

[0175]

[0176] Where Φ s and Φ a represent the sampling results of the linear layer for the input state s and output a respectively, and D is the dimension of the output vector. The attention weight for each pair of input s and output a is

[0177]

[0178] Where α s,a The larger it is, the higher the matching degree between s and a. The attention mechanism enables the policy network to perceive the resource competition of task offloading and learn the state mapping between vehicle tasks and task queues in the MEC server. In addition, the adopted masked attention mechanism can ensure that the absence of vehicle states does not affect the update of the policy network. Algorithm 2 details the joint decision-making process of the double-layer DDPG.

[0179]

[0180] 3.4 Computational Complexity Analysis

[0181] The DRL-based scheduling algorithm can process a batch of task requests simultaneously in a time slot, and this advantage improves the task processing speed and environmental adaptability. This section analyzes the complexity of the double-layer DDPG network by comparing DDPG with DDQN. Assume that the computational complexity of training N training episodes for DDQN is O(N). DDPG adds a Critic network on top of DDQN to reduce the negative impact brought by the continuous action space, but at the same time increases the computational workload of network updates. Due to the fewer parameters in the Critic network, the computational complexity of DDPG is approximately O(cN) (1 < c < 1.5). However, the double-layer DDPG needs to alternately calculate the resource allocation and task offloading strategies, and its computational complexity is approximately O(c(N + M)), where M is the number of training episodes of the resource slicing network. Although the complexity of DDPG is slightly higher than that of DDQN, DDPG can make more refined resource configurations to compensate for the computational cost.

[0182] 4 Simulation Experiments and Result Analysis

[0183] In this section, simulation experiments are used to evaluate the effectiveness and superiority of the proposed method. The simulation scenario considers a four-lane urban road with a length of 2 kilometers. Four base stations and one MEC server are deployed on both sides of the road. The effective coverage radius of each base station is 500 meters, and the MEC server provides computing services for vehicles on the road. The transmission power of the base station and the vehicle is 40dBm and 0.2w respectively. The average speed of the vehicle is 60 km / h. The traffic flow trajectory is selected from the open road traffic data platform OpenITS. We use the Lighthill-Whitham-Richards (LWR) model [1 5] to convert traffic flow information into vehicle density, as Figure 5 As shown in Figure 2, Resnet 34 is used as the backbone network for dynamic model segmentation, and CIFAR10 is used to test the effectiveness of the segmentation scheme.

[0184] The model training and testing platform is equipped with an AMD 24-core 32-thread 13th Gen Intel(R)Core(TM)i9-13900CPU and an NVIDIA GeForce RTX 4090GPU. Table 2 gives the simulation parameters of DDPG. The data volume and calculation result size of the task are randomly generated within a given range. Five methods are selected as benchmarks, each of which includes two functional modules: resource allocation and task scheduling. Table 3 lists the implementation details of the five benchmark methods.

[0185] Table 1 DDPG training parameters

[0186] parameter value Epochs (training rounds) 800 Timesteps (number of time steps in each round) 2000 Discount factor 0.99 Epsilon_minimum / maximum (minimum and maximum values ​​of epsilon in the ε-greedy strategy) 0.01 / 0.99 Epsilon_decay (epsilon decay value) 0.995 Memory capacity 5000 Learning rate 0.01 Num of hidden layers 3 Num of neurons [64,128,32] Background noise power -70dBm

[0187] Table 2 Simulation parameters

[0188]

[0189] Table 3 Implementation of Baseline method

[0190] Baseline approach Resource allocation Task scheduling Baseline-1 No Only Local computation Baseline-2 Average DDPG (Dynamic split point) Baseline-3 DDPG DDPG (Static split point) Baseline-4 DDQ DDQN (Dynamic split point) Baseline-5 DDPG All offloading Proposed DDPG DDPG (Dynamic split point)

