Dual-stage resource reservation and priority scheduling system for accidents of Internet of Vehicles

Abstract

The invention discloses a dual-stage resource reservation and scheduling system for vehicle networking sudden accident emergency calculation, and relates to the technical field of edge calculation and intelligent traffic. The system comprises a traffic situation sensing module, a risk prediction decision module, a resource pre-freezing execution module, a task grading analysis module and a hybrid scheduling optimization module. The system innovatively divides a traffic accident handling process into an early warning stage and a response stage: in the early warning stage, the system adjusts a resource reservation ratio of edge nodes based on traffic flow density and historical task load conditions; in a response stage, the system performs hierarchical scheduling according to task types and dependency relationships by using a mixed integer programming model and a priority perception strategy. Through an end-side cloud collaborative architecture and a resource reservation mechanism, the system can significantly improve the task processing efficiency under emergencies and reduce the response time delay of high-priority tasks, has good expansibility and practicability, and is suitable for various vehicle networking application scenes such as intelligent expressways and urban traffic.

Description

Technical Field

[0001] The present invention relates to the field of edge computing and intelligent resource scheduling technologies for connected vehicles (IoVs), and more specifically, to an end-edge-cloud collaborative task scheduling system for traffic accident scenarios. In particular, it relates to a system that integrates a preventative resource reservation strategy, a subtask dependency resolution mechanism, and a mixed integer programming modeling method. The system is suitable for use in intelligent transportation systems, including emergency response to emergencies, rapid on-board task response services, and distributed edge computing resource management. Background Art

[0002] With the continued maturity of fifth-generation mobile communication technology (5G) and the Internet of Vehicles (IoV) technology, an increasing number of in-vehicle intelligent applications, such as autonomous driving, path planning, and obstacle avoidance and warning, are being widely deployed in smart cars. These applications have greatly improved travel safety and comfort. However, the rapid growth of in-vehicle applications has also led to an explosive growth in the variety and number of computing tasks generated by in-vehicle terminals, placing higher performance requirements on existing communication and computing infrastructure.

[0003] Currently, mainstream computing offload architectures mostly adopt a "terminal-edge server" model, offloading tasks generated by in-vehicle terminals to nearby edge nodes for processing to reduce latency and save local energy. Although this architecture is highly efficient in typical scenarios, in emergencies (such as traffic accidents), due to the sudden increase in task types and urgency, existing offload solutions face significant challenges in task scheduling and resource management.

[0004] Specifically, traffic accidents usually cause the following changes:

[0005] 1. A sharp increase in high-priority tasks: Low-latency tasks such as emergency obstacle avoidance and accident warnings require rapid processing, and the system must prioritize ensuring their response speed to avoid more serious chain reactions.

[0006] 2. Increased computationally intensive tasks: After an accident, vehicles often need to perform complex path replanning calculations, further increasing the system load.

[0007] 3. Increased overall workload: Due to accidents, vehicle speeds are reduced and vehicles gather together, resulting in a sudden increase in terminal computing demand in the same geographical area, which in turn creates congestion pressure on edge nodes.

[0008] Furthermore, in real-world scenarios, computing tasks are often broken down into multiple subtasks, which are then processed collaboratively by multiple computing nodes. While this task division improves resource utilization and processing efficiency to a certain extent, it also introduces data dependencies between subtasks, increasing the complexity of offloading strategy formulation.

[0009] When dealing with burst traffic and multi-priority task scheduling, most existing technologies adopt static resource allocation mechanisms or single-stage scheduling strategies, lacking fine-grained management of resource tension and task dependencies in emergencies. Therefore, how to balance the urgency of tasks, computing requirements, data dependencies, and the limited nature of edge computing resources in dynamic scenarios has become a key challenge in current task offloading research. Summary of the Invention

[0010] The present invention discloses a resource management strategy based on early warning-emergency two-stage collaboration for dynamic scheduling of computing resources and task offloading in the scenario of sudden traffic accidents in the vehicle-to-everything (V2X) environment. This strategy divides the traffic accident handling process into an early warning stage and an emergency stage, and designs corresponding scheduling mechanisms for each stage to coordinate the computing resources of terminals, edge nodes, and the cloud, so as to achieve low latency, high reliability, and maximized utilization of resources.

