Intelligent automobile cloud parallel acceleration test method based on timing agent scheduling

Abstract

The application belongs to the technical field of intelligent automobile testing, and particularly relates to an intelligent automobile cloud parallel acceleration testing method based on time sequence agent scheduling. The method comprises the following steps: step one, constructing an initial cloud computing module; step two, constructing a time sequence enhancement type; step three, establishing an agent-driven scene search mechanism; step four, designing an online convex optimization resource scheduling mechanism; step five, constructing a public network environment transmission architecture; and step six, executing an end-to-end testing closed loop process. The application deeply integrates four core components of a cloud computing module, a time sequence agent module, a scene search module and a resource planning and scheduling module, takes the time sequence agent as a driving core, and realizes a testing closed loop system of "evaluation-decision-execution-feedback".

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle testing technology, specifically a cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling. Background Technology

[0002] High-fidelity simulation testing for intelligent vehicles faces a fundamental contradiction: high-fidelity simulations must be executed completely to obtain authoritative verification results, but critical failure scenarios are extremely sparse in a high-dimensional parameter space, resulting in over 90% of computing resources being wasted on simulating non-critical scenarios. Traditional cloud platforms treat simulation tasks as uninterrupted "black boxes," unable to dynamically assess scenario value during execution, thus missing opportunities to terminate low-value tasks early and dynamically reallocate computing resources. Existing solutions mainly adopt a batch testing model, i.e., a discrete "submit-wait-collect" process: testers submit a large number of scenarios at once, the cloud platform statically allocates them to computing nodes, and collects the results after all simulations are completed. This model has three fatal flaws: first, cloud devices idle for a long time while waiting for simulation results, with equipment utilization less than 40%; second, it cannot utilize intermediate states during the simulation process for dynamic decision-making, missing early termination opportunities; and finally, the discovery of critical scenarios is inefficient, requiring traversing massive numbers of scenarios to capture only a few dangerous cases. Although some studies have attempted to introduce surrogate models for pre-screening, they only use static scene parameters as input and cannot perceive the dynamic evolution patterns in the simulation process. The prediction accuracy is highly dependent on large-scale labeled data and is difficult to support reliable decision-making in the early stages of testing.

[0003] Cloud-based testing is typically deployed in heterogeneous public network environments, facing inherent challenges such as firewall blocking, unreachable IP addresses, drastic bandwidth fluctuations, and high packet loss rates. Traditional TCP / UDP transport protocols are prone to control command loss or delays in public network environments, especially at critical moments when the proxy model triggers interrupt commands. If control packets are lost, simulation tasks will continuously consume computing resources, severely impacting testing efficiency. Furthermore, different simulation scenarios exhibit significant differences in computing resource consumption: complex traffic scenarios require substantial GPU memory and computing units, while simple scenarios primarily consume storage I / O. Static task allocation strategies can easily lead to overload and lag of some high-performance nodes, while other nodes idle and wait. Existing technologies lack fine-grained resource awareness and cannot dynamically migrate tasks based on real-time device load. More seriously, when hardware failures or network interruptions occur, ongoing simulation tasks are directly lost, wasting computing power and compromising the reproducibility of test results. Current solutions have not yet established a complete fault tolerance mechanism, failing to guarantee highly reliable synchronization of control commands and simulation data in heterogeneous public network environments. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling. By deeply integrating four core components—cloud computing module, time-series agent module, scene search module, and resource planning and scheduling module—and using the time-series agent as the driving core, a closed-loop testing system of "evaluation-decision-execution-feedback" is achieved.

[0005] The technical solution of this invention is described below in conjunction with the accompanying drawings:

[0006] This invention provides a cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling, comprising the following steps:

[0007] S1. Construct and initialize the cloud computing module;

[0008] The initialization cloud computing module includes:

[0009] The scenario parameter decoupling module is used to deconstruct complex test scenarios into structured JSON parameter files;

[0010] The automated simulation configuration module is used to achieve fully automated configuration from scene description to simulation environment;

[0011] A millisecond-level interrupt mechanism module is used to integrate runtime interrupt interfaces in the Simulink simulation environment;

[0012] A containerized environment encapsulation module is used to achieve second-level deployment and environment consistency;

[0013] The four-dimensional load monitoring system module is used to build a real-time monitoring stack based on the Prometheus open-source monitoring system.

[0014] S2. Construct a time-enhanced agent model;

[0015] The time-enhanced proxy model includes:

[0016] The temporal feature modeling module is used to capture the dynamic evolution patterns during the simulation process by using sliding window temporal features;

[0017] Incremental learning architecture module, used to reduce training costs;

[0018] The hierarchical decision-making mechanism module is used for hierarchical decision-making.

[0019] S3. Establish an agent-driven scenario search mechanism;

[0020] Search optimization for mirages;

[0021] Agent-guided fitness assessment;

[0022] Dynamic population management;

[0023] Search-simulation collaborative closed loop;

[0024] S4. Design an online convex optimization resource scheduling mechanism;

[0025] Problem modeling;

[0026] Objective function design;

[0027] Constraint settings;

[0028] Execute a deterministic solution process;

[0029] Interrupt-driven resource reclamation;

[0030] S5. Construct a public network environment transmission architecture;

[0031] Virtual intranet construction;

[0032] Redundant routing design;

[0033] Coding mechanism design;

[0034] Multi-level feedback mechanism design;

[0035] S6. Execute the end-to-end test closed-loop process.

[0036] Furthermore, the specific method of S1 is as follows:

[0037] S11. Design a scenario parameter decoupling module;

[0038] Using the ISO 34501 scenario standard as a bridge, complex test scenarios are deconstructed into structured JSON parameter files. The JSON parameter files fully cover the initial state of the master vehicle and the transport vehicle, road geometry features, weather conditions, and behavior trigger sequences.

[0039] S12, Design an automated simulation configuration module;

[0040] Using MATLAB scripts, the system automatically parses JSON parameter files, dynamically modifies scene structure parameters in the PreSim project, and synchronously updates vehicle attributes, tire models, and braking system parameters in the CarSim vehicle dynamics model, achieving fully automated configuration from scene description to simulation environment.

