A vehicle-road cooperation control delay suppression method based on edge computing

Abstract

The application discloses a vehicle-road cooperative control delay suppression method based on edge computing, which comprises the following steps: acquiring a global multi-target weight vector issued by a cloud management and control platform and multi-dimensional state parameters collected by a vehicle-end intelligent agent and an edge computing node in real time; constructing a hierarchical decision adaptation degree function, calculating a decision adaptation degree value matched with each candidate decision level at the current time through the hierarchical decision adaptation degree function; matching a preset adaptation degree threshold interval to determine a target decision level effective at the current time; calling a target reward function matched with the target decision level in a pre-constructed hierarchical multi-target game reward function suite; and taking the target reward function as an optimization target to execute hierarchical reinforcement learning decision reasoning. The application realizes full-link suppression of vehicle-road cooperative control delay from four dimensions of transmission delay, calculation delay, reasoning delay and iteration delay, while guaranteeing the full-scene adaptability, safety, convergence and feasibility of the decision.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and vehicle-road cooperative technology, specifically to a method for suppressing delays in vehicle-road cooperative control based on edge computing. Background Technology

[0002] With the rapid development of autonomous driving technology, vehicle-road cooperative systems (V2X) have overcome the limitations of single-vehicle perception and decision-making in terms of vision and computing power through information interaction and collaborative decision-making between vehicles, edge computing, and the cloud, becoming a core supporting technology for the implementation of high-level autonomous driving. Edge computing, as the core computing power carrier of V2X systems, enables local low-latency processing of roadside perception data and the issuance of decision commands, representing a core technical path for solving the latency problem in V2X control. Hierarchical reinforcement learning (HRL), by decomposing complex autonomous driving decision-making tasks into high-level sub-objective planning and low-level action execution, solves the problems of low sample efficiency, poor generalization ability, and uncontrollable decision-making in end-to-end reinforcement learning. It can be deeply integrated with edge computing to effectively suppress latency in V2X control.

[0003] Existing vehicle-road cooperative control schemes based on edge computing generally suffer from the following core technical deficiencies in latency suppression: The static and fixed decision-making architecture cannot adapt to the latency suppression requirements of dynamic scenarios: Existing solutions all adopt a fixed two-layer architecture of "edge layer sub-target planning + vehicle layer action execution". The layer granularity, decision permissions and interaction rules of the architecture are fixed in the design stage. This cannot adapt to the real-world scenarios of dynamic fluctuations in V2X communication quality, changes in traffic scenario complexity and heterogeneity of vehicle computing power. In degraded scenarios such as communication interruption and high latency, problems such as sub-target failure, decision timeout and sharp increase in control latency are very likely to occur, and it is impossible to achieve latency adaptive suppression in dynamic scenarios.

[0004] Deep decoupling of communication and decision-making, only peripheral delay compensation: Those skilled in the art generally regard V2X communication as a data transmission link before decision-making, rather than an endogenous variable for controlling delay suppression. Communication problems can only be compensated for by peripheral means such as data retransmission and timeout retries. It is impossible to achieve synergistic optimization of communication quality and decision logic from the bottom layer of decision-making, and it is impossible to solve the problem of excessive transmission delay and control loop failure caused by communication degradation at the root.

[0005] The disconnect between upper and lower reward functions leads to slow decision convergence and high inference latency: Existing solutions only weakly correlate the upper-level global reward with the vehicle-side local reward through the completion of sub-goals. They generally use manually fixed-weight multi-objective reward functions, which cannot balance the conflicts between multiple objectives such as safety, efficiency, and energy consumption. This easily leads to the problem of "lower-level sub-goals being completed but upper-level global optimization objectives not being met," resulting in a large number of decision iterations and slow convergence speed, which significantly increases decision inference latency. At the same time, it also suffers from the common problems of sparse rewards and poor generalization ability.

[0006] The architecture and reward are designed independently, resulting in a fundamental mismatch and making it impossible to achieve closed-loop optimization of latency: The existing solution follows a fixed process of "first defining a fixed layered architecture, and then designing a reward function to match the architecture". The architecture switching logic is completely disconnected from the global latency optimization goal. The reward function and the decision level cannot be dynamically adapted. Optimizing the layered architecture or the reward function separately cannot solve the problem of decreased decision performance and high control latency caused by the mismatch between the two, and it is impossible to form a complete closed-loop optimization for latency suppression. Summary of the Invention

