A method for rapidly generating digital scenes for data-real fusion testing

Abstract

This invention discloses a method for rapidly generating digital scenarios for data-real fusion testing, belonging to the field of digital testing and verification of high-end equipment. The method includes setting overall testing objectives, the functions and performance indicators of the object to be tested; preparing various basic unit scenario data required for realizing digital test scenarios in data-real fusion testing based on the testing objectives and the characteristics of the object to be tested, combined with the object testing process; when the data is insufficient to meet the requirements for generating high-precision scenario models, data augmentation operations are required; using datasets and two-dimensional diffusion models, a unit scenario generation model based on deep learning is constructed; based on the generated unit scenarios, a multi-element database of test scenarios and a set of test rule constraints are constructed; a comprehensive scenario test model is rapidly generated using deep reinforcement learning, and the complexity and coverage of the generated comprehensive test scenarios are evaluated and verified to meet the requirements of the overall testing objectives.

Description

Technical Field

[0001] This invention belongs to the field of digital testing and verification of high-end equipment, and specifically relates to a method for rapidly generating digital scenarios for data-real fusion testing. Background Technology

[0002] With the continuous development of digitalization and automation technologies, simulation testing has become a crucial link in modern system design and verification. Especially in complex scenarios such as unmanned systems, autonomous driving, and aerospace, real-world testing faces high costs and potential risks, making digital scenario simulation an important alternative. Traditional digital scenario generation methods are typically based on simple geometric modeling and physical simulation, which struggle to accurately capture the various elements and their interactions within complex scenarios, failing to meet the demands of high-precision simulation. Furthermore, existing methods for assessing the coverage and complexity of test scenarios are limited, making it difficult to effectively verify whether the generated scenarios fully meet testing requirements, further impacting the reliability and comprehensiveness of test results. Therefore, there is an urgent need for a method that can rapidly generate high-fidelity, high-coverage complex digital test scenarios, adaptable to changing testing needs, and automatically verifying the rationality and effectiveness of the scenarios, thereby accelerating the verification process for the functionality, performance, safety, and reliability of the test objects. Summary of the Invention

[0003] This invention aims to address the problems of insufficient accuracy, complexity, and difficulty in evaluating coverage in existing digital test scene generation technologies, and proposes a rapid method for generating digital test scenes that integrates data and reality. This method enables the automatic generation of high-fidelity 3D scenes from multiple data sources, and optimizes the scene generation process through deep learning and deep reinforcement learning algorithms, thereby meeting the functional and performance requirements of different types of test objects. The technical problem solved by this invention is achieved through the following technical solution: A rapid method for generating digital test scenes that integrates data and reality, comprising the following steps:

[0004] Step 1: Design the test scenario data construction module. This module, based on the purpose and requirements analysis of data-real fusion testing, generates various basic data required for model training. Specifically, this includes: Test objective setting: Based on the type, function, and performance characteristics of the test object, setting the overall test objective to ensure that the generated test scenario covers all key requirements of the test object; Basic data preparation and enhancement: Based on the test objective and the characteristics of the test object, and combined with the test process, preparing the basic unit scenario data required for data-real fusion testing, including geographic, terrain, meteorological, physical performance, model data, and road network data. When existing data is insufficient to meet the needs of generating a high-precision scenario model, data augmentation operations, such as combining test rule embedding, data annotation, and data expansion, are used to expand the diversity and accuracy of the dataset.

[0005] Step 2: Design a unit-level scene model generation module. This module generates unit-level models for various test scenarios required during the testing process. The specific implementation is as follows: Unit Scene Generation: Based on the prepared dataset, feature injection for the test scenario is performed on the 2D diffusion model to achieve transfer training; multi-angle images of each unit-level scene are generated using prompts input to the 2D image diffusion model, and images lacking consistency are filtered out; the SFM algorithm is used to process multi-view image data for light field reconstruction, encoding the 3D point cloud space and viewpoint relationship of the scene; the scene is divided into multi-scale voxel grids using an adaptive meshing method, dynamically adjusting the resolution to adapt to complex areas; a layered volumetric rendering method is used to render optical features at different depths to achieve high-fidelity scene generation; density calculation and rendering methods in 3D Gaussian splashing are used to generate each unit-level 3D scene.

