Intelligent networked automobile digital twinning evaluation method and system oriented to extreme scene

Abstract

The invention relates to the technical field of automobile intelligent driving, in particular to an intelligent network connection automobile digital twinning evaluation method and system for extreme scenes. The evaluation method comprises the following steps: constructing a knowledge graph; generating an extreme scene; the automatic driving system is tested in a virtual environment generated by the digital twin test platform based on an extreme scene; predicting a safety risk score in real time, and if the safety risk score exceeds a safety threshold, sending an intervention instruction to the digital twin test platform; performing anomaly detection on the full-quantity test log, deriving a scene and injecting the scene into a scene library if a performance inflection point is found; and calculating a root cause based on the weight vector, the system fault cause and effect graph and the test data, and generating a diagnosis result. According to the method, extreme scenes can be covered, the system can be scored from multiple dimensions, the performance bottleneck can be accurately positioned, and the closed loop of the test scene can be realized. According to the system, two-way closed loop and common evolution of physical test data and a virtual test environment are realized through a cloud side end architecture, and continuous synchronization and two-way interaction with real world data are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving of automobiles, and in particular to an intelligent networked automobile digital twin evaluation method and system for extreme scenarios. BACKGROUND

[0002] Intelligent networked automobiles are comprehensive systems that deeply integrate vehicle engineering, artificial intelligence, information communication and other multidisciplinary technologies. Intelligent networked automobiles are based on traditional vehicle platforms, collect environmental data through on-board sensors, and use artificial intelligence and machine learning algorithms to achieve accurate perception and driving decisions in complex environments. At the information interaction level, the system integrates C-V2X direct communication and 5G cellular network technology to build a comprehensive communication link between vehicles, roads, people and cloud platforms, enabling low-latency, high-reliability information sharing and collaborative perception.

[0003] Existing virtual testing technologies can cover a large number of test scenarios at low cost and high efficiency, but still face three major challenges: existing test scenarios are mostly based on standard regulations and common working conditions, making it difficult to effectively cover low-probability, high-risk long-tail extreme scenarios that lead to accidents in the real world; existing evaluation systems can score autonomous driving systems from multiple dimensions, but the weights of each dimension are fixed, making it difficult to accurately identify performance bottlenecks and guide optimization; and the vast amount of test data is mostly used to generate test reports in actual applications, without automatically and intelligently feeding back to the test scenario library to form a self-improving closed loop, resulting in long test iteration cycles and high costs.

[0004] CN112015164A discloses an intelligent networked automobile complex test scenario implementation system based on digital twinning, which obtains a virtual world based on a real world twin, but lacks continuous synchronization and bidirectional interaction with real world data, and the test scenario evolution capability is insufficient, lacking root cause diagnosis. CN120633216A discloses an automatic driving test scenario automatic derivation method, system, device and medium, which constructs a test scenario based on test requirements, selects and reorganizes the most scene set that meets the constraints in the initial scene fragment set, and sorts based on causal logic relationships to obtain multiple safety verification automatic driving test scenarios. Although the scenario construction considers the causal logic relationship of time, the test scenario evolution capability is still insufficient, and cannot continuously synchronize and bidirectionally interact with real world data.

[0005] Therefore, it is necessary to develop an evaluation system that can construct a complete digital twin closed loop, automatically generate high-value test scenarios, perform deep diagnosis, and self-iterate and optimize. SUMMARY

[0006] The application aims to provide an intelligent networked vehicle digital twin evaluation method and system for extreme scenarios, which constructs a knowledge graph of accident scenarios based on massive data, optimizes high-value test scenarios based on performance evaluation of generated test scenarios, and realizes scenario evolution and root cause diagnosis by sending intervention instructions in real time based on test data, reconstructing seed scenarios based on performance inflection points, and injecting scenarios into the scenario library after derivation.

[0007] To achieve this purpose, the application adopts the following technical solutions:

[0008] An intelligent networked vehicle digital twin evaluation method for extreme scenarios,

[0009] The evaluation method comprises the following steps:

[0010] S1: Construct a knowledge graph, which is based on unstructured accident reports in existing databases to construct a knowledge graph of accident scenarios;

[0011] S2: Generate extreme scenarios, generate test scenarios with clues sampled from the knowledge graph as seeds, generate test scenarios based on a functional copy of the autonomous driving system, run the generated test scenarios in the digital twin test platform and output their performance, evaluate the performance and optimize the test scenarios based on the evaluation results, and save the obtained test scenarios in the scenario library;

[0012] S3: The digital twin test platform generates a virtual environment based on the extreme scenarios, tests the autonomous driving system, and outputs test data;

[0013] S4: Real-time acquisition of vehicle state data from the digital twin test platform for predicting safety risk score, if the safety risk score exceeds the safety threshold, send intervention instructions to the digital twin test platform;

[0014] S5: Abnormal detection of full-amount test logs of the digital twin test platform, if a performance inflection point is found, reconstruct the seed scenario, generate a derivative scenario through parameter perturbation and inject it into the scenario library;

[0015] S6: Real-time acquisition of feature encoding of the previous test scenario in the digital twin test platform, calculation of the weight vector of the evaluation index of the current test scenario, calculation of the root cause based on the weight vector, system fault cause and effect diagram and test data, and generation of a diagnosis result.

[0016] Further, the step S1 comprises:

[0017] S101: Extract scenario elements from the existing database and convert them into standardized scenario element vectors;

[0018] S102: Based on standardized scene element vectors, construct the relationship between environment, entity, behavior and consequence to form a knowledge graph of accident scenarios.

[0019] Furthermore, step S2 includes:

[0020] S201: Using clues sampled from the knowledge graph as seeds, based on policy... Generate test scenarios, These are weight parameters;

[0021] S202: Based on a functional copy of the autonomous driving system under test, the generated test scenario is run in a digital twin test platform and its performance is output. The performance includes whether the autonomous driving system under test violates safety constraints and the types and number of safety constraints violated.