[0011] The system includes a traffic situation awareness module, a risk prediction and decision-making module, a resource pre-freezing execution module, a task classification and analysis module, and a hybrid scheduling optimization module. These modules operate in coordination, relying on edge nodes, vehicle-mounted terminals, and cloud infrastructure to build a complete intelligent response framework.

[0012] In the early warning stage, the system first receives real-time traffic data by the traffic situation awareness module to construct a spatio-temporal fusion depth model.

[0013] This model extracts short-term historical data features through a sliding time window mechanism. The input variables include key indicators such as upstream and downstream traffic flow and flow velocity difference, and the dimensionality influence is eliminated through standardization processing.

[0014] Subsequently, the system performs feature engineering, extracts statistical and spatio-temporal correlation indicators, uses the Boruta algorithm and PCA to achieve feature selection and dimensionality reduction, and introduces the SMOTE algorithm to enhance the representativeness of minority class samples.

[0015] In terms of the model structure, a bidirectional LSTM network is used to complete risk level prediction. The network contains three layers of LSTM units, with 64 nodes in each layer, Dropout set to 0.2, and finally the potential accident probability is output through Sigmoid.

[0016] When the prediction result exceeds the preset risk threshold, the system activates the resource pre-freezing subsystem.

[0017] This subsystem dynamically divides the edge node resource pool based on the current traffic density and network load to form a regular resource area and an emergency resource area, and adaptively adjusts the emergency resource configuration through proportional adjustment.

[0018] To optimize the resource freezing ratio and task allocation strategy, a genetic algorithm is introduced to perform evolutionary search on resource configuration.

[0019] The resource coding scheme uses a 12-bit binary string. The first 4 bits identify the frozen resource ratio, and the last 8 bits define the task category weight allocation.

[0020] The fitness function combines the delay cost and the energy consumption cost, and obtains the optimal resource allocation scheme through strategies such as tournament selection, two-point crossover, and bit-flip mutation.

[0021] In the emergency stage, when the system receives the confirmation signal of a sudden traffic accident event, the emergency resource scheduling subsystem immediately activates the frozen resources and switches the scheduling strategy.

[0022] Resources are divided according to the running state. Some resources enter the high-priority scheduling channel and can skip the queuing process to achieve zero-waiting response.

[0023] The key in this stage lies in the priority sorting of tasks. The task priority sorting module uses the b-level sorting mechanism to hierarchically divide the task sub-structure according to indicators such as task urgency, resource requirements, and system impact, and performs internal sorting according to the priority factor.

[0024] The task scoring model dynamically adjusts the sorting weight according to the real-time input to ensure that core tasks (such as accident identification and obstacle avoidance control) are scheduled and executed first.

[0025] The task priority sorting result will be used as the scheduling input and passed to the mixed-integer programming scheduling engine.

[0026] This module takes minimizing the scheduling cost as the optimization goal, constructs decision variables to represent whether a task is offloaded to a certain node, and at the same time introduces multiple sets of constraint controls, including task reachability, maximum acceptable delay, task dependency order, etc.

[0027] The objective function comprehensively considers the task execution delay and the transmission and computing energy consumption on different nodes, distinguishes three modes of terminal computing, edge offloading, and cloud collaboration, and takes into account both stability and flexibility.

[0028] The scheduling engine finely allocates the edge computing resources through the mixed-integer programming model, and at the same time uses the queuing steady-state model established in the early warning stage to estimate the possible resource waiting time to achieve fast response and maximize task throughput.

[0029] The beneficial effects produced by adopting the above technical solutions are as follows:

[0030] Improve the response speed to emergencies: Through the pre-freezing of resources in the early warning stage and the priority sorting of tasks in the emergency stage, the system can quickly respond when a sudden traffic accident occurs, ensure that high-priority tasks (such as emergency obstacle avoidance and accident warning) are processed in a timely manner, thereby reducing the accident chain reaction and improving the overall traffic safety.