[0006] Step 3: Design a comprehensive scenario model generation module. This module uses rule constraints to form a comprehensive scenario model that conforms to real-world situations and patterns, based on various unit-level scenario models required during the testing process. Specifically, this is implemented as follows: Multi-element database and rule constraint set construction: Based on the generated unit scenarios, a multi-element database for the test scenario is constructed. This database covers scenario environment variables, test object parameters, and dynamically changing data. Simultaneously, a test rule constraint set is constructed, including expert knowledge sets and explicit and implicit constraint sets, to ensure that the test scenario meets multi-faceted testing standards. Deep reinforcement learning optimization: Using deep reinforcement learning algorithms, a preliminary comprehensive test scenario with various complex scenario combinations is generated through an adaptive sampler. Through multiple iterations of evaluation and adjustment, the preliminary scenario model is further optimized, ultimately generating a comprehensive test scenario model including standard test scenarios, extreme test scenarios, rare test scenarios, and edge test scenarios. Complexity and coverage verification: The complexity and coverage of the generated comprehensive test scenario are verified. Verification is performed using complexity and coverage evaluation algorithms. Evaluation criteria include the complexity of the scenario's environment type, the complexity of object behavior, and the breadth and depth of the test scenario's coverage.

[0007] This invention offers the following beneficial technical effects: The system architecture of this invention includes a test scenario data construction model module, a unit-level scenario model generation module, a unit scenario generation module, and a comprehensive scenario model generation module. These modules collaborate to achieve efficient generation and verification of digital scenarios. Through this invention, high-fidelity, high-complexity digital test scenarios can be rapidly constructed, meeting the testing needs of various complex systems and improving the accuracy, breadth, and reliability of the testing environment. Attached Figure Description

[0008] Figure 1This is a schematic diagram of a rapid digital scene generation system for data-real fusion testing according to the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0010] This invention proposes a method for rapidly generating digital scenes for data-real fusion testing, such as... Figure 1 As shown, the method includes:

[0011] Step 1: Design the test scenario data construction module. Based on the purpose and requirements analysis of the data-real fusion test, generate various basic data required for model training. The specific implementation is as follows:

[0012] Analyze the test objects and set overall test objectives based on their type, function, and performance characteristics to ensure that the generated test scenarios cover all key requirements of the test objects. This includes: proposing test requirements / objectives and performance indicators, and developing a preliminary test implementation plan; preparing various basic unit scene data required for digital test scenarios in data-real fusion testing based on the test objectives and characteristics of the test objects, combined with the object testing process, including geographic, terrain, meteorological, physical performance, model data, and road network data; when the data is insufficient to meet the needs of generating high-precision scene models, data augmentation operations are required, such as combining test rule embedding, data annotation, and data expansion to expand the diversity and accuracy of the dataset and construct a dataset.

[0013] Step 2: Design a unit-level scene model generation module to quickly generate unit-level models of various test scenarios required by the test object during the testing process. The specific implementation is as follows:

[0014] Using the constructed dataset, feature injection in test scenarios was performed on the image diffusion model for image generation to achieve transfer training. Multi-angle images of each unit-level scene were generated using prompts input to the image diffusion model, and images lacking consistency were filtered out. The SFM algorithm was used to process the multi-view image data for preliminary scene reconstruction, obtaining the 3D point cloud space and camera extrinsic parameters of the coded scene. Based on the characteristics of the point cloud scene, an adaptive point cloud partitioning method was used to divide the scene into multi-scale point cloud regions, dynamically adjusting the resolution to adapt to complex regions. A layered volumetric rendering method was used to render optical features at different depths, achieving high-fidelity scene generation. Density calculation and rendering methods from 3D Gaussian splashing were used to generate each 3D unit-level scene.

[0015] Step 3: Design a comprehensive scene model generation module. This module uses rule constraints to form a comprehensive scene model that conforms to real-world situations and patterns, based on the various unit-level models required during the testing process of the test object. The specific implementation is as follows:

[0016] Based on the generated unit-level scenarios and test requirements, a multi-element database of test scenarios and a set of test rule constraints are constructed. This database covers scenario environment variables, test object parameters, and dynamically changing data. The set of test rule constraints includes expert knowledge sets, explicit and implicit constraint sets. Using deep reinforcement learning, combined with the set of test rule constraints and test objectives, an adaptive sampler selects the optimal scenario elements and their relative positional relationships to generate preliminary comprehensive test scenarios with various complex scenario combinations. Through multiple iterations of evaluation and adjustment, a comprehensive scenario test model is rapidly generated. Through further iterations of evaluation and adjustment, the preliminary scenario model is further optimized, ultimately generating a comprehensive test scenario model that includes standard test scenarios, extreme test scenarios, rare test scenarios, and edge test scenarios. Finally, the complexity and coverage of the comprehensive test scenarios are evaluated and verified using complexity and coverage evaluation algorithms. The evaluation criteria include the complexity of the scenario's environment type, the complexity of object behavior, and the breadth and depth of the test scenario's coverage.