[0022] S203: Using reward function Calculate the reward value, where, For the scene The overall reward value, Indicates a performance degradation reward. For the system under test in the scenario The success rate in This indicates that the security constraint violates the reward. For the first A violation indication function for a safety constraint. For the corresponding weights; Indicates a reward for the novelty of the scenario. For the scene With historical scene library Mid-scene Similarity, weight parameters , , Online optimization using the REINFORCE policy gradient algorithm satisfies the following requirements. ;

[0023] S204: Update weight parameters based on overall reward value ,Strategy From weight parameters Parameterization , For learning rate, This is an estimate of the policy gradient calculated using the Monte Carlo method. It maximizes long-term expected reward by iteratively optimizing and generating extreme scenarios that can continuously challenge the tested system.

[0024] Furthermore, step S4 includes:

[0025] S401: Acquire vehicle status data from the digital twin testing platform in real time and calculate a comprehensive risk score. , , Indicates the risk of collision time. This is the collision time threshold; Indicates the risk of deceleration. This represents the deceleration requirement at time t. Indicates the maximum permissible deceleration; Indicates comfort risk, This represents the lateral acceleration of the vehicle at time t. This indicates the maximum permissible lateral acceleration threshold; These are the normalized weight parameters;

[0026] S402: Judgment Does it exceed the preset security threshold? When the preset safety threshold is exceeded, a real-time intervention command is sent to the digital twin testing platform based on a predefined standardized "intervention command protocol".

[0027] Furthermore, step S5 includes:

[0028] S501: Perform streaming processing on the full test data, and use an anomaly detection algorithm based on isolated forest to calculate the anomaly score of the multi-dimensional performance index vector in the full test log data. , For the sample Abnormal scores; For the sample Path length in the isolation tree; Let be the expected value of the path length among multiple isolated trees. For a given subsample size Standardized path length at time For the first One harmonic number;

[0029] S502: When abnormal scores are found When the performance inflection point detection threshold is exceeded, the data point is determined to be a performance anomaly.

[0030] S503: Extract scene parameters within the time period before and after the corresponding performance anomaly point, reconstruct a reproducible seed scene, perform targeted perturbation and mutation on the key parameters of the seed scene, synthesize a derived scene, and inject the derived scene into the basic scene library.

[0031] Furthermore, step S6 includes:

[0032] S601: Standardize the current test scenario to obtain the scenario feature encoding vector. , the weight vector of the evaluation index corresponding to the test scene output by the MLP network, the weight mapping function of the output layer of the MLP network , is the weight vector of the evaluation index , is the feature encoding vector of the scene , is a multi-layer perception network , ensure weight normalization

[0033] S602: The nodes of the constructed system fault causal diagram are the performance indicators of each level of the autonomous driving system. When the top-level evaluation index is abnormal, the Bayesian inference algorithm uses the dynamic weight vector as the observation condition, and performs probabilistic backtracking along the causal dependence relationship defined in the system fault causal diagram. When the abnormal performance of the high-level and intermediate-level is calculated, each potential root cause node at the bottom level is the probability of the true fault source: , is the posterior probability of the root cause when the abnormal evidence is observed is the prior probability of the root cause , is the likelihood probability of observing evidence when the root cause occurs is the total number of candidate root causes

[0034] S603: Sort all candidate root causes by probability to generate a quantitative diagnostic report.

[0035] Further, the historical test log set and the expert data set are used as the training data set. The test scene in the training data set is standardized to a scene feature encoding vector. The scene feature encoding vector is composed of environment features, road topology features, traffic flow features, and key event identification sub-vectors. The historical test logs in the training data set are constructed into target weight vectors by regression backtracking method. The target weight vectors of the expert data in the training data set are taken from the normalized expert assignment

[0036] The MLP network is trained using the training data set. Training is performed by minimizing the composite loss function, which is:

[0037] ,

[0038] ,

[0039] ,

[0040] wherein a composite loss function; a mean square error loss, a cosine similarity loss, the cosine similarity is in the range of , , is a hyperparameter and satisfies ; is a weight vector predicted by the model; is a target weight vector; is the number of performance indicators.

[0041] An intelligent networked vehicle digital twin evaluation system for extreme scenarios is used to implement the intelligent networked vehicle digital twin evaluation method for extreme scenarios described above, and the evaluation system includes a digital twin test platform, an edge calculator and a cloud server.

[0042] The digital twin test platform is used to generate a virtual environment based on extreme scenarios, test the autonomous driving system, and output test data.

[0043] The cloud server includes:

[0044] A multi-source accident data fusion module extracts scene elements from an unstructured accident report database and converts them into standardized scene element vectors.

[0045] An accident scene knowledge graph construction module constructs the correlation between environment, entity, behavior and consequence based on the standardized scene element vectors, and forms a knowledge graph of the accident scene.

[0046] An adversarial scene generation engine generates test scenes using the sampled clues in the knowledge graph as seeds, obtains the performance of the tested autonomous driving system running in the test scenes of the digital twin test platform, evaluates the performance, and optimizes the test scenes based on the evaluation results.

[0047] A scene library update module performs anomaly detection on the full test logs of the digital twin test platform, and if a performance inflection point is found, the seed scene is reconstructed, derivative scenes are generated by parameter perturbation and injected into the scene library.

[0048] A cross-dimension root cause positioning module obtains the feature encoding of the previous test scenes in the digital twin test platform in real time, calculates the weight vector of the evaluation indicators of the current test scene, calculates the root cause based on the weight vector, the system fault causal graph and the test data, and generates a diagnosis result.

[0049] The edge calculator is used to obtain vehicle state data of the digital twin test platform in real time for predicting a safety risk score, and if the safety risk score exceeds a safety threshold, an intervention instruction is sent to the digital twin test platform.

