A Trustworthy Accelerated Generation Method for Test Scenarios Based on Multi-Simulator Consistency Verification

Abstract

This invention belongs to the field of autonomous driving testing technology, specifically a reliable accelerated generation method for test scenarios based on multi-simulator consistency verification. First, the traffic scene to be tested is parameterized, and simulators with different dynamics, sensor, and behavioral models are selected. The same scene parameters are synchronously mapped to each simulator for execution, collecting results such as the main vehicle's response, collision state, and minimum distance. Then, the consistency deviation of the multiple simulators in risk scoring, failure judgment, and trajectory evolution is calculated to construct a reliable risk-reward index and screen reliable high-risk scenarios. Sampling weights are allocated according to the index, the accelerated generation distribution is updated, and constraint projection ensures the compliance of new scenarios. Finally, the reliable failure risk of the autonomous driving system is estimated by utilizing the reweighted relationship between the natural scene and the generated distribution, combined with the consistency results. This invention can effectively eliminate pseudo-high-risk scenarios caused by the deviation of a single simulator, improving the reliability and stability of test scenario generation.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving testing technology, specifically a reliable accelerated generation method for test scenarios based on multi-simulator consistency verification. Background Technology

[0002] With the continuous development of autonomous driving technology, the need for verification of autonomous driving systems in closed-field testing, open road testing, and simulation testing is constantly increasing. Due to the high cost, long cycle, and low probability of dangerous scenarios in real-world testing, relying solely on natural driving data or open road testing is insufficient to fully expose the potential failure modes of autonomous driving systems in complex traffic environments within a limited timeframe. Therefore, constructing large-scale test scenarios based on simulation environments and increasing the frequency of high-risk scenarios through accelerated testing methods has become an important technical approach for the safety verification of autonomous driving systems.

[0003] Existing methods for generating autonomous driving test scenarios are mostly based on a single simulation platform, generating potentially risky test scenarios through parameter perturbation, combined testing, optimization search, or reinforcement learning. While these methods can improve testing efficiency to some extent, the generated results are highly dependent on the simulator's vehicle dynamics model, sensor model, traffic participant behavior model, collision detection model, and road environment modeling method. When the same test scenario is executed in different simulators, inconsistencies may arise in risk results, collision states, driver behavior responses, and traffic participant trajectory evolution. This means that some high-risk scenarios are merely results under specific simulator model biases, making them unreliable samples for autonomous driving system performance testing.

[0004] Furthermore, existing accelerated testing methods often focus on increasing the frequency of hazardous scenarios, neglecting the stability and transferability of generated scenarios across different simulation platforms. If scenario selection is based solely on collision, minimum distance, or collision time metrics from a single simulator, simulation errors, numerical anomalies, or platform-specific behaviors may be misidentified as real high-risk scenarios. This not only reduces the credibility of autonomous driving testing but may also lead to the concentration of testing resources in low-credibility scenarios, impacting the efficiency of accelerated testing and the accuracy of failure probability estimation. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a reliable accelerated generation method for test scenarios based on multi-simulator consistency verification. The method constructs a parameterized representation of the traffic scenario to be tested and selects multiple simulators with differentiated modeling capabilities. Parameters of the same scenario are synchronously mapped to multiple simulators for execution, collecting data on the driver's response, traffic participant trajectories, collision states, minimum distance, minimum collision time, and simulation validity status. The method calculates the consistency of risk results, failure results, and trajectory evolution across multiple simulators. Based on the consistency evaluation results, a reliable risk-benefit index is constructed, and high-reliability, high-risk scenarios are selected. The accelerated test generation distribution is updated according to the reliable risk-benefit. Finally, the simulation reliability weights are used to estimate the failure risk of the autonomous driving system.

[0006] The technical solution of this invention is described below in conjunction with the accompanying drawings:

[0007] This invention provides a reliable and accelerated generation method for test scenarios based on multi-simulator consistency verification, including:

[0008] S1, Scene parameterization and multi-simulator configuration;

[0009] The road structure, traffic objects, environmental conditions, and vehicle tasks are converted into computable parameter vectors, and simulators with different dynamic models, sensor models, and traffic behavior models are selected.

[0010] S2, Multi-emulator Testing;

[0011] The same scenario parameters were input into different simulators and run to collect test results on collision state, minimum distance, minimum collision time, main vehicle braking intensity, lateral avoidance intensity, and trajectory response.

[0012] S3. Construction of Consistency Evaluation Indicators;

[0013] The difference between the output risk score, failure judgment and main vehicle trajectory response is calculated. By the risk score dispersion, failure result consistency deviation and trajectory evolution consistency deviation, it is determined whether the test scenario has cross-simulator stability.

[0014] S4. Screening of Trusted High-Risk Scenarios;

[0015] Invalid simulation results are excluded, and the consistency of multiple simulators, simulation effectiveness, and scenario risk level are integrated into a credible risk-reward indicator.

[0016] S5, Scene Distribution Update and Accelerated Generation;

[0017] Sampling weights are assigned based on the risk-reward magnitude of credible high-risk scenarios, while constrained projection ensures that newly generated scenarios meet the requirements of road structure, traffic rules, vehicle dynamics, and environmental parameters.

[0018] S6. Trustworthy Failure Risk Assessment;

[0019] By utilizing the proportional relationship between the natural scene distribution and the accelerated generation distribution, the accelerated generation samples are reweighted and corrected. Combined with the consistency and validity results of multiple simulators, the reliable failure risk of the autonomous driving system in natural traffic scenarios is estimated.

[0020] Furthermore, the specific method of S1 is as follows:

[0021] S11. Structured representation of autonomous driving test scenarios;

[0022] First, the traffic scenario to be tested is structurally represented. Road information, traffic rule information, dynamic object information, environmental information, and vehicle task information related to the autonomous driving system test are uniformly written into the scenario state, resulting in the... The structured representation of each test scenario is shown in Equation (1):

[0023] (1)

[0024] In the formula, For the first One autonomous driving test scenario; Number the test scenarios; For the first The road structure information for each test scenario is used to represent the road geometry, number of lanes, lane width, road boundaries, and road curvature. For the first The road facilities and traffic rules information for each test scenario are used to represent lane line type, speed limit information, traffic signs, traffic signals, and traffic priority. For the first The dynamic object information of each test scenario is used to represent the state of vehicles, pedestrians, non-motorized vehicles and other moving objects within the scenario; For the first Environmental condition information for each test scenario, used to represent weather, lighting, visibility, road surface adhesion conditions, and sensor degradation conditions; For the first The main vehicle task information for each test scenario is used to represent the main vehicle's starting point, ending point, target lane, expected speed, and driving task;

[0025] S12, Construct the test scenario parameter vector;

[0026] The scene structure representation is mapped to a unified parameter vector, as shown in equation (2):

[0027] (2)

[0028] In the formula, For the first A unified parameter vector for each test scenario; For the first Each test scenario has a road parameter subvector representing road curvature, lane width, slope, intersection structure, and road boundary parameters. For the first Each test scenario has a traffic object parameter subvector, which represents the number of traffic objects, initial position, initial speed, acceleration, target lane, and behavior type. For the first Each test scenario has an environmental parameter subvector representing light intensity, rainfall intensity, fog concentration, road surface adhesion coefficient, and sensor noise intensity. For the first The main vehicle task parameter subvectors for each test scenario represent the main vehicle's initial state, target speed, target position, and planned task; superscript... This represents the vector transpose operation;

[0029] S13. Determine the multi-emulator test environment;

[0030] Select The simulators constitute a multi-simulator test environment, as shown in equation (3):

[0031] (3)

[0032] In the formula, It is a collection of multiple simulators; For the first One simulator; Number the simulator; The total number of simulators, and It is an integer greater than or equal to 2.