[0041] S13. Design a millisecond-level interrupt mechanism module;

[0042] Integrate a runtime interrupt interface into the Simulink simulation environment. When the current scenario is determined to be non-critical, the simulation process can be forcibly terminated within milliseconds through the interface, immediately releasing the occupied computing resources.

[0043] S14. Design a containerized environment encapsulation module;

[0044] The complete simulation environment is encapsulated using Docker container technology, including the operating system, MATLAB Runtime, PreSim, CarSim, and Simulink, ensuring second-level deployment and environment consistency on 30 homogeneous RTX 4090 nodes;

[0045] S15. Design a four-dimensional load monitoring system module;

[0046] A real-time monitoring stack was built based on the Prometheus open-source monitoring system, defining device load evaluation metrics covering four dimensions: computing resources, hardware status, network performance, and storage I / O, providing fine-grained data support for resource scheduling.

[0047] Furthermore, the specific method of S2 is as follows:

[0048] S21, Design timing feature modeling module;

[0049] A sliding window temporal feature is introduced to capture the dynamic evolution during the simulation process. For each simulation task being executed, the surrogate model receives four types of input features in real time: the first type is the initial scene parameters, including static information; the second type is the sliding window temporal feature, specifically the mean and standard deviation of the relative position, relative speed, and relative acceleration of the master vehicle and the commuter vehicle within the current 2-second time window; the third type is the risk quantification index, i.e., the minimum TTC value within the 2-second window; and the fourth type is the behavior label for the next stage, which uses one-hot encoding to represent the actions of the commuter vehicle based on the ISO 34502 standard.

[0050] S22, Design the incremental learning architecture module;

[0051] XGBoost gradient boosting tree is used as the base model. An incremental learning strategy is designed: when the amount of new simulation data exceeds 20% of the current training set size, 5% of new decision trees are dynamically added to the existing model. An adaptive depth control mechanism is adopted: for shallow nodes with a depth of no more than 3, the feature splitting threshold is relaxed to quickly capture the main pattern in the data; for deep nodes, the splitting threshold is tightened to suppress overfitting.

[0052] S23, Design a hierarchical decision-making mechanism module;

[0053] In the initial stage of simulation, if the confidence level calculated by the model based on the historical mean square error is less than 80%, all scenarios are forced to enter the complete simulation sequence. During the simulation execution, the proxy model evaluates the current simulation segment every 1 second. If the prediction uncertainty is less than 20% and the predicted final minimum TTC is higher than the preset safety threshold, an interrupt command is immediately triggered, and the current scenario is marked as a "non-critical scenario". Otherwise, the simulation continues to be executed, and the complete results are used for subsequent model updates.

[0054] Furthermore, the specific method of S3 is as follows:

[0055] S31, Mirage search optimization;

[0056] A mirage search optimization algorithm is introduced as an efficient search engine. By simulating the optical and physical mechanisms of "superimposed mirage" (global exploration) and "subimposed mirage" (local development) in nature, an efficient balance between exploration and development is achieved in a high-dimensional parameter space. The "superimposed mirage" strategy generates diverse candidates through random perturbation in the global scope, avoiding getting trapped in local optima. The "subimposed mirage" strategy performs a fine local search in the discovered high-risk areas to uncover critical boundary scenes.

[0057] In the mirage stage, for individuals Generate global perturbation candidates as follows:

[0058] ;

[0059] In the formula, The current optimal individual; The attraction coefficient towards the optimal solution; To explore the intensity; It is a random vector; For individuals Global perturbation candidate solutions generated after performing the "superimposed mirage" operation;

[0060] During the mirage phase, local sampling is performed, as shown below:

[0061] ;

[0062] In the formula, , For nearby high-risk individuals; This represents the local disturbance coefficient; These are local sampling candidate solutions generated through the "subsequent mirage" mechanism;

[0063] S32. Agent-guided fitness assessment;

[0064] Use the predicted output of the time-series enhanced surrogate model as the real-time fitness signal of MSO to achieve differential processing of three scenarios: for scenarios where the simulation has been completed, use the real measured minimum TTC as the accurate fitness; for scenarios interrupted by the surrogate model, use the TTC predicted by the surrogate as the approximate fitness; for newly generated candidate scenarios, conduct a quick pre-evaluation by the surrogate model before submitting the simulation, and only send high-risk individuals with a predicted TTC lower than the safety threshold into the simulation queue;

[0065] S33. Dynamic population management;

[0066] Adopt a load-aware dynamic population mechanism; the population size is dynamically adjusted according to the number of devices and the number of individuals whose evaluation has been completed. When the number of remaining tasks in the simulation queue is lower than the preset threshold, immediately trigger MSO to generate a new generation of candidate scenarios, ensuring that there are always candidate tasks with a predicted minimum TTC lower than the safety threshold and not less than the preset queue length threshold for the computing device to execute, realizing seamless connection of "search - execution";

[0067] Let the current number of available devices be , the length of the simulation queue be Q, and the preset threshold be Qmin. If Q < Qmin, trigger the generation of a new generation as follows:

[0068] ;

[0069] In the formula, is the average simulation duration; is the average time-consuming for a single simulation;<000​​​​​​​​​​​​​​​​​The resource allocation problem is modeled as an online convex optimization problem; let the current candidate task set be J, and the device set be D. For each task j∈J, the time-enhanced surrogate model provides a prediction of the simulation duration. Harmony Mission Key scores For each device ∈D, the monitoring module provides a normalized overall load;

[0075] S42. Objective function design;

[0076] Define decision variables ∈[0,1] represents the task Assigned to device The objective function is designed to minimize the weighted completion time, which is equivalent to prioritizing the scheduling of high-risk, short-duration tasks while considering device load balancing; as shown in the following formula:

[0077] ;

[0078] In the formula, For smoothing terms; This is the set of tasks currently awaiting scheduling; For the task Assigned to device ;

[0079] S43. Constraint Setting: Constraints include three types:

[0080] The task integrity constraint requires that each task must be completely assigned to a single device, as shown below:

[0081] ;

[0082] The load limit constraint means that the device load cannot exceed a preset threshold, as shown below:

[0083] ;

[0084] In the formula, This refers to the simulation duration. This is the load threshold; This represents the current base load of device i;

[0085] The non-negativity constraint means that the allocation ratio is non-negative, as shown below:

[0086] ;

[0087] S44. Execute the deterministic solution process;

[0088] For each scheduling moment, the system executes the following deterministic process: First, based on the current candidate task set, device set, agent prediction, and device load, an optimization sub-problem is constructed. This problem is an integer linear programming problem, employing a relaxation + rounding strategy:

[0089] Relaxation: Transform the problem into a linear programming problem; solve it using the interior-point method within milliseconds to obtain the optimal allocation scheme; then perform deterministic rounding on each task and allocate it to the device that maximizes the objective function gain, as shown below:

[0090] ;

[0091] In the formula, For load balancing weights; The optimal device index that minimizes the objective function;

[0092] Finally, immediately update the device load status to prepare for the next round of scheduling, as shown below:

[0093] ;

[0094] S45, interrupt-driven resource reclamation;

[0095] When the time-enhanced proxy model triggers a simulation interrupt instruction, the scheduler immediately performs resource reclamation operations: marking the interrupted task as "non-critical" and no longer retrying; releasing the occupied device load; and triggering a new round of online optimization to allocate idle resources to candidate tasks in real time.

[0096] Furthermore, the specific method of S5 is as follows:

[0097] S51, virtual intranet construction;

[0098] A unified virtual intranet is constructed using software-defined networking technology based on ZeroTier; all computing nodes participating in the test automatically join the same virtual network upon startup, and the system assigns fixed private IP addresses, making each node logically appear to be in the same local area network.

[0099] S52, redundant routing design;

[0100] The system designs a hierarchical tree-like redundant routing architecture. The entire communication network is rooted at the scheduler, with multiple primary relay nodes. Each primary node is then connected to several simulation computing nodes as secondary terminals. The system continuously monitors the health status of each communication path and detects link latency and connectivity through periodic heartbeat packets. Once a path becomes abnormal, the system will automatically switch the affected nodes to backup paths.

[0101] S53. Design the coding mechanism;

[0102] A four-segment data frame format is defined; each data packet consists of four parts in sequence: first, a 4-byte target node identifier, used to precisely specify the receiver; second, a 1-byte task type field, indicating whether it is starting a simulation, requesting an interruption, or reporting a status; third, a variable-length payload content, using a type-length-value structure, carrying the task ID and parameter configuration; and finally, a 4-byte CRC-32 checksum, covering all the preceding fields; the receiving end first verifies the checksum before processing, and if it does not match, it is discarded directly and a retransmission is requested, and suspicious instructions are never executed;

[0103] S54. Design a multi-level feedback mechanism;

[0104] The system is designed with an acknowledgment and retransmission mechanism. When the scheduler issues an interrupt command, the receiving node is required to return an acknowledgment (ACK) within 5 milliseconds. If the sender does not receive the ACK within 10 milliseconds, the first retransmission is triggered immediately, and a maximum of two retransmissions are allowed. All retransmissions carry the same task identifier, and the receiver has deduplication capabilities to avoid repeated execution. At the same time, the system opens a high-priority transmission channel for critical instructions to ensure that they can still be processed with priority when the network is congested.

[0105] Furthermore, the specific method of S6 is as follows:

[0106] S61, Initialization phase;

[0107] Load system configuration parameters, start all compute nodes, initialize the time-series enhanced agent model, initialize the MSO population, and clear the task queue;

[0108] S62, Scene Generation Stage;

[0109] MSO generates 50 candidate scenarios based on the current population state, quickly pre-evaluates them using a time-enhanced agent model, and selects scenarios whose predicted minimum TTC is lower than a preset safety threshold to add to the task queue.

[0110] S63, Task Allocation Phase;

[0111] The resource scheduling module dynamically allocates tasks in the queue to computing nodes that minimize the weighted completion time and meet the load limit constraint based on the real-time load of the equipment and the criticality of the tasks.

[0112] S64, Simulation Execution Phase;

[0113] The computing nodes perform high-fidelity simulations, and the timing-enhanced agent module extracts timing features every second for dynamic evaluation.

[0114] S65, Interruption of decision-making stage;

[0115] If the time-enhanced proxy model is ready and meets the interruption conditions, a millisecond-level interrupt is triggered to release resources; otherwise, execution continues until completion.

[0116] S66, Feedback and Update Phase;

[0117] The simulation results are fed back to the MSO and time-series enhanced surrogate model to update the population fitness and model parameters;

[0118] S67, Iterative Looping;

[0119] Repeat steps S62-S66 until the preset total number of scenarios or the maximum test time is reached.

[0120] The beneficial effects of this invention are as follows:

[0121] 1. This invention significantly improves the efficiency and coverage of cloud-based simulation testing for intelligent vehicles; by introducing a time-series proxy model to dynamically evaluate the simulation process, non-critical scenarios can be terminated in advance within milliseconds, avoiding invalid calculations, significantly improving the utilization rate of cluster devices, and greatly reducing the computing power cost brought by high-fidelity simulation.

[0122] 2. This invention constructs a closed-loop collaborative mechanism of "search-simulation-feedback," deeply integrating the mirage optimization algorithm with online agent prediction to achieve proactive guidance and targeted discovery of high-risk scenarios. Compared with traditional random or grid search, it effectively solves the verification problem of extremely sparse dangerous cases in high-dimensional parameter space;

[0123] 3. This invention designs a highly reliable transmission architecture for heterogeneous public network environments, combining virtual intranets, redundant routing, four-segment encoding, and multi-level ACK / NACK mechanisms to ensure that interrupt commands are reliably delivered within 10ms. Even under harsh network conditions such as high packet loss and NAT traversal failures, it can still guarantee the continuity, reproducibility, and real-time resource scheduling of test tasks, providing industrial-grade robustness support for large-scale cloud-based parallel testing. Attached Figure Description

[0124] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0125] Figure 1 This is a schematic diagram of the process of the present invention;

[0126] Figure 2 Flowchart for applying the proxy model;

[0127] Figure 3 This is a schematic diagram showing the load distribution of the equipment.