[0017] Furthermore, the basic unit-level scene data in step 1 includes, but is not limited to, geographic environment data, terrain data, meteorological data, physical characteristic data, object function and performance data, model data, road network data, and related historical test data.

[0018] Furthermore, the data augmentation operations in step 1 include combining test rule embedding, data annotation, and data expansion to expand the diversity and accuracy of existing data and build a dataset that covers a wide range of scenario variables.

[0019] Furthermore, the consistency in step 2 includes geometric consistency, lighting consistency, and material consistency, ensuring that the generated 3D scene conforms to physical laws.

[0020] Furthermore, the test rule constraint set in step 3 includes an expert knowledge set, a test rule set, an explicit constraint set, and an implicit constraint set to ensure that the test scenario meets various test requirements and standards.

[0021] Furthermore, the comprehensive scenarios in step 3 include various comprehensive test scenario models, such as standard test scenarios, extreme test scenarios, rare test scenarios, and edge test scenarios.

[0022] Furthermore, the complexity and coverage in step 3 include the complexity of the scenario's environment type, the complexity of object behavior, and the breadth and depth of the test scenario's coverage.

[0023] More specifically, the present invention provides the following illustrative embodiments. The overall test objective is set according to the type, function, and performance of the test object.

[0024] The test subjects may be autonomous vehicles, drones, or other equipment systems with complex functions.

[0025] The overall objective defines the scenario conditions required for testing, including normal and extreme operating scenarios, testing requirements for different environmental variables (such as weather, terrain, etc.), and performance evaluation of the system under various conditions.

[0026] Setting test objectives is fundamental to the scenario generation process, ensuring that the final generated test scenarios meet the requirements for expected functionality and performance verification.

[0027] According to the requirements of the testing objectives, the system needs to prepare various basic unit scene data. These data include, but are not limited to, geographic data, terrain data, meteorological data, model data, road network data, and building data.

[0028] By analyzing the functional and performance requirements of the test objects, a suitable dataset is selected for preliminary processing.

[0029] Terrain data can come from actual surveying data, meteorological data can include historical meteorological data and future weather forecast information, and model data may cover three-dimensional models of objects such as vehicles and drones in the physical scene and their dynamic characteristics.

[0030] In some cases, existing datasets may not be able to fully meet the requirements for generating high-fidelity scene models.

[0031] Therefore, data augmentation is necessary. Data augmentation can be achieved through methods such as embedding test rules, data annotation, and data expansion.

[0032] For example, in scenarios where data is scarce, the system can incorporate rule embedding, such as simulating scenarios without actual data through physical modeling or synthetic data generation methods; it can also improve the accuracy and diversity of scenario generation by labeling existing data and expanding the dataset.

[0033] Data augmentation ensures that highly complex test scenarios can still be generated even with incomplete or low-quality data.

[0034] Using a two-dimensional dataset in a flight scenario, the diffusion model is transferred and trained, and drone test-related features (such as wind speed, temperature, and terrain) are injected into the model.

[0035] Multi-angle images of unit-level scenes are generated based on prompt words, such as simulating the flight status of drones in different time periods and weather conditions.

[0036] The generated images are filtered using a uniform distillation method to ensure consistency in geometry, lighting, and materials, making the scene conform to physical laws.

[0037] After data preparation and augmentation, an algorithm is used to generate a high-precision 3D unit scene. NeRF generates a 3D representation of the scene by inputting 2D images from multiple perspectives into a neural network.

[0038] Specifically, the algorithm uses a volumetric rendering method to encode the geometry and lighting of the scene.

[0039] By predicting the density and color of each voxel in the voxel mesh, high-precision reconstruction of various objects and their details in the scene can be achieved.

[0040] For example, in complex urban environments, NeRF can capture the appearance of buildings, the details of roads, and the occlusion relationships between objects, and generate a 3D scene that is highly consistent with the actual scene.

[0041] Based on the generated unit scenarios, the system constructs a multi-factor database and creates a set of rule constraints for the test scenarios. The multi-factor database includes scene environment variables, dynamic data (such as the speed and acceleration of moving objects), the model of the test object, and its physical parameters. These elements form the basis for the subsequent generation of comprehensive scenarios.

[0042] In addition, the rule constraint set includes the expert knowledge set, the test rule set, the explicit constraint set, and the implicit constraint set.