[0031] Optimize resource scheduling efficiency: Combining the dual-stage scheduling mechanism in the early warning and emergency stages, rationally allocate edge computing resources, avoiding waste and over-occupation of resources. The resource reservation and freezing mechanism ensures the priority activation of emergency resources, enabling the system to maintain efficient task processing under high load.

[0032] Reduce the latency of high-priority tasks: Through the optimized resource pre-freezing and scheduling mechanism, zero-wait scheduling is achieved in the processing of emergency tasks, significantly reducing the response latency of high-priority tasks, thus ensuring that emergency tasks can be executed in the shortest time. Description of the Drawings

[0033] Figure 1 It is a schematic diagram of the end-edge-cloud collaborative processing structure of the resource management system based on early warning-emergency dual-stage collaboration according to the present invention in the scenario of sudden traffic accidents in the vehicle network.

[0034] Figure 2 It is a flow chart of the early warning-emergency dual-stage resource management method in the process of handling sudden traffic accidents according to the present invention;

[0035] Figure 3 It is a performance comparison chart of the average cost of the dual-stage scheduling strategy and the comparison scheme in the scenario of sudden events according to the present invention;

[0036] Figure 4 It is a performance comparison chart of the average delay of the dual-stage scheduling strategy and the comparison scheme in the scenario of sudden events according to the present invention;

[0037] Figure 5 It is a flow schematic diagram of a dual-stage resource reservation and priority scheduling system for sudden traffic accidents in the vehicle network according to the present invention. Detailed Embodiments

[0038] The following will describe the detailed embodiments of the present invention in conjunction with the drawings. It should be understood that the following embodiments are only used to illustrate the present invention and do not limit the protection scope of the present invention.

[0039] Figure 1 It is a schematic diagram of the architecture of the end-edge-cloud collaborative scheduling system according to some embodiments of this specification. As Figure 1 shown, the end-edge-cloud collaborative scheduling system may include the following parts:

[0040] On-vehicle terminal nodes (Vehicles): Having sensing and preliminary computing capabilities, responsible for collecting original task data and sending offloading requests to the edge server.

[0041] Edge Computing Node (Edge Server): Deployed at locations such as traffic intersections and service areas, it undertakes most of the task processing and has a partial resource pre-freezing mechanism.

[0042] Cloud Computing Center (Cloud Server): As a supplement to computing resources, it is mainly used to assist in executing computing tasks when edge computing resources are insufficient or the task processing is complex.

[0043] Scheduling Control Unit (Scheduling Controller): Deployed at edge nodes or working in coordination with the cloud computing center, it is responsible for early warning detection, resource planning, formulation of scheduling strategies, and core decision-making.

[0044] Figure 2 It is a schematic flowchart of the edge-cloud collaborative scheduling system shown in some embodiments of this specification. As Figure 2 shown, the present invention adopts a two-stage scheduling mechanism, including an early warning stage and a response stage. This scheduling mechanism may include the following steps:

[0045] Step 1, in the traffic situation perception module, the current traffic flow conditions are jointly collected by vehicles and roadside edge servers.

[0046] The information collected may include upstream and downstream traffic flow, upstream and downstream vehicle speeds, and upstream and downstream lane occupancy rates over a period of time, etc.

[0047] Step 2, according to the collected data, min-max transformation is performed on features such as upstream traffic flow, downstream traffic flow, and flow velocity difference to eliminate the dimension difference, and a sliding time series dataset is constructed with a 5-minute time window.

[0048] Based on the current data, spatio-temporal features of the data are extracted, including:

[0049] Statistical features: road section type, traffic flow mean, average speed, speed difference between lanes, traffic flow / speed variance, autocorrelation coefficient, etc.

[0050] Spatio-temporal correlation features: upstream and downstream traffic flow difference, upstream and downstream traffic flow occupancy difference, upstream and downstream average speed difference, upstream and downstream traffic flow / occupancy, speed and difference from 10 minutes to 5 minutes before an accident, etc.

[0051] Step 3, in the risk prediction and decision-making module, first the Boruta algorithm is used for feature selection, the principal component analysis is used to reduce the dimension of features, and finally the SMOTE algorithm is used for data augmentation to solve the problem of class imbalance.