[0128] Figure 4 A graphical illustration for training the surrogate model;

[0129] Figure 5 This is a schematic diagram of the collision time data during the test process;

[0130] Figure 6 This is a schematic diagram of the discovery curve for key test scenarios. Detailed Implementation

[0131] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0132] See Figures 1-6 This invention provides a cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling, comprising the following steps:

[0133] S1. Construct the initial cloud computing module, the specific method is as follows:

[0134] The initialization cloud computing module includes a scene parameter decoupling module, an automated simulation configuration module, a millisecond-level interrupt mechanism module, a containerized environment encapsulation module, and a four-dimensional load monitoring system module.

[0135] S11. Design a scenario parameter decoupling module;

[0136] Using the ISO 34501 scenario standard as a bridge, complex test scenarios are deconstructed into structured JSON parameter files. These files fully cover the initial states of the host vehicle and other vehicles (including position, speed, acceleration, and heading angle), road geometry features (including curvature, slope, and lane width), weather conditions (including rainfall, visibility, and road surface adhesion coefficient), and behavior trigger sequences (including timestamps and parameters of events such as cutting in, cutting out, and emergency braking).

[0137] S12, Design an automated simulation configuration module;

[0138] A dedicated MATLAB script is used to automatically parse the JSON parameter file, dynamically modify the scene structure parameters in the PreSim project, and synchronously update key parameters such as vehicle attributes, tire models, and braking systems in the CarSim vehicle dynamics model, achieving fully automatic configuration from scene description to simulation environment.

[0139] S13. Design a millisecond-level interrupt mechanism module;

[0140] Integrating a runtime interrupt interface into the Simulink simulation environment allows the simulation process to be forcibly terminated within milliseconds when the current scenario is determined to be non-critical, immediately releasing the occupied computing resources and avoiding unnecessary computation.

[0141] S14. Design a containerized environment encapsulation module;

[0142] The complete simulation environment is encapsulated using Docker container technology, including all dependent components such as the operating system, MATLAB Runtime, PreSim, CarSim, and Simulink, ensuring second-level deployment and environment consistency on 30 homogeneous RTX 4090 nodes, and avoiding reproduction deviations caused by heterogeneous environments;

[0143] S15. Design a four-dimensional load monitoring system module;

[0144] A real-time monitoring stack is built based on the Prometheus open-source monitoring system, defining device load evaluation metrics covering four dimensions: computing resources (GPU utilization, video memory usage, CPU utilization), hardware status (device temperature, power consumption, fan speed), network performance (bandwidth utilization, packet loss rate, latency jitter), and storage I / O (disk read / write speed, IOPS, queue depth), providing fine-grained data support for resource scheduling.

[0145] S2. Construct a time-series enhanced proxy model, the specific method is as follows:

[0146] The temporal enhanced agent model includes a temporal feature modeling module, an incremental learning architecture module, and a hierarchical decision-making mechanism module.

[0147] S21, Design timing feature modeling module;

[0148] Unlike traditional proxy models that only use initial scene parameters, this invention introduces sliding window temporal features to capture the dynamic evolution during the simulation process. For each simulation task being executed, the proxy model receives four types of input features in real time: the first type is initial scene parameters, including static information such as road curvature, vehicle speed, spacing, and weather conditions; the second type is sliding window temporal features, specifically the mean and standard deviation of the relative position, relative speed, and relative acceleration of the master vehicle and the traffic vehicle within the current 2-second time window; the third type is risk quantification indicators, namely the minimum TTC (Time To Collision) value within the aforementioned 2-second window; and the fourth type is the behavior label for the next stage, based on the ISO 34502 standard, which uses one-hot encoding to represent the actions of the traffic vehicle.

[0149] S22, Design the incremental learning architecture module;

[0150] XGBoost gradient boosting trees are used as the base model, which has lower training costs compared to deep learning models in small-sample, high-dimensional numerical regression tasks and supports uncertainty quantification. An incremental learning strategy is designed: when the amount of newly added simulation data exceeds 20% of the current training set size, 5% of new decision trees are dynamically added to the existing model. An adaptive depth control mechanism is adopted: for shallow nodes with a depth of no more than 3, the feature splitting threshold is relaxed to quickly capture the dominant patterns in the data; for deep nodes, the splitting threshold is tightened to suppress overfitting.

[0151] S23, Design a hierarchical decision-making mechanism module;

[0152] The availability of the surrogate model's prediction results is closely related to its own confidence level; therefore, a strict hierarchical decision-making process is designed. At the initial stage of simulation, if the confidence level calculated by the model based on historical mean squared error (MSE) is below 80%, all scenarios are forced into a complete simulation sequence to avoid early misjudgments due to model immaturity. During simulation execution, the surrogate model evaluates the current simulation segment every second. If the prediction uncertainty (calculated based on a combination of inter-tree standard deviation and historical error) is below 20% and the predicted final minimum TTC is above a preset safety threshold (usually 2.0 seconds), an interrupt command is immediately triggered, and the current scenario is marked as a "non-critical scenario." Otherwise, the simulation continues, and the complete results are used for subsequent model updates.

[0153] S3. Establish an agent-driven scenario search mechanism, the specific method is as follows:

[0154] S31, Mirage Search Optimization (MSO);

[0155] A mirage search optimization algorithm is introduced as a high-efficiency search engine. This algorithm achieves an efficient balance between exploration and exploitation in a high-dimensional parameter space by simulating the optical and physical mechanisms of "superimposed mirage" (global exploration) and "subimposed mirage" (local exploitation) in nature. The "superimposed mirage" strategy generates diverse candidates through random perturbation in the global scope, avoiding getting trapped in local optima; the "subimposed mirage" strategy performs a fine-grained local search in the discovered high-risk areas to uncover critical boundary scenes.