[0033] Furthermore, the specific method of S2 is as follows:

[0034] S21. Execute the same test scenario in multiple simulators;

[0035] For the Each test scenario uses the same parameter vector. Input multiple simulator sets separately Each simulator in the dataset, and the first one is collected. The test result vector under each simulator is shown in Equation (4):

[0036] (4)

[0037] In the formula, For the first The test scenario is in the first... A vector of test results from each simulator; For the first The test scenario is in the first... The collision status indicator in the simulator takes a value of 1 when a collision occurs and a value of 0 otherwise. For the first The test scenario is in the first... The minimum distance between the main vehicle and other traffic participants in a simulator; For the first The test scenario is in the first... The minimum collision time between the main vehicle and other traffic participants in a simulator; For the first The test scenario is in the first... The absolute value of the maximum braking deceleration of the main vehicle in each simulator; For the first The test scenario is in the first... The lateral avoidance strength in a simulator is used to represent the maximum lateral maneuvering strength of the main vehicle in order to avoid risks;

[0038] S22. Construct a risk normalization index for a single simulator;

[0039] The minimum distance, minimum collision time, and braking intensity are normalized as shown in equations (5), (6), and (7):

[0040] (5)

[0041] (6)

[0042] (7)

[0043] In the formula, For the first The test scenario is in the first... Distance risk normalization value in each simulator; For the first The test scenario is in the first... Normalized collision time risk values ​​in each simulator; For the first The test scenario is in the first... Braking intensity risk normalization value in a simulator; This is a reference threshold for safe distance; This is a reference threshold for dangerous distance, and Greater than ; This serves as a reference threshold for safe collision time. This serves as a reference threshold for the time of a dangerous collision, and Greater than ; This serves as a reference threshold for safe braking intensity. This is a reference threshold for dangerous braking intensity, and Greater than ; This is the distance-normalized stable term; This is the collision time normalized stable term; This is the normalized stability term for braking intensity; This is a function to find the maximum value. The function is for finding the minimum value;

[0044] S23. Calculate the risk score for a single simulator;

[0045] Based on the normalized risk index, the calculation of the first... The test scenario is in the first... The risk score in each simulator is shown in Equation (8):

[0046] (8)

[0047] In the formula, For the first The test scenario is in the first... Risk scoring in a simulator; These are the collision state weighting coefficients; This is the distance risk weighting coefficient; This refers to the collision time risk weighting coefficient. This refers to the risk weighting coefficient for braking intensity. This refers to the lateral avoidance strength weighting coefficient.

[0048] S24. Determine the failure status of the single simulator;

[0049] Determine the first based on the risk score. The test scenario is in the first... The failure states in the simulator are shown in equation (9):

[0050] (9)

[0051] In the formula, For the first The test scenario is in the first... The failure status indicator in the simulator is 1 when it means that the autonomous driving system has failed in the test in the simulator, and 0 when it means that no failure has occurred. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. For the first The test scenario is in the first... Risk scoring in a simulator; To determine the risk score threshold for the failure of an autonomous driving system test.

[0052] Furthermore, the specific method of S3 is as follows:

[0053] S31. Construct an average risk score for multiple simulators;

[0054] Calculate the first The average risk score of multiple simulators for each test scenario is shown in Equation (10):

[0055] (10)

[0056] In the formula, For the first Average risk score of multiple simulators in a test scenario; This represents the total number of simulators. Number the simulator; For the first The test scenario is in the first... Risk scoring in a simulator;

[0057] S32. Construct a risk scoring dispersion index;

[0058] The risk score dispersion is calculated as shown in equation (11):

[0059] (11)

[0060] In the formula, For the first Risk score dispersion for each test scenario; For the first The test scenario is in the first... Risk scoring in a simulator; For the first The average risk score of a test scenario across multiple simulators. The greater the dispersion of the risk score, the more significant the difference in risk assessment results for that scenario across different simulators.

[0061] S33. Construct a consistency deviation index for failure results;

[0062] The consistency deviation index of the failure results is calculated as shown in equations (12) and (13):

[0063] (12)

[0064] (13)

[0065] In the formula, For the first Average failure state of multiple simulators in a test scenario; For the first The test scenario is in the first... Failure status indications in a simulator; For the first Consistency deviation index of failure results for each test scenario; For absolute value operators, the larger the consistency deviation index of failure results, the more inconsistent the judgments of different simulators on whether the scenario has failed.

[0066] S34. Construct a trajectory evolution consistency deviation index;

[0067] Compare the differences in the trajectory evolution of the main vehicle in multiple simulators, and let the first... The test scenario is in the first... The first simulator The vehicle state vector at each discrete moment is: The trajectory evolution consistency deviation index is calculated as shown in equation (14):

[0068] (14)

[0069] In the formula, For the first The consistency deviation index of trajectory evolution in each test scenario; This represents the total number of simulators. and All of these are simulator serial numbers, and Greater than ; This represents the total number of discrete sampling moments during the execution of the test scenario. Number the discrete time points; For the first The test scenario is in the first... The first simulator The main vehicle state vector at each discrete moment includes the main vehicle position, velocity, heading angle, and acceleration; For the first The test scenario is in the first... The first simulator The main vehicle state vector at each discrete moment; It is a two-dimensional or multi-dimensional Euclidean norm; The normalization scaling parameter for trajectory differences; The stability term for trajectory consistency calculation is as follows: the larger the trajectory evolution consistency deviation index, the more obvious the difference in the main vehicle response induced by the same scenario in multiple simulators.

[0070] S35. Construct a multi-simulator consistency score;

[0071] Taking into account the dispersion of the comprehensive risk score, the consistency deviation of failure results, and the consistency deviation of trajectory evolution, we obtain the first... The consistency score of multiple simulators for each test scenario is shown in Equation (15):

[0072] (15)

[0073] In the formula, For the first Consistency scoring of multiple simulators in a test scenario; It is a natural exponential function; This is the penalty coefficient for the dispersion of the risk score; The penalty coefficient for consistency deviation of failure results; The penalty coefficient for trajectory evolution consistency deviation; For the first Risk score dispersion for each test scenario; For the first Consistency deviation index of failure results for each test scenario; For the first The consistency deviation index of trajectory evolution for each test scenario. The higher the consistency score of multiple simulators, the more stable the test results of the scenario are in different simulators, and the higher the credibility of the scenario.

[0074] Furthermore, the specific method of S4 is as follows:

[0075] S41. Construct simulation execution effectiveness indicators;

[0076] The validity of the scene execution in each simulator is determined as shown in Equation (16):

[0077] (16)

[0078] In the formula, For the first The test scenario is in the first... The simulation execution validity indicator in each simulator is set to 1 when the simulation execution is valid and 0 when the simulation execution is invalid. For indicator functions; For the first The test scenario is in the first... The simulation completion status indicator in each simulator takes a value of 1 when the simulation has been fully executed to the termination time, and a value of 0 otherwise. For the first The test scenario is in the first... The numerical anomaly indicator in the simulator takes a value of 1 when numerical divergence, non-physical abrupt change in vehicle state, or simulation program abnormality occurs; otherwise, it takes a value of 0. For the first The test scenario is in the first... The geometric anomaly indicator in the simulator takes a value of 1 when a vehicle passes through a mold, an anomaly is generated outside the road, or an illegal overlap of the initial position occurs; otherwise, it takes a value of 0. This represents the logical AND operation;

[0079] S42. Construct a multi-simulator effectiveness scoring system;

[0080] Calculate the first The effectiveness score of multiple simulators for each test scenario is shown in Equation (17):

[0081] (17)

[0082] In the formula, For the first Effectiveness scoring of multiple simulators in a single test scenario; This represents the total number of simulators. Number the simulator; For the first The test scenario is in the first... The simulation execution effectiveness indicator in each simulator is as follows: the higher the multi-simulator effectiveness score, the more stable, complete, and reasonable the scenario can be executed in multiple simulators.

[0083] S43. Construct credible risk-return indicators;

[0084] The consistency score of multiple simulators, the effectiveness score of multiple simulators, the average risk score, and the average failure state are fused to obtain the credible risk-benefit index, as shown in Equation (18):

[0085] (18)

[0086] In the formula, For the first Credible risk-reward metrics for each test scenario; For the first Consistency scoring of multiple simulators in a test scenario; For the first Effectiveness scoring of multiple simulators in a single test scenario; This refers to the average risk score weighting coefficient. This is the average failure state weighting coefficient; For the first Average risk score of multiple simulators in a test scenario; For the first The higher the credible risk-reward index of the average failure state of multiple simulators in a test scenario, the higher the risk level of the scenario, and the more stable and reliable the test results are in multiple simulators.