[0191] 4.1 Task offloading performance analysis of segmentation model

[0192] The effect of learning rate on performance during training was evaluated. The learning speed and training effect of the segmentation model are affected by the learning rate and training cycle. Figure 6(a) shows the convergence of the proposed method when the learning rate is 0.1, 0.01, and 0.001. The model training is performed offline, and it takes about 40 minutes to execute 300 training sets. In the first 50 training cycles, the loss first decreases rapidly and then decreases slowly. Due to the random selection of parameters, the segmentation model is difficult to make correct judgments in the initial stage. Only after learning a large amount of data can the model capture the data features and update the parameters. When the learning rate is 0.1, the loss shows obvious jitter, and the loss reaches the lowest value of 0.045 at 230 rounds. When the learning rate is 0.001, the loss converges to 0.042 and reaches the local optimum at 200 rounds. The convergence performance of the learning rate of 0.01 is better than the other two learning rates. Not only does the loss decrease the fastest in the early stage of training, but it can also obtain the minimum loss of 0.038 at 180 training rounds. Correspondingly, Figure 6(b) shows the convergence curve of the reward value under different learning rates. DDPG adds a Critic network based on DDQN, which can more effectively handle the problem of continuous action space. Therefore, the reward value of the proposed method is higher than that of Baseline-4 under different learning rates, and has the highest reward value when the learning rate is 0.1. In subsequent experiments, the learning rate of DDPG is set to 0.1.

[0193] Study the effect of the choice of split point on model performance. Assume that a day is divided into 72 time windows, each window lasting 20 minutes. Figure 2 Figure 7(a) shows the impact of different task offloading methods on target classification accuracy for o1, o2 and l of the model segmentation. The classification accuracy of all offloading at o1 is 93.31%, the classification accuracy of local calculation at o2 is 82.21%, and the classification accuracy from segmentation point 1 to segmentation point 4 is 92.54%, 91.44%, 85.23% and 66.32% respectively. It can be seen that offloading tasks to the MEC server from the first three segmentation points can ensure the accuracy of vehicle target recognition; however, the recognition accuracy drops significantly at segmentation point 4, which is the effect of freezing too many shallow parameters. Figure 7(b) shows the data volume of intermediate features under different task offloading methods. Although local calculation of tasks does not require data transmission, it consumes local computing power and limits the target recognition accuracy. The backward shift of the segmentation point gradually reduces the amount of data of intermediate features that need to be transmitted. Although offloading tasks from split point 4 only requires the transmission of fewer intermediate features, it will lead to insufficient classification accuracy, and its performance is even lower than local calculation; while offloading tasks from split points 1, 2, and 3 only consumes a small amount of local computing power and can complete the computing tasks at a lower communication cost by using the computing power of the MEC server. Considering the accuracy and cost comprehensively, subsequent experiments implement task offloading at o1, o2, and the first three split points.

[0194] Although the experimental results of Figure 7(a) and Figure 7(b) exclude the segmentation points with low target classification accuracy, the reasoning accuracy of the segmentation model still needs to be improved. We further study the impact of knowledge distillation on the target classification performance based on model segmentation. Figure 8 It can be seen that compared with the case without knowledge distillation, the vehicle model as a teacher provides local data features to the MEC model as a student and enhances its generalization ability. This improves the target classification accuracy of the three unloading queues of the proposed method in the MEC server by 1.50%, 1.70% and 1.62% respectively. Knowledge distillation helps alleviate the adverse effects of parameter freezing on inference results during model segmentation training.

[0195] The effect of a window duration on the task failure rate was evaluated. When a window duration gradually changed from 10 minutes to 60 minutes, Fig. 9 It shows that the average task failure rate of Baseline-4 and Proposed is linearly correlated with the window duration. This shows that a long window time will lead to a low resource allocation rate, which is difficult to cope with the real-time fluctuations of traffic flow, especially during peak hours in the morning and evening; however, a shorter window time will frequently call the slice allocation algorithm and increase unnecessary computational burden. The dual decision of Proposed is better than Baseline-4, so its task failure rate is reduced by 4% to 13.3%. In order to balance the task completion rate and computational overhead, 20 minutes is used as the window duration in subsequent experiments.