[0007] This invention proposes a vehicle-road cooperative control latency suppression method based on edge computing, which can at least solve the core pain points of existing technologies such as static architecture, decoupling of communication and decision-making, mismatch between reward and decision-making levels, and conflict between global and local optimization objectives. It achieves end-to-end latency suppression of vehicle-road cooperative control from four dimensions: transmission latency, computation latency, inference latency, and iteration latency, while ensuring the adaptability, security, convergence, and feasibility of decision-making across all scenarios.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A delay suppression method for vehicle-road cooperative control based on edge computing, applied to a vehicle-road cooperative system including a vehicle-side intelligent agent, edge computing nodes, and a cloud-based management and control platform, includes the following steps: S1. Obtain the global multi-objective weight vector issued by the cloud management platform, as well as the multi-dimensional state parameters collected in real time by the vehicle-side intelligent agent and edge computing nodes; the global multi-objective weight vector contains the dynamic weights of three optimization objectives: regional traffic efficiency, driving safety, and energy consumption economy, and satisfies the normalization constraint that the weight sum is 1; the multi-dimensional state parameters include the communication quality factor, scenario complexity factor, vehicle computing power factor, and safety risk factor that directly affect the delay of vehicle-road cooperative control, and all parameters are normalized values ​​in the range of 0 to 1; S2. Using the global multi-objective weight vector as an endogenous driving variable, construct a hierarchical decision fitness function, and calculate the decision fitness value that matches each candidate decision level at the current time through the hierarchical decision fitness function; the hierarchical decision fitness function couples the global multi-objective weight vector with the hierarchical basic weight matrix of multi-dimensional state parameters to match the optimal decision level that can minimize control delay; S3. Based on the decision fit value at the current moment, match the preset fit threshold range to determine the target decision level that is effective at the current moment; the candidate decision level includes four preset decision levels, each decision level corresponds to a unique fit threshold range, vehicle-edge-cloud decision permission allocation rules and information interaction rules, and suppress end-to-end control delay through dynamic adaptation of permissions and interaction rules. S4. Based on the target decision level, call the target reward function that matches the target decision level from the pre-built hierarchical multi-objective game reward function suite; the hierarchical multi-objective game reward function suite contains reward functions that correspond one-to-one with the four candidate decision levels. Each reward function is set with multi-objective weights, upper and lower layer reward alignment coefficients and safety constraint penalty terms that match the target decision level, which are used to accelerate decision convergence and suppress decision reasoning delay; S5. Based on the decision authority allocation rules corresponding to the target decision level, and with the target reward function as the optimization objective, perform hierarchical reinforcement learning decision reasoning, output the autonomous driving control command of the vehicle-side intelligent agent, and reduce computation latency and transmission latency through the collaborative allocation of computing power between edge computing nodes and vehicle-side intelligent agents. S6. Collect execution feedback data of autonomous driving control commands, update the global multi-objective weight vector and multi-dimensional state parameters based on the execution feedback data, and complete the dynamic iterative optimization of decision-making closed-loop optimization and delay suppression effect.

[0009] As a preferred embodiment of the edge computing-based vehicle-road cooperative control latency suppression method of the present invention, the global multi-objective weight vector is dynamically generated by the cloud management platform based on the real-time regional traffic situation, traffic control rules and emergency event status, or configured by a preset fixed control strategy; when an emergency event or high-risk scenario is detected, the weight ratio corresponding to the driving safety target is automatically increased to above the preset safety threshold, and the priority configuration of the low-latency decision mode is triggered at the same time.

[0010] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method of the present invention, the calculation formula of the hierarchical decision-fit function is:

[0011] In the formula for Time of the first The decision fit value corresponding to the candidate decision level. for The global multi-objective weight vector at time step [time]. The sequence number of the candidate decision-making level and , For the first The hierarchical base weight matrix corresponding to the candidate decision level. For communication quality factor, As the scene complexity factor, For vehicle computing power factor, As a safety risk factor.

[0012] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method of the present invention, wherein: the fitness threshold ranges of the four candidate decision levels and the decision authority allocation rules are arranged in ascending order of fitness value as follows: Level 1: The adaptability value is in the first threshold range, corresponding to the pure vehicle-side decision-making mode. The vehicle-side intelligent agent takes over the decision-making authority of the entire link. The edge computing nodes and the cloud management platform only perform perception data backup and security monitoring. Transmission latency is suppressed by closing cross-end data interaction. Level 2: The adaptation value is in the second threshold range, corresponding to the vehicle-dominated - edge-assisted mode. The edge computing node only issues global traffic constraint rules, and the vehicle-side intelligent agent autonomously completes sub-objective planning and action execution decisions, thereby reducing control latency by reducing the frequency of edge-vehicle interaction. Level 3: The adaptability value is in the third threshold range, corresponding to the standard collaborative mode. The edge computing node is responsible for the planning of the passage sub-goal, the vehicle-side intelligent agent is responsible for the action execution decision, and the cloud management platform issues global optimization constraints. The computing power of the edge and the vehicle is split to reduce the computing latency of a single node. Level 4: The adaptability value is in the fourth threshold range, corresponding to the edge full control mode. The edge computing node directly issues trajectory-level decision commands, the vehicle-side intelligent agent only executes commands and performs safety redundancy checks, and the cloud management platform issues global access policies. The overall control latency of multi-vehicle collaboration is reduced by optimizing the centralized computing power of the edge nodes.

