Automobile data intelligent processing method

Abstract

The application discloses an intelligent processing method for automobile data and relates to the technical field of automatic driving.The method comprises the following steps: S1: constructing a simulation model: a dynamic parallel simulation model is constructed by using multi-source data obtained by an automatic driving vehicle fleet; S2: determining a test case: at least one adversarial test agent is set in the dynamic parallel simulation model, interactive testing is performed by using the adversarial test agent, and an effective test case is determined; S3: case verification: according to the test result of the effective test case, a corresponding comprehensive score is obtained, the effective test case is screened according to the comprehensive score, and the screened effective test case is verified in a real scene. Thus, a simulation-real scene closed-loop iterative optimization is formed, and the performance of an automatic driving system can be continuously improved, and the safe application of the automatic driving system is accelerated.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a method for intelligent processing of vehicle data. Background Technology

[0002] With the rapid development of the automotive industry and the continuous advancement of intelligent transportation systems, modern automobiles have gradually evolved into mobile intelligent terminals integrating perception, computing, communication, and control. The widespread application of onboard sensors (such as radar, cameras, lidar, GPS, and IMU), in-vehicle infotainment systems, and vehicle-to-everything (V2X) devices enables vehicles to collect massive amounts of multi-source heterogeneous data in real time during operation. This includes vehicle status data (such as speed, RPM, fuel consumption, and battery status), environmental perception data (such as road conditions, pedestrian recognition, and traffic sign recognition), driving behavior data, and location and trajectory information.

[0003] This data not only provides fundamental support for advanced functions such as autonomous driving, smart cockpits, remote diagnostics, and fleet management, but also creates important conditions for automakers to optimize product design, improve service quality, and conduct big data analysis and artificial intelligence modeling. However, faced with the ever-increasing scale of data, the diverse types of data, and the high requirements for real-time performance, accuracy, and security, traditional automotive data processing methods are no longer sufficient to meet practical application needs.

[0004] Chinese invention patent application CN117666785A discloses a training method and system for human-machine interaction takeover in autonomous driving based on digital twins. First, multiple static 2D images of the object to be modeled are acquired from multiple angles using a camera, and the camera's intrinsic and extrinsic parameters for each image are estimated using tools such as COLMAP. Then, segmentation cues are provided for the target object in each image, and the image is segmented using the SAM large model to remove background noise, resulting in a segmented image of the target object. Next, the segmented image and camera parameters are input into the NeRF neural radiation field algorithm for 3D reconstruction and rendering. Point cloud reconstruction technology is used to convert the implicit 3D model into a visible 3D model, thus obtaining a high-precision digital twin geometric model. This method effectively utilizes the powerful semantic understanding capabilities of the SAM large model and the efficient 3D reconstruction capabilities of the NeRF neural radiation field algorithm to achieve automation and accuracy in digital twin modeling.

[0005] While the aforementioned technical solutions can automate and improve the accuracy of digital twin modeling, the construction of their virtual environments relies on data collection and reconstruction of real-world scenarios. However, autonomous driving, due to the high cost and risks of real-world testing, is mostly tested and validated in simulation environments, making real-world data collection difficult. Furthermore, the inherent gap between simulation environments and the real world makes it difficult to directly transfer autonomous driving algorithms trained and validated in simulations to the real world. Once deployed in a real environment, performance degradation or even decision-making failures may occur due to subtle, unperceived differences or long-tail scenarios, potentially constituting a core technical obstacle to large-scale application. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent processing method for automotive data to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent processing of automotive data, comprising:

[0008] S1: Constructing a simulation model: Constructing a dynamic parallel simulation model using multi-source data obtained from an autonomous driving fleet;

[0009] S2: Determine test cases: Set up at least one adversarial test agent in the dynamic parallel simulation model, and conduct interactive tests through the adversarial test agent to determine valid test cases, including:

[0010] S2.1: Set up the intelligent agent: Generate at least one controllable vehicle model through simulation software, and import the controllable vehicle model into a dynamic parallel simulation model for adversarial training to obtain an adversarial test intelligent agent;

[0011] S2.2: Monitoring Records: The autonomous driving algorithm to be tested is set up on the cloud platform of the dynamic parallel simulation model through the simulation task scheduling cluster, and the adversarial test agent is injected into the dynamic parallel simulation model as background traffic flow for interactive testing. At the same time, the test results of each test case are obtained through the set failure criteria.

