A Simulation Test Acceleration Method for Estimating Mileage Accident Rate of Autonomous Vehicles

Abstract

This application discloses a simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles, belonging to the field of autonomous driving technology. The method includes: acquiring natural driving data and constructing a probabilistic representation model of the time-series driving process, the probabilistic representation model including the natural driving distribution; constructing a deep generative model based on standardized flow, learning the dangerous driving distribution through the natural driving distribution; conducting simulation tests using the components of the dangerous driving distribution as importance sampling distributions until a dangerous event occurs or a preset duration is reached, and outputting the test results; calculating the mileage accident rate estimate by combining the natural driving distribution and the dangerous driving distribution, and repeating the simulation test until the relative half-width of the estimate converges to a preset threshold. This application can handle high-dimensional time-series driving distributions and achieve unbiased accelerated estimation of the mileage accident rate.

Description

Technical Field

[0001] This application belongs to the field of autonomous driving technology, and in particular relates to a simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles. Background Technology

[0002] Current estimates of autonomous driving accident rates face three main challenges: The estimation of autonomous driving accident rates is limited by the characteristics of "rare events". Traditional methods based on the accumulation of natural driving mileage rely on massive samples and have a long testing cycle. There is an urgent need to introduce accelerated testing methods to improve efficiency.

[0003] Although existing simulation scenario acceleration testing methods are relatively mature, the lack of a unified definition of the spatiotemporal span of scenarios leads to inconsistencies between the units of "scenario accident rate" and the commonly used "mileage accident rate" and makes it difficult to benchmark. There is a lack of acceleration evaluation methods based on a unified mileage benchmark.

[0004] Traditional acceleration methods struggle to cope with the dynamic complexity of high-dimensional time-series driving processes and cannot accurately calculate the probability density of generated samples, making it difficult to achieve an equivalent unbiased estimate of mileage accident rate. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a simulation test acceleration method for mileage accident rate estimation of autonomous vehicles. Based on the processing of high-dimensional time series features, a bidirectional mapping between dangerous driving distribution and natural driving distribution is established to obtain accurate statistical weights. This method can handle high-dimensional time series driving distribution and achieve unbiased accelerated estimation of mileage accident rate.

[0006] The objective of this application is achieved through the following technical solution: A simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles, the method comprising: Acquire natural driving data and construct a probabilistic representation model of the time-series driving process, wherein the probabilistic representation model includes the natural driving distribution; A deep generative model based on normalized flow is constructed, and the dangerous driving distribution is learned through the natural driving distribution; The dangerous driving distribution components are used as importance sampling distributions for simulation testing until a dangerous event occurs or a preset time is reached, and the test results are output. The mileage accident rate estimate is calculated by combining the natural driving distribution and the dangerous driving distribution. The simulation test is repeated until the relative half-width of the estimate converges to a preset threshold.

[0007] Furthermore, the probabilistic representation model for constructing the temporal driving process includes: Based on the Markov assumption, the driving process is decomposed into conditional probabilistics, which decomposes the high-dimensional temporal natural driving process into a low-dimensional natural driving state-behavior distribution, thus obtaining a probabilistic representation model of the temporal driving process.

[0008] Furthermore, the dangerous driving distribution learned through the natural driving distribution includes: A hazard index function is introduced into the loss function of the deep generative model, and the dangerous driving distribution is obtained by targeted learning from natural driving distribution data without the need for a warm start.

[0009] Furthermore, targeted learning yields dangerous driving distributions including: The training set data is sampled from the natural driving distribution. The loss function is transformed by the correlation between the natural driving distribution and the dangerous driving distribution, which guides the deep generative model to learn the probability density of the dangerous driving distribution.

[0010] Furthermore, the simulation test using the dangerous driving distribution components as an importance sampling distribution includes: The rejection-acceptance sampling principle is adopted to sample background vehicle behavior. The correlation between background vehicle behavior variables and state variables is expressed by the probability of rejection or acceptance sampling, so as to achieve targeted evolution of the time-series driving process.