[0050] Further, the cross-dimension root cause positioning module is equipped with an MLP network, the MLP network takes a scene feature encoding vector of a current test scene as input, and outputs a weight vector of an evaluation index under the test scene.

[0051] A computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the evaluation method.

[0052] The technical solution provided by the application can include the following beneficial effects:

[0053] The evaluation method of the application can effectively cover the extreme scenarios such as long tail, low probability and high risk that cause accidents in the real world based on the performance evaluation of the test scenarios generated based on the knowledge graph of the accident scene constructed based on the massive data. The intervention instruction is sent in real time based on the test data, the seed scene is reconstructed based on the performance inflection point, and the scene evolution is realized by injecting the scene into the scene library after derivation. The test scene library forms a self-perfecting closed loop, and the scene iteration speed is fast. Based on the system fault causal graph constructed, the root cause diagnosis is realized according to the test scene weight vector and the test data, and a quantitative diagnosis report is output.

[0054] The system of the application realizes the bidirectional closed loop and co-evolution of physical test data and virtual test environment through the cloud edge end architecture, realizes the continuous synchronization and bidirectional interaction with the real world data. The system constitutes a complete evaluation ecological system which is closely coupled with the real world data and can realize bidirectional closed loop and self-evolution. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a flowchart of an intelligent networked vehicle digital twin evaluation method for extreme scenarios according to an embodiment of the application;

[0056] Figure 2 is a flowchart of knowledge graph construction and extreme scenario generation;

[0057] Figure 3 is a flowchart of generating extreme scenarios;

[0058] Figure 4 is a flowchart of a test cycle;

[0059] Figure 5 is a flowchart of MLP network training;

[0060] Figure 6 is a system fault causal graph. DETAILED DESCRIPTION

[0061] Embodiments of the present application are described below in detail, examples of which are shown in the drawings, wherein the same or similar reference numbers represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary, only for the purpose of explaining the present application, and cannot be understood as a limitation of the present application.

[0062] The embodiments of the present application are described below in detail Figures 1 to 6 , a smart connected vehicle digital twin evaluation method for extreme scenarios, comprising the following steps:

[0063] S1: Constructing a knowledge graph, constructing a knowledge graph of accident scenarios based on unstructured accident reports in an existing database;

[0064] S2: Generating extreme scenarios, generating test scenarios with the clues sampled in the knowledge graph as seeds, generating test scenarios based on a functional copy of the autonomous driving system, running the generated test scenarios in the digital twin test platform and outputting their performance, evaluating the performance and optimizing the test scenarios based on the evaluation results, and saving the resulting test scenarios in the scenario library;

[0065] S3: The digital twin test platform generates a virtual environment based on the extreme scenarios, tests the autonomous driving system, and outputs test data;

[0066] S4: Real-time acquisition of vehicle state data of the digital twin test platform for predicting safety risk score, if the safety risk score exceeds the safety threshold, send intervention instructions to the digital twin test platform;

[0067] S5: Abnormal detection of full-quantity test logs of the digital twin test platform, if a performance inflection point is found, reconstruct the seed scenario, generate derivative scenarios through parameter perturbation and inject them into the scenario library;

[0068] S6: Real-time acquisition of feature encoding of the previous test scenarios in the digital twin test platform, calculation of the weight vector of the evaluation index of the current test scenario, calculation of the root cause based on the weight vector, system fault causal diagram and test data, generation of a diagnosis result.

[0069] The evaluation method of the present application, based on massive data, constructs a knowledge graph of accident scenarios and generates test scenarios based on performance evaluation to optimize the generation of high-value test scenarios, and based on test data, real-time intervention instructions, reconstruction of seed scenarios based on performance inflection points and injection of derivative scenarios into the scenario library to realize scenario evolution, and realize root cause diagnosis, greatly improving the reliability and authenticity of the evaluation of the autonomous driving system. The existing database in the present application includes NHTSA and CIDAS.

[0070] Referring to Figure 2 and Figure 3 The step S1 of the present application comprises S101 and S102.

[0071] S101: Extract scene elements from existing databases and convert them into standardized scene element vectors (data source and processing). Specifically, analyze unstructured accident reports in NHTSA, CIDAS, etc. databases through natural language processing technology (e.g. BERT-based NLP model) to extract key scene elements, including weather, road geometry, traffic participant types and behaviors, accident morphology, and convert them into standardized scene element vectors (such as JSON format standardized scene element vectors) and store them in the cloud database, preparing for the next stage of graph construction.

[0072] S102: Based on the standardized scene element vector, construct the association relationship between environment, entity, behavior and consequence, and form the knowledge graph of accident scene. The nodes in the association relationship include environmental elements, traffic participants, driving behaviors, and edges represent their causal, temporal or statistical associations. The accident scene knowledge graph construction module uses a graph database to organize the discrete scene element vectors extracted in the previous stage into a mutually connected and semantically rich network structure, where nodes represent various extracted scene elements and edges define the complex relationships between elements. In this way, countless isolated accident reports are integrated into a large and interconnected knowledge network. The knowledge graph no longer just piles up data, but encapsulates deep logic and patterns about how accidents occur, enabling the system to understand under what conditions, what behaviors may lead to what consequences. The value of the knowledge graph is ultimately realized through the adversarial scenario generation engine. Refer to Figure 2 , which constitutes a knowledge graph with an environment of "rainy-wet road", an entity of "highway-truck", a behavior of "emergency braking", and a consequence of "rear-end collision".

[0073] The step S2 (scenario generation and output) includes S201-S204.