[0012] S3: Test Case Verification: Based on the test results of the valid test cases, obtain the corresponding comprehensive score, and filter the valid test cases according to the comprehensive score, and then perform real-world verification on the filtered valid test cases.

[0013] Furthermore, a dynamic parallel simulation model is constructed, including:

[0014] S1.1: Data Acquisition: Fixed nodes are set on the driving path of the intelligent vehicle, and mobile nodes are set on the intelligent vehicle. Through the fixed nodes and mobile nodes, multi-source data during the driving process of the intelligent vehicle is collected and acquired. The multi-source data includes perception data, dynamic data, pose data and scene data.

[0015] S1.2: Environment Reconstruction: The perceived data is processed by data association and alignment to obtain fused perceived data. The fused perceived data is then fused with the neural representation model to construct an implicit 3D model. At the same time, the dynamic data, pose data, and scene data are injected into the implicit 3D model to construct a dynamic parallel simulation model.

[0016] Furthermore, adversarial test agents are obtained, including:

[0017] S2.1.1: Set adversarial targets: By setting a reward function, monitor the behavior of the test agents in the dynamic parallel simulation model and obtain the total reward score of each test agent's operation;

[0018] S2.1.2: Agent Training: In the dynamic parallel simulation model, the corresponding reconstructed scenarios are loaded according to multi-source data, and in each reconstructed scenario, the main vehicle and the test agent are simulated and deduced. At the same time, the test agent is trained and tested through a deep reinforcement learning algorithm to obtain an adversarial test agent.

[0019] Furthermore, through the simulation engine in the dynamic parallel simulation model, the core reward item scores and constraint penalty item scores are obtained during the simulation and deduction process of the test agent. The total operation reward score is determined through the core reward item scores and constraint penalty item scores. The core reward item scores include the score for maximizing the uncertainty of the master vehicle, the score for minimizing the safety margin of the master vehicle, and the score for inducing control failure. The constraint penalty item scores include the collision penalty score and the traffic rule penalty score.

[0020] Furthermore, the total reward score for each test agent's actions is obtained, including:

[0021] S2.1.1.1: Trajectory monitoring: Based on the simulated trajectory of the test agent, the prediction module of the main vehicle sets the probability value of each simulated trajectory, and determines the change in trajectory entropy value based on the probability value of the simulated trajectory.

[0022] S2.1.1.2: Collision Detection: The collision time between the main vehicle and the test intelligent agent is determined through the main vehicle's safety core module, and the change in the risk factor is determined based on the collision time;

[0023] S2.1.1.3: Braking Monitoring: The braking response corresponding to the main vehicle's planning control module is compared with the set comfort threshold, and the planning control failure value is determined based on the comparison result. Specifically:

[0024] When the deceleration of the braking response is greater than the comfort threshold, the control failure value is the preset control failure value; otherwise, the control failure value is 0.

[0025] S2.1.1.4: Consequence monitoring: The simulation engine is used to detect the behavior of the main vehicle and the test agent after the simulation is completed, and collision penalty scores and traffic rule penalty scores are set according to the detection results.

[0026] S2.1.1.5: Determine the total reward score: Combine the changes in trajectory entropy, risk factor, regulatory failure, collision penalty score, and traffic rule penalty score to determine the total reward score.

[0027] Furthermore, when the host vehicle and / or the test agent experience a physical collision, the collision penalty score is the preset collision penalty threshold; otherwise, the collision penalty score is 0.

[0028] When the host vehicle and / or the test agent violates traffic rules, the traffic rule penalty score is the preset traffic rule penalty threshold; otherwise, the traffic rule penalty score is 0.

[0029] Furthermore, by using the failure criteria, the recorded result at the moment of the main vehicle failure is determined, and the recorded result is matched with the vulnerability type, severity level, and scenario label to obtain the vulnerability report of the adversarial testing agent.