[0087] S44. Screen for trustworthy high-risk scenarios;

[0088] The set of credible high-risk scenarios is selected based on multi-simulator consistency score, multi-simulator effectiveness score and average risk score, as shown in Equation (19):

[0089] (19)

[0090] In the formula, A collection of trustworthy high-risk scenarios; For the first One autonomous driving test scenario; For the first Consistency scoring of multiple simulators in a test scenario; The lower limit threshold for multi-simulator consistency scoring; For the first Effectiveness scoring of multiple simulators in a single test scenario; The lower limit threshold for multi-simulator effectiveness scoring; For the first Average risk score of multiple simulators in a test scenario; This is the lower limit threshold of the average risk score that must be met for a credible high-risk scenario.

[0091] Furthermore, the specific method of S5 is as follows:

[0092] S51. Construct trusted scene sampling weights;

[0093] In the During the accelerated generation process, sampling weights are assigned to credible high-risk scenarios based on the credible risk-reward index, as shown in equation (20):

[0094] (20)

[0095] In the formula, For the first The first round of accelerated generation process Sampling weights for each credible high-risk scenario; To expedite the generation of round numbers; For the first Wheel of Life Credible risk-reward metrics for each test scenario; This is a stable term for sampling weights, used to prevent the weights from being uncalculated when the credible risk-return indicator is zero; For the first A set of trustworthy high-risk scenarios that can be obtained in turn; The first in the set of trusted high-risk scenarios One test scenario; For the first Wheel of Life Credible risk-reward metrics for each test scenario; The scenario number is located within the set of trusted high-risk scenarios.

[0096] S52, Update the trusted accelerated generation distribution;

[0097] The previous round of scene generation distribution is fused with the kernel density distribution guided by the trusted high-risk scene to obtain the next round of trusted accelerated generation distribution, as shown in Equation (21):

[0098] (twenty one)

[0099] In the formula, For the first +1 round of trusted accelerated distribution generation; This is the parameter vector for the test scenario to be generated; The step size coefficient is updated for the distribution, and its value ranges from 0 to 1. For the first Wheel scene generation distribution; For the first One credible high-risk scenario; For the first Wheel of Life Sampling weights for each credible high-risk scenario; The bandwidth matrix is The kernel function; This is the kernel function bandwidth matrix, used to control the range of expansion of newly generated parameters around the parameters of the credible high-risk scenario; For the first A parameter vector for a credible high-risk scenario;

[0100] S53. Generate a new test scenario that satisfies the parameter constraints;

[0101] According to the After obtaining candidate parameters through +1 rounds of trusted accelerated generation distribution sampling, the candidate parameters are constrained and projected to ensure that the generated scene meets the constraints of road structure, traffic rules, vehicle dynamics, and environmental value range, as shown in Equation (22):

[0102] (twenty two)

[0103] In the formula, For the first +1 round generated the first A test scenario parameter vector that satisfies the constraints; For the first +1 round of candidate scene number; This is a parameter-constrained projection operator used to project candidate parameters that do not meet the constraints to the feasible parameter space. The feasible parameter space represents the set of parameters that meet the requirements of road structure, traffic rules, vehicle dynamics, and environmental values. For the first +1 rounds of sampling from the trusted accelerated generation distribution. A vector of candidate parameters; symbol This indicates that the probability follows a certain probability distribution; For the first +1 round of trusted accelerated distribution generation.

[0104] Furthermore, the specific method of S6 is as follows:

[0105] S61. Construct the reweighting coefficients between the natural distribution and the accelerated generation distribution;

[0106] To enable the accelerated scenario generation to be used for failure risk estimation of autonomous driving systems, the calculation of the first... Wheel of Life The reweighting coefficients for each test scenario are shown in equation (23):

[0107] (twenty three)

[0108] In the formula, For the first Wheel of Life The reweighting coefficients for each test scenario; For the first The probability density of a test scenario parameter vector under the natural scene distribution; For the first A parameter vector for each test scenario; For the first The test scenario parameter vector is at the... The probability density under the accelerated generation distribution is given by equation (21). of Values The value at time; This is a weighted stabilizing term used to avoid the denominator being zero;

[0109] S62. Construct a reliable failure indicator;

[0110] By combining the multi-simulator consistency score, the multi-simulator effectiveness score, and the average failure state, a reliable failure indicator is obtained, as shown in equation (24):

[0111] (twenty four)

[0112] In the formula, For the first The number of reliable failure indicators for each test scenario; For the first Consistency scoring of multiple simulators in a test scenario; For the first Effectiveness scoring of multiple simulators in a single test scenario; For the first The larger the credible failure indication quantity of the multi-simulator average failure state of a test scenario, the more likely the scenario is to be a credible test sample that stably triggers the failure of the autonomous driving system in multiple simulators.

[0113] S63. Estimate the risk of reliable failure in autonomous driving systems;

[0114] Based on the reweighting coefficients and the credible failure indicator, the credible failure risk of the autonomous driving system under natural scene distribution is estimated, as shown in Equation (25):

[0115] (25)

[0116] In the formula, For the first The reliable failure risk of the autonomous driving system obtained by round estimation; For the first The number of test scenarios in which risk estimation is involved; Number the test scenarios; For the first Wheel of Life The reweighting coefficients for each test scenario; For the first The number of reliable failure indicators for each test scenario;

[0117] S64. Calculate and accelerate the efficiency of effective test samples;

[0118] To determine whether the accelerated test results are stable, the effective sample efficiency of the g-th round of accelerated testing is calculated, as shown in equation (26):

[0119] (26)

[0120] In the formula, For the first Wheel acceleration test of effective sample efficiency; For the first The number of test scenarios in which risk estimation is involved; Number the test scenarios; For the first Wheel of Life The reweighting coefficients for each test scenario indicate that the higher the efficiency of the effective samples, the more balanced the weights of the accelerated test samples and the more stable the failure risk estimation.

[0121] S65. Construct accelerated generation termination conditions;

[0122] When the credible failure risk estimation results tend to stabilize and the efficiency of the effective sample meets the requirements, the credible accelerated generation process is terminated, as shown in equation (27):

[0123] (27)

[0124] In the formula, For the first The reliable failure risk of the autonomous driving system obtained by round estimation; For the first -1 round of estimation yields the reliable failure risk of the autonomous driving system; This is for absolute value operations; To estimate the convergence threshold for failure risk; For logical AND operation; For the first Wheel acceleration test of effective sample efficiency; The effective sample efficiency lower limit threshold is defined as follows: when equation (27) holds, it means that the credible failure risk estimation has been stabilized and the sample weight distribution meets the requirements, and it is possible to stop generating new test scenarios.

[0125] The beneficial effects of this invention are as follows:

[0126] This invention provides a reliable accelerated generation method for test scenarios based on multi-simulator consistency verification. It constructs a parameterized representation of the traffic scenario to be tested and selects multiple simulators with differentiated modeling capabilities. Parameters of the same scenario are synchronously mapped to multiple simulators for execution, collecting data on the driver vehicle's response, traffic participant trajectories, collision states, minimum distance, minimum collision time, and simulation validity status. The method calculates the consistency of risk results, failure results, and trajectory evolution across multiple simulators. Based on the consistency evaluation results, a reliable risk-benefit index is constructed, and high-reliability, high-risk scenarios are selected. The accelerated test generation distribution is updated according to the reliable risk-benefit. Finally, simulation reliability weights are used to estimate the failure risk of the autonomous driving system. This invention avoids false high-risk scenarios caused by the bias of a single simulator from entering the test set, improving the reliability and stability of generated autonomous driving test scenarios. Attached Figure Description

[0127] 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.

[0128] Figure 1 This is a flowchart of the present invention;

[0129] Figure 2 A diagram illustrating the execution of the same test scenario in multiple simulators and the verification of result consistency.

[0130] Figure 3 A flowchart for accelerating the generation of distributions based on credible risk-return metrics;

[0131] Figure 4 This is a flowchart for estimating the credible failure risk of an autonomous driving system based on reweighted coefficients. Detailed Implementation

[0132] 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.