[0074] S201: Seed the sampled clues in the knowledge graph, generate test scenarios based on the strategy , where is the weight parameter;

[0075] S202: Based on the functional copy of the tested autonomous driving system, the generated test scenarios are run in the digital twin test platform and output their performance, including whether the tested autonomous driving system violates safety constraints and the types and number of violated safety constraints;

[0076] S203: Calculate the reward value with the reward function , where is the comprehensive reward value of the scenario , and represents the performance degradation reward, For the system under test in the scenario The success rate in This indicates that the security constraint violates the reward. For the first A violation indication function for a safety constraint. For the corresponding weights; Indicates a reward for the novelty of the scenario. For the scene With historical scene library Mid-scene Similarity, weight parameters , , Online optimization using the REINFORCE policy gradient algorithm satisfies the following requirements. ;

[0077] S204: Update weight parameters based on overall reward value ,Strategy From weight parameters Parameterization , For learning rate, This is an estimate of the policy gradient calculated using the Monte Carlo method. It maximizes long-term expected reward by iteratively optimizing and generating extreme scenarios that can continuously challenge the tested system.

[0078] Step S201 of this invention does not randomly generate scenarios, but rather involves guided sampling from a pre-constructed accident scenario knowledge graph. This process can start from a known high-risk causal chain or discover new potential hazard combinations through graph reasoning. The sampled element relationship chain is used as an outline, and specific parameter values ​​are assigned to the abstract elements within it, thereby instantiating a complete test scenario file that can be executed by the digital twin testing platform. After the generated test scenario is executed in the digital twin environment, its performance is output, and this performance is evaluated. The weight parameters are then updated based on the comprehensive reward value obtained through a reward model. This feedback is then fed back to the scene generation strategy. If a pattern sampled from the knowledge graph effectively exposes system weaknesses, the scene generator is rewarded and will tend to explore more graph regions and their variants related to that pattern in the future. Conversely, ineffective patterns are suppressed. This allows for the efficient generation of extreme scenarios that accurately expose system weaknesses.

[0079] The scheme sets n safety constraints for extreme scenarios, and the setting principle is mainly based on the following three aspects: (1) the causal association extraction based on the accident knowledge graph, the high-risk scene elements are extracted by analyzing the real accident report, and these elements are organized into the accident scene knowledge graph, and the setting of safety constraints is derived from these causal chains; (2) mapping of multi-dimensional performance and safety indicators, a plurality of performance indicators are monitored in the test, such as decision delay, and the part closely related to safety in these indicators is quantified as a safety constraint; (3) standardization and expert knowledge fusion, the system describes the scene and behavior in accordance with the ASAM OpenX series standards, which already contains a series of safety-related behaviors and state constraints, and the system also integrates expert annotated data to clearly define the relative importance of each performance indicator in extreme scenarios, and then set the constraint weight. The weight of the safety constraint is optimized online through the REINFORCE policy gradient algorithm, so that it can be dynamically adjusted according to the risk characteristics of different scenes, realizing dynamic weight and scene adaptation. The performance in step S202 is whether the measured autonomous driving system violates the above safety constraints and the type and number of violated safety constraints.

[0080] In an embodiment of the present application, the "highway construction zone merging" capability test of the measured autonomous driving system is taken as the background. First, sampling is performed from the constructed accident scene knowledge graph. The initial clues obtained by sampling are "highway construction zone merging", and relevant environment and behavior nodes are attached, such as "sunny weather", "dry road surface", "medium traffic density", based on these clues, a set of initial scene parameters are generated, and the scene editing API provided by the digital twin test platform is called to generate an initial test scene file in the OpenSCENARIO 2.0 format. The initial scene file is loaded into the digital twin test platform, the platform starts high-fidelity simulation, runs the functional copy of the measured autonomous driving system, and real-time monitors and records the multi-dimensional performance of the measured system in this simulation, including but not limited to: trajectory tracking error, minimum distance to construction cone and adjacent vehicle, merging decision delay, smoothness of planned path, etc. This simulation runs smoothly, and the measured system successfully completes the merging task, and the output comprehensive performance score is high. Subsequently, the reward model calculates the reward value of this scene according to the preset comprehensive reward function of the formula, wherein, since the measured system performs well this time, there is no performance degradation, the reward value is low; no collision or serious violation occurs during running, the reward value is zero; the similarity between the current scene and all scenes in the historical scene library is calculated, since the initial scene is ordinary, the novelty is general, The reward value is medium. Overall, the initial scene obtains a medium-low comprehensive reward value, which is fed back to step S201 through the REINFORCE policy gradient algorithm, and then the "scene-reward" pair is learned as experience data.

[0081] Under the driving of the reward signal, step S201 aims to generate a scene that can obtain a higher reward. Again, nodes associated with high risk, low visibility, and sudden behavior are sampled from the knowledge graph. Subsequently, the second round of scene parameters is generated: the weather is changed from sunny to heavy rain, the time is changed from daytime to nighttime, and a new dynamic element "pedestrian suddenly rushes out from the construction isolation zone visual blind area" is introduced. The digital twin test platform loads this new and extremely high scene, and after simulation, the performance evaluator monitors that the measured system has significant performance degradation, and the heavy rain and night cause the detection distance of the perception module to pedestrians to be greatly shortened, and the delay is high; the decision module hesitates between emergency braking and turning to avoid, causing the vehicle trajectory to oscillate sharply. Recalculate the comprehensive reward value based on these performance degradation data (performance). Among them, the vehicle has obvious performance degradation, The reward value is high; the minimum distance between the vehicle and the pedestrian is below the safety threshold, triggering a safety constraint violation, The reward value is high; the complex extreme scene is very different from most scenes in the scene library, and the novelty reward is high. Therefore, the second round of scenes obtains a high comprehensive reward value, and the high reward value strongly reinforces the strategy of step S201 to generate such complex and dangerous scenes, and continues to iterate multiple rounds, constantly optimizing and adjusting the parameters. After several rounds of iteration, a strategy that can stably and efficiently generate high-value extreme scenes is converged, and a series of extreme test scene files that can accurately expose the deep weaknesses of the measured system in perception, decision-making, and control systems are successfully generated. These scene files are automatically stored in the scene library, not only for in-depth analysis of this evaluation, but also provide crucial data support for subsequent system optimization, realizing the closed loop of evaluation and improvement.