[0052] Step 4, a bidirectional LSTM network is used for modeling. The structure of the bidirectional LSTM network includes: an input layer (with 16 neurons corresponding to the feature dimension), a hidden layer (3 layers of LSTM structure, with 64 units in each layer and a dropout rate of 0.2), and an output layer (outputting the probability of a traffic accident risk through a Sigmoid activation function).

[0053] When the prediction result is that no traffic accident will occur, the status quo will be maintained and no measures will be taken, and it will return to step 120. When the prediction result is that a traffic accident will occur, the resource pre-freezing mechanism will be triggered and the emergency resource management process will be started.

[0054] Step 5, if the prediction result indicates the existence of potential accident risks, the task queuing situation at the steady state before the accident is calculated through queuing theory. The specific steps are as follows:

[0055] S51. Assume that there are K types of heterogeneous data streams in the system, and the arrival processes of various tasks follow a Poisson distribution with parameter λ i (i ∈ [1, k]). Define the allocation probabilities of tasks at the terminal device, edge node, and cloud center as ρ v , ρ e , ρ c , where the probability normalization constraint is satisfied: ρ v + ρ e + ρ c = 1. Based on this, the task arrival rates of each computing node can be obtained:

[0056] λ i local = ρ v λ i , λ i edge = N v ρ e λ i , λ i cloud = N v ρ c λ i

[0057] where N v is the number of vehicles within the service coverage area. μ m i (m ∈ {local, edge, cloud}) represents the average service duration of tasks at node m (including transmission and computing delays).

[0058] S52. Based on the queuing theory model, the resource utilization rate of each node can be characterized as:

[0059] ρ i m = λ im / μ i m

[0060] S53. When the priority i task arrives, the expected remaining waiting time for the task currently being served by the current device is:

[0061]

[0062] According to the non-preemptive priority queuing theory, the waiting time of the task with priority k is:

[0063]

[0064] S54, thus the total delay of the task with priority k is:

[0065]

[0066] The resource pre-freeze ratio θ(t) is determined based on the current traffic flow information, and the available computing resources are reconstructed as:

[0067] C m freeze =(1-θ)C m max , where C m max is the original resource capacity of node m.

[0068] Step 6: After determining the calculation method of resource pre-freeze ratio and task waiting time, a genetic algorithm is used to solve the allocation ratio of tasks of different priorities to each device under steady state conditions.

[0069] Step 7: Calculate the length of time that different tasks stay in the queue in a steady state, which is used as the waiting time required for tasks of different priorities to be executed on each device in an emergency scenario.

[0070] When an emergency occurs and is confirmed, immediately switch to the response phase.

[0071] Step 8: Task priority classification: Based on the task type and service urgency, the tasks are classified into high, medium, and low priorities. The details are as follows:

[0072] At each moment, the vehicle terminal generates tasks, which can be divided into a series of subtasks through the segmentation algorithm. These subtasks usually have data dependencies, but they can still be executed on different devices, as long as the execution order is guaranteed. This allows the offloading of subtasks to more efficiently utilize computing device resources. The data dependencies between subtasks can be represented by a directed acyclic graph G i (N i ,E i) indicates, where N i represents the subtask node of task i, E i represents the dependency relationship between subtasks.

[0073] To effectively schedule related tasks between different devices, the most important issue is to determine the priority of tasks in the system. For task i, its task priority is related to its own task attributes, task deadline, and required resources. For subtask j of task i, the priority is more related to the dependency relationship between subtasks. This paper uses bottomlevel (b-level) to sort the subtasks of tasks. Task v i,j 's b-level is defined as the longest path from task v i,n+1 to task v i,j , where v i,n+1 is the virtual final node, and the path length represents the sum of the computing cost and transmission cost of the nodes on the path.

[0074] For task i, its task priority is mainly related to its own importance, deadline, and required resources.

[0075] The priority of task i where ω i represents the attribute of the task itself, d i represents the maximum tolerable delay of task i, c i , s i and c max , s max represent the computing resources and transmission resources required by task i, as well as the maximum computing resources and transmission resources, respectively.

[0076] Based on this, for each task i in the system, first determine the scheduling order of the task according to the overall priority of task i, and then determine the scheduling order of each subtask in turn according to the subtask priority.