[0030] Furthermore, the test results of the effective test cases are mapped to a risk matrix. The comprehensive score of each effective test case is determined through the risk matrix. Based on the comprehensive score, the effective test cases are sorted in descending order. At the same time, a preset number of effective test cases are selected by sorting in descending order according to a preset quantity requirement.

[0031] Furthermore, the risk matrix is ​​constructed based on the probability and severity of the test cases, with both the probability score and severity score set to 1-5 points.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] Firstly, this invention constructs a simulation environment by collecting multi-source data from real vehicle fleets and roadside equipment, and performs high-precision environment reconstruction through a neural representation model, making the generated dynamic parallel simulation model closer to the real world. Then, high-risk use cases discovered in the simulation are verified through real-world scenarios, thereby forming a closed-loop iterative optimization of simulation and real-world scenarios. This can continuously promote the improvement of the performance of autonomous driving systems and accelerate the large-scale safety application.

[0034] Secondly, this invention constructs a dynamic parallel simulation model using multi-source data collected from autonomous driving fleets, and uses an adversarial testing agent to conduct automated testing on the dynamic parallel simulation model. This not only shortens the testing cycle but also avoids the high costs and safety hazards of real-world testing.

[0035] Thirdly, this invention uses a risk matrix to quantify and score the probability and severity of each test case, thereby identifying the test cases with the highest risk and the most urgent need for resolution, i.e., high-risk test cases, and conducting real-world verification, which can systematically improve the safety and reliability of autonomous driving algorithms. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the intelligent vehicle data processing method of the present invention;

[0037] Figure 2 This is a schematic diagram of the process for obtaining the adversarial test agent in this invention;

[0038] Figure 3 This is a risk score distribution diagram for test cases in this invention;

[0039] Figure 4 This is a schematic diagram of the risk matrix in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Because real-world testing of autonomous driving is costly and risky, it is mostly conducted and verified through simulation environments, making real-world data acquisition difficult. Furthermore, the inherent gap between simulation and the real world makes it difficult to directly transfer autonomous driving algorithms trained and verified in simulations to the real world. Once deployed in a real environment, performance degradation or even decision failures may occur due to subtle, unperceived differences or long-tail scenarios, potentially becoming a core technical obstacle to large-scale application. This application's technical solution constructs a dynamic parallel simulation model using multi-source data collected from an autonomous driving fleet. An adversarial test agent is then set up and trained within this model, interacting with the autonomous driving algorithm under test to generate valid test cases. These test cases are scored and filtered based on a risk matrix, with high-risk cases being selected for real-world verification. This systematically improves the safety and reliability of autonomous driving algorithms.

[0042] Example 1

[0043] refer to Figures 1-4 This embodiment provides a method for intelligent processing of automotive data, which includes the following steps:

[0044] Step S1: Construct the simulation model. This involves building a corresponding dynamic parallel simulation model using multi-source data obtained from a real autonomous vehicle fleet. Details are as follows:

[0045] Step S1.1: Data Acquisition. This involves setting up fixed nodes along the driving path of intelligent vehicles (such as autonomous vehicles or a fleet of test vehicles) and mobile nodes on the intelligent vehicles to collect multi-source data during the vehicle's operation. Specifically, smart cameras, millimeter-wave radars, and roadside units are installed at key locations along the intelligent vehicle's driving path (such as merging points, traffic islands, and streetlights), and onboard data recorders and inertial measurement units are installed on the intelligent vehicles. In other words, fixed nodes are set up based on the smart cameras, millimeter-wave radars, and roadside units, and mobile nodes are set up based on the onboard data recorders and inertial measurement units.

[0046] Furthermore, intelligent cameras and millimeter-wave radar continuously scan the surrounding environment along the intelligent vehicle's driving path to collect corresponding perception data. An onboard data recorder records the intelligent vehicle's control signals to collect corresponding dynamic data. An inertial measurement unit on the intelligent vehicle records its latitude, longitude, and altitude coordinates to collect corresponding pose data. Simultaneously, real-time weather data (such as rainfall and fog concentration) is obtained through a meteorological API interface to collect corresponding scene data.