[0013] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method described in this invention, wherein: in the hierarchical multi-objective game reward function suite, the reward function corresponding to each level includes a basic multi-objective reward term, a hierarchical matching reward alignment term, and a safety constraint penalty term; wherein: The reward function corresponding to level 1 retains only the basic multi-objective reward items and safety constraint penalty items on the vehicle side, and disables the upper and lower layer reward alignment items, so as to achieve fast decision convergence of single node on the vehicle side and suppress inference latency. The reward function corresponding to level 2 sets a one-way reward alignment item between the vehicle end and the edge, disables the cloud reward alignment item, reduces the computational overhead of cross-level alignment, and reduces the delay of decision iteration. The reward function corresponding to level 3 sets a bidirectional reward alignment term between the vehicle end and the edge, with the cloud reward serving as a soft constraint alignment term to balance the accuracy and convergence speed of collaborative decision-making and avoid the increase in latency caused by invalid iterations; The reward function corresponding to level 4 sets the full-link reward alignment term from vehicle to edge to cloud. The global reward in the cloud is used as a hard constraint alignment term to achieve one-time convergence of global collaborative decision-making and suppress the cumulative delay caused by multiple rounds of interaction.

[0014] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method described in this invention, when a switch in the target decision level is detected, a reward smooth transition mechanism is activated. Within two preset decision cycles, the weight coefficients and alignment coefficients of the reward function are smoothly adjusted through linear gradient interpolation. At the same time, a strategy consistency check is performed to ensure that the rate of change of the control quantity of the autonomous driving control command output before and after the switch does not exceed a preset threshold, thereby avoiding control jitter and increased closed-loop iteration delay caused by the level switch.

[0015] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method of the present invention, the hierarchical reinforcement learning decision model is constructed by jointly training an architecture switching policy network and a multi-objective reward optimization network, specifically including: The architecture switching strategy network and the multi-objective reward optimization network share the global multi-objective weight vector and multi-dimensional state parameters as input layers; Construct a joint loss function, which includes a multi-objective reward achievement penalty term for architecture switching, an architecture switching stability constraint term for reward optimization, and a control latency exceeding penalty term; A multi-agent reinforcement learning framework is adopted to complete joint training based on the interaction data of multiple vehicle-end agents in the same road network, and to simultaneously optimize the parameters of the two networks, thereby improving the decision response speed in dynamic scenarios and suppressing control delay fluctuations in sudden scenarios.

[0016] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method described in this invention, hard safety constraint rules are embedded throughout the entire process of hierarchical decision adaptation function calculation, target decision level matching, and target reward function invocation. When the safety risk factor exceeds the preset risk threshold, the target decision level is forcibly switched to level 1 or level 2, and the weight ratio of driving safety target in the target reward function is forcibly increased to filter all decision outputs that violate safety constraints and avoid control delay exceeding the standard due to safety verification iteration.

[0017] As a preferred embodiment of the edge computing-based vehicle-road cooperative control delay suppression method of the present invention, wherein: the step of updating the global multi-objective weight vector and multi-dimensional state parameters based on execution feedback data specifically includes: Based on the execution feedback data, calculate the global optimization goal achievement rate, security risk status, communication quality change value, and end-to-end control delay measured value for the current decision cycle; The global optimization target achievement and the measured control delay are fed back to the cloud management platform to dynamically adjust the global multi-objective weight vector and low-latency decision priority for the next cycle. The security risk status and communication quality change values ​​are synchronized to the vehicle-side intelligent agent and edge computing nodes to update the multi-dimensional state parameters for the next cycle, thereby achieving dynamic optimization of latency suppression capabilities.