[0133] Example 1

[0134] See Figures 1-4 This invention provides a reliable and accelerated generation method for test scenarios based on multi-simulator consistency verification, comprising the following steps:

[0135] S1. Construct a parameterized representation of the autonomous driving test scenario to be tested, and determine the multi-simulator test environment;

[0136] The road structure, traffic objects, environmental conditions, and vehicle tasks in the traffic scenario under test are uniformly converted into computable parameter vectors. Simultaneously, multiple simulators with different dynamic models, sensor models, and traffic behavior models are selected to provide a foundational environment for subsequent consistency verification. The specific method is as follows:

[0137] S11. Structured representation of autonomous driving test scenarios;

[0138] This invention first presents a structured representation of the traffic scenario to be tested. It integrates road information, traffic rule information, dynamic object information, environmental information, and the main vehicle's task information related to the autonomous driving system test into the scenario state, thus obtaining the first... The structured representation of each test scenario is shown in Equation (1):

[0139] (1)

[0140] In the formula, For the first One autonomous driving test scenario; Number the test scenarios; For the first The road structure information for each test scenario is used to represent the road geometry, number of lanes, lane width, road boundaries, and road curvature. For the first The road facilities and traffic rules information for each test scenario are used to represent lane line type, speed limit information, traffic signs, traffic signals, and traffic priority. For the first The dynamic object information of each test scenario is used to represent the state of vehicles, pedestrians, non-motorized vehicles and other moving objects within the scenario; For the first Environmental condition information for each test scenario, used to represent weather, lighting, visibility, road surface adhesion conditions, and sensor degradation conditions; For the first The main vehicle task information for each test scenario is used to represent the main vehicle's starting point, ending point, target lane, expected speed, and driving task;

[0141] S12, Construct the test scenario parameter vector;

[0142] To facilitate the synchronous reproduction of experiments across multiple simulators, this invention further maps the structured representation of the scene into a unified parameter vector, as shown in equation (2):

[0143] (2)

[0144] In the formula, For the first A unified parameter vector for each test scenario; For the first Each test scenario has a road parameter subvector representing road curvature, lane width, slope, intersection structure, and road boundary parameters. For the first Each test scenario has a traffic object parameter subvector, which represents the number of traffic objects, initial position, initial speed, acceleration, target lane, and behavior type. For the first Each test scenario has an environmental parameter subvector representing light intensity, rainfall intensity, fog concentration, road surface adhesion coefficient, and sensor noise intensity. For the first The main vehicle task parameter subvectors for each test scenario represent the main vehicle's initial state, target speed, target position, and planned task; superscript... This represents the vector transpose operation;

[0145] S13. Determine the multi-emulator test environment;

[0146] To avoid bias in test conclusions caused by a single simulator, this invention selects... The simulators constitute a multi-simulator test environment, as shown in equation (3):

[0147] (3)

[0148] In the formula, It is a collection of multiple simulators; For the first One simulator; Number the simulator; The total number of simulators, and It is an integer greater than or equal to 2; different simulators have different vehicle dynamics models, sensor models, traffic participant behavior models, collision detection models, and road environment modeling methods;

[0149] S2. Map the same test scenario to multiple simulators for execution and collect test results from multiple simulators;

[0150] The same scenario parameters are input into different simulators to ensure that each simulator performs the same test task. During the simulation, test results such as collision state, minimum distance, minimum collision time, main vehicle braking intensity, lateral avoidance intensity, and trajectory response are collected. The specific method is as follows:

[0151] S21. Execute the same test scenario in multiple simulators;

[0152] For the In a test scenario, this invention uses the same parameter vector. Input multiple simulator sets separately Each simulator in the dataset, and the first one is collected. The test result vector under each simulator is shown in Equation (4):

[0153] (4)

[0154] In the formula, For the first The test scenario is in the first... A vector of test results from each simulator; For the first The test scenario is in the first... The collision status indicator in the simulator takes a value of 1 when a collision occurs and a value of 0 otherwise. For the first The test scenario is in the first... The minimum distance between the main vehicle and other traffic participants in a simulator; For the first The test scenario is in the first... The minimum collision time between the main vehicle and other traffic participants in a simulator; For the first The test scenario is in the first... The absolute value of the maximum braking deceleration of the main vehicle in each simulator; For the first The test scenario is in the first... The lateral avoidance strength in a simulator represents the maximum lateral maneuvering strength of the main vehicle in order to avoid a risk; superscript This represents the vector transpose operation;

[0155] S22. Construct a risk normalization index for a single simulator;

[0156] To unify different risk dimensions, this invention normalizes the minimum distance, minimum collision time, and braking intensity, as shown in equations (5), (6), and (7):

[0157] (5)

[0158] (6)

[0159] (7)

[0160] In the formula, For the first The test scenario is in the first... Distance risk normalization value in each simulator; For the first The test scenario is in the first... Normalized collision time risk values ​​in each simulator; For the first The test scenario is in the first... Braking intensity risk normalization value in a simulator; This is a reference threshold for safe distance; This is a reference threshold for dangerous distance, and Greater than ; This serves as a reference threshold for safe collision time. This serves as a reference threshold for the time of a dangerous collision, and Greater than ; This serves as a reference threshold for safe braking intensity. This is a reference threshold for dangerous braking intensity, and Greater than ; This is the distance-normalized stable term; This is the collision time normalized stable term; This is the normalized stability term for braking intensity; This is a function to find the maximum value. The function is for finding the minimum value;

[0161] S23. Calculate the risk score for a single simulator;

[0162] Based on the normalized risk index, this invention calculates the first... The test scenario is in the first... The risk score in each simulator is shown in Equation (8):

[0163] (8)

[0164] In the formula, For the first The test scenario is in the first... Risk scoring in a simulator; These are the collision state weighting coefficients; This is the distance risk weighting coefficient; This refers to the collision time risk weighting coefficient. This refers to the risk weighting coefficient for braking intensity. This refers to the lateral avoidance strength weighting coefficient. , , , and The meaning is the same as in equation (4); , and The meaning is the same as in equations (5), (6), and (7);

[0165] S24. Determine the failure status of the single simulator;

[0166] This invention determines the first based on risk scoring. The test scenario is in the first... The failure states in the simulator are shown in equation (9):

[0167] (9)

[0168] In the formula, For the first The test scenario is in the first... The failure status indicator in the simulator is 1 when it indicates that the autonomous driving system has failed in the test in the simulator, and 0 when it indicates that no failure has occurred. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. For the first The test scenario is in the first... Risk scoring in a simulator; To determine the risk score threshold for the failure of an autonomous driving system test;

[0169] S3. Construct a consistency evaluation index for results from multiple simulators to quantify the stability of the same scenario in different simulators;

[0170] The differences in risk scores, failure judgments, and vehicle trajectory responses output by different simulators are calculated. By analyzing the risk score dispersion, failure result consistency deviation, and trajectory evolution consistency deviation, the cross-simulator stability of the test scenario is determined. The specific method is as follows:

[0171] S31. Construct an average risk score for multiple simulators;

[0172] To evaluate the overall risk level of the same scenario across multiple simulators, this invention calculates the first... The average risk score of multiple simulators for each test scenario is shown in Equation (10):

[0173] (10)

[0174] In the formula, For the first Average risk score of multiple simulators in a test scenario. This represents the total number of simulators. Number the simulator. For the first The test scenario is in the first... Risk scoring in a simulator;

[0175] S32. Construct a risk scoring dispersion index;

[0176] To measure the difference in risk outcomes for the same scenario across different simulators, this invention calculates the risk score dispersion, as shown in equation (11):

[0177] (11)

[0178] In the formula, For the first Risk score dispersion for each test scenario For the first The test scenario is in the first... Risk scoring in a simulator For the first The average risk score of a test scenario across multiple simulators. The greater the dispersion of the risk score, the more significant the difference in risk assessment results for that scenario across different simulators.

[0179] S33. Construct a consistency deviation index for failure results;

[0180] To measure whether the same scenario triggers the failure of the autonomous driving system in different simulators, this invention calculates the consistency deviation index of failure results, as shown in equations (12) and (13):

[0181] (12)

[0182] (13)

[0183] In the formula, For the first Average failure state of multiple simulators in a test scenario. For the first The test scenario is in the first... Failure status indications in a simulator For the first Consistency deviation index of failure results in each test scenario For absolute value operators, the larger the consistency deviation index of failure results, the more inconsistent the judgments of different simulators on whether the scenario has failed.