[0156] In the mirage (global exploration) phase, for individuals Generate global perturbation candidates:

[0157] ;

[0158] In the formula, The current optimal individual; The attraction coefficient towards the optimal solution; is the exploration intensity; is a random vector; is for an individual is the global perturbation candidate solution generated after performing the "looming" operation on the individual

[0159] In the inferior mirage (local development) stage, local sampling is performed:

[0160] ;

[0161] In the formula, , is the neighboring high-risk individual; is the local perturbation coefficient; is the local sampling candidate solution generated through the "inferior mirage" mechanism;

[0162] S32. Agent-guided fitness evaluation;

[0163] Traditional evolutionary algorithms rely on complete simulation results as fitness, resulting in serious evaluation delays; the present invention uses the predicted output of a time-series enhanced surrogate model as the real-time fitness signal of the MSO, realizing differential processing of three scenarios: for scenarios where simulation has been completed, the true measured minimum TTC is used as the exact fitness; for scenarios interrupted by the surrogate model, the TTC predicted by the surrogate is used as an approximate fitness, which, although having a certain error, is sufficient to guide the search direction; for newly generated candidate scenarios, a quick pre-evaluation is first performed by the surrogate model before submitting the simulation, and only high-risk individuals with a predicted TTC lower than the safety threshold are sent into the simulation queue, greatly reducing the ineffective simulation overhead;

[0164] S33. Dynamic population management;

[0165] To match the streaming execution characteristics of the cloud platform, the design of a traditional fixed population size is abandoned, and a load-aware dynamic population mechanism is adopted. The population size is dynamically adjusted according to the number of devices and the number of individuals whose evaluation has been completed. When the number of remaining tasks in the simulation queue is lower than the preset threshold, the MSO is immediately triggered to generate a new generation of candidate scenarios, ensuring that the computing devices always have sufficient high-value tasks to execute, and realizing the seamless connection of "search - execution";

[0166] Let the current number of available devices be M, the length of the simulation queue be Q, and the preset threshold be Qmin. If Q < Qmin, then a new generation is triggered:

[0167] ;

[0168] In the formula, is the average simulation duration; is the average time-consuming for a single simulation; is the maximum population size; The number of individuals that will be generated for the next generation;

[0169] S34, Search-Simulation Cooperative Closed Loop;

[0170] MSO works in deep collaboration with cloud simulation and time-series agents to form a complete online closed loop: In the exploration phase, MSO uses the "upper mirage" strategy to sample a wide area in high-dimensional space, generating diverse candidate scenarios; in the verification phase, candidate scenarios are submitted to the cloud computing module for parallel simulation, and the time-series agent dynamically evaluates the scenario value during execution; in the feedback phase, the agent outputs the actual or predicted TTC back to MSO to update the fitness of individuals in the population; in the development phase, MSO uses the "lower mirage" strategy to perform local fine-grained searches in high-risk areas, gradually converging to the dense area of ​​critical failure scenarios.

[0171] S4. Design an online convex optimization resource scheduling mechanism, the specific method of which is as follows:

[0172] S41. Problem Modeling;

[0173] The resource allocation problem is modeled as an Online Convex Optimization (OCO) problem. Its scientific goal is not to pursue a theoretical global optimum, but rather to achieve real-time coordination among three key aspects: prioritizing critical tasks, balancing cluster load, and promptly recovering interrupted resources. Specifically, let the current set of candidate tasks be J, and the set of devices be D. For each task j∈J, the time-enhanced proxy model provides a predicted simulation duration. Harmony Mission Key scores ( (Negatively correlated with predicted TTC, the lower the TTC, the higher the score) for each device. ∈D, the monitoring module provides a normalized overall load;

[0174] S42. Objective function design;

[0175] Define decision variables ∈[0,1] represents the task Assigned to device The objective function is designed to minimize the weighted completion time, which is equivalent to prioritizing scheduling high-risk (high-risk) tasks. ), short duration (small) The task involves considering both equipment load balancing and the following formula:

[0176] ;

[0177] In the formula, For smoothing terms (avoiding division by zero); For the task Assigned to device This formula is equivalent to priority scheduling. ,Low The task;

[0178] S43. Constraint Setting: Constraints include three types:

[0179] Task integrity constraint (each task must be fully assigned to a single device):

[0180] ;

[0181] Load limit constraint (device load cannot exceed a preset threshold):

[0182] ;

[0183] In the formula, This refers to the simulation duration. This is the load threshold; This represents the current base load of device i;

[0184] Non-negativity constraint (non-negative allocation ratio):

[0185] ;

[0186] S44. Execute the deterministic solution process;

[0187] For each scheduling moment, the system executes the following deterministic process: First, based on the current candidate task set, device set, agent prediction and device load, an optimization subproblem is constructed. The problem is an integer linear programming (ILP) problem, but the scale is small, and a relaxation + rounding strategy is adopted.

[0188] Relaxation: Transform the problem into a linear programming (LP) problem; solve the problem using interior-point methods within milliseconds to obtain the optimal allocation scheme; then perform deterministic rounding on each task to allocate it to the device that maximizes the objective function gain.

[0189] ;

[0190] In the formula, For load balancing weights; The optimal device index that minimizes the objective function;

[0191] Finally, immediately update the device load status to prepare for the next round of scheduling:

[0192] ;

[0193] S45, interrupt-driven resource reclamation;

[0194] When the time-enhanced proxy model triggers a simulation interrupt command, the scheduler immediately performs resource reclamation operations: marking the interrupted task as "non-critical" and preventing retrying; releasing the device load it occupies; and triggering a new round of online optimization to allocate idle resources to candidate tasks in real time. This mechanism ensures that interruption decisions are instantly translated into resource optimization, avoiding the scheduling lag problem of "tasks completed but resources not released" in traditional methods.