[0196] 4.3 Impact of Resource Slicing and Task Offloading

[0197] Figure 10(a) shows the impact of MEC computing resources on the task failure rate when the number of BS subchannels is fixed at 15. The task failure rate of all methods shows a decreasing trend. The failure rate of the DRL strategy (Baseline-4, Proposed) eventually stabilizes at around 20%, while the failure rate of the average allocation strategy (Baseline-2) can only be reduced to 29.92%. This shows that DRL is able to learn from the real-time traffic environment and formulate better resource allocation strategies. Sufficient MEC computing resources give vehicles more offloading options when processing tasks, although this is not the only condition for performance improvement. Next, consider the impact of spectrum resources on the task completion rate when the MEC computing resources are fixed at 12. In Figure 10(b), the task failure rate drops rapidly in the early stage. However, when the number of subchannels reaches 18, increasing spectrum resources no longer brings significant performance improvement. This shows that MEC computing resources are one of the system bottlenecks.

[0198] The following experiments evaluate the impact of different resource slicing methods on performance. As shown in Figures 11(a) and 11(b), when the number of tasks is small (λ = 0.8), the task failure rate and average task processing delay of the three methods are similar. However, the vehicle groups during the morning and evening peak hours will generate a large number of task requirements, which leads to the rapid consumption of MEC computing resources. Increasing the task arrival rate λ causes a significant increase in the task failure rate and average delay of the three methods. Compared with the average allocation of computing resources by Baseline-2, the task failure rate and average task processing delay of Baseline-4 and Proposed, which are obtained through training and learning, are 25.6% to 44.1% and 10.6% to 37.4% lower. When the number of tasks is doubled (λ = 1.6), the task failure rate and average delay of Proposed are still within the controllable range, which are 21.6% and 0.412s. This shows that DDPG has a stronger ability to solve coupling problems than DDQN.

[0199] Finally, the impact of different offloading strategies on performance was evaluated. As the task arrival rate λ increases from 0.8 to 1.6, Figures 12(a) and 12(b) show that the average task processing delay of Baseline-1, which only performs local calculations, mainly comes from the queuing delay of the offloading queue, so it has been maintained at a high level of 0.61s. The corresponding task failure rate is higher than 30%. The limited computing resources of the vehicle are the bottleneck for completing tasks with low latency. Baseline-5 offloads tasks to the MEC server to help alleviate the computing pressure of the vehicle. When λ increases to 1.2, Baseline-5 and Proposed benefit from the assistance of the MEC server to the vehicle, and the task failure rate and average task processing delay do not increase significantly. However, when λ increases from 1.2 to 1.6, Baseline-5 completely offloads tasks to the MEC server, causing serious queuing in the processing queue, which makes many tasks unable to meet the time constraints. Proposed can adaptively adjust the offloading strategy according to task requirements, and the task failure rate and average task processing delay remain low, at 21.6% and 0.439s. This effectively alleviates the resource competition problem caused by fluctuations in the number of tasks and enables smooth unloading decisions under complex traffic conditions.

[0200] 4.4 Scalability Analysis

[0201] The scalability analysis is performed below to demonstrate the efficiency and robustness of the proposed method.

[0202] (1) Scalability of model segmentation. Segmentation models that support different task types can be added to support more vehicle services. However, creating more types of segmentation models will inevitably increase the pressure on computing and communication resource allocation, which will squeeze the resources obtained by each type of segmentation model. At the same time, more task types require more stable and robust services. For the former, referring to the results of Figure 10, the proposed method can maintain an acceptable task completion delay when computing resources and spectrum resources are insufficient. For the latter, referring to the results of adaptive scheduling resources and decision offloading in Figures 11 and 12, the efficiency and robustness of the proposed method can be verified.

[0203] (2) Scalability of vehicle service scope. The edge structure managed by MEC can be replicated in autonomous driving scenarios. The proposed method is centered on MEC and expands the vehicle service scope by deploying it to more areas. This is conducive to supporting multi-type vehicle services in large-scale scenarios. Resource allocation and task scheduling in each service area can be adjusted according to the task type. With the help of differentiated local data of vehicles, each MEC server can customize a scheduling strategy that is highly matched with the local environment through learning knowledge, so as to provide high QoS services to passing vehicles.