[0082] In an embodiment of the present application, the test scene is input into the digital twin test platform, which continuously tests the autonomous driving system and outputs vehicle state data in real time. When the safety index exceeds the preset safety threshold in a certain test scene, a real-time intervention instruction is sent to the digital twin test platform, ensuring that the most critical safety bottom line is achieved in the simulation environment. The full-amount test log data includes the intervention instruction data. The step S4 includes S401-S02.

[0083] S401: Real-time acquisition of vehicle state data of the digital twin test platform, and calculation of a comprehensive risk score , , Indicates the time-to-collision risk, calculated based on the current vehicle speed and the distance to the obstacle. This is the collision time threshold; Indicates the risk of deceleration. This represents the deceleration requirement at time t. Indicates the maximum permissible deceleration; The risk to comfort is represented by the rate of change of acceleration, where This represents the lateral acceleration of the vehicle at time t. This indicates the maximum permissible lateral acceleration threshold; These are the normalized weight parameters;

[0084] S402: Judgment Does it exceed the preset security threshold? When the preset safety threshold is exceeded, a real-time intervention command is sent to the digital twin testing platform based on the predefined standardized "Intervention Command Protocol" (ROS standard).

[0085] The method of this invention also includes a step of automatically mining and synthesizing high-value scenarios. It performs streaming processing on all test data from a test cycle, reconstructs seed scenarios based on scenarios with performance anomalies, and automatically synthesizes a batch of new, challenging derived scenarios, injecting them into the scenario library to achieve self-evolution of the evaluation system. Step S5 includes:

[0086] S501: Perform streaming processing on the full test data, and use an anomaly detection algorithm based on isolated forest to calculate the anomaly score of the multi-dimensional performance index vector in the full test log data. , For the sample Abnormal scores; For the sample Path length in the isolation tree; Let be the expected value of the path length among multiple isolated trees. For a given subsample size Standardized path length at time For the first One harmonic number;

[0087] S502: When abnormal scores are found When the performance inflection point detection threshold is exceeded, the data point is determined to be a performance anomaly.

[0088] S503: Extract the scene parameters within the time period (e.g., 3 seconds) before and after the corresponding performance anomaly point, reconstruct a reproducible seed scene, perform targeted perturbation and mutation on the key parameters of the seed scene, synthesize a batch of new and challenging derivative scenes, inject the derivative scenes into the basic scene library, and will be given priority in the next round of testing, so that the test can continuously and accurately attack the known weaknesses of the system and drive its continuous optimization.

[0089] After a test scenario is completed, such as in an extreme scenario where "a pedestrian suddenly rushes out from a blind spot during a nighttime downpour," if the digital twin of the tested autonomous driving system ultimately fails to avoid a collision and the test is marked as a failure, then the cross-dimensional root cause localization process is immediately initiated. Step S6 includes S601-S603.

[0090] S601: Standardize the current test scenario to obtain the scenario feature encoding vector. The MLP network outputs a weight vector corresponding to the evaluation metric in the test scenario, and the MLP network output layer weight mapping function is used. , for The weight vector of each evaluation indicator For the scene The feature encoding vector, It is a multilayer sensing network. Ensure weight normalization. The MLP network is used to dynamically output an evaluation weight vector optimized for the current scenario. After establishing the core evaluation dimensions in the current scenario, cross-dimensional diagnosis is performed based on the system fault cause-effect graph.

[0091] S602: Constructing a system fault cause-effect graph ( Figure 6 The nodes represent the performance metrics (i.e., evaluation metrics) of each level of the autonomous driving system. When the top-level evaluation metric is abnormal, the Bayesian inference algorithm uses a dynamic weight vector as the observation condition to perform probabilistic backward inference along the causal dependencies defined in the system fault cause-effect graph, and calculates the probability that each potential root cause node at the bottom level is the true source of the fault when the high-level and intermediate-level abnormal performance occurs. , In order to observe abnormal evidence At that time, the root cause The posterior probability, root cause The prior probability, For the root cause When it happened, evidence was observed. The likelihood probability, The total number of candidate root causes;

[0092] S603: Sort all candidate root causes by probability size, generate a quantitative diagnosis report.

[0093] For the test scene that fails in bad weather, the weights of "perception confidence" and "brake response time" in the evaluation index are increased. Then, combined with the system fault cause and effect diagram and the specific data of this test, the root cause is calculated through the Bayesian inference algorithm, and a diagnosis report is generated. In the embodiment of the application, a three-layer The network outputs the weight vector of the evaluation index, the number of input layer nodes corresponds to the 64-dimensional scene feature encoding vector, the hidden layer is a fully connected layer with 128 nodes, the ReLU activation function is used, the number of output layer nodes is equal to the number of performance indicators M, and the Softmax function is used to ensure that the output weight is normalized. The model uses a large amount of historical test data to predict the cross entropy between the weight and the target weight as the loss function, and uses the Adam optimizer for pre-training.

[0094] Referring to Figure 5 , the MLP network is learned by a large amount of test data, so that the evaluation weight can be adaptively adjusted according to the scene risk. The training process is as follows:

[0095] The historical test log set and the expert data set are used as the training data set, the test scene in the training data set is standardized to a scene feature encoding vector, the scene feature encoding vector is composed of an environment feature, a road topology feature, a traffic flow feature, and a key event identification sub-vector, the historical test log in the training data set is constructed into a target weight vector by a regression back-propagation method, and the target weight vector of the expert data in the training data set is taken from the normalized expert assignment;

[0096] The MLP network is trained by using the training data set, and is trained by minimizing the composite loss function. The loss function is:

[0097] ,

[0098] ,

[0099] ,

[0100] wherein, is a composite loss function; is a mean square error loss, which ensures the numerical accuracy of the predicted weight, and minimizes the absolute error of each index weight; is a cosine similarity loss, and the cosine similarity has a value range of , and the closer the value is to 1, the more consistent the direction is, which ensures that the relative distribution of the predicted weight and the target weight is consistent; , is a hyperparameter and satisfies ; a weight vector predicted for the model; a target weight vector; a number of performance indicators.