[0018] As a preferred embodiment of the edge computing-based vehicle-road cooperative control latency suppression method described in this invention, different target decision levels correspond to different vehicle-edge-cloud information interaction rules. Transmission latency is suppressed through adaptive adjustment of data transmission volume and interaction frequency, specifically: When the target decision level is level 1, only the heartbeat information interaction between the vehicle and the edge is retained, and unnecessary autonomous driving and vehicle-road cooperative data transmission is turned off to minimize the control latency caused by the transmission link. When the target decision level is level 2, the vehicle and the edge only interact with traffic constraint rules and global perception data, reducing the latency overhead caused by invalid data transmission. When the target decision level is level 3 or level 4, the perception data, sub-target instructions and global strategy data required for full interaction decision-making between the vehicle, edge and cloud are ensured. At the same time, the local data processing of edge nodes avoids long-link transmission delay in the cloud.

[0019] The beneficial effects of this invention are: This invention breaks away from the conventional design logic of "first defining a fixed architecture, then designing the reward function" in the field. It uses a global multi-objective weight vector as an endogenous driving variable for hierarchical architecture switching, while allowing the reward function to dynamically adapt to the decision-making level, forming a complete closed loop of "objective-driven architecture switching - architecture matching reward system - reward-optimized decision-making strategy - strategy feedback to goal achievement". This fundamentally solves the core pain points of slow decision convergence and high control latency caused by the disconnect between the existing technical architecture and the optimization objective, and the mismatch between rewards and the decision-making level, and achieves full-link closed-loop suppression of control latency.

[0020] This invention uses the communication quality factor as the core input of the hierarchical decision adaptation function, transforming V2X communication quality from an external constraint of decision-making into an endogenous variable for controlling latency suppression. Through adaptive switching of four types of dynamic decision levels, coupled with differentiated vehicle-edge-cloud decision permissions and information interaction rules, it perfectly adapts to the needs of all scenarios, from pure vehicle-side scenarios with communication interruptions to full-domain collaborative scenarios with excellent communication quality. It achieves dynamic adaptive suppression of latency control from four dimensions: transmission, computation, inference, and iteration. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the steps of the vehicle-road cooperative control delay suppression method based on edge computing of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0023] like Figure 1 As shown, this embodiment presents a vehicle-road cooperative control latency suppression method based on edge computing, which is applied to a vehicle-road cooperative system including a vehicle-side intelligent agent, an edge computing node, and a cloud-based management and control platform. Vehicle-mounted intelligent agent: Intelligent connected vehicles equipped with L4-level autonomous driving systems, featuring onboard LiDAR, cameras, millimeter-wave radar and other perception devices, onboard computing platforms, and onboard units (OBU) that support C-V2X communication. Edge computing nodes: Multi-access edge computing (MEC) devices deployed at urban road intersections, equipped with roadside LiDAR, cameras and other all-domain perception devices, roadside units (RSUs), can realize full perception and low-latency computing of traffic participants within the communication domain; Cloud-based management and control platform: A city-level intelligent traffic management and control cloud platform that can acquire traffic situation data of the entire road network, issue traffic management and control rules and global optimization strategies, and achieve global optimization of regional traffic flow.

[0024] The core utilizes edge computing as its core computing power carrier. Through dynamic adaptive switching of a hierarchical decision-making architecture, it achieves latency suppression across the entire vehicle-road cooperative control link. In this embodiment, the decision cycle is 100ms, meaning that a complete decision-making and latency optimization process is executed every 100ms. The specific implementation steps are as follows: Step 1: Obtain the global multi-objective weight vector and multi-dimensional state parameters First, the vehicle-side intelligent agent communicates with the edge computing node via V2X to obtain the global multi-objective weight vector issued by the cloud-based management platform. ; In this embodiment, the global multi-objective weight vector ,in The weighting for regional traffic efficiency targets. As a weight for driving safety objectives, As the weight of energy consumption economic targets, satisfy ; In typical urban road scenarios, the default weight vector distributed from the cloud is: When high-risk scenarios such as intersection accidents or emergency vehicle passage are detected, the cloud automatically increases the weight of the driving safety objective to above 0.8. For example... It meets the preset security threshold requirements and triggers the priority configuration of the low-latency decision mode.

[0025] Meanwhile, the vehicle-side intelligent agent and edge computing nodes collect multi-dimensional state parameters in real time. All parameters are normalized to the range of 0 to 1. The specific calculation method is as follows: 1. Communication quality factor The formula for calculating one-way latency, packet loss rate, and available bandwidth based on V2X communication is as follows:

[0026] In the formula, This is the one-way communication delay (unit: ms), with a maximum tolerable delay of 200ms. For communication packet loss rate; Available communication bandwidth, This refers to the total bandwidth; for example, when the latency is 40ms, the packet loss rate is 3%, and the bandwidth is sufficient, At that time, with a latency of 180ms and a packet loss rate of 35%, .