[0011] Furthermore, the calculation of the mileage accident rate estimate by combining the natural driving distribution and the dangerous driving distribution includes: In the process of calculating the expected mileage accident rate, an importance sampling probability density function is introduced. By sampling in the importance sampling probability density function, an estimated value expression for the mileage accident rate is obtained. The relative half-width is used to determine whether the estimated value of the mileage accident rate converges.

[0012] Furthermore, the deep generative model is constructed based on an explicit probability density transformation architecture, including multiple cascaded invertible affine coupling layers, which are configured to establish a bijective mapping relationship between the initial latent variable distribution and the target data distribution.

[0013] Furthermore, the method also includes: By using importance sampling and mileage accident rates obtained based on dangerous driving distributions, and by applying probability measure changes through importance weights, an equivalent unbiased estimate of the mileage accident rate is obtained.

[0014] The beneficial effects of this application are as follows: This application makes Markov assumptions about the driving process and performs conditional probability decomposition, decomposing the high-dimensional temporal natural driving distribution into a low-dimensional continuous distribution, describing the driving state-behavior probability relationship, which is easier to learn and sample.

[0015] This application modifies the loss function of the standardized flow depth generation model using a hazard index function, enabling it to achieve targeted learning of dangerous driving distributions without relying on a warm start.

[0016] This application utilizes a combination of series distributions to prioritize sampling the dangerous behaviors of background vehicles through the rejection-acceptance sampling principle, guiding the gradual generation of dangerous events such as collisions during testing. It derives an accelerated mileage accident rate estimation statistic using importance sampling theory, ensuring its equivalence and unbiasedness while accelerating the mileage accident rate estimation. Attached Figure Description

[0017] Figure 1 This is a flowchart of a simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles, as proposed in this application. Figure 2 This is a flowchart illustrating the application of the accelerated testing method according to an embodiment of this application; Figure 3 This is a comparison diagram of natural driving distribution and dangerous driving distribution in an embodiment of this application; Figure 4 This is a graph showing the acceleration estimation results of the collision accident rate in an embodiment of this application. Detailed Implementation

[0018] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0019] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Traditional acceleration methods struggle to cope with the dynamic complexity of high-dimensional time-series driving processes and cannot accurately calculate the probability density of generated samples, making it difficult to achieve an equivalent unbiased estimate of mileage accident rate.

[0021] To address the aforementioned technical problems, the present application proposes the following embodiments of a simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles.

[0022] Reference Figure 1 ,like Figure 1 The diagram shown is a flowchart of a simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles according to this embodiment. The specific steps of the method are as follows: Step S1: Probabilistic Representation and Decomposition of High-Dimensional Temporal Driving Process. Acquire natural driving data, perform conditional probability decomposition of the driving process based on the Markov assumption, decompose the high-dimensional temporal natural driving process into low-dimensional natural driving state-behavior distribution components, and construct a probabilistic representation model of the temporal driving process; Step S2: Construction of a deep generative model for dangerous driving distribution. A deep generative model based on normalized flow is constructed. A hazard index function is defined and incorporated into the model's loss function. Dangerous driving distribution components are obtained through targeted learning from natural driving distribution component data without requiring a warm start. Step S3: Evolution and sampling of the time-series driving process. During the simulation test, the dangerous driving distribution component is used as the importance sampling distribution; the rejection-acceptance sampling method is adopted to sample the background vehicle driving behavior of the next time step from the dangerous driving distribution component according to the vehicle state of the current time step, and dynamically update the state of all vehicles until a dangerous event occurs or the preset duration is reached, thus completing one test and outputting the test results; Step S4: Accelerated estimation of hazard rate based on importance sampling. Establish a statistical model for mileage-based accident rate and derive a weighted estimate of the mileage-based accident rate based on importance sampling theory; Step S5: Mileage Accident Rate Estimation. Calculate the mileage accident rate estimate by combining the natural driving distribution and the dangerous driving distribution, determine whether the relative half-width meets the confidence requirement, and repeat step S3 until the relative half-width of the estimate converges to the preset threshold.