[0043] The expert knowledge set sets the physical constraints for scenario generation based on expert experience, such as the maximum centrifugal force when a vehicle turns; the test rule set defines the operational behavior of the test object in the scenario, such as the route planning and reaction time of an autonomous vehicle; the explicit constraint set sets explicit conditions in the scenario, while the implicit constraint set covers those scenario rules that cannot be directly observed, such as implicit system vulnerabilities.

[0044] After the comprehensive test scenario is generated, the system verifies its complexity and coverage.

[0045] First, complexity verification involves analyzing the comprehensive scene element model and calculating the weights of each element in the scene, then using the information entropy method to assess the scene's complexity. The specific formula for information entropy is as follows:

[0046] ,

[0047] ,

[0048] ,

[0049] Where H(X) represents the information entropy of information source X, and H(XY) represents the joint entropy of X and Y. This represents the information entropy of the system.

[0050] This process can quantify the interrelationships and uncertainties among multiple variables in a scenario, ensuring that the scenario complexity meets the needs of actual applications.

[0051] The coverage verification process includes test requirements analysis, selection of effective test scenarios, and coverage calculation.

[0052] The system first analyzes the test objectives and selects test scenarios closely related to the objectives. Then, by evaluating the capabilities of each test scenario, it calculates the coverage of the scenario tests to ensure that the generated test scenarios can effectively cover all functional and performance requirements, thus guaranteeing the comprehensiveness and reliability of the tests.

[0053] The system architecture of this invention includes a data preparation module, a data augmentation module, a unit scene generation module, a rule constraint module, a deep reinforcement learning module, and an evaluation and verification module. These modules cooperate with each other to achieve efficient generation and verification of digital scenes.

[0054] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0055] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for quickly generating a digital scene for a number-real hybrid test, the method comprising: receiving a digital scene description; and generating a number-real hybrid test digital scene based on the digital scene description. Comprise: Step 1, design test scene data construction module, the module for the purpose of numerical and real fusion test and demand analysis, realize the generation of various basic data required for model training, the specific implementation is as follows: analyze the test object, put forward the test demand or target, performance index, and formulate the preliminary test implementation scheme; according to the test target and the characteristics of the object to be tested, combining with the object test process, realize the preparation of various basic unit scene data required for digital test scene in numerical and real fusion test; when the data is not enough to meet the demand of scene model generation, carry out data enhancement operation; Step 2, design unit level scene model generation module, the module generates various unit level models required for the test object in the test process, the specific implementation is as follows: using the constructed data set, the image diffusion model for realizing image generation is tested under the characteristics injection of test scene to realize migration training; use the prompt words input to the image diffusion model to generate multi-angle pictures of each unit level scene, and exclude the pictures lacking of consistency; adopt SFM algorithm to process multi-view image data, carry out preliminary scene reconstruction, get three-dimensional point cloud space and camera external parameter of coded scene; according to the characteristics of point cloud scene, the scene is divided into multi-scale point cloud area by using adaptive point cloud division method, and the resolution is dynamically adjusted to adapt to complex area; use the density calculation and rendering method in three-dimensional Gaussian splashing to realize the generation of each three-dimensional unit level scene; Step 3, design integrated level scene model generation module, the module uses rule constraint to form integrated level scene model conforming to the reality and law by using the unit level scene model required for the test object in the test process, the specific implementation is as follows: on the basis of the generated unit level scene and test demand, carry out test scene multi-element database construction and test rule constraint set construction; use deep reinforcement learning, combine with test rule constraint set and test target, carry out rapid generation of integrated level scene test model conforming to demand; finally, evaluate and verify the complexity and coverage of integrated level test scene; The basic unit level scene data in step 1 includes but is not limited to geographic environment data, terrain data, meteorological data, physical property data, object function and performance data, model data, road network data and related historical test data; The data enhancement operation in step 1 includes embedding combined with test rules, data labeling, data expansion method, expanding the diversity and accuracy of existing data, and constructing data set covering a wide range of scene variables.

2. The method of claim 1, wherein, The consistency in step 2 includes geometric consistency, lighting consistency and material consistency, which ensures that the generated three-dimensional scene conforms to the physical law.

3. The method of claim 1, wherein, The test rule constraint set in step 3 includes expert knowledge set, test rule set, explicit constraint set and implicit constraint set, so as to ensure that the test scene conforms to the test requirements and standards in many aspects.

4. The method of claim 1, wherein, The integrated level test scene in step 3 includes various integrated test scene models such as standard test scene, extreme test scene, rare test scene and edge test scene.

5. The method of claim 1, wherein, The complexity and coverage of the integrated level test scene in step 3 include the complexity of the environment type of the scene, the complexity of the object behavior, and the coverage breadth and coverage depth of the test scene.

Images