[0047] Furthermore, the collected perception data, dynamic data, pose data, and scene data undergo data synchronization and standardization processing, including but not limited to time synchronization processing, position synchronization processing, and data formatting processing, to obtain processed multi-source data. It is worth noting that the time synchronization processing, position synchronization processing, and data formatting processing in this embodiment are all conventional technical processing methods, and therefore are not specifically described in this embodiment.

[0048] Step S1.2: Environment Reconstruction. This involves using the processed multi-source data (images and point cloud data) obtained in Step S1.1, and performing data association and alignment based on the acquisition time and location of the multi-source data. This fuses the perception data uploaded by different intelligent vehicles at the same time and location, resulting in fused perception data. For example, the image acquired by the first intelligent vehicle at time T is matched and fused with the image acquired by the second intelligent vehicle at time T+1 to obtain the fused image.

[0049] Furthermore, the acquired fused perception data is fused with a neural representation model, such as NeRF or a 3D Gaussian Splatting model. This means the fused perception data is used as input to the NeRF or 3D Gaussian Splatting model to construct an implicit 3D model of the corresponding driving path. Simultaneously, the dynamic data, pose data, and scene data from the processed multi-source data obtained in step S1.1 are injected into the implicit 3D model to construct the corresponding dynamic parallel simulation model.

[0050] Step S2: Determine test cases. This involves setting up at least one adversarial test agent in the dynamic parallel simulation model constructed in step S1.2. Simultaneously, the adversarial test agent performs interactive testing within the dynamic parallel simulation model, and during this interactive testing process, valid test cases are determined. Specifically:

[0051] Step S2.1: Set up the intelligent agent. This involves generating at least one controllable vehicle model using simulation software (such as CARLA and Baidu Apollo simulation platforms) and importing this model into the dynamic parallel simulation model constructed in step S1.2 to serve as the test intelligent agent. Simultaneously, within the dynamic parallel simulation model, the test intelligent agent is endowed with driving capabilities, such as following simple paths, controlling the accelerator, brakes, and steering, and undergoes adversarial training to obtain an adversarial test intelligent agent. Details are as follows:

[0052] Step S2.1.1: Set the adversarial target. This involves monitoring the behavior of the test agents in the dynamic parallel simulation model using a set reward function to obtain the operation reward score for each test agent. Specifically, in this embodiment, based on the combined actions executed by the test agents in the dynamic parallel simulation model, the simulation engine within the model obtains the core reward item score and constraint penalty item score corresponding to the execution of these combined actions. Based on these core reward item scores and constraint penalty item scores, the total operation reward score for each test agent is determined.

[0053] Furthermore, in this embodiment, the core reward scores include scores for maximizing the uncertainty of the master vehicle, minimizing the safety margin of the master vehicle, and inducing regulatory control failure. The constraint penalty scores include collision penalty scores and traffic rule penalty scores. In other words, the scores for each item are obtained through the simulation engine in the dynamic parallel simulation model, thereby determining the total reward score for each test agent's operation.

[0054] Step S2.1.2: Agent Training. Based on the multi-source data acquired in Step S1.1, the corresponding reconstructed scenarios, including but not limited to weather, lighting, and traffic flow conditions, are loaded into the dynamic parallel simulation model. Simultaneously, in each reconstructed scenario, the main vehicle and the test agent are simulated and tested using deep reinforcement learning algorithms (such as the PPO algorithm and DDPG algorithm), thereby obtaining the corresponding adversarial test agent.

[0055] Step S2.2: Monitoring and Recording. This involves using the adversarial test agent obtained in step S2.1.2 to perform interactive testing on the autonomous driving algorithm under test within the dynamic parallel simulation model. During this interactive testing process, corresponding valid test cases are identified. Specifically, a simulation task scheduling cluster is set up on the cloud platform of the dynamic parallel simulation model, and all autonomous driving algorithms under test are placed in the simulation task scheduling cluster. Simultaneously, in each simulation instance on the cloud platform, the obtained adversarial test agent is injected as background traffic flow for simulation testing.