[0023] Joint probability density function of natural driving process for: (1) Based on the Markov property assumption and conditional probability, it can also be written as: (2) in, Represents the state variables of the tested autonomous vehicle and the background vehicle at time t; The variable representing the driving behavior of the background vehicle at time t; Let be the probability density of the states of all vehicles at the initial moment; Let be the joint probability density of the state variables and the background vehicle driving behavior variables at time t; and Natural driving data and Gaussian mixture models can be used for fitting.

[0024] The mapping method for each affine coupling layer in the normalized flow model is as follows: (3) (4) study and The loss functions are defined as follows: (5) (6)

[0025] in, ; These are latent variables of S, and the corresponding Jacobian matrix reflects the relationship between S and... The mapping relationship; yes The latent variables, and the corresponding Jacobian matrix, represent and The mapping relationship.

[0026] The expected value of the mileage-based accident rate is expressed as: (7)

[0027] The equivalent unbiased estimate of the mileage-based accident rate is written as: (8)

[0028] in, The first one generated in the simulation test One test case (containing all vehicle states and behavior sequences from start to finish); Judge the first Each test case indicates whether an accident, collision / danger occurred, is assigned a value of 1; if safe, the value is 0. Indicates the first The mileage traveled by the tested autonomous vehicle in each test case; This represents the probability density of the natural driving distribution of the test cases; This represents the probability density of test cases in the distribution of dangerous scenarios; This indicates the importance weight.

[0029] The relative half-width is: (9)

[0030] As one implementation method, this embodiment takes the following process of an autonomous vehicle in a highway environment as an example, and applies the simulation mileage test acceleration method for autonomous vehicle accident rate estimation proposed in this embodiment. The core idea is as follows: model the driving process based on Markov decision process, and fit the natural driving distribution components from natural driving data; shift the distribution towards the dangerous driving distribution area through normalization flow to generate a dangerous driving test environment; combine importance sampling to ensure that the evaluation results are unbiased to the natural driving environment; finally, generate the dangerous behavior of background vehicles through acceptance-rejection sampling, calculate the mileage accident rate estimate, and complete the accelerated evaluation of the accident rate of autonomous vehicles. The steps include: (I) Probabilistic Representation and Decomposition of High-Dimensional Temporal Driving Processes: In this embodiment, a probabilistic representation model of the driving process is first established. For a natural driving time-series test case with a time series length of T, it is defined as a vehicle state variable. driving behavior of the background vehicle Serial sequence. Considering the high-dimensionality of the continuous driving process and the temporal dependencies between variables, the joint probability density function is analyzed based on the Markov assumption. Perform conditional probability decomposition: (10)

[0031] in, Let represent the probability density function of all vehicle state variables at time t; Let represent the joint probability density function of all vehicle state variables and background vehicle driving behavior variables at time t. Thus, the high-dimensional time-series joint probability density function is obtained by multiplying a series of probability densities. .

[0032] This embodiment will , In short, it is a component of the natural driving distribution. When fitting the distribution, known data at all time points can be used; only when solving for the probability is it necessary to multiply by time t.

[0033] make Dangerous accident The indicator function represents the sample Is it a dangerous test case? It can be represented as: (11)

[0034] According to importance sampling theory, the speedup depends on the variance of the estimator. Minimizing the number of tests (i.e., maximizing the speedup ratio) essentially involves solving for an optimal sampling distribution. This minimizes the variance under this distribution.

[0035] By solving the functional extremum problem of minimizing variance, the theoretically optimal importance sampling probability density function can be obtained. Must meet: (12)

[0036] In the formula, c is the normalization constant. According to its practical meaning, the optimal generation distribution should only occur when... There is a probability value at this point, and this probability value is related to the probability of natural driving. Proportional.

[0037] (II) Construction of a deep generative model for the distribution of dangerous driving: To accelerate testing, a hazardous driving distribution needs to be constructed to increase the sampling frequency of hazardous events. This embodiment uses a normalized flow model to learn the components of the hazardous driving distribution. , .