[0027] 2. Scene complexity factor The formula is calculated based on the number of traffic participants, scenario type, and interaction complexity within the communication domain:

[0028] In the formula, The number of traffic participants in the current scenario. The preset maximum quantity threshold (set to 50 in this embodiment); This is a scenario type coefficient: 0.2 for highway cruising, 0.5 for ordinary urban roads, and 0.1 for intersections without traffic lights. For example, at an intersection without traffic lights with 20 traffic participants... .

[0029] 3. Vehicle computing power factor The formula for calculating the remaining computing power percentage based on the onboard computing platform is as follows:

[0030] In the formula, For the remaining available computing power of the vehicle platform, This refers to the total computing power of the vehicle platform; for example, when the vehicle platform has 60% of its computing power remaining, .

[0031] 4. Safety risk factors Based on the collision probability, following distance, and relative speed between the vehicle and surrounding traffic participants, calculations are performed. The higher the collision probability and the closer the following distance, the better. The higher the value, for example, when there is sufficient following distance and no risk of collision. When following another vehicle at close range, where the probability of a collision is high, .

[0032] Step 2: Calculate the decision fit value for each candidate decision level. The global multi-objective weight vector obtained in step 1 As an endogenous driving variable, a hierarchical decision fit function is constructed to calculate the decision fit values ​​corresponding to the four candidate decision levels at the current time.

[0033] In this embodiment, the formula for calculating the hierarchical decision fitness function is as follows:

[0034] Among them, the hierarchical basic weight matrix corresponding to the four types of candidate decision levels As shown in Table 1 below: Table 1 ; Taking a typical scenario as an example, , , , , Substituting into the formula, we get: Level 1 fit score: 0.278; Level 2 fit score: 0.376; Level 3 fit score: 0.474; Level 4 fit score: 0.572; Step 3: Match threshold ranges to determine the target decision level In this embodiment, the preset fit threshold ranges, arranged from low to high fit values, are as follows: First threshold range (Level 1): 0~0.3 Second threshold range (Level 2): ​​0.3~0.5 Third threshold range (Level 3): 0.5~0.7 Fourth threshold range (level 4): 0.7~1.0 Based on the fitness value calculated in step 2, the fitness value of level 4, 0.572, falls within the third threshold range. Therefore, the target decision level at the current moment is determined to be level 3 (standard collaborative mode). The corresponding decision authority allocation rules are as follows: edge computing nodes are responsible for planning traffic sub-objectives (such as target vehicle speed, target lane, and intersection passage sequence) for the vehicle-side intelligent agent. The vehicle-side intelligent agent is responsible for outputting autonomous driving control commands for longitudinal acceleration / deceleration and lateral steering based on the sub-objectives. The cloud management platform issues global optimization constraints. By splitting the computing power between the edge and the vehicle, the computational latency of a single node is reduced, while avoiding long-link transmission latency in the cloud.

[0035] When a communication interruption occurs The calculated fitness value of level 1 is the highest and is within the first threshold range. Therefore, the target decision level is switched to level 1 (pure vehicle-side decision mode). The vehicle-side intelligent agent takes over the decision-making authority of the entire link, closes unnecessary cross-end data transmission, completely suppresses the control delay caused by the transmission link, and avoids decision failure and delay exceeding the standard caused by communication interruption.

[0036] Step 4: Call the matching target reward function Based on the target decision level determined in step 3, the target reward function that matches level 3 is called from the pre-built hierarchical multi-objective game reward function suite.

[0037] In this embodiment, the reward function suite for hierarchical multi-objective game theory includes a basic multi-objective reward term, a hierarchical matching reward alignment term, and a safety constraint penalty term in the reward function of each level, specifically: 1. The edge layer reward function corresponding to level 3:

[0038] In the formula, For collision risk penalties, As a reward item for sub-goal completion, This is the normalized value of communication resource usage. This is the global reward value in the cloud. Assign a cloud-based reward alignment coefficient to align edge decisions with global optimization goals; 2. The vehicle-side reward function corresponding to level 3:

[0039] In the formula, For sub-target tracking accuracy reward item, This is the normalized value of longitudinal acceleration. This represents the normalized value of single-vehicle energy consumption. This is the edge reward alignment coefficient, used to align vehicle-side actions with edge sub-targets.

[0040] When the target decision level is switched to level 1, the reward function corresponding to level 1 is automatically called. Only the basic multi-objective reward items and safety constraint penalty items on the vehicle side are retained, and all upper and lower level reward alignment items are disabled. This adapts to pure vehicle side decision-making scenarios, enables fast decision convergence of single node on the vehicle side, and minimizes inference latency.