[0204] 5. Conclusion

[0205] In the present invention, a solution for autonomous driving resource orchestration and task offloading based on model segmentation is adopted to support vehicle services with low latency and high completion rate. Different from traditional task offloading methods, the proposed method integrates model segmentation, DDPG-based resource slicing and task scheduling, and adapts to task intensity fluctuations without predicting future information. Simulation results confirm that the proposed method can dually optimize task offloading latency and target classification accuracy, especially in the morning and evening peak hours. In terms of adaptability, through the segmentation model method, the offloading method can be dynamically selected according to the fluctuation of task intensity, balancing the system's computational and communication overhead while ensuring the segmentation model reasoning accuracy. In terms of efficiency, a two-layer DDPG framework for joint resource allocation and task offloading is adopted, which comprehensively considers the interaction between resource allocation and task scheduling to minimize the total latency of task completion. In terms of scalability, the proposed edge framework can be deployed on roads in multiple regions to expand the scope of vehicle services.

[0206] References

[0207] [1]Xia X,Meng Z,Han X,et al.An Automated Driving Systems DataAcquisition and Analytics Platform[J].Transportation Research Part C:EmergingTechnologies,2023,151:104120.

[0208] [2]Chu W,Jia X,Yu Z,et al.Joint Service Caching,Resource Allocationand Task Offloading for MEC-based Networks:AMulti-layer Optimization Approach[J].IEEE Transactions on Mobile Computing,2023,23(4):2958-2975.

[0209] [3]Cui Y,Yang X,He P,et al.O-RAN Slicing for Multi-service ResourceAllocation in Vehicular Networks[J].IEEE Transactions on VehicularTechnology,2023.

[0210] [4]Lotfi F,Afghah F,Ashdown J.Attention-based Open Ran SliceManagement Using Deep Reinforcement Learning[C].2023IEEE GlobalCommunications Conference.IEEE,2023:6328-6333.

[0212] [5]Amiri M,Rohban M H,Hessabi S.ARobust Heterogeneous OffloadingSetup Using Adversarial Training[J].IEEE Transactions on Mobile Computing,2023.

[0213] [6]Consul P,Budhiraja I,Garg D,et al.SFL-TUM:Energy Efficient SFRLMethod for Large Scale AI Model's Task Offloading in UAV-assisted MECNetworks[J].Vehicular Communications,2024,48:100790.

[0214] [7]Luo C,Chen J,Feng X,et al.Sustainable Collaborative Inference inIntelligent Transportation Systems[J].IEEE Transactions on IntelligentTransportation Systems,2023. [8]Fan W,Su Y,Liu J,et al.Joint Task Offloading and ResourceAllocation for Vehicular Edge Computing Based on V2I and V2V Modes[J].IEEETransactions on Intelligent Transportation Systems,2023,24(4):4277-4292.

[0215] [9]Hazarika B,Saikia P,Singh K,et al.Enhancing Vehicular NetworksWith Hierarchical O-RAN Slicing and Federated DRL[J].IEEE Transactions onGreen Communications and Networking,2024.

[0216]

[10] Narmeen R,Mach P,Becvar Z,et al.Joint Exit Selection andOffloading Decision for Applications Based on Deep Neural Networks[J].IEEEInternet of Things Journal,2024.

[0217]

[11] Sun Y,Wu Z,Meng K,et al.Vehicular Task Offloading and JobScheduling Method Based on Cloud-edge Computing[J].IEEE Transactions onIntelligent Transportation Systems,2023.

[0219]

[12] Wu C L,Chiu T C,Wang C Y,et al.Mobility-aware Deep ReinforcementLearning with Seq2seq Mobility Prediction for Offloading and Allocation inEdge Computing[J].IEEE Transactions on Mobile Computing,2023.

[0220]

[13] Yan M,Xiong R,Wang Y,et al.Edge Computing Task OffloadingOptimization for a UAV-Assisted Internet of Vehicles via Deep ReinforcementLearning[J].IEEE Transactions on Vehicular Technology,2023.