[0038] Importance sampling probability density function construction: because The rarity, It will exhibit a very sparse, dispersed discrete form, therefore It lacks learnability. Therefore, it needs to be replaced. By introducing a new hazard indicator function, a continuous and easily learnable importance sampling probability density function is obtained. The new hazard indicator function needs to possess the following characteristics: ① The value range is between [0,1], and the value increases monotonically as the risk of the test case increases; ② The closer the function value is to 1, the larger the absolute value of the derivative; ③ It can be decomposed in time, that is, the corresponding value can be obtained at any time. Based on the above constraints, we introduce replace , and The formal similarity is extremely high and the above conditions are satisfied. Therefore, we obtain: (13) In the formula, c is the normalization constant. To align with the natural driving distribution and facilitate learning, the following form can be obtained: (14) (15) (16) As a component of dangerous driving distribution, The probability density function representing the driving state. This represents the joint probability density function of the driving state variable and the background vehicle driving behavior variable. Although it is difficult to obtain the theoretically optimal importance sampling probability density function, the correlation form between the natural driving distribution and the dangerous driving distribution, as well as the approximate optimal expression of the dangerous driving distribution, can be analytically derived from the above formula.

[0039] Design of loss function for standardized flow model: This embodiment uses a standardized flow model to learn the state distribution components separately. and background vehicle driving behavior distribution components Flow model for latent variables Apply a series of invertible transformations F to map them into latent variables. .in, . It is a known simple basic distribution. It is a complex distribution of dangerous driving characteristics that needs to be learned.

[0040] Can be disassembled into and Therefore, they will be studied separately. and .by For example, let ,in , (17) (18) To learn the dangerous driving distribution directly from natural driving data without prior warm-start training, this embodiment reconstructs the model's loss function by utilizing the functional relationship between the natural driving distribution and the dangerous driving distribution. In the learning of the flow model, the core objective is to transform the simple latent variable distribution into the target distribution through a series of invertible mappings. The learning objective of the standardized flow model is defined in this section as minimizing the target distribution. Learning distribution with the model KL divergence between: (19) Skipping the warm start approach involves using training set data from the natural driving distribution. Mid-sampling is used, and the loss function is transformed by the correlation between the natural driving distribution and the dangerous driving distribution. , and It is a constant independent of model training, therefore learning The loss function is written as follows: (20) Similarly, learning The loss function is written as: (twenty one) in, ; It is a latent variable of s, and the corresponding Jacobian matrix reflects the relationship between s and... The mapping relationship; yes The latent variables, and the corresponding Jacobian matrix, represent and The mapping relationship.

[0041] (III) Evolution and Sampling of the Temporal Driving Process: Natural driving distribution components and The distribution components of dangerous driving can be fitted by a Gaussian mixture model. and These four distribution components can be learned from a standardized flow model. They are all explicitly expressed and can be directly sampled and their probability densities calculated.

[0042] The key to targeted evolution in the temporal driving process lies in the behavior of the background vehicle. This is because the probability density function of the background vehicle's natural driving behavior cannot be directly obtained. Or the probability density function of dangerous driving behavior in the background vehicle. In this embodiment, the background vehicle behavior sampling is performed using the rejection-acceptance sampling principle. The strength of the correlation between the background vehicle behavior variables and the state variables is implicitly expressed by the probability of rejection or acceptance of sampling, indirectly reflecting the dependency between the background vehicle behavior and the relative speed, relative distance, and other states of each vehicle. Taking the evolution of dangerous behavior as an example, the specific steps of the sampling process are as follows: exist Initial state of random sampling test cases .

[0043] Candidates for the next time step are sampled from a uniform distribution. : (twenty two) These represent the minimum and maximum values ​​of the background vehicle driving behavior variable in the natural driving distribution, respectively.

[0044] Sampling candidates in a uniform distribution Candidate probability : (twenty three) According to the rejection-acceptance sampling principle, candidates that satisfy the inequality are selected for acceptance. As a background vehicle driving behavior variable : (twenty four) calculate : (25) Repeat step two until the current time. If a dangerous event occurs, the test is stopped.