[0184] S34. Construct a trajectory evolution consistency deviation index;

[0185] In addition to risk and failure results, this invention further compares the differences in trajectory evolution of the main vehicle in multiple simulators, assuming the first... The test scenario is in the first... The first simulator The vehicle state vector at each discrete moment is: The present invention calculates the trajectory evolution consistency deviation index, as shown in equation (14):

[0186] (14)

[0187] In the formula, For the first The consistency deviation index of trajectory evolution in each test scenario. This represents the total number of simulators. and All of these are simulator serial numbers, and Greater than , This represents the total number of discrete sampling moments during the execution of the test scenario. Numbering discrete time points. For the first The test scenario is in the first... The first simulator A discrete-time vehicle state vector, which includes the vehicle's position, velocity, heading angle, and acceleration. For the first The test scenario is in the first... The first simulator The main vehicle state vector at each discrete moment. For two-dimensional or multi-dimensional Euclidean norms, The normalization scaling parameter for trajectory differences. The stability term for trajectory consistency calculation is as follows: the larger the trajectory evolution consistency deviation index, the more obvious the difference in the main vehicle response induced by the same scenario in multiple simulators.

[0188] S35. Construct a multi-simulator consistency score;

[0189] This invention integrates the dispersion of risk scoring, the consistency deviation of failure results, and the consistency deviation of trajectory evolution to obtain the first... The consistency score of multiple simulators for each test scenario is shown in Equation (15):

[0190] (15)

[0191] In the formula, For the first Consistency scoring of multiple simulators in a test scenario; It is a natural exponential function; This is the penalty coefficient for the dispersion of the risk score; The penalty coefficient for consistency deviation of failure results; The penalty coefficient for trajectory evolution consistency deviation; For the first Risk score dispersion for each test scenario; For the first Consistency deviation index of failure results in each test scenario For the first The consistency deviation index of trajectory evolution of each test scenario. The larger the consistency score of multiple simulators, the more stable the test results of the scenario are in different simulators, and the higher the credibility of the scenario.

[0192] S4. Construct a credible risk-return index based on consistency evaluation and simulation effectiveness, and screen credible high-risk scenarios;

[0193] Invalid simulation results, such as simulation interruptions, numerical anomalies, geometric clipping, or illegal initial states, are excluded. Then, the consistency of multiple simulators, simulation effectiveness, and scenario risk level are integrated into a reliable risk-reward index, which is used to screen stable and reliable high-risk test scenarios. The specific method is as follows:

[0194] S41. Construct simulation execution effectiveness indicators;

[0195] To avoid issues such as numerical anomalies, simulation interruptions, map loading failures, or unreasonable geometric clipping being misjudged as high-risk results, this invention determines the validity of scene execution in each simulator, as shown in equation (16):

[0196] (16)

[0197] In the formula, For the first The test scenario is in the first... The simulation execution validity indicator in each simulator is set to 1 when the simulation execution is valid and 0 when the simulation execution is invalid. For indicator functions; For the first The test scenario is in the first... The simulation completion status indicator in each simulator takes a value of 1 when the simulation has been fully executed to the termination time, and a value of 0 otherwise. For the first The test scenario is in the first... The numerical anomaly indicator in the simulator takes a value of 1 when numerical divergence, non-physical abrupt change in vehicle state, or simulation program abnormality occurs; otherwise, it takes a value of 0. For the first The test scenario is in the first... The geometric anomaly indicator in the simulator takes a value of 1 when a vehicle passes through a mold, an anomaly is generated outside the road, or an illegal overlap of the initial position occurs; otherwise, it takes a value of 0. This represents the logical AND operation;

[0198] S42. Construct a multi-simulator effectiveness scoring system;

[0199] After obtaining the execution validity of each simulator, the present invention calculates the... The effectiveness score of multiple simulators for each test scenario is shown in Equation (17):

[0200] (17)

[0201] In the formula, For the first Effectiveness scoring of multiple simulators in a single test scenario; This represents the total number of simulators. Number the simulator; For the first The test scenario is in the first... The simulation execution effectiveness indicator in each simulator is as follows: the higher the multi-simulator effectiveness score, the more stable, complete, and reasonable the scenario can be executed in multiple simulators.

[0202] S43. Construct credible risk-return indicators;

[0203] This invention integrates multi-simulator consistency score, multi-simulator effectiveness score, average risk score, and average failure state to obtain a reliable risk-benefit index, as shown in equation (18):

[0204] (18)

[0205] In the formula, For the first Credible risk-reward metrics for each test scenario; For the first Consistency scoring of multiple simulators in a test scenario; For the first Effectiveness scoring of multiple simulators in a single test scenario; This refers to the average risk score weighting coefficient. This is the average failure state weighting coefficient; For the first Average risk score of multiple simulators in a test scenario; For the first The higher the credible risk-reward index of the average failure state of multiple simulators in a test scenario, the higher the risk level of the scenario, and the more stable and reliable the test results are in multiple simulators.

[0206] S44. Screen for trustworthy high-risk scenarios;

[0207] This invention filters a set of credible high-risk scenarios based on multi-simulator consistency scoring, multi-simulator effectiveness scoring, and average risk scoring, as shown in equation (19):

[0208] (19)

[0209] In the formula, A collection of trustworthy high-risk scenarios; For the first One autonomous driving test scenario; For the first Consistency scoring of multiple simulators in a test scenario; The lower limit threshold for multi-simulator consistency scoring; For the first Effectiveness scoring of multiple simulators in a single test scenario; The lower limit threshold for multi-simulator effectiveness scoring; For the first Average risk score of multiple simulators in a test scenario; The lower limit threshold of the average risk score that must be met in a credible high-risk scenario;

[0210] S5. Update the scenario generation distribution based on credible risk and return to achieve credible and accelerated generation of autonomous driving test scenarios;

[0211] Sampling weights are assigned based on the risk-reward ratio of credible high-risk scenarios, making the subsequent generation process more inclined to generate high-value test scenarios. At the same time, constraint projection ensures that the newly generated scenarios meet the requirements of road structure, traffic rules, vehicle dynamics, and environmental parameters, thereby achieving credible accelerated generation. The specific method is as follows:

[0212] S51. Construct trusted scene sampling weights;

[0213] In the During the accelerated generation process, this invention assigns sampling weights to credible high-risk scenarios based on the credible risk-reward index, as shown in equation (20):

[0214] (20)

[0215] In the formula, For the first The first round of accelerated generation process Sampling weights for trusted high-risk scenarios. To speed up the generation of round numbers, For the first Wheel of Life Credible risk-reward metrics for each test scenario. This is a sampling weight stabilization term, used to prevent the weights from being uncalculated when the credible risk-return metric is zero. For the first The set of trustworthy high-risk scenarios that can be obtained in turn The first in the set of trusted high-risk scenarios One test scenario, For the first Wheel of Life Credible risk-reward metrics for each test scenario. The scenario number is located within the set of trusted high-risk scenarios.

[0216] S52, Update the trusted accelerated generation distribution;

[0217] This invention fuses the previous round of scene generation distribution with the kernel density distribution guided by trusted high-risk scenes to obtain the next round of trusted accelerated generation distribution, as shown in equation (21):

[0218] (twenty one)

[0219] In the formula, For the first +1 round of trusted accelerated distribution generation; This is the parameter vector for the test scenario to be generated; The step size coefficient is updated for the distribution, and its value ranges from 0 to 1. For the first Wheel scene generation distribution; For the first A set of trustworthy high-risk scenarios that can be obtained in turn; For the first One credible high-risk scenario; For the first Wheel of Life Sampling weights for each credible high-risk scenario; The bandwidth matrix is The kernel function; This is the kernel function bandwidth matrix, used to control the range of expansion of newly generated parameters around the parameters of the credible high-risk scenario; For the first A parameter vector for a credible high-risk scenario;

[0220] S53. Generate a new test scenario that satisfies the parameter constraints;

[0221] According to the After obtaining candidate parameters through +1 rounds of trusted accelerated generation distribution sampling, this invention performs constrained projection on the candidate parameters to ensure that the generated scene meets the constraints of road structure, traffic rules, vehicle dynamics, and environmental value range, as shown in equation (22):

[0222] (twenty two)

[0223] In the formula, For the first +1 round generated the first A test scenario parameter vector that satisfies the constraints; For the first +1 round of candidate scene number; This is a parameter-constrained projection operator used to project candidate parameters that do not meet the constraints to the feasible parameter space. The feasible parameter space represents the set of parameters that meet the requirements of road structure, traffic rules, vehicle dynamics, and environmental values. For the first +1 rounds of sampling from the trusted accelerated generation distribution. A vector of candidate parameters; symbol This indicates that the probability follows a certain probability distribution. For the first +1 round of trusted accelerated distribution generation;

[0224] S6. Estimate the reliable failure risk of autonomous driving system based on the reweighted relationship between the generated distribution and the natural scene distribution;