[0056] Furthermore, during the simulation test, the moment of failure of the main vehicle is recorded using pre-set failure criteria. In this embodiment, the failure criteria include control failure, functional failure, performance degradation, and critical failure, which are respectively: the main vehicle triggering automatic emergency braking, deviating from the predetermined navigation route, a decrease in the confidence level of the perception module, the uncertainty entropy value of the prediction module exceeding a preset danger threshold, and a collision occurring. It is worth noting that the setting of these failure criteria uses conventional technical methods, therefore, they are not specifically described in this embodiment.

[0057] Furthermore, based on the recorded results of the moment the main vehicle fails, these results are matched with the corresponding vulnerability type, severity level, and scenario tag to obtain the corresponding vulnerability report. Vulnerability types include, but are not limited to, perception errors, prediction misjudgments, and planning defects; the severity level can be specifically assessed based on the consequences of the failure; and scenario tags include, but are not limited to, rainy days, ramp merging, and nighttime. In other words, the test results for each test case can be obtained through the obtained vulnerability report.

[0058] Step S3: Test Case Verification. Based on the test cases and test results obtained in Step S2.2, a comprehensive score is obtained for each test case. The test cases are then filtered based on the comprehensive scores, and the filtered test cases are subjected to real-world verification to collect corresponding real-world verification feedback results.

[0059] In this embodiment, the test results of each test case are mapped to the corresponding risk matrix to obtain the comprehensive score for each test case, as shown in Table 1 below:

[0060] Table 1: Risk Matrix

[0061]

[0062] Specifically, based on the probability of occurrence and severity of each test case, it is mapped onto a risk matrix, from which a corresponding comprehensive score can be obtained. Then, based on the comprehensive score, the test cases are sorted in descending order of magnitude. A predetermined number of test cases are selected from this list according to actual testing needs; that is, the number of test cases selected can be determined based on actual needs, and there is no fixed quantity requirement. It is worth noting that during the selection process, the results are sorted in descending order, from largest to smallest.

[0063] refer to Figure 3 and Figure 4Based on the obtained comprehensive score and the preset score threshold, the risk levels corresponding to the test cases are divided into low risk, medium risk, and high risk. In other words, in this embodiment, test cases corresponding to high risk, medium risk, and low risk are selected in sequence.

[0064] Example 2

[0065] This embodiment provides an intelligent processing method for automotive data. The specific implementation method is the same as in Embodiment 1, except that the core reward item score and constraint penalty item score are obtained through the simulation engine in the dynamic parallel simulation model. Based on the core reward item score and constraint penalty item score, the total operation reward score corresponding to the test agent is determined. The invention will be illustrated below with specific examples of this embodiment.

[0066] In this embodiment, the prediction module of the master vehicle in the dynamic parallel simulation model collects behavioral data of the test agents within a preset radius centered on the master vehicle. Specifically, the test agents are simulated in the dynamic parallel simulation model according to their assigned driving capabilities. Simultaneously, the prediction module of the master vehicle monitors the simulation process and results of each test agent in real time to determine the total reward score for each test agent's operation, as detailed below:

[0067] Step S2.1.1.1: Trajectory Monitoring. This involves the main vehicle's prediction module predicting the simulated trajectory of the test agent and determining the corresponding uncertainty score based on the predicted trajectory probability distribution entropy value. Specifically, the main vehicle's prediction module sets the probability value for each simulated trajectory based on the test agent's simulated trajectory, and determines the corresponding trajectory entropy value change based on the probability value of the simulated trajectory and the corresponding set weights. Specifically:

[0068]

[0069] in: The change in trajectory entropy value. Let be the probability value of the i-th simulated trajectory at time point t. Let be the probability value of the i-th simulated trajectory at time point t-1.

[0070] In the specific implementation process, before the simulation test of the intelligent agent, the prediction module of the main vehicle obtained three trajectory probabilities of 80%, 15%, and 5%, respectively, with a corresponding trajectory entropy value of 0.62. After the simulation test, the prediction module obtained three trajectory probabilities of 45%, 10%, and 45%, respectively, with a corresponding trajectory entropy value of 0.95. That is, the change in trajectory entropy value is 0.95 - 0.62 = 0.33.