[0101] Specifically, the original scene description in the mass of historical test logs is converted into a machine-readable format, and through standardization and vectorization processing, the environmental conditions, road topology, traffic flow density, key event types and other multi-dimensional information are fused into a unified, high-dimensional scene feature encoding vector. For expert annotated data, the target weight vector directly adopts the normalized expert assignment; for historical test data, the target weight vector is constructed by regression back-propagation method, first, a proxy model is trained using historical logs, the task of the model is to predict the final success or failure of this test according to the performance indicators of the system in the scene. Subsequently, the SHAP attribution analysis method is used to analyze this proxy model, and the contribution of each performance indicator to the prediction of failure is quantified, and these contribution degrees are normalized to form the target weight vector corresponding to the historical scene.

[0102] In an embodiment of the present application, the MLP network includes an input layer, one or more hidden layers and an output layer. To ensure that the weight of the model output is accurate and reasonable, the training process uses a composite loss function composed of numerical precision loss and distribution consistency loss. The prepared training data is input into the network, and the model iteratively minimizes the above composite loss function through the back propagation algorithm and the gradient descent optimizer, and the training process continues until the model performs stably on the independent validation data set, indicating that it has learned the complex, nonlinear mapping relationship from scene features to indicator weights. The trained and validated model is exported in a standard deployment format and integrated into the test server of the evaluation system. At the beginning of each test, the features of the current test scene are encoded and input into the model, and the model outputs the evaluation weight vector of the scene instantaneously. The dynamic weight is then used for root cause analysis, so that the entire evaluation process has intelligent scene adaptation.

[0103] Correspondingly, the present application also provides an intelligent networked vehicle digital twin evaluation system for extreme scenes, which is used to realize the intelligent networked vehicle digital twin evaluation method for extreme scenes. The evaluation system includes a digital twin test platform, an edge calculator and a cloud server.

[0104] The digital twin test platform is used to generate a virtual environment based on extreme scenes, test the autonomous driving system, and output test data.

[0105] The cloud server includes:

[0106] Multi-source accident data fusion module: extract scene elements from unstructured accident report database, and convert them into standardized scene element vectors;

[0107] Accident scene knowledge graph construction module: based on the standardized scene element vectors, the correlation between environment, entity, behavior and consequence is constructed to form the knowledge graph of the accident scene;

[0108] Adversarial scene generation engine, with the sampled clues in the knowledge graph as seeds to generate test scenes, obtain the performance of the tested autonomous driving system in the test scene of the digital twin test platform, evaluate the performance and optimize the test scene based on the evaluation results;

[0109] Scene library update module, abnormality detection is performed on the full test log of the digital twin test platform, if a performance inflection point is found, the seed scene is reconstructed, the derived scene is generated by parameter disturbance and injected into the scene library;

[0110] Cross-dimension root cause positioning module, real-time acquisition of feature encoding of the previous test scene in the digital twin test platform, calculation of the weight vector of the evaluation index of the current test scene, calculation of the root cause based on the weight vector, system fault causal graph and test data, generation of a diagnosis result;

[0111] The edge calculator is used for real-time acquisition of vehicle state data of the digital twin test platform for predicting a safety risk score, and if the safety risk score exceeds a safety threshold, an intervention instruction is sent to the digital twin test platform.

[0112] The system of the application realizes the bidirectional closed loop and co-evolution of physical test data and virtual test environment through cloud edge architecture, realizes the continuous synchronization and bidirectional interaction with real world data. The test method of the application is deployed in a distributed computing architecture composed of a digital twin test platform, an edge computing node and a cloud server. In order to realize efficient cooperation, the system internally uses ROS 2 middleware based on DDS protocol for communication to ensure low delay and high reliability. The data interaction with the digital twin test platform complies with the ASAM OpenX standard, and the edge real-time intervention instruction complies with the signal and service interface standard defined in the AUTOSAR adaptive platform. In terms of hardware configuration, the edge node uses the NVIDIA Jetson AGX Orin computing platform, and the cloud server cluster is configured with high-performance GPU and PB-level distributed storage system to support massive scene data analysis and storage. The system constitutes a complete evaluation ecosystem that is tightly coupled with real world data and can realize bidirectional closed loop and self-evolution.

[0113] The cross-dimension root cause positioning module in the test system of the application is equipped with an MLP network, which takes the scene feature encoding vector of the current test scene as input and outputs the weight vector of the evaluation index under the test scene.

[0114] The test system of the application accesses various data sources including NHTSA, CIDAS traffic accident database, and in-depth accident investigation report. Before testing, the multi-source accident data fusion module parses the unstructured accident report in the NHTSA, CIDAS and other databases, extracts key scene elements based on the BERT NLP model, and converts them into standardized scene element vectors in JSON format, and stores them in the system database. The value of the knowledge graph obtained by the accident scene knowledge graph construction module is finally realized through the adversarial scene generation engine.

[0115] The adversarial scene generation engine includes a scene generator, a performance evaluator, and a reward model. The scene generator first samples from the constructed accident scene knowledge graph, obtains initial clues and related environmental and behavior nodes, generates a set of initial scene parameters based on these clues, and generates an initial test scene file. The initial scene file is loaded into the digital twin test platform, the platform starts high-fidelity simulation, runs a functional copy of the tested autonomous driving system, and the performance evaluator starts simultaneously, monitors and records the multi-dimensional performance of the tested system in this simulation in real time. The performance evaluator sends the full set of performance data of this run to the reward model, calculates the comprehensive reward value, and the comprehensive reward value is fed back to the scene generator through the REINFORCE policy gradient algorithm. Driven by the reward signal, the scene generator aims to generate scenes that can obtain higher rewards. The scene generator samples from the knowledge graph again, but this time it tends to select nodes associated with high-risk, low-visibility, and sudden behavior. After several rounds of iteration, the adversarial scene generation engine converges to a strategy that can stably and efficiently generate high-value extreme scenes, and it has successfully generated a series of extreme test scene files that can accurately expose the deep weaknesses of the tested system in perception, decision-making, and control systems.