[0221]

[14] Wang T,Zhang Y,Shen H,et al.Task Partitioning and SchedulingBased on Stochastic Policy Gradient in Mobile Crowdsensing[J].IEEETransactions on Computational Social Systems,2024.

[0222]

[15] Kachroo P, K.Feedback Control Theory for Dynamic TrafficAssignment[M].London:Springer,1999。

Claims

1. A method for resource scheduling and task offloading of autonomous driving based on model segmentation, which is applied to an autonomous driving system, which consists of a collection of base stations. Vehicle collection and an MEC server; in the autonomous driving system, each vehicle continuously generates on-board tasks, and the vehicle task calculation model makes real-time decisions based on the collected perception data, and then the vehicle processor controls the vehicle operation; each base station performs task offloading scheduling for multiple vehicles within the coverage area; the MEC server communicates with the base station and performs edge-side computing resource orchestration and offload task calculation; Its characteristics are The model in the model segmentation is the vehicle task calculation model. The model segmentation method is as follows: the vehicle deploys a lightweight shallow network as the vehicle model, and the MEC server deploys a deep network as the edge model; the residual network Resnet is used as the backbone network to achieve model segmentation; the residual blocks in Resnet are dynamically segmented by the segmentation points, so that the vehicle has a task processing queue Including the shallow model corresponding to the first L segmentation points and 1 local model; the MEC server has a task offloading queue Includes deep models corresponding to the first L segmentation points and a full unloading model; Resource orchestration, i.e. resource allocation and task offloading, is solved by using a two-layer DDPG model. Specifically, the collaborative decision-making of resource orchestration and task offloading is modeled as a constrained optimization problem P1 that minimizes the long-term task completion delay. The problem P1 is then decoupled into resource allocation and task offloading sub-problems, and a two-layer DDPG network is constructed to solve the sub-problems, where: (I) The modeling process of problem P1 is: Resource orchestration and allocation: The computing resources of the MEC server and the spectrum resources of the base station BS are arranged into L+1 service slices for full or partial offloading; the division of service slices is periodically adaptively adjusted according to the network situation; the time domain of the autonomous driving system is divided into a series of slice windows of equal length, each slice window contains multiple scheduling slots of equal length; the duration of the slice window w is recorded as f (w) , the set of scheduling slots contained in w is denoted as T (w) ; Computing resources are allocated in units of VM instances; the set of all VM instances held by the MEC server is represented as Its cardinality is O. At the beginning of slice window w, the computing resources of the MEC server are rearranged according to the task scheduling decision of window w-1. Let O k represents the number of VM instances assigned to offload queue k by the MEC server, where This resource slicing strategy continues until the slicing window w ends; (w) At the beginning of each scheduling time slot, base station i is the set of covered vehicles according to the received task information. Determine the task offloading strategy; the vehicle and the MEC server collaborate to complete the computing task according to the strategy; at the end of each slice window, the MEC server collects the set of scheduling decisions in window w for the next window; Task offloading: Task offloading and processing are completed by the collaboration between the vehicle and the MEC server. According to the model segmentation results, each vehicle deploys a series of processing queues to cache two types of tasks: local processing and partial offloading. The MEC server correspondingly deploys a series of offloading queues to cache two types of tasks: partial offloading and full offloading. The MEC server allocates resources for the offloading task and transmits the final execution result back to the vehicle via the base station. Based on the current resource usage, the base station selects different offloading strategies for the vehicle. The task completion delay is derived according to the M / M / 1 queuing theory, including task unloading delay, processing delay and result return delay; the vehicle task h