[0225] By utilizing the proportional relationship between the natural scene distribution and the accelerated generation distribution, the accelerated generation samples are reweighted and corrected. Based on this, and combining the consistency and effectiveness results of multiple simulators, the reliable failure risk of the autonomous driving system in natural traffic scenarios is estimated. The specific method is as follows:

[0226] S61. Construct the reweighting coefficients between the natural distribution and the accelerated generation distribution;

[0227] To enable the accelerated scenario generation to be used for failure risk estimation of autonomous driving systems, this invention calculates the first... Wheel of Life The reweighting coefficients for each test scenario are shown in equation (23):

[0228] (twenty three)

[0229] In the formula, For the first Wheel of Life The reweighting coefficients for each test scenario. For the first The probability density of a test scenario parameter vector under the natural scene distribution. For the first Parameter vectors for each test scenario For the first The test scenario parameter vector is at the... The probability density under the accelerated generation distribution is given by equation (21). of Values The value of time, This is a weighted stabilizing term used to avoid the denominator being zero;

[0230] S62. Construct a reliable failure indicator;

[0231] This invention combines multi-simulator consistency score, multi-simulator effectiveness score and average failure state to obtain a reliable failure indicator, as shown in equation (24):

[0232] (twenty four)

[0233] In the formula, For the first The number of reliable failure indicators for each test scenario. For the first Consistency scoring of multiple simulators in a test scenario. For the first Effectiveness scoring of multiple simulators in each test scenario. For the first The larger the credible failure indication quantity of the multi-simulator average failure state of a test scenario, the more likely the scenario is to be a credible test sample that stably triggers the failure of the autonomous driving system in multiple simulators.

[0234] S63. Estimate the risk of reliable failure in autonomous driving systems;

[0235] Based on the reweighting coefficients and the credible failure indicator, this invention estimates the credible failure risk of the autonomous driving system under natural scene distribution, as shown in Equation (25):

[0236] (25)

[0237] In the formula, For the first The reliable failure risk of the autonomous driving system obtained by round estimation For the first The number of test scenarios involved in risk estimation. Assign test scenario numbers, For the first Wheel of Life The reweighting coefficients for each test scenario. For the first The number of reliable failure indicators for each test scenario;

[0238] S64. Calculate and accelerate the efficiency of effective test samples;

[0239] To determine whether the accelerated test results are stable, this invention calculates the effective sample efficiency of the g-th round of accelerated testing, as shown in equation (26):

[0240] (26)

[0241] In the formula, For the first Wheel acceleration test effective sample efficiency, For the first The number of test scenarios involved in risk estimation. Assign test scenario numbers, For the first Wheel of Life The reweighting coefficients for each test scenario indicate that the higher the efficiency of the effective samples, the more balanced the weights of the accelerated test samples and the more stable the failure risk estimation.

[0242] S65. Construct accelerated generation termination conditions;

[0243] When the credible failure risk estimation result tends to stabilize and the efficiency of the effective sample meets the requirements, the present invention terminates the credible accelerated generation process, as shown in equation (27):

[0244] (27)

[0245] In the formula, For the first The reliable failure risk of the autonomous driving system obtained by round estimation; For the first -1 round of estimation yields the reliable failure risk of the autonomous driving system; This is for absolute value operations; To estimate the convergence threshold for failure risk; This represents the logical AND operation; For the first Wheel acceleration test of effective sample efficiency; The effective sample efficiency lower limit threshold is defined as follows: when equation (27) is true, it means that the credible failure risk estimation has been stabilized and the sample weight distribution meets the requirements, and the generation of new test scenarios can be stopped.

[0246] Example 2

[0247] To verify the effectiveness of this invention, this embodiment constructs 80 autonomous driving following and preceding vehicle cut-in test scenarios, and sets up three simulators with different vehicle dynamics biases, perception noise biases, and traffic behavior biases to test the same scenario parameters. The parameter vector for each test scenario includes the initial distance between the driver vehicle and the target vehicle, the relative speed, and the sensor noise intensity. The three simulators output the collision state, minimum distance, minimum collision time, maximum braking deceleration, lateral avoidance strength, simulation effectiveness, and driver vehicle trajectory response, respectively, and calculate the risk score for each simulator.

[0248] This embodiment sets the following screening criteria for high-confidence test scenarios: the average risk score across multiple simulators is not less than 0.55, the consistency score across multiple simulators is not less than 0.72, the effectiveness score across multiple simulators is equal to 1, and the credible risk-reward index is not less than 0.40. Test scenarios that simultaneously meet all four conditions are selected as high-confidence test scenarios. If a scenario only exhibits high risk in a single simulator, but shows significant differences in risk scores, inconsistent failure judgments, or obvious differences in trajectory responses in other simulators, then the scenario is judged as a pseudo-high-risk scenario and is eliminated.

[0249] The results show that in the first round of natural sampling, only 6 out of 80 test scenarios were identified as high-risk by a single simulator. Among these, 3 scenarios met the high-confidence screening criteria, while 3 were eliminated due to insufficient consistency. After accelerated generation guided by the credible risk-reward index, 56 high-risk scenarios were identified by a single simulator in the 12th round. Of these, 31 scenarios simultaneously met the requirements of average risk score, consistency score, validity score, and credible risk-reward index, and were selected as high-confidence test scenarios. The remaining 25 scenarios, although exhibiting high risk in at least one simulator, were identified as pseudo-high-risk scenarios and eliminated due to insufficient cross-simulator consistency or insufficient credible risk-reward. Thus, the proportion of high-confidence test scenarios increased from 3.8% in the first round to 38.8% in the 12th round. This demonstrates that the present invention not only improves the generation efficiency of high-risk test scenarios but also eliminates pseudo-high-risk scenarios caused by single simulator bias through multi-simulator consistency verification, thereby achieving effective screening of high-confidence test scenarios.

[0250] In summary, this invention constructs a parameterized representation of the traffic scenario under test and selects multiple simulators with differentiated modeling capabilities. Parameters of the same scenario are synchronously mapped to multiple simulators for execution, collecting data on the driver's response, traffic participant trajectories, collision states, minimum distance, minimum collision time, and simulation effectiveness. The consistency of risk results, failure results, and trajectory evolution among the multiple simulators is calculated. A credible risk-reward index is constructed based on the consistency evaluation results, and high-credibility, high-risk scenarios are selected. The accelerated test generation distribution is updated based on the credible risk-reward index. The simulation credibility weight is used to complete the failure risk estimation of the autonomous driving system. This invention can avoid false high-risk scenarios caused by the bias of a single simulator from entering the test set, improving the credibility and stability of autonomous driving test scenario generation.

[0251] 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 reliable and accelerated generation method for test scenarios based on multi-simulator consistency verification, characterized in that, include: S1, Scene parameterization and multi-simulator configuration; The road structure, traffic objects, environmental conditions, and vehicle tasks are converted into computable parameter vectors, and simulators with different dynamic models, sensor models, and traffic behavior models are selected. S2, Multi-emulator Testing; The same scenario parameters were input into different simulators and run to collect test results on collision state, minimum distance, minimum collision time, main vehicle braking intensity, lateral avoidance intensity, and trajectory response. S3. Construction of Consistency Evaluation Indicators; The difference between the output risk score, failure judgment and main vehicle trajectory response is calculated. By the risk score dispersion, failure result consistency deviation and trajectory evolution consistency deviation, it is determined whether the test scenario has cross-simulator stability. S4. Screening of Trusted High-Risk Scenarios; Invalid simulation results are excluded, and the consistency of multiple simulators, simulation effectiveness, and scenario risk level are integrated into a credible risk-reward indicator. S5, Scene Distribution Update and Accelerated Generation; Sampling weights are assigned based on the risk-reward magnitude of credible high-risk scenarios, while constrained projection ensures that newly generated scenarios meet the requirements of road structure, traffic rules, vehicle dynamics, and environmental parameters. S6. Trustworthy Failure Risk Assessment; By utilizing the proportional relationship between the natural scene distribution and the accelerated generation distribution, the accelerated generation samples are reweighted and corrected. Combined with the consistency and validity results of multiple simulators, the reliable failure risk of the autonomous driving system in natural traffic scenarios is estimated.