[0116] Reference Figure 4, the digital twin test platform loads an extreme scenario at the beginning of the test, generates high-fidelity vehicle state, environmental data and performance data, and sends them to the edge calculator and the cloud server in parallel through a low-latency communication link. The edge calculator continuously obtains key dynamic parameters from the data stream at a very high frequency. For each calculation period, the module calls its built-in comprehensive risk scoring function to generate a real-time, comprehensive safety risk score. Once the calculated comprehensive risk score exceeds the preset safety threshold, it means that the risk of collision or loss of control is imminent. The module will no longer wait for remote instructions from the cloud, but will directly generate intervention instructions and send them to the digital twin test platform. At the same time of implementing the intervention, the edge calculator synchronously uploads the key pre-and-post data of this intervention event to the cloud server as a high-value data package. The scenario library update module of the cloud server runs continuously. The scenario library update module performs anomaly detection based on the uploaded full test logs. If a performance inflection point is found, it automatically extracts and reconstructs the complete test scenario that led to the inflection point from the data storage as a seed scenario, and automatically synthesizes a batch of derivative scenarios by perturbing and mutating the key parameters of the seed scenario and injecting them into the scenario library.

[0117] During the test, the cloud-edge collaborative test real-time intervention and scenario library iterative optimization system starts to work. The edge calculator obtains vehicle state data from the digital twin test platform bus at a frequency of 100Hz, and calculates the comprehensive risk score. When the comprehensive risk score is greater than the safety threshold, the module no longer relies on the cloud decision, but directly sends instructions to the digital twin test platform according to the pre-defined ROS standardized intervention instruction protocol, while all test data are synchronously uploaded to the cloud in real time through the Kafka message queue.

[0118] ​The cross-dimension root cause positioning module includes a scene feature-index weight mapping model and a cross-dimension diagnosis engine based on a causal diagram. When the test of the measured autonomous driving system in an extreme scene is marked as a failure, the cross-dimension root cause positioning module is started, the scene feature-index weight mapping model in the cross-dimension root cause positioning module is activated first, the feature encoding of the current test scene is input, and an evaluation weight vector optimized for the current scene is dynamically output. After the core evaluation dimension in the current scene is established, the diagnosis process enters the core link, and the cross-dimension diagnosis engine based on the causal diagram starts to work, which is mainly dependent on the pre-constructed system fault causal diagram stored in the system memory. The diagnosis engine injects the performance data collected in this test and weighted by the dynamic weight into the system fault causal diagram as observation evidence. These evidences show the abnormal state of some nodes in the graph. Then the diagnosis engine starts the Bayesian inference algorithm, which takes the data adjusted by the dynamic weight as the observation condition, and performs probabilistic reverse deduction along the causal dependency relationship defined in the system fault causal diagram, and calculates the probability of each potential root cause node at the bottom level being the real fault source when the abnormal performance of the high-level and intermediate-level is calculated. After the inference calculation is completed, the diagnosis engine sorts all candidate root causes according to their probabilities and generates a quantitative diagnosis report.

[0119] Correspondingly, the application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the above evaluation method.

[0120] Other configurations and operations of the intelligent networked vehicle digital twin evaluation method and system for extreme scenes according to the embodiments of the application are known to those skilled in the art, and will not be described in detail here.

[0121] In the description of the present application, the description of the terms "embodiment", "example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0122] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A digital twin evaluation method for intelligent connected vehicles in extreme scenarios, characterized in that, The assessment method includes the following steps: S1: Construct a knowledge graph based on unstructured accident reports in the existing database to build a knowledge graph of accident scenarios; S2: Generate extreme scenarios. Using clues sampled from the knowledge graph as seeds, generate test scenarios. Based on the functional copy of the autonomous driving system under test, the generated test scenarios run in the digital twin test platform and output their performance. Evaluate the performance and optimize the test scenarios based on the evaluation results. The resulting test scenarios are stored in the scenario library. S3: The digital twin testing platform generates a virtual environment based on extreme scenarios to test autonomous driving systems and output test data; S4: Real-time acquisition of vehicle status data from the digital twin testing platform to predict safety risk scores; if the safety risk score exceeds the safety threshold, an intervention command is sent to the digital twin testing platform. S5: Perform anomaly detection on the full test logs of the digital twin testing platform. If a performance inflection point is found, reconstruct the seed scenario, generate a derivative scenario through parameter perturbation, and inject it into the scenario library. S6: Real-time acquisition of feature codes of previous test scenarios in the digital twin test platform, calculation of weight vectors of evaluation indicators of the current test scenario, calculation of root causes based on weight vectors, system fault cause-effect graphs and test data, and generation of diagnostic results.

2. The evaluation method according to claim 1, characterized in that, Step S1 includes: S101: Extract scene elements from the existing database and convert them into standardized scene element vectors; S102: Based on standardized scene element vectors, construct the relationship between environment, entity, behavior and consequence to form a knowledge graph of accident scenarios.

3. The evaluation method according to claim 1, characterized in that, Step S2 includes: S201: Using clues sampled from the knowledge graph as seeds, based on policy... Generate test scenarios, These are weight parameters; S202: Based on a functional copy of the autonomous driving system under test, the generated test scenario is run in a digital twin test platform and its performance is output. The performance includes whether the autonomous driving system under test violates safety constraints and the types and number of safety constraints violated. S203: Using reward function Calculate the reward value, where, For the scene The overall reward value, Indicates a performance degradation reward. For the system under test in the scenario The success rate in This indicates that the security constraint violates the reward. For the first A violation indication function for a safety constraint. For the corresponding weights; Indicates a reward for the novelty of the scenario. For the scene With historical scene library Mid-scene Similarity, weight parameters , , Online optimization using the REINFORCE policy gradient algorithm satisfies the following requirements. ; S204: Update weight parameters based on overall reward value ,Strategy From weight parameters Parameterization , For learning rate, This is an estimate of the policy gradient calculated using the Monte Carlo method. It maximizes long-term expected reward by iteratively optimizing and generating extreme scenarios that can continuously challenge the tested system.