is composed of a four-tuple {τ h ,z h ,y h ,ε h } represents, where τ h ,z h ,y h ,ε h They represent the amount of task calculation data, the size of intermediate features, the size of final processing results, and the delay constraint of the task respectively; Problem modeling: Problem P1 is to allocate spectrum and computing resources to each slice, aiming to meet the resource requirements of each offload queue and minimize the long-term average task completion delay; In slice window w, the service delay of task h is Vehicle j unloads queue k and task h under the coverage of base station i. represents the task offloading delay, Indicates task processing delay; Indicates the delay in returning the task processing result; Define a binary variable in Indicates that the calculation result of task h is successfully delivered within the specified time; The average service delay of all tasks is the sum of the unloading delay, processing delay and return delay. represents the unloading delay of task h, including the sum of the average queuing delay of all vehicles covered by base station i and the average delay of feature extraction and uploading; represents the average processing delay of all tasks in the offload queue k, including the sum of the average queuing delay and the average calculation delay of all tasks in the offload queue k; represents the average return delay of the calculation results in unloading queue k to the vehicle; The task incomplete rate during offloading is defined as in express The base number, is the set of vehicle tasks received by base station i; The strategy sets for computing resource allocation and spectrum resource allocation in the MEC server are and represents the strategy for allocating computing resources, Indicates the strategy for allocating spectrum resources; The slice window w consists of a series of continuous scheduling slots t. In time slot t, let the set of tasks initiated by vehicle j be And the task set received by base station i is The corresponding uninstallation strategy is Therefore, a set of scheduling slots T in the window w (w) The set of task offloading strategies is Given a slice window index set The optimization goal of system task scheduling is to minimize the average task service delay. Under long-term accumulation, this optimization problem is modeled as The constraints include: limiting the sum of the spectrum and computing resources allocated to the offloading queue to not exceed the total amount; maintaining the stability of each offloading queue; the number of base station subchannels does not exceed its own resources; each task can only select one base station to return the processing result; each task is only offloaded through one base station and enters the designated offloading queue; the uncompleted rate of offloading tasks is lower than the expected threshold to meet the set performance standards; (II) Solution process of problem P1 Problem P1 is decoupled into three sub-problems: 1) model segmentation; 2) resource allocation; 3) task offloading scheduling; Deep deterministic policy gradient DDPG networks are deployed on MEC servers and base stations respectively. These sub-problems are solved alternately by the two-layer DDPG network to form a closed loop that runs continuously. The short-term scheduling behaviors accumulated by each base station in the lower-layer DDPG network provide the long-term resource status of the system for the MEC server in the upper-layer DDPG network. When the continuous operation of the two-layer DDPG network reaches a balance, the autonomous driving system will be able to handle a large number of vehicle-mounted tasks. 1) Training of the segmentation model: The vehicle model and edge model are regarded as the teacher model and the student model respectively, and the distillation method is used to train them to achieve collaborative reasoning of the two; 2) Resource allocation is solved by the upper-layer DDPG network: The MEC server designs calculation and spectrum resource allocation strategies for the current window by analyzing historical scheduling data The resource allocation strategy of the slice continues until the end of window w, until the next window is reallocated; 3) Task offloading scheduling is solved by the lower-layer DDPG network: At the beginning of the scheduling time slot t, Send to the lower-level Actor network to determine the task offloading schedule At the end of the last time slot of each slice window, all offloading decisions within that window All will be recorded and used to determine resource allocation for the next window.