2. The method for accelerating the generation of a reliable test scenario based on multi-simulator consistency verification according to claim 1, characterized in that, The specific method of S1 is as follows: S11. Structured representation of autonomous driving test scenarios; First, the traffic scenario to be tested is structurally represented. Road information, traffic rule information, dynamic object information, environmental information, and vehicle task information related to the autonomous driving system test are uniformly written into the scenario state, resulting in the... The structured representation of each test scenario is shown in Equation (1): (1) In the formula, For the first One autonomous driving test scenario; Number the test scenario; For the first The road structure information for each test scenario is used to represent the road geometry, number of lanes, lane width, road boundaries, and road curvature. For the first The road facilities and traffic rules information for each test scenario are used to represent lane line type, speed limit information, traffic signs, traffic signals, and traffic priority. For the first The dynamic object information of each test scenario is used to represent the state of vehicles, pedestrians, non-motorized vehicles and other moving objects within the scenario; For the first Environmental condition information for each test scenario, used to represent weather, lighting, visibility, road surface adhesion conditions, and sensor degradation conditions; For the first The main vehicle task information for each test scenario is used to represent the main vehicle's starting point, ending point, target lane, expected speed, and driving task; S12, Construct the test scenario parameter vector; The scene structure representation is mapped to a unified parameter vector, as shown in equation (2): (2) In the formula, For the first A unified parameter vector for each test scenario; For the first Each test scenario has a road parameter subvector representing road curvature, lane width, slope, intersection structure, and road boundary parameters. For the first Each test scenario has a traffic object parameter subvector, which represents the number of traffic objects, initial position, initial speed, acceleration, target lane, and behavior type. For the first Each test scenario has an environmental parameter subvector representing light intensity, rainfall intensity, fog concentration, road surface adhesion coefficient, and sensor noise intensity. For the first The main vehicle task parameter subvectors for each test scenario represent the main vehicle's initial state, target speed, target position, and planned task; superscript... This represents the vector transpose operation; S13. Determine the multi-emulator test environment; Select The simulators constitute a multi-simulator test environment, as shown in equation (3): (3) In the formula, It is a collection of multiple simulators; For the first One simulator; Number the simulator; The total number of simulators, and It is an integer greater than or equal to 2.

3. The method for accelerating the generation of a reliable test scenario based on multi-simulator consistency verification according to claim 1, characterized in that, The specific method of S2 is as follows: S21. Execute the same test scenario in multiple simulators; For the Each test scenario uses the same parameter vector. Input multiple simulator sets separately Each simulator in the dataset, and the first one is collected. The test result vector under each simulator is shown in Equation (4): (4) In the formula, For the first The test scenario is in the first... A vector of test results from each simulator; For the first The test scenario is in the first... The collision status indicator in the simulator takes a value of 1 when a collision occurs and a value of 0 otherwise. For the first The test scenario is in the first... The minimum distance between the main vehicle and other traffic participants in a simulator; For the first The test scenario is in the first... The minimum collision time between the main vehicle and other traffic participants in a simulator; For the first The test scenario is in the first... The absolute value of the maximum braking deceleration of the main vehicle in each simulator; For the first The test scenario is in the first... The lateral avoidance strength in a simulator is used to represent the maximum lateral maneuvering strength of the main vehicle in order to avoid risks. S22. Construct a risk normalization index for a single simulator; The minimum distance, minimum collision time, and braking intensity are normalized as shown in equations (5), (6), and (7): (5) (6) (7) In the formula, For the first The test scenario is in the first... Distance risk normalization value in each simulator; For the first The test scenario is in the first... Normalized collision time risk values ​​in each simulator; For the first The test scenario is in the first... Normalized values ​​of braking intensity risk in each simulator; This is a reference threshold for safe distance; This is a reference threshold for dangerous distance, and Greater than ; This serves as a reference threshold for safe collision time. This serves as a reference threshold for the time of a dangerous collision, and Greater than ; This serves as a reference threshold for safe braking intensity. This is a reference threshold for dangerous braking intensity, and Greater than ; This is the distance-normalized stable term; This is the collision time normalized stable term; This is the normalized stability term for braking intensity; This is a function to find the maximum value. The function is for finding the minimum value; S23. Calculate the risk score for a single simulator; Based on the normalized risk index, the calculation of the first... The test scenario is in the first... The risk score in each simulator is shown in Equation (8): (8) In the formula, For the first The test scenario is in the first... Risk scoring in a simulator; These are the collision state weighting coefficients; This is the distance risk weighting coefficient; This refers to the collision time risk weighting coefficient. This refers to the risk weighting coefficient for braking intensity. This refers to the lateral avoidance strength weighting coefficient; S24. Determine the failure status of the single simulator; Determine the first based on the risk score. The test scenario is in the first... The failure states in the simulator are shown in equation (9): (9) In the formula, For the first The test scenario is in the first... The failure status indicator in the simulator is 1 when it means that the autonomous driving system has failed in the test in the simulator, and 0 when it means that no failure has occurred. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. For the first The test scenario is in the first... Risk scoring in a simulator; To determine the risk score threshold for the failure of an autonomous driving system test.

4. The method for accelerating the generation of a reliable test scenario based on multi-simulator consistency verification according to claim 1, characterized in that, The specific method of S3 is as follows: S31. Construct an average risk score for multiple simulators; Calculate the first The average risk score of multiple simulators for each test scenario is shown in Equation (10): (10) In the formula, For the first Average risk score of multiple simulators in a test scenario; This represents the total number of simulators. Number the simulator; For the first The test scenario is in the first... Risk scoring in a simulator; S32. Construct a risk scoring dispersion index; The risk score dispersion is calculated as shown in equation (11): (11) In the formula, For the first Risk score dispersion for each test scenario; For the first The test scenario is in the first... Risk scoring in a simulator; For the first The average risk score of a test scenario across multiple simulators. The greater the dispersion of the risk score, the more significant the difference in risk assessment results for that scenario across different simulators. S33. Construct a consistency deviation index for failure results; The consistency deviation index of the failure results is calculated as shown in equations (12) and (13): (12) (13) In the formula, For the first Average failure state of multiple simulators in a test scenario; For the first The test scenario is in the first... Failure status indications in a simulator; For the first Consistency deviation index of failure results for each test scenario; For absolute value operators, the larger the consistency deviation index of failure results, the more inconsistent the judgments of different simulators on whether the scenario has failed. S34. Construct a trajectory evolution consistency deviation index; Compare the differences in the trajectory evolution of the main vehicle in multiple simulators, and let the first... The test scenario is in the first... The first simulator The vehicle state vector at each discrete moment is: The trajectory evolution consistency deviation index is calculated as shown in equation (14): (14) In the formula, For the first The consistency deviation index of trajectory evolution in each test scenario; This represents the total number of simulators. and All of these are simulator serial numbers, and Greater than ; This represents the total number of discrete sampling moments during the execution of the test scenario. Number the discrete time points; For the first The test scenario is in the first... The first simulator The main vehicle state vector at each discrete moment includes the main vehicle position, velocity, heading angle, and acceleration; For the first The test scenario is in the first... The first simulator The main vehicle state vector at each discrete moment; It is a two-dimensional or multi-dimensional Euclidean norm; The normalization scaling parameter for trajectory differences; The stability term for trajectory consistency calculation is as follows: the larger the trajectory evolution consistency deviation index, the more obvious the difference in the main vehicle response induced by the same scenario in multiple simulators. S35. Construct a multi-simulator consistency score; Taking into account the dispersion of the comprehensive risk score, the consistency deviation of failure results, and the consistency deviation of trajectory evolution, we obtain the first... The consistency score of multiple simulators for each test scenario is shown in Equation (15): (15) In the formula, For the first Consistency scoring of multiple simulators in a test scenario; It is a natural exponential function; This is the penalty coefficient for the dispersion of the risk score; The penalty coefficient for consistency deviation of failure results; The penalty coefficient for trajectory evolution consistency deviation; For the first Risk score dispersion for each test scenario; For the first Consistency deviation index of failure results for each test scenario; For the first The consistency deviation index of trajectory evolution for each test scenario. The higher the consistency score of multiple simulators, the more stable the test results of the scenario are in different simulators, and the higher the credibility of the scenario.