[0195] S5. Construct a public network environment transmission architecture, the specific methods are as follows:

[0196] S51, virtual intranet construction;

[0197] To address communication interruptions caused by firewall policies, NAT type differences, or dynamic IP address changes in public network environments, this invention employs ZeroTier-based software-defined networking technology to construct a unified virtual intranet. All participating computing nodes (including the scheduling server and simulation execution nodes) automatically join the same virtual network upon startup, and the system assigns them fixed private IP addresses, making each node logically appear to be on the same local area network. This virtualized network layer completely shields the complexity of the underlying public network, enabling point-to-point direct communication regardless of whether nodes are deployed in an enterprise intranet, public cloud, or hybrid environment. Even if the physical network path changes, the virtual IP remains constant, thus ensuring the continuity and stability of upper-layer control commands and simulation data transmission.

[0198] S52, redundant routing design;

[0199] Considering the risk of single-point failures in public network links (such as the downtime of a relay server or network congestion in a certain area), this invention designs a hierarchical tree-like redundant routing architecture. The entire communication network is rooted at the scheduler, with multiple primary relay nodes (typically deployed in different availability zones or different cloud service providers). Each primary node then connects to several simulation computing nodes as secondary terminals. The system continuously monitors the health status of each communication path, detecting link latency and connectivity through periodic heartbeat packets. Once a path becomes abnormal (such as three consecutive heartbeat timeouts), the system automatically switches the affected node to a backup path—for example, if a simulation node originally communicated through primary node A, when A fails, it immediately switches to primary node B for data relay. This multi-path redundancy mechanism ensures that even if some network nodes or links fail, control commands can still be reliably delivered through other paths, preventing tasks from being lost or stuck due to network interruptions.

[0200] S53, reliable coding mechanism;

[0201] To prevent instruction parsing errors caused by packet loss, tampering, or out-of-order delivery during public network transmission, this invention defines a clearly structured and rigorously validated four-segment data frame format. Each data packet consists of four parts in sequence: first, a 4-byte target node identifier to precisely specify the receiver; second, a 1-byte task type field indicating whether the packet is for starting a simulation, requesting an interrupt, or reporting a status; third, a variable-length payload content using a Type-Length-Value (TLV) structure, which can flexibly carry specific information such as task ID and parameter configuration; and finally, a 4-byte CRC-32 checksum, covering all the preceding fields. The receiving end verifies the checksum before processing; if it does not match, the packet is discarded and a retransmission is requested, and any suspicious instructions are never executed. This encoding method ensures both data integrity and parsability while also possessing good scalability, adapting to the needs of different types of control commands.

[0202] S54, Multi-level feedback mechanism;

[0203] To address the extremely high timeliness and reliability requirements of critical control commands (especially interrupt commands), this invention designs a lightweight yet efficient acknowledgment and retransmission mechanism. When the scheduler issues an interrupt command, the receiving node is required to return an acknowledgment (ACK) within 5 milliseconds. If the sender does not receive an ACK within 10 milliseconds, a first retransmission is immediately triggered, with a maximum of two retransmissions allowed. All retransmissions carry the same task identifier, and the receiver has deduplication capabilities to avoid duplicate execution. Simultaneously, the system allocates a high-priority transmission channel for critical commands, ensuring they are processed preferentially even during network congestion. Through this mechanism, even under conditions of high packet loss rates or large latency fluctuations on the public network, interrupt commands can reliably and accurately act on the target simulation process within 10 milliseconds, truly achieving real-time control capabilities of "commands delivered immediately, execution stopped immediately," providing a solid guarantee for efficient resource recovery and rapid iteration of the test loop.

[0204] S6. Execute the end-to-end test closed-loop process, the specific method is as follows:

[0205] S61, Initialization phase;

[0206] Load system configuration parameters, start all computing nodes, initialize the enhanced temporal agent model (initial state has no training data), initialize the MSO population (randomly generate 100 individuals), and clear the task queue;

[0207] S62 scene generation stage;

[0208] MSO generates 50 candidate scenarios based on the current population state, quickly pre-evaluates them using a time-enhanced agent model, and selects scenarios whose predicted minimum TTC is lower than a preset safety threshold to add to the task queue.

[0209] S63, Task Allocation Phase;

[0210] The resource scheduling module dynamically allocates tasks in the queue to computing nodes that minimize the weighted completion time and meet the load limit constraint, based on the real-time load of the equipment and the criticality of the tasks.

[0211] S64, Simulation Execution Phase;

[0212] The computing nodes perform high-fidelity simulations, and the timing-enhanced agent module extracts timing features every second for dynamic evaluation.

[0213] S65, Interruption of decision-making stage;

[0214] If the time-enhanced proxy model is ready and meets the interruption conditions, a millisecond-level interrupt is triggered to release resources; otherwise, execution continues until completion.

[0215] S66, Feedback and Update Phase;

[0216] The simulation results (real TTC or predicted TTC) are fed back to the MSO and the time-series augmented surrogate model to update the population fitness and model parameters;

[0217] S67, Iterative Looping;

[0218] Repeat steps S62-S66 until the preset total number of scenarios (e.g., 1000) or the maximum test time is reached.

[0219] In summary, this invention achieves a closed-loop testing system of "evaluation-decision-execution-feedback" by deeply integrating four core components: cloud computing module, time-series agent module, scene search module, and resource planning and scheduling module, with time-series agent as the driving core.

[0220] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling, characterized in that, Includes the following steps: S1. Construct and initialize the cloud computing module; The initialization cloud computing module includes: The scenario parameter decoupling module is used to deconstruct complex test scenarios into structured JSON parameter files; The automated simulation configuration module is used to achieve fully automated configuration from scene description to simulation environment; A millisecond-level interrupt mechanism module is used to integrate runtime interrupt interfaces in the Simulink simulation environment; A containerized environment encapsulation module is used to achieve second-level deployment and environment consistency; The four-dimensional load monitoring system module is used to build a real-time monitoring stack based on the Prometheus open-source monitoring system. S2. Construct a time-enhanced agent model; The time-enhanced proxy model includes: The temporal feature modeling module is used to capture the dynamic evolution patterns during the simulation process by using sliding window temporal features; Incremental learning architecture module, used to reduce training costs; The hierarchical decision-making mechanism module is used for hierarchical decision-making. S3. Establish an agent-driven scenario search mechanism; Search optimization for mirages; Agent-guided fitness assessment; Dynamic population management; Search-simulation collaborative closed loop; S4. Design an online convex optimization resource scheduling mechanism; Problem modeling; Objective function design; Constraint settings; Execute a deterministic solution process; Interrupt-driven resource reclamation; S5. Construct a public network environment transmission architecture; Virtual intranet construction; Redundant routing design; Coding mechanism design; Multi-level feedback mechanism design; S6. Execute the end-to-end test closed-loop process.