4. The evaluation method according to claim 1, characterized in that, Step S4 includes: S401: Acquire vehicle status data from the digital twin testing platform in real time and calculate a comprehensive risk score. , , Indicates the risk of collision time. This is the collision time threshold; Indicates the risk of deceleration. This represents the deceleration requirement at time t. Indicates the maximum permissible deceleration; Indicates comfort risk. This represents the lateral acceleration of the vehicle at time t. This indicates the maximum permissible lateral acceleration threshold; These are the normalized weight parameters; S402: Judgment Does it exceed the preset security threshold? When the preset safety threshold is exceeded, a real-time intervention command is sent to the digital twin testing platform based on a predefined standardized "intervention command protocol".

5. The evaluation method according to claim 4, characterized in that, Step S5 includes: S501: Perform streaming processing on the full test data, and use an anomaly detection algorithm based on isolated forest to calculate the anomaly score of the multi-dimensional performance index vector in the full test log data. , For the sample Abnormal scores; For the sample Path length in the isolation tree; Let be the expected value of the path length among multiple isolated trees. For a given subsample size Standardized path length at time For the first One harmonic number; S502: When abnormal scores are found When the performance inflection point detection threshold is exceeded, the data point is determined to be a performance anomaly. S503: Extract scene parameters within the time period before and after the corresponding performance anomaly point, reconstruct a reproducible seed scene, perform targeted perturbation and mutation on the key parameters of the seed scene, synthesize a derived scene, and inject the derived scene into the basic scene library.

6. The evaluation method according to claim 1, characterized in that, Step S6 includes: S601: Standardize the current test scenario to obtain the scenario feature encoding vector. The MLP network outputs a weight vector corresponding to the evaluation metric in the test scenario, and the MLP network output layer weight mapping function is used. , for The weight vector of each evaluation indicator For the scene The feature encoding vector, It is a multilayer sensing network. Ensure weight normalization; S602: The nodes of the constructed system fault cause-effect graph represent the performance indicators of each level of the autonomous driving system. When the top-level evaluation indicator is abnormal, the Bayesian inference algorithm uses a dynamic weight vector as the observation condition to perform probabilistic backward inference along the causal dependencies defined in the system fault cause-effect graph, and calculates the probability that each potential root cause node at the bottom level is the true source of the fault when the high-level and intermediate-level abnormal performance is as follows: , In order to observe abnormal evidence At that time, the root cause The posterior probability, root cause The prior probability, For the root cause When it happened, evidence was observed. The likelihood probability, The total number of candidate root causes; S603: Sort all candidate root causes by probability and generate a quantitative diagnostic report.

7. The evaluation method according to claim 6, characterized in that, Using historical test log sets and expert datasets as training datasets, the test scenarios in the training datasets are standardized into scene feature encoding vectors. The scene feature encoding vectors are composed of sub-vectors of environmental features, road topology features, traffic flow features, and key event identifiers. The historical test logs in the training datasets are used to construct target weight vectors using regression back-inference method. The target weight vectors of the expert data in the training datasets are taken from normalized expert assignments. The MLP network is trained using the training dataset by minimizing a composite loss function, which is: ， ， ， in, It is a composite loss function; For mean square error loss, The loss is based on cosine similarity, and the cosine similarity value ranges from 1 to 2. , , It is a hyperparameter and satisfies ; This is the weight vector predicted by the model; The target weight vector; The number of performance metrics.

8. A digital twin evaluation system for intelligent connected vehicles in extreme scenarios, characterized in that, The evaluation system is used to implement the digital twin evaluation method for intelligent connected vehicles in extreme scenarios as described in any one of claims 1 to 7, wherein the evaluation system includes a digital twin testing platform, an edge calculator, and a cloud server; The digital twin testing platform is used to generate virtual environments based on extreme scenarios to test autonomous driving systems and output test data. The cloud server includes: Multi-source accident data fusion module: Extracts scene elements from unstructured accident report databases and transforms them into standardized scene element vectors; Accident Scene Knowledge Graph Construction Module: Based on standardized scene element vectors, it constructs the relationship between environment, entity, behavior and consequence to form a knowledge graph of accident scenes; The adversarial scenario generation engine uses clues sampled from the knowledge graph as seeds to generate test scenarios, obtain the performance of the autonomous driving system under test running in the test scenarios of the digital twin test platform, evaluate the performance, and optimize the test scenarios based on the evaluation results. The scenario library update module performs anomaly detection on the full test logs of the digital twin testing platform. If a performance inflection point is found, the seed scenario is reconstructed, and a derivative scenario is generated by parameter perturbation and injected into the scenario library. The cross-dimensional root cause localization module acquires the feature codes of the previous test scenarios in the digital twin test platform in real time, calculates the weight vector of the evaluation indicators of the current test scenario, calculates the root cause based on the weight vector, the system fault cause-effect graph and test data, and generates diagnostic results. The edge calculator is used to acquire vehicle status data from the digital twin testing platform in real time to predict safety risk scores. If the safety risk score exceeds the safety threshold, an intervention command is sent to the digital twin testing platform.

9. The evaluation system according to claim 8, characterized in that, The cross-dimensional root cause localization module is equipped with an MLP network. The MLP network takes the scene feature encoding vector of the current test scene as input and outputs the weight vector of the evaluation index under the corresponding test scene.

10. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the evaluation method according to any one of claims 1-7.

Images