[0041] When a switch in the target decision level is detected, a reward smooth transition mechanism is activated. Within two decision cycles, the reward weight and alignment coefficient are smoothly adjusted through linear gradient interpolation. At the same time, a strategy consistency check is performed to ensure that the rate of change of acceleration and deceleration of the autonomous driving control commands before and after the switch does not exceed 10%, thereby avoiding vehicle control jitter and increased closed-loop iteration delay.

[0042] Step 5: Execute hierarchical reinforcement learning decision reasoning and output control commands. Based on the decision-making authority allocation rules corresponding to level 3, and with the target reward function determined in step 4 as the optimization objective, hierarchical reinforcement learning decision reasoning is executed to output the autonomous driving control instructions of the vehicle-side intelligent agent. Through the collaborative allocation of computing power between edge computing nodes and the vehicle-side intelligent agent, computing latency and transmission latency are reduced.

[0043] The specific execution process is as follows: The upper-layer policy network of the edge computing node, based on global perception data and cloud global constraints, takes the edge layer reward function as the optimization objective and outputs the passage sub-objective of the vehicle-side intelligent agent. Relying on the local computing power of the edge computing node, it realizes low-latency processing of sub-objective planning and avoids long link latency caused by cloud transmission. The lower-level policy network of the vehicle-side intelligent agent, based on real-time perception data from the vehicle and sub-objectives sent from the edge, takes the vehicle-side reward function as the optimization objective and outputs autonomous driving control commands for longitudinal acceleration and deceleration and lateral steering. It relies on the local computing power of the vehicle to achieve rapid response in action execution and reduce computation latency. The vehicle-side intelligent agent's safety verification module performs hard safety constraint verification on the output control commands, filters out commands that violate traffic rules or pose a collision risk, ensures that the final output control commands meet safety requirements, and avoids excessive closed-loop delay caused by multiple rounds of safety verification.

[0044] In this embodiment, the hierarchical reinforcement learning decision model is constructed by jointly training an architecture switching policy network and a multi-objective reward optimization network. The specific training process is as follows: The two networks share the global multi-objective weight vector and the input layer of multi-dimensional state parameters; Construct the joint loss function: ,in For multi-objective reward achievement penalty items during architecture switching, Switch stability constraints for reward-optimized architecture; A multi-agent reinforcement learning framework is adopted, and joint training is completed based on SUMO / CARLA joint simulation data of 100+ vehicle-side agents in the same road network and real vehicle test data. The parameters of the two networks are optimized simultaneously to ensure the synergy between architecture switching and reward optimization.

[0045] Step 6: Perform closed-loop optimization The vehicle-side intelligent agent collects execution feedback data of autonomous driving control commands, including the vehicle's actual driving trajectory, travel time, energy consumption data, collision risk status, communication quality changes, and measured end-to-end control latency. Based on the execution feedback data, it performs dynamic iteration of closed-loop optimization and latency suppression effects. Calculate the global optimization goal achievement rate, security risk status, communication quality change value, and end-to-end control delay measured value for the current decision cycle; The global optimization target achievement and control delay measured value are fed back to the cloud management platform. Based on the overall traffic situation and delay suppression effect in the region, the cloud dynamically adjusts the global multi-objective weight vector and low-latency decision priority for the next decision cycle. The system synchronizes the security risk status and communication quality changes to the vehicle-side intelligent agent and edge computing nodes, updates the multi-dimensional status parameters for the next decision cycle, and enters the next round of decision-making process, thereby continuously optimizing latency suppression capabilities.

[0046] Meanwhile, in this embodiment, different target decision-making levels correspond to different vehicle-edge-cloud information interaction rules. Transmission latency is suppressed through adaptive adjustment of data transmission volume and interaction frequency, specifically as follows: When the target decision level is level 1, only the 1Hz heartbeat information interaction between the vehicle and the edge is retained, unnecessary V2X data transmission is turned off, bandwidth usage is reduced, and transmission latency is suppressed to the maximum extent. When the target decision level is level 2, the vehicle and the edge only interact with traffic constraint rules and global perception data at 10Hz, reducing the latency overhead caused by invalid data transmission. When the target decision level is level 3 or level 4, the perception data, sub-target instructions and global policy data required for the full interaction decision between the vehicle, edge and cloud at 20Hz are ensured. At the same time, the local data processing of the edge node avoids the long link transmission delay in the cloud.