2. The cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling according to claim 1, characterized in that, The specific method of S1 is as follows: S11. Design a scenario parameter decoupling module; Using the ISO 34501 scenario standard as a bridge, complex test scenarios are deconstructed into structured JSON parameter files. The JSON parameter files fully cover the initial state of the master vehicle and the transport vehicle, road geometry features, weather conditions, and behavior trigger sequences. S12, Design an automated simulation configuration module; Using MATLAB scripts, the system automatically parses JSON parameter files, dynamically modifies scene structure parameters in the PreSim project, and synchronously updates vehicle attributes, tire models, and braking system parameters in the CarSim vehicle dynamics model, achieving fully automated configuration from scene description to simulation environment. S13. Design a millisecond-level interrupt mechanism module; Integrate a runtime interrupt interface into the Simulink simulation environment. When the current scenario is determined to be non-critical, the simulation process can be forcibly terminated within milliseconds through the interface, immediately releasing the occupied computing resources. S14. Design a containerized environment encapsulation module; The complete simulation environment is encapsulated using Docker container technology, including the operating system, MATLAB Runtime, PreSim, CarSim, and Simulink, ensuring second-level deployment and environment consistency on 30 homogeneous RTX 4090 nodes; S15. Design a four-dimensional load monitoring system module; A real-time monitoring stack was built based on the Prometheus open-source monitoring system, defining device load evaluation metrics covering four dimensions: computing resources, hardware status, network performance, and storage I / O, providing fine-grained data support for resource scheduling.

3. The cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling according to claim 1, characterized in that, The specific method of S2 is as follows: S21, Design timing feature modeling module; A sliding window temporal feature is introduced to capture the dynamic evolution pattern during the simulation process; for each simulation task being executed, the proxy model receives four types of input features in real time: the first type is the initial scene parameters, including static information; The second category is the sliding window time series features, specifically the mean and standard deviation of the relative position, relative speed, and relative acceleration of the main vehicle and the transport vehicle within the current 2-second time window; The third category is risk quantification indicators, namely the minimum TTC value within a 2-second window; The fourth category is the next stage of behavior labeling, which uses one-hot encoding to represent the actions of the vehicle based on the ISO 34502 standard. S22, Design the incremental learning architecture module; XGBoost gradient boosting tree is used as the base model. An incremental learning strategy is designed: when the amount of new simulation data exceeds 20% of the current training set size, 5% of new decision trees are dynamically added to the existing model. An adaptive depth control mechanism is adopted: for shallow nodes with a depth of no more than 3, the feature splitting threshold is relaxed to quickly capture the main pattern in the data; for deep nodes, the splitting threshold is tightened to suppress overfitting. S23, Design a hierarchical decision-making mechanism module; In the initial stage of simulation, if the confidence level calculated by the model based on the historical mean square error is less than 80%, all scenarios are forced to enter the complete simulation sequence. During the simulation execution, the proxy model evaluates the current simulation segment every second. If the prediction uncertainty is less than 20% and the predicted final minimum TTC is higher than the preset safety threshold, an interrupt command is immediately triggered, and the current scenario is marked as a "non-critical scenario". Otherwise, the simulation continues to be executed, and the complete results are used for subsequent model updates.

4. The cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling according to claim 1, characterized in that, The specific method of S3 is as follows: S31, Mirage search optimization; A mirage search optimization algorithm is introduced as an efficient search engine. By simulating the optical and physical mechanisms of "superimposed mirage" (global exploration) and "subimposed mirage" (local development) in nature, an efficient balance between exploration and development is achieved in a high-dimensional parameter space. The "superimposed mirage" strategy generates diverse candidates through random perturbation in the global scope, avoiding getting trapped in local optima. The "subimposed mirage" strategy performs a fine local search in the discovered high-risk areas to uncover critical boundary scenarios. In the mirage stage, for individuals Generate global perturbation candidates as follows: ； In the formula, The current optimal individual; The attraction coefficient towards the optimal solution; To explore the intensity; It is a random vector; For individuals Global perturbation candidate solutions generated after performing the "superimposed mirage" operation; During the mirage phase, local sampling is performed, as shown below: ； In the formula, , For nearby high-risk individuals; This represents the local disturbance coefficient; These are local sampling candidate solutions generated through the "subsequent mirage" mechanism; S32. Agent-guided fitness assessment; The predicted output of the time-enhanced surrogate model is used as the real-time fitness signal of the MSO to achieve differentiated processing for three scenarios: for scenarios that have been simulated, the minimum time-to-competition (TTC) measured in reality is used as the exact fitness; for scenarios interrupted by the surrogate model, the TTC predicted by the surrogate model is used as the approximate fitness; for newly generated candidate scenarios, the surrogate model performs a rapid pre-evaluation before submitting to the simulation, and only high-risk individuals with predicted TTC below the safety threshold are sent to the simulation queue. S33, Dynamic Population Management; A load-aware dynamic population mechanism is adopted; the population size is dynamically adjusted according to the number of devices and the number of individuals that have completed evaluation. When the number of remaining tasks in the simulation queue is lower than the preset threshold, the MSO is immediately triggered to generate a new generation of candidate scenarios, ensuring that the computing devices always have candidate tasks that are not lower than the preset queue length threshold and whose predicted minimum TTC is lower than the safety threshold, thus achieving a seamless connection between "search and execution". Let the current number of available devices be , the simulation queue length be Q, and the preset threshold be Qmin. If Q < Qmin, a new generation is triggered as follows: ； In the formula, This represents the average simulation duration. This represents the average time taken for a single simulation. The maximum population size; The number of individuals that will be generated for the next generation; S34, Search-Simulation Cooperative Closed Loop; MSO works in deep collaboration with cloud simulation and time-series agents to form a complete online closed loop: In the exploration phase, MSO uses the "superimposed mirage" strategy to sample a wide area in high-dimensional space and generate diverse candidate scenarios; During the verification phase, candidate scenarios are submitted to the cloud computing module for parallel simulation, and the time-series agent dynamically evaluates the value of the scenarios during execution. During the feedback phase, the agent outputs the actual or predicted TTC back to the MSO to update the fitness of individuals in the population. During the development phase, MSO uses the "underlying mirage" strategy to conduct localized fine-grained searches in high-risk areas, gradually converging to areas with a high density of critical failure scenarios.