[0071] Step S2.1.1.2: Collision Detection. This involves using the main vehicle's safety core module to determine the safe distance and collision time between the main vehicle and the test agent. Based on the collision time corresponding to each simulation trajectory, a corresponding safety margin reward score is determined. Specifically:

[0072]

[0073] in: This represents the change in the risk factor. The collision time of the simulated trajectory monitored by the main vehicle safety core module at time point t. This refers to the collision time of the simulated trajectory monitored by the main vehicle safety core module at time point t-1.

[0074] In the specific implementation process, the collision time before the simulation test of the test agent was 5 seconds, and the corresponding risk factor was 0.2. After the simulation test of the test agent, the collision time was 1.2 seconds, and the corresponding risk factor was approximately 0.833. That is to say, the corresponding change in risk factor was 0.633.

[0075] Step S2.1.1.3: Braking Monitoring. This involves determining the corresponding planning and control failure value based on the braking response of the main vehicle's planning and control module. Specifically, based on the set comfort threshold (which can be set according to actual data requirements, and therefore not specifically described in this embodiment), when the main vehicle's automatic emergency braking system is activated, the deceleration during the activation process of the automatic emergency braking system is compared with the comfort threshold, and the planning and control failure value is determined based on the comparison result. Specifically:

[0076] When the deceleration is greater than the comfort threshold, the control failure value is the preset control failure value (which can be specifically set according to actual data requirements, so it is not specifically described in this embodiment). Conversely, when the deceleration is not greater than the comfort threshold, the control failure value is 0.

[0077] Step S2.1.1.4: Consequence Monitoring. After the simulation, the behavior of the main vehicle and the test agent is monitored using the simulation engine. Specifically, it determines whether a physical collision or traffic violation occurs between the main vehicle and the test agent. When a physical collision occurs between the main vehicle and / or the test agent, the collision penalty score is the preset collision penalty threshold (which can be set according to actual data requirements, but is not specifically described in this embodiment); otherwise, the collision penalty score is 0. Similarly, when the main vehicle and / or the test agent violates traffic rules, the traffic rule penalty score is the preset traffic rule penalty threshold (which can be set according to actual data requirements, but is not specifically described in this embodiment); otherwise, the traffic rule penalty score is 0.

[0078] Step S2.1.1.5: Determine the total reward score. This involves determining the corresponding total reward score based on the trajectory entropy change obtained in Step S2.1.1.1, the hazard coefficient change obtained in Step S2.1.1.2, the traffic control failure value obtained in Step S2.1.1.3, and the traffic rule penalty score and collision penalty score obtained in Step S2.1.1.4. Specifically:

[0079]

[0080] in: For the total reward score, The change in trajectory entropy value. This represents the change in the risk factor. To control failure values, Points are awarded for collision penalties. Points are awarded for traffic rule violations.

[0081] During the specific implementation process, the change in trajectory entropy is 0.33, so the corresponding uncertainty reward score is 3.3; the change in risk coefficient is 0.633, so the corresponding safety margin reward score is 31.65; at the same time, the control failure value is 500, the collision penalty score is 0, and the traffic rule penalty score is 0. Therefore, the corresponding total reward score is 3.3 + 31.65 + 500 + 0 + 0 = 534.95.

[0082] 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 embodiments and their equivalents.