[0077] Step 9, formulate the task offloading scheduling strategy. Among them, the scheduler determines the optimal offloading strategy for each task in the queue according to the real-time system state of the end-edge-cloud architecture to reduce the task execution delay and improve the user experience.

[0078] When vehicle v submits task i, the execution time of its subtask j is defined as: where c i,j represents the computing resources required by task j

[0079] S91, energy consumption modeling and analysis.

[0080] When the task is determined to be executed locally, computational energy consumption will be generated. According to the research of [Zhan], the average power consumption (in Joule / slot) can be modeled as a super-linear function of the CPU frequency, and the expression is: p l = ξ·(f l ) v , where ξ and ν are constants. Therefore, the energy consumed by task j when executed locally can be calculated by the following formula:

[0081] e v,i,j l = p l T v,i,j com,local

[0082] When it is determined to be executed at the edge node and in the cloud, transmission energy consumption will be generated. The communication rate between vehicle i and the edge server and the cloud can be calculated by the following formula:

[0083] where B is the available bandwidth between the RSU and the vehicle, p v is the transmission power of vehicle v, h vm represents the channel gain between vehicle v and the RSU, and σ 2 is Gaussian noise.

[0084] The transmission delay required for subtask j to be transmitted from the vehicle to device m is:

[0085] where s i,j represents the data volume size of subtask j.

[0086] The energy required to transmit subtask j is given by the following formula:

[0087] e v,i,j o = p v ·T v,i,j trans

[0088] S92, construction of the task delay calculation model.

[0089] The scheduler decides which server task j is assigned to execute according to the offloading strategy. Before task j is executed, the following two conditions need to be met: 1. All the preceding subtasks of task j have been executed; 2. The current device is idle. Then the earliest execution time of task j is defined as:

[0090]

[0091] where preQ m <j>Tasks on device m that are before j in order, pre <j>The subtask preceding j.

[0092] Then the completion time of task j on device m is:

[0093]

[0094] S93, Implementation of the dynamic resource scheduling mechanism

[0095] In the emergency phase, the system allocates computing resources through a mixed-integer programming (MIP) model to ensure the rapid response of tasks. The specific model is as follows:

[0096]

[0097] s.t.

[0098]

[0099]

[0100] (1),(2)

[0101] x v,i,j,m ∈{0,1},

[0102] where x v,i,j,m is a decision variable that identifies whether task j is offloaded to node m. The objective function consists of two parts. The first part is the latency cost from the generation of the task to its final execution, and the second part is the computing energy consumption when the task is computed at the terminal and the transmission energy consumption when the task is decided to be offloaded to the edge or the cloud. The constraint conditions ensure task integrity and latency feasibility. The device waiting time is pre-computed through the steady-state queuing model established in the early warning phase, and zero-wait scheduling can be achieved with frozen resources.

[0103] IV. Description of Preferred Embodiments

[0104] The following further illustrates the application effect of the scheduling method of the present invention in the vehicle networking environment in combination with specific application scenarios. This embodiment should not be regarded as a limitation on the protection scope of the present invention.

[0105] In this embodiment, the method of the present invention is deployed on a certain highway section, and multiple vehicle-mounted terminals, an edge computing node (Edge Node E), and a cloud computing node (Cloud Node C) are deployed within the coverage area. Among them:

[0106] · The vehicle-mounted terminal is responsible for collecting perception data such as vehicle speed, acceleration, camera images, and radar scans;

[0107] · The edge node E is deployed near the service area or transportation hub and has medium computing resources and task caching capabilities;

[0108] · The cloud node C is deployed in a remote data center, with strong computing capabilities but relatively high communication latency.

[0109] Under different task load levels, the "pre-freezing - post-scheduling" two-stage scheduling strategy (WECO) proposed in the present invention is compared and tested with three existing baseline strategies (such as the static average allocation strategy, the shortest queue first strategy, and the priority-only strategy). The test results are as Figure 1 and Figure 2 shown.

[0110] In the experimental setup, the computing tasks generated by the vehicles are divided into three categories of priorities (high Priority 3, medium Priority 2, low Priority 1), and each task has different latency sensitivities and resource requirements.