5. The cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling according to claim 1, characterized in that, The specific method of S4 is as follows: S41. Problem Modeling; The resource allocation problem is modeled as an online convex optimization problem; let the current candidate task set be J, and the device set be D. For each task j∈J, the time-enhanced surrogate model provides a prediction of the simulation duration. and tasks Key scores For each device ∈D, the monitoring module provides a normalized overall load; S42. Objective function design; Define decision variables ∈[0,1] represents the task Assigned to device The objective function is designed to minimize the weighted completion time, which is equivalent to prioritizing the scheduling of high-risk, short-duration tasks while also considering device load balancing. As shown in the following formula: ； In the formula, For smoothing terms; This is the set of tasks currently awaiting scheduling; For the task Assigned to device ; S43. Constraint Setting: Constraints include three types: The task integrity constraint requires that each task must be completely assigned to a single device, as shown below: ； The load limit constraint means that the device load cannot exceed a preset threshold, as shown below: ； In the formula, This refers to the simulation duration. This is the load threshold; This represents the current base load of device i; The non-negativity constraint means that the allocation ratio is non-negative, as shown below: ； S44. Execute the deterministic solution process; For each scheduling moment, the system executes the following deterministic process: First, based on the current candidate task set, device set, agent prediction, and device load, an optimization sub-problem is constructed. This problem is an integer linear programming problem, employing a relaxation + rounding strategy: Relaxation: Transform the problem into a linear programming problem; solve it using the interior-point method within milliseconds to obtain the optimal allocation scheme; then perform deterministic rounding on each task and allocate it to the device that maximizes the objective function gain, as shown below: ； In the formula, For load balancing weights; The optimal device index that minimizes the objective function; Finally, immediately update the device load status to prepare for the next round of scheduling, as shown below: ； S45, interrupt-driven resource reclamation; When the time-enhanced agent model triggers a simulation interrupt instruction, the scheduler immediately performs resource reclamation operations: marking the interrupted task as "non-critical" and no longer retrying; releasing the occupied device load; and triggering a new round of online optimization to allocate idle resources to candidate tasks in real time.

6. The cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling according to claim 1, characterized in that, The specific method of S5 is as follows: S51, virtual intranet construction; A unified virtual intranet is constructed using software-defined networking technology based on ZeroTier; all computing nodes participating in the test automatically join the same virtual network upon startup, and the system assigns fixed private IP addresses, making each node logically appear to be in the same local area network. S52, redundant routing design; The system designs a hierarchical tree-like redundant routing architecture. The entire communication network is rooted at the scheduler, with multiple primary relay nodes. Each primary node is then connected to several simulation computing nodes as secondary terminals. The system continuously monitors the health status of each communication path and detects link latency and connectivity through periodic heartbeat packets. If a path becomes abnormal, the system will automatically switch the affected nodes to the backup path; S53. Design the coding mechanism; A four-segment data frame format is defined; each data packet consists of four parts in sequence: first, a 4-byte target node identifier, used to precisely specify the receiver; second, a 1-byte task type field, indicating whether it is starting a simulation, requesting an interruption, or reporting a status; third, a variable-length payload content, using a type-length-value structure, carrying the task ID and parameter configuration; and finally, a 4-byte CRC-32 checksum, covering all the preceding fields; the receiving end first verifies the checksum before processing, and if it does not match, it is discarded directly and a retransmission is requested, and suspicious instructions are never executed; S54. Design a multi-level feedback mechanism; The system is designed with an acknowledgment and retransmission mechanism. When the scheduler issues an interrupt command, the receiving node is required to return an acknowledgment (ACK) within 5 milliseconds. If the sender does not receive the ACK within 10 milliseconds, the first retransmission is triggered immediately, and a maximum of two retransmissions are allowed. All retransmissions carry the same task identifier, and the receiver has deduplication capabilities to avoid repeated execution. At the same time, the system opens a high-priority transmission channel for critical instructions to ensure that they can still be processed with priority when the network is congested.

7. The cloud-based parallel acceleration testing method for intelligent vehicles based on time-series agent scheduling according to claim 1, characterized in that, The specific method of S6 is as follows: S61, Initialization phase; Load system configuration parameters, start all compute nodes, initialize the time-series enhanced agent model, initialize the MSO population, and clear the task queue; S62, Scene Generation Stage; MSO generates 50 candidate scenarios based on the current population state, quickly pre-evaluates them using a time-enhanced agent model, and selects scenarios whose predicted minimum TTC is lower than a preset safety threshold to add to the task queue. S63, Task Allocation Phase; The resource scheduling module dynamically allocates tasks in the queue to computing nodes that minimize the weighted completion time and meet the load limit constraint based on the real-time load of the equipment and the criticality of the tasks. S64, Simulation Execution Phase; The computing nodes perform high-fidelity simulations, and the timing-enhanced agent module extracts timing features every second for dynamic evaluation. S65, Interruption of decision-making stage; If the time-enhanced proxy model is ready and the interrupt conditions are met, a millisecond-level interrupt is triggered to release resources; Otherwise, continue execution until completion; S66, Feedback and Update Phase; The simulation results are fed back to the MSO and time-series enhanced surrogate model to update the population fitness and model parameters; S67, Iterative Looping; Repeat steps S62-S66 until the preset total number of scenarios or the maximum test time is reached.

Images