[0045] To compare the risk rate estimation with the Monte Carlo method based on natural driving distribution, the comparative experiment needs to generate test cases that follow the natural driving distribution. Similarly, the background vehicle driving behavior variables can be obtained. The following conditions must be met: (26) In the test, to avoid extreme values ​​in the rate of change of acceleration, candidate samples were sampled in step (2) starting from the second time step. At that time, it will accelerate in the previous step. Centered on, in Uniform sampling within the range, The acceptable rate of change of acceleration.

[0046] (iv) Implementation of the hazard rate acceleration estimation method based on importance sampling: Based on the above derivation process, the flow of test case evolution generation and accelerated testing methods is as follows: Figure 2 As shown, the model used for distribution generation is trained before testing, and the sampling of dangerous behaviors of background vehicles is carried out progressively during testing.

[0047] The natural driving distribution was fitted from natural driving data, while the dangerous driving distribution was learned through a standardized flow depth generation model. In the acceleration test, the dangerous driving distribution component... The first frame's state is obtained through sampling, including the initial speed and position of each vehicle. Subsequently, following the rejection-acceptance sampling principle and process described in this embodiment, the background vehicle behavior for the next frame is obtained, and the states of both the background vehicle and the tested autonomous vehicle are updated. This process of "background vehicle action sampling – all vehicle state update" is then repeated until a dangerous event occurs or the time limit is reached. This embodiment uses the collision accident rate as a statistical objective; if a collision event occurs, the probability of that test case in the natural driving distribution and the probability in the dangerous driving distribution are calculated.

[0048] (V) Construction and Implementation of Mileage Accident Rate Estimation Framework: Mileage accident rate is written as: (27) In the above formula, the denominator is the cumulative simulation test mileage across n test cases. Equation (27) is an estimate of the mileage failure rate, and the expected mileage failure rate can be expressed as: (28) In existing technologies, although the "accident rate based on number of times" (accident scenario / number of tests) and the "accident rate based on mileage" (accident scenario / test mileage) are theoretically only related by a conversion of measurement benchmarks, there are significant technical obstacles to this conversion in the engineering practice of accelerated testing. The fundamental reason is that traditional scenario representation methods are mainly designed for discrete, specific functional scenarios, lacking universality and making it difficult to establish an effective mapping relationship with mileage tests that represent continuous driving processes. This dimensional decoupling leads to significant deviations in the estimation of statistics during accelerated testing, making the accurate derivation of mileage accident rates based on scenario testing a technical bottleneck in this field. Due to the change in statistical units, it is necessary to re-derive the statistics and the relative half-width of the statistics. By applying the idea of ​​distribution measure change to equation (28), an importance sampling probability density function is introduced. ,get: (29) Therefore, through The equivalent unbiased estimate of the mileage accident rate for medium sampling can be written as: (30) To judge Whether it converges or not, relatively half-width writing: (31) In one embodiment, the threshold for the relative half-width is set to 0.2.

[0049] (vi) Training and optimization of deep generative models: In one embodiment, the normalized flow model is built on an architecture based on a real-valued non-volume preserving transformation. This model comprises multiple cascaded affine coupling layers configured to establish a bijective mapping from the initial latent variable distribution to the target data distribution.

[0050] Specifically, the affine coupling layer is configured to simplify the calculation of the Jacobian determinant in probability density estimation while maintaining computational invertibility, specifically by avoiding the computation of the complete Jacobian matrix. For example, each affine coupling layer in Real NVP needs to be learned through a neural network. and The mapping method is as follows: (32) (33) In practice, a binary swap is used, that is, for an odd number of mappings, the first... Keeping the dimensional data unchanged, transforming the remaining dimensional data, and swapping the relationships before and after an even number of mappings, is equivalent to... The dimensional data remains unchanged. This operation allows Real NVP to efficiently generate complex distributions of high-dimensional data while maintaining computational reversibility.

[0051] A comparison of the dangerous driving distribution generated using Real NVP and the natural driving distribution is shown in the figure. Figure 3 .