2. The method for autonomous driving resource scheduling and task offloading based on model segmentation according to claim 1, characterized in that 1) The training of the segmentation model is a frozen distillation training method, and the steps include: First, pre-train the vehicle model using local data; Next, for a given split point l, the vehicle model parameters are copied to the corresponding part of the edge model and the parameters are frozen to ensure that the parameters of the two models are consistent before the split point l; Subsequently, the knowledge distillation method is used to transfer the knowledge of the vehicle model to the edge model; The temperature factor H is used to soften the output target of the vehicle model and control the smoothness of the soft target: When the temperature H>1, the calculation result Y of the vehicle model i,j,m Softened by Softmax function and converted into probability distribution On this basis, KL divergence is used as the cross entropy loss function where Y i,j,m ,Y i,j,k Represent the logarithmic vectors generated by the vehicle model and edge model respectively; Multiple rounds of frozen distillation training are performed according to the cross entropy loss function, and then the edge model corresponding to the split point l is deployed to the lth unloading queue.

3. The method for autonomous driving resource scheduling and task offloading based on model segmentation according to claim 1, characterized in that 2) The resource allocation sub-problem is to minimize the long-term average task completion delay of the system by allocating resources across MEC slices. According to the task scheduling strategy in window w-1 and the vehicle density in window w, the MEC server predicts the computing resources required for each unloading queue in window w and the spectrum resources required by the base station; The resource allocation problem in a single window is defined as a Markov decision process and solved by the deep deterministic policy gradient DDPG algorithm. Then, the long-term optimization problem of P1.1 is decomposed into a short-term optimization problem for a single slice window, and the MEC server is set as an intelligent agent. By observing the current environment state s (w) , the agent uses the Actor network to infer the optimal allocation strategy a (w) , and get feedback from the environment (w) ;Agent based on the transition probability distribution Pr(s (w+1) |s (w) ,a (w) ) makes the environment state transfer to the next state s (w+1) ; The action space, state space and reward function of DDPG considering resource allocation are described as follows: State space: The computational resource allocation strategy of the slice window w depends on the vehicle density ρ of the current window (w) and the previous window's uninstall decision set The spectrum resource allocation strategy depends on the available subchannels q in the current window. (w) , so the environmental state is Action space: For the optimization variables in subproblem P1.1, the actions are the allocation strategies of computation and spectrum resources. Each action element uses a positive integer to represent the number of VM instances and subchannels assigned to the task queue; Reward function: The reward is the system's evaluation of the quality of the action performed; the system goal is transformed from minimizing the average task service delay to maximizing the reward; the reward function is r (w) =-logD (w) +Λ The first term in the reward function expression is the incentive given by the system when task offloading scheduling is feasible; the second term is the vehicle density exceeding the system capacity, i.e., ρ (w) >ρ (max) The penalty given by the system when a task fails.

4. The method for autonomous driving resource scheduling and task offloading based on model segmentation according to claim 3 is characterized in that 3) The task offloading sub-problem is to enable the vehicle to select the best offloading strategy to minimize the total delay of the system task The long-term optimization problem in P1.2 is decomposed into scheduling slots t∈T (w) The Markov decision process is solved using the DDPG algorithm; The base station acts as an agent to allocate resources and vehicle status s t Input the DDPG network to obtain the task offloading strategy. The action space, state space and reward function of the DDPG are described as follows: Action space: The action goal of the task offloading subproblem should be the task offloading decision State space: Task offloading needs to consider real-time information related to tasks, vehicles, and MEC servers; Indicates vehicle speed and d i,j represents the distance between vehicle i and the currently communicating base station j, and the system state is Reward function: The effect of the unloading decision can be reflected by the total system delay; the reward function is 5. The method for autonomous driving resource scheduling and task offloading based on model segmentation according to claim 4 is characterized in that A two-layer DDPG network is used to jointly solve problem P1; The upper DDPG network (i.e., resource allocation network) allocates resources according to the current state s (w) Generate a strategy for allocating computing and spectrum resources (w) ; Then, at each scheduling slot t, the lower DDPG network (i.e., the task offloading network) calculates the vehicle status according to the vehicle state. and a (w) Generating task offloading decisions Joint decision making Applied to edge collaborative computing, the reward value is used to guide the parameter update of the two-layer DDPG network; An additional policy network and linear layers are used to adapt to different action space dimensions; the policy network determines relevance by calculating attention scores between the input and output of the network. where Φ s and Φ a They represent the sampling results of the linear layer for the input state s and the output a, respectively, and D is the dimension of the output vector; the attention weight of each pair of input s and output a is where α s,a The larger the value, the better the match between s and a.

6. The method for autonomous driving resource scheduling and task offloading based on model segmentation according to claim 1 is characterized in that in the model segmentation, the residual block of Resnet F n (x), then the Resnet containing N residual blocks can be expressed as in represents the residual block set, Indicates function compound operation. Set split points according to dynamic split strategy The vehicle calculates the first l residual blocks of Resnet and obtains the intermediate feature representation The intermediate feature z is transmitted from the vehicle to the MEC server, which continues to calculate the remaining Nl residual blocks and obtains the final output In particular, the vehicle sets split point o1 to indicate that all tasks are offloaded to the MEC server, and split point o2 to indicate that the tasks are completed locally.

Images