[0047] The technical solution of this embodiment fully realizes the end-to-end latency suppression of vehicle-road cooperative control with edge computing as the core. In real vehicle tests on urban open roads, compared with the existing fixed architecture solution, the end-to-end average control latency is reduced by more than 45%, the latency exceedance rate in communication degradation scenarios is reduced by 90%, the decision convergence speed is improved by 60%, the intersection traffic efficiency is improved by 30%, and zero collision accidents are achieved, fully achieving the expected latency suppression and safety management effects.

[0048] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0049] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0050] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for suppressing delay in vehicle-road cooperative control based on edge computing, applied to a vehicle-road cooperative system comprising a vehicle-side intelligent agent, an edge computing node, and a cloud-based management and control platform, characterized in that... Includes the following steps: S1. Obtain the global multi-objective weight vector issued by the cloud management platform, as well as the multi-dimensional state parameters collected in real time by the vehicle-side intelligent agent and edge computing nodes; the global multi-objective weight vector contains the dynamic weights of three optimization objectives: regional traffic efficiency, driving safety, and energy consumption economy, and satisfies the normalization constraint that the weight sum is 1; the multi-dimensional state parameters include the communication quality factor, scenario complexity factor, vehicle computing power factor, and safety risk factor that directly affect the delay of vehicle-road cooperative control, and all parameters are normalized values ​​in the range of 0 to 1; S2. Using the global multi-objective weight vector as an endogenous driving variable, construct a hierarchical decision fitness function, and calculate the decision fitness value that matches each candidate decision level at the current time through the hierarchical decision fitness function; the hierarchical decision fitness function couples the global multi-objective weight vector with the hierarchical basic weight matrix of multi-dimensional state parameters to match the optimal decision level that can minimize control delay; S3. Based on the decision fit value at the current moment, match the preset fit threshold range to determine the target decision level that is effective at the current moment; The candidate decision levels include four preset decision levels. Each decision level corresponds to a unique adaptation threshold range, vehicle-edge-cloud decision permission allocation rules, and information interaction rules. The dynamic adaptation of permissions and interaction rules suppresses end-to-end control latency. S4. Based on the target decision level, call the target reward function that matches the target decision level from the pre-built hierarchical multi-objective game reward function suite; the hierarchical multi-objective game reward function suite contains reward functions that correspond one-to-one with the four candidate decision levels. Each reward function is set with multi-objective weights, upper and lower layer reward alignment coefficients and safety constraint penalty terms that match the target decision level, which are used to accelerate decision convergence and suppress decision reasoning delay; S5. Based on the decision authority allocation rules corresponding to the target decision level, and with the target reward function as the optimization objective, perform hierarchical reinforcement learning decision reasoning, output the autonomous driving control instructions of the vehicle-side intelligent agent, and reduce computational latency and transmission latency through the collaborative allocation of computing power between edge computing nodes and the vehicle-side intelligent agent. S6. Collect execution feedback data of autonomous driving control commands, update the global multi-objective weight vector and multi-dimensional state parameters based on the execution feedback data, and complete the dynamic iterative optimization of decision-making closed-loop optimization and delay suppression effect.

2. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 1, characterized in that: The global multi-objective weight vector is dynamically generated by the cloud-based management platform based on the real-time regional traffic situation, traffic control rules, and emergency event status, or configured by a preset fixed management strategy. When an emergency event or high-risk scenario is detected, the weight ratio corresponding to the driving safety objective is automatically increased to above the preset safety threshold, and the priority configuration of the low-latency decision mode is triggered.

3. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 1, characterized in that: The formula for calculating the hierarchical decision fit function is as follows: ； In the formula for Time of the first The decision fit value corresponding to the candidate decision level. for The global multi-objective weight vector at time step [time]. The sequence number of the candidate decision level and , For the first The hierarchical base weight matrix corresponding to the candidate decision level. For communication quality factor, As the scene complexity factor, For vehicle computing power factor, As a safety risk factor.

4. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 3, characterized in that: The fit threshold ranges and decision authority allocation rules for the four candidate decision levels, ranked from lowest to highest fit value, are as follows: Level 1: The adaptability value is in the first threshold range, corresponding to the pure vehicle-side decision-making mode. The vehicle-side intelligent agent takes over the decision-making authority of the entire link. The edge computing nodes and the cloud management platform only perform perception data backup and security monitoring. Transmission latency is suppressed by closing cross-end data interaction. Level 2: The adaptation value is in the second threshold range, corresponding to the vehicle-dominated - edge-assisted mode. The edge computing node only issues global traffic constraint rules, and the vehicle-side intelligent agent autonomously completes sub-objective planning and action execution decisions, thereby reducing control latency by reducing the frequency of edge-vehicle interaction. Level 3: The adaptability value is in the third threshold range, corresponding to the standard collaborative mode. The edge computing node is responsible for the planning of the passage sub-goal, the vehicle-side intelligent agent is responsible for the action execution decision, and the cloud management platform issues global optimization constraints. The computing power of the edge and the vehicle is split to reduce the computing latency of a single node. Level 4: The adaptability value is in the fourth threshold range, corresponding to the edge full control mode. The edge computing node directly issues trajectory-level decision commands, the vehicle-side intelligent agent only executes commands and performs safety redundancy checks, and the cloud management platform issues global access policies. The overall control latency of multi-vehicle collaboration is reduced by optimizing the centralized computing power of the edge nodes.

5. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 4, characterized in that: In the hierarchical multi-objective game reward function suite, the reward function corresponding to each level includes a basic multi-objective reward term, a hierarchical matching reward alignment term, and a safety constraint penalty term; wherein: The reward function corresponding to level 1 retains only the basic multi-objective reward items and safety constraint penalty items on the vehicle side, and disables the upper and lower level reward alignment items. The reward function corresponding to level 2 sets the one-way reward alignment item between the vehicle end and the edge, and disables the cloud reward alignment item; The reward function corresponding to level 3 sets a two-way reward alignment term between the vehicle end and the edge, with the cloud reward serving as a soft constraint alignment term; The reward function corresponding to level 4 sets the end-to-end reward alignment item from vehicle to edge to cloud, with the cloud global reward serving as a hard constraint alignment item.

6. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 5, characterized in that: When a switch in the target decision level is detected, a reward smooth transition mechanism is activated. Within two preset decision cycles, the weight coefficients and alignment coefficients of the reward function are smoothly adjusted through linear gradient interpolation. At the same time, a strategy consistency check is performed to ensure that the rate of change of the control quantity of the autonomous driving control command output before and after the switch does not exceed a preset threshold, thereby avoiding control jitter and increased closed-loop iteration delay caused by the switch.

7. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 1, characterized in that: The hierarchical reinforcement learning decision model is constructed by jointly training an architecture switching policy network and a multi-objective reward optimization network, specifically including: The architecture switching strategy network and the multi-objective reward optimization network share the global multi-objective weight vector and multi-dimensional state parameters as input layers; Construct a joint loss function, which includes a multi-objective reward achievement penalty term for architecture switching, an architecture switching stability constraint term for reward optimization, and a control latency exceeding penalty term; A multi-agent reinforcement learning framework is adopted to complete joint training based on the interaction data of multiple vehicle-end agents in the same road network, and to simultaneously optimize the parameters of the two networks, thereby improving the decision response speed in dynamic scenarios and suppressing control delay fluctuations in sudden scenarios.

8. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 4, characterized in that: Hard safety constraint rules are embedded throughout the entire process of calculating the hierarchical decision adaptation function, matching the target decision level, and calling the target reward function. When the safety risk factor exceeds the preset risk threshold, the target decision level is forcibly switched to level 1 or level 2. At the same time, the weight of the driving safety target in the target reward function is forcibly increased to filter all decision outputs that violate safety constraints and avoid excessive control delay caused by safety verification iteration.

9. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 1, characterized in that: The process of updating the global multi-objective weight vector and multi-dimensional state parameters based on execution feedback data specifically includes: Based on the execution feedback data, calculate the global optimization goal achievement rate, security risk status, communication quality change value, and end-to-end control delay measured value for the current decision cycle; The global optimization target achievement and the measured control delay are fed back to the cloud management platform to dynamically adjust the global multi-objective weight vector and low-latency decision priority for the next cycle. The security risk status and communication quality change values ​​are synchronized to the vehicle-side intelligent agent and edge computing nodes to update the multi-dimensional state parameters for the next cycle, thereby achieving dynamic optimization of latency suppression capabilities.

10. The delay suppression method for vehicle-road cooperative control based on edge computing according to claim 4, characterized in that: Different target decision-making levels correspond to different vehicle-edge-cloud information interaction rules. Transmission latency is suppressed through adaptive adjustment of data transmission volume and interaction frequency, specifically: When the target decision level is level 1, only the heartbeat information interaction between the vehicle and the edge is retained, and unnecessary autonomous driving and vehicle-road cooperative data transmission is turned off to minimize the control latency caused by the transmission link. When the target decision level is level 2, the vehicle and the edge only interact with traffic constraint rules and global perception data, reducing the latency overhead caused by invalid data transmission. When the target decision level is level 3 or level 4, the perception data, sub-target instructions and global strategy data required for full interaction decision-making between the vehicle, edge and cloud are ensured. At the same time, the local data processing of edge nodes avoids long-link transmission delay in the cloud.

Images