[0052] (vii) System Implementation and Testing Verification: The experimental section uses the car-following process as an example. This is because the car-following process has a lower probability of collision than other complex driving processes; therefore, verifying the acceleration test effect during car-following can intuitively verify the acceleration advantage of the proposed method. In this experiment, the autonomous vehicle's control algorithm is characterized by an Intelligent Driver Model (IDM). An IDM is a traffic flow model used to simulate vehicle car-following behavior, designed to describe how a following vehicle adjusts its speed based on the position and speed of the preceding vehicle to maintain a safe distance and avoid collisions.

[0053] The acceleration effect of using acceleration testing methods to estimate the accident rate over following distance is shown in the figure. Figure 4 If sampled from the natural driving distribution, the estimated mileage accident rate converges to 4.44 × 10⁻⁶. -3 The relative half-width reached the required threshold after 9663.5 km of testing. Applying the acceleration testing method proposed in this embodiment, the relative half-width reached the required threshold after 262.4 km of testing, and the mileage accident rate converged to 4.73 × 10⁻⁶. -3 This indicates that compared to testing from a natural driving distribution, the accelerated testing method is 36.8 times faster in estimating the mileage accident rate, reducing the test mileage by 9401 kilometers. Assuming an average speed of 25 mph (approximately 40.23 km / h) for a test vehicle, based on this assumption, if we assume the test vehicle is driven continuously around the clock, the reduction of 9401 kilometers in test mileage is equivalent to saving 234 hours of testing time; if we assume 8 hours of testing per day, the reduction of 9401 kilometers in test mileage is equivalent to saving 30 days of testing time.

[0054] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles, characterized in that, The method includes: Acquire natural driving data and construct a probabilistic representation model of the time-series driving process, wherein the probabilistic representation model includes the natural driving distribution; A deep generative model based on normalized flow is constructed, and the dangerous driving distribution is learned from the natural driving distribution. The dangerous driving distribution components are used as importance sampling distributions for simulation testing until a dangerous event occurs or a preset time is reached, and the test results are output. The mileage accident rate estimate is calculated by combining the natural driving distribution and the dangerous driving distribution. The simulation test is repeated until the relative half-width of the estimate converges to a preset threshold.

2. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 1, characterized in that, The probabilistic representation model for constructing the time-series driving process includes: Based on the Markov assumption, the driving process is decomposed into conditional probabilistics, which decomposes the high-dimensional temporal natural driving process into a low-dimensional natural driving state-behavior distribution, thus obtaining a probabilistic representation model of the temporal driving process.

3. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 1, characterized in that, The dangerous driving distribution learned through the natural driving distribution includes: A hazard index function is introduced into the loss function of the deep generative model, and the dangerous driving distribution is obtained by targeted learning from natural driving distribution data without the need for a warm start.

4. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 3, characterized in that, The dangerous driving distribution obtained through targeted learning of natural driving distribution data includes: The training set data is sampled from the natural driving distribution. The loss function is transformed by the correlation between the natural driving distribution and the dangerous driving distribution, which guides the deep generative model to learn the probability density of the dangerous driving distribution.

5. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 1, characterized in that, The simulation test using the dangerous driving distribution components as importance sampling distributions includes: The rejection-acceptance sampling principle is adopted to sample background vehicle behavior. The correlation between background vehicle behavior variables and state variables is expressed by the probability of rejection or acceptance sampling, so as to achieve targeted evolution of the time-series driving process.

6. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 1, characterized in that, The estimated mileage accident rate calculated by combining natural driving distribution and dangerous driving distribution includes: In the process of calculating the expected mileage accident rate, an importance sampling probability density function is introduced. By sampling in the importance sampling probability density function, an estimated value expression for the mileage accident rate is obtained. The relative half-width is used to determine whether the estimated value of the mileage accident rate converges.

7. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 1, characterized in that, The deep generative model is built on an explicit probability density transformation architecture, which includes multiple cascaded invertible affine coupling layers and is configured to establish a bijective mapping relationship between the initial latent variable distribution and the target data distribution.

8. The simulation test acceleration method for estimating the mileage accident rate of autonomous vehicles as described in claim 1, characterized in that, The method further includes: By using importance sampling and mileage accident rates obtained based on dangerous driving distributions, and by applying probability measure changes through importance weights, an equivalent unbiased estimate of the mileage accident rate is obtained.

Images