[0111] Figure 1 shows the comparison of the average task costs (including the weighted sum of latency and energy consumption) of the four scheduling strategies under different numbers of tasks. The results show that under different load conditions where the number of tasks increases from 36 to 132, the model proposed in the present invention achieves the lowest average task cost at all task scales. Compared with the baseline model with the best performance (Normal strategy), the WECO model reduces the cost by more than 10% under each load condition, verifying the effectiveness of the resource pre-freezing and phased scheduling mechanisms in task offloading and resource allocation.

[0112] Figure 2 Further shows the average processing latency of tasks with each priority under different numbers of tasks:

[0113] · For high-priority tasks (Priority 3), when the number of tasks is 36, the average latency of the WECO model is 0.130 seconds, much lower than 0.225 seconds of the Normal strategy, with a reduction rate of 42.2%. When the number of tasks increases to 132, WECO can still maintain a low latency of 0.266 seconds, better than 0.296 seconds of the Normal strategy. Through linear regression analysis, the latency growth rate of high-priority tasks under different loads is 0.0010 seconds / task, only 60% of the Normal strategy, indicating that the pre-freezing mechanism in the present invention can effectively alleviate the latency accumulation effect caused by task accumulation.< / j> < / j>

Claims

1. A two-stage resource reservation and task scheduling system for vehicle networking emergency accidents, characterized in that It includes a traffic situation awareness module, a risk prediction and decision-making module, a resource pre-freezing execution module, a task classification and parsing module, and a hybrid scheduling optimization module. The traffic situation awareness module is used to collect and analyze vehicle operation status, road traffic data, and traffic flow density information in real time, and construct a road traffic situation awareness map. The risk prediction and decision-making module is used to predict the risk probability of traffic accidents based on the traffic situation awareness results and historical accident data, and determine whether to enter the early warning stage. The resource pre-freezing execution module is used to freeze a certain proportion of computing and communication resources according to the adaptive strategy based on the prediction results and the resource load status of the current edge node during the early warning stage. The task classification and parsing module is used to divide the priority of tasks and decompose them into subtasks according to the urgency, resource requirements, and dependencies of tasks after the tasks enter the queue. The hybrid scheduling optimization module is used to receive the task classification results during the response stage, and generate an optimal scheduling strategy under the end-edge-cloud architecture through a mixed-integer programming model to achieve the priority scheduling of high-priority tasks. The scheduling execution and feedback module is used to schedule the execution of tasks according to the decision instructions output by the hybrid scheduling optimization module, and feedback the execution status to the system management center in real time for closed-loop optimization.

2. The method according to claim 1, wherein The obtaining of traffic state data and feature generation processing includes: Collecting vehicle density, traffic flow change rate, and communication quality indicators to generate original traffic data. Performing fusion processing on the original traffic data, constructing a feature vector containing multi-dimensional elements, and generating traffic feature data.

3. The method according to claim 1, wherein The accident risk prediction based on traffic feature data includes: Inputting the traffic feature data into a risk assessment model to output the continuous value range risk probability of accident occurrence. Judging the risk level according to the set dynamic threshold, and triggering the resource pre-freezing process when the threshold is exceeded.

4. The method according to claim 1, characterized in that, The execution of resource pre-freezing includes: Dividing the edge node computing resources into a regular resource pool and an emergency resource pool. Dynamically adjusting the proportion of the emergency resource pool according to the change of real-time traffic flow density to complete the resource freezing configuration.

5. The method according to claim 1, wherein The parsing of tasks to be processed and the generation of a scheduling graph include: Dividing the priority according to the urgency of the task, and calibrating it into three levels: high, medium, and low. Constructing a directed acyclic graph according to the structural dependencies of tasks to generate a task dependency constraint graph.

6. The method according to claim 1, characterized in that, The generation of a task scheduling plan and resource coordination include: Generating a multi-level scheduling plan for end nodes, edge nodes, and cloud nodes based on task priorities and dependency graphs. Coordinating the available resources of each node to ensure that high-priority tasks preferentially occupy the computing resources in the emergency resource pool.

Images