Claims

1. A method for intelligent processing of automotive data, characterized in that, Including: S1: Constructing the simulation model: Using multi-source data obtained from the autonomous driving fleet, a dynamic parallel simulation model is constructed, including: S1.1: Data Acquisition: Fixed nodes are set on the driving path of the intelligent vehicle, and mobile nodes are set on the intelligent vehicle. Through the fixed nodes and mobile nodes, multi-source data during the driving process of the intelligent vehicle is collected and acquired. The multi-source data includes perception data, dynamic data, pose data and scene data. S1.2: Environment Reconstruction: The perceived data is processed by data association and alignment to obtain fused perceived data, and the fused perceived data is fused with the neural representation model to construct an implicit 3D model. At the same time, the dynamic data, pose data and scene data are injected into the implicit 3D model to construct a dynamic parallel simulation model. S2: Determine test cases: Set up at least one adversarial test agent in the dynamic parallel simulation model, and conduct interactive tests through the adversarial test agent to determine valid test cases, including: S2.1: Setting up the intelligent agent: Generate at least one controllable vehicle model using simulation software, and import the controllable vehicle model into a dynamic parallel simulation model for adversarial training to obtain an adversarial test intelligent agent, including: S2.1.1: Setting the adversarial target: By setting the reward function, the behavior of the test agent in the dynamic parallel simulation model is monitored to obtain the total operation reward score of each test agent. Through the simulation engine in the dynamic parallel simulation model, the core reward item score and constraint penalty item score of the test agent in the simulation process are obtained. The total operation reward score is determined by the core reward item score and constraint penalty item score. The core reward item score includes the score of maximizing the uncertainty of the master vehicle, the score of minimizing the safety margin of the master vehicle, and the score of inducing control failure. The constraint penalty item score includes the collision penalty score and the traffic rule penalty score. S2.1.2: Agent Training: In the dynamic parallel simulation model, the corresponding reconstructed scene is loaded according to the multi-source data, and in each reconstructed scene, the main vehicle and the test agent are simulated and deduced. At the same time, the test agent is trained and tested through a deep reinforcement learning algorithm to obtain an adversarial test agent. S2.2: Monitoring Records: The autonomous driving algorithm to be tested is set up on the cloud platform of the dynamic parallel simulation model through the simulation task scheduling cluster, and the adversarial test agent is injected into the dynamic parallel simulation model as background traffic flow for interactive testing. At the same time, the test results of each test case are obtained through the set failure criteria. S3: Test Case Verification: Based on the test results of the valid test cases, obtain the corresponding comprehensive score, and filter the valid test cases according to the comprehensive score, and then perform real-world verification on the filtered valid test cases.

2. The intelligent processing method for automotive data according to claim 1, characterized in that, Obtain the total reward score for each test agent's actions, including: S2.1.1.1: Trajectory monitoring: Based on the simulated trajectory of the test agent, the prediction module of the main vehicle sets the probability value of each simulated trajectory, and determines the change in trajectory entropy value based on the probability value of the simulated trajectory. S2.1.1.2: Collision Detection: The collision time between the main vehicle and the test intelligent agent is determined through the main vehicle's safety core module, and the change in the risk factor is determined based on the collision time; S2.1.1.3: Braking Monitoring: The braking response corresponding to the main vehicle's planning control module is compared with the set comfort threshold, and the planning control failure value is determined based on the comparison result. Specifically: When the deceleration of the braking response is greater than the comfort threshold, the control failure value is the preset control failure value; otherwise, the control failure value is 0. S2.1.1.4: Consequence monitoring: The simulation engine is used to detect the behavior of the main vehicle and the test agent after the simulation is completed, and collision penalty scores and traffic rule penalty scores are set according to the detection results. S2.1.1.5: Determine the total reward score: Combine the changes in trajectory entropy, risk factor, regulatory failure, collision penalty score, and traffic rule penalty score to determine the total reward score.

3. The intelligent processing method for automotive data according to claim 2, characterized in that, When the host vehicle and / or the test agent experience a physical collision, the collision penalty score is the preset collision penalty threshold; otherwise, the collision penalty score is 0. When the host vehicle and / or the test agent violates traffic rules, the traffic rule penalty score is the preset traffic rule penalty threshold; otherwise, the traffic rule penalty score is 0.

4. The intelligent processing method for automotive data according to claim 1, characterized in that, The failure criteria are used to determine the recorded results at the moment of the main vehicle failure, and the recorded results are matched with the vulnerability type, severity level and scenario label to obtain the vulnerability report of the adversarial testing agent.

5. The intelligent processing method for automotive data according to claim 1, characterized in that, The test results of the valid test cases are mapped to a risk matrix. The comprehensive score of each valid test case is determined through the risk matrix. The valid test cases are then sorted in descending order based on the comprehensive score. At the same time, a preset number of valid test cases are selected by sorting in descending order according to a preset quantity requirement.

6. The intelligent processing method for automotive data according to claim 5, characterized in that, The risk matrix is ​​constructed based on the probability and severity of the test cases, with both the probability score and severity score set to 1-5.

Images