5. The method for accelerating the generation of a reliable test scenario based on multi-simulator consistency verification according to claim 1, characterized in that, The specific method of S4 is as follows: S41. Construct simulation execution effectiveness indicators; The validity of the scene execution in each simulator is determined as shown in Equation (16): (16) In the formula, For the first The test scenario is in the first... The simulation execution validity indicator in each simulator is set to 1 when the simulation execution is valid and 0 when the simulation execution is invalid. For indicator functions; For the first The test scenario is in the first... The simulation completion status indicator in each simulator takes a value of 1 when the simulation has been fully executed to the termination time, and a value of 0 otherwise. For the first The test scenario is in the first... The numerical anomaly indicator in the simulator takes a value of 1 when numerical divergence, non-physical abrupt change in vehicle state, or simulation program abnormality occurs; otherwise, it takes a value of 0. For the first The test scenario is in the first... The geometric anomaly indicator in the simulator takes a value of 1 when a vehicle passes through a mold, an anomaly is generated outside the road, or an illegal overlap of the initial position occurs; otherwise, it takes a value of 0. This represents the logical AND operation; S42. Construct a multi-simulator effectiveness scoring system; Calculate the first The effectiveness score of multiple simulators for each test scenario is shown in Equation (17): (17) In the formula, For the first Effectiveness scoring of multiple simulators in a single test scenario; This represents the total number of simulators. Number the simulator; For the first The test scenario is in the first... The simulation execution effectiveness indicator in each simulator is as follows: the higher the multi-simulator effectiveness score, the more stable, complete, and reasonable the scenario can be executed in multiple simulators. S43. Construct credible risk-return indicators; The consistency score of multiple simulators, the effectiveness score of multiple simulators, the average risk score, and the average failure state are fused to obtain the credible risk-benefit index, as shown in Equation (18): (18) In the formula, For the first Credible risk-reward metrics for each test scenario; For the first Consistency scoring of multiple simulators in a test scenario; For the first Effectiveness scoring of multiple simulators in a single test scenario; This refers to the average risk score weighting coefficient. This is the average failure state weighting coefficient; For the first Average risk score of multiple simulators in a test scenario; For the first The higher the credible risk-reward index of the average failure state of multiple simulators in a test scenario, the higher the risk level of the scenario, and the more stable and reliable the test results are in multiple simulators. S44. Screen trusted high-risk scenarios; The set of credible high-risk scenarios is selected based on multi-simulator consistency score, multi-simulator effectiveness score and average risk score, as shown in Equation (19): (19) In the formula, A collection of trustworthy high-risk scenarios; For the first One autonomous driving test scenario; For the first Consistency scoring of multiple simulators in a test scenario; The lower limit threshold for multi-simulator consistency scoring; For the first Effectiveness scoring of multiple simulators in a single test scenario; The lower limit threshold for multi-simulator effectiveness scoring; For the first Average risk score of multiple simulators in a test scenario; This is the lower limit threshold of the average risk score that must be met for a credible high-risk scenario.

6. The method for accelerating the generation of a reliable test scenario based on multi-simulator consistency verification according to claim 1, characterized in that, The specific method of S5 is as follows: S51. Construct trusted scene sampling weights; In the During the accelerated generation process, sampling weights are assigned to credible high-risk scenarios based on the credible risk-reward index, as shown in equation (20): (20) In the formula, For the first The first round of accelerated generation process Sampling weights for each credible high-risk scenario; To expedite the generation of round numbers; For the first Wheel of Life Credible risk-reward metrics for each test scenario; This is a stable term for sampling weights, used to prevent the weights from being uncalculated when the credible risk-return indicator is zero; For the first A set of trustworthy high-risk scenarios that can be obtained in turn; The first in the set of trustworthy high-risk scenarios One test scenario; For the first Wheel of Life Credible risk-reward metrics for each test scenario; The scenario number is located within the set of trusted high-risk scenarios. S52, Update the trusted accelerated generation distribution; The previous round of scene generation distribution is fused with the kernel density distribution guided by the trusted high-risk scene to obtain the next round of trusted accelerated generation distribution, as shown in Equation (21): (21) In the formula, For the first +1 round of trusted accelerated distribution generation; This is the parameter vector for the test scenario to be generated; The step size coefficient is updated for the distribution, and its value ranges from 0 to 1. For the first Wheel scene generation distribution; For the first One credible high-risk scenario; For the first Wheel of Life Sampling weights for each credible high-risk scenario; The bandwidth matrix is Kernel function; This is the kernel function bandwidth matrix, used to control the range of expansion of newly generated parameters around the parameters of the credible high-risk scenario; For the first A parameter vector for a credible high-risk scenario; S53. Generate a new test scenario that satisfies the parameter constraints; According to the After obtaining candidate parameters through +1 rounds of trusted accelerated generation distribution sampling, the candidate parameters are constrained and projected to ensure that the generated scene meets the constraints of road structure, traffic rules, vehicle dynamics, and environmental value range, as shown in Equation (22): (22) In the formula, For the first +1 round generated the first A test scenario parameter vector that satisfies the constraints; For the first +1 round of candidate scene number; This is a parameter-constrained projection operator used to project candidate parameters that do not meet the constraints to the feasible parameter space. The feasible parameter space represents the set of parameters that meet the requirements of road structure, traffic rules, vehicle dynamics, and environmental values. For the first +1 rounds of sampling from the trusted accelerated generation distribution. A vector of candidate parameters; symbol This indicates that the probability follows a certain probability distribution; For the first +1 round of trusted accelerated distribution generation.

7. The method for accelerating the generation of a reliable test scenario based on multi-simulator consistency verification according to claim 1, characterized in that, The specific method of S6 is as follows: S61. Construct the reweighting coefficients between the natural distribution and the accelerated generation distribution; To enable the accelerated scenario generation to be used for failure risk estimation of autonomous driving systems, the calculation of the first... Wheel of Life The reweighting coefficients for each test scenario are shown in equation (23): (23) In the formula, For the first Wheel of Life The reweighting coefficients for each test scenario; For the first The probability density of a test scenario parameter vector under the natural scene distribution; For the first A parameter vector for each test scenario; For the first The test scenario parameter vector is at the... The probability density under the accelerated generation distribution is given by equation (21). of Values The value at time; This is a weighted stabilizing term used to avoid the denominator being zero; S62. Construct a reliable failure indicator; By combining the multi-simulator consistency score, the multi-simulator effectiveness score, and the average failure state, a reliable failure indicator is obtained, as shown in equation (24): (24) In the formula, For the first The number of reliable failure indicators for each test scenario; For the first Consistency scoring of multiple simulators in a test scenario; For the first Effectiveness scoring of multiple simulators in a single test scenario; For the first The larger the credible failure indication quantity of the multi-simulator average failure state of a test scenario, the more likely the scenario is to be a credible test sample that stably triggers the failure of the autonomous driving system in multiple simulators. S63. Estimate the risk of reliable failure in autonomous driving systems; Based on the reweighting coefficients and the credible failure indicator, the credible failure risk of the autonomous driving system under natural scene distribution is estimated, as shown in Equation (25): (25) In the formula, For the first The reliable failure risk of the autonomous driving system obtained by round estimation; For the first The number of test scenarios involved in risk estimation; Number the test scenario; For the first Wheel of Life The reweighting coefficients for each test scenario; For the first The number of reliable failure indicators for each test scenario; S64. Calculate and accelerate the efficiency of effective test samples; To determine whether the accelerated test results are stable, the effective sample efficiency of the g-th round of accelerated testing is calculated, as shown in equation (26): (26) In the formula, For the first Wheel acceleration test of effective sample efficiency; For the first The number of test scenarios involved in risk estimation; Number the test scenario; For the first Wheel of Life The reweighting coefficients for each test scenario indicate that the higher the efficiency of the effective samples, the more balanced the weights of the accelerated test samples and the more stable the failure risk estimation. S65. Construct accelerated generation termination conditions; When the credible failure risk estimation results tend to stabilize and the efficiency of the effective sample meets the requirements, the credible accelerated generation process is terminated, as shown in equation (27): (27) In the formula, For the first The reliable failure risk of the autonomous driving system obtained by round estimation; For the first -1 round of estimation yields the reliable failure risk of the autonomous driving system; This is for absolute value operations; To estimate the convergence threshold for failure risk; For logical AND operation; For the first Wheel acceleration test of effective sample efficiency; The effective sample efficiency lower limit threshold; When equation (27) holds, it means that the credible failure risk estimation has stabilized and the sample weight distribution meets the requirements, and it is possible to stop generating new test scenarios.

Images