Method and system for creating a user-specific virtual dummy model

Abstract

The invention relates to a method for creating a user-specific virtual dummy model (50) for carrying out computer-aided crash test simulations for the certification and improvement of design parameters of parts and components of a vehicle, comprising the following method steps: - Recording time series of stress values ​​of a real dummy (10) and of time series of an impactor (20) using sensors (250) and storing the time series as input data (350); - Processing the input data (350) with a software application (450) for dimension reduction in order to convert the input data (350) into simpler numerical values ​​(features); - Transferring the converted input data (350) to a trained AI model (430), wherein the AI ​​model (430) is at least designed as an AI regression model; - Calculating a user-specific virtual dummy model (50) with at least one fixed value for at least one input parameter (55), wherein an input parameter (55) relates to at least one property of the real dummy (10) or at least one experimental boundary condition of a certification test of the dummy (10), - Outputting the user-specific virtual dummy model (50) as output data (550).

Description

[0001] The invention relates to a method, a system and a computer program product for creating a user-specific virtual dummy model for carrying out computer-aided crash test simulations for the certification and improvement of design parameters of parts and components of a vehicle.

[0002] Crash tests are an essential part of vehicle safety certification and assessment, used to ensure that a vehicle provides an adequate level of safety for occupants and other road users in the event of a collision. A specific vehicle model is selected for the crash test. The vehicle is equipped with sensors and dummies to record relevant data during the test. The dummies are positioned in the seats to represent human occupants in various configurations (adults, children, driver, front passenger, etc.).

[0003] There are various types of crash tests, such as frontal impact, side impact, rear impact, and rollover. Each test simulates different accident scenarios. The crash test is conducted at specific speeds and angles that correspond to realistic traffic accident conditions. The vehicle collides with a barrier or another vehicle under precisely defined impact conditions. Sensors measure the forces, accelerations, and movements of the vehicle and the dummies during and after the impact, while high-speed cameras visually record the events.

[0004] After the impact, the damage to the vehicle, such as deformations, the behavior of safety devices like airbags, and the crash test dummies are analyzed. Based on the test data and analysis, a decision is made as to whether the vehicle meets the safety requirements. If not, modifications to the design must be made and further tests conducted. Upon successful evaluation, the vehicle receives a certificate confirming its conformity with the relevant safety standards.

[0005] A vehicle crash test dummy is a specially designed test device used in automotive crash tests to simulate the effects of an accident on the human body. The dummy has a human-like shape and corresponds to an average person in size and weight. There are different types of dummies for adults, children, and infants. The dummy is made of a combination of different materials that simulate various body parts. Typically, the outer layers are made of durable plastics that behave similarly to human skin and muscles. The internal structures are made of metal, plastic, or rubber to replicate the body's bones, joints, and soft tissues. The dummies are equipped with joints similar to those of the human body. They can move at the shoulders, elbows, wrists, hips, knees, and ankles.These joints are designed to exhibit realistic mobility and flexibility.

[0006] An essential component of the crash test dummy is its integrated sensors. These measure various forces and stresses exerted on the dummy during a crash test. The most commonly used sensors include accelerometers, strain gauges, and pressure sensors. These sensors provide detailed data on the stresses placed on the head, chest, spine, limbs, and other body parts. The sensors in the dummy are connected to an electronic system that collects and stores the data during the crash test. The wiring and electronics are designed to withstand the extreme stresses of a crash test. Crash test dummies are used in various scenarios to simulate different types of accidents, such as frontal, side, rear, and rollover impacts. The dummies are installed in vehicle seats and secured accordingly to simulate realistic scenarios.The data collected during crash tests with dummies is used to improve vehicle safety. The data is analyzed to identify weaknesses in the vehicle and to optimize safety features such as airbags, seat belts, and body structures.

[0007] However, differences (variances) occur in the properties of the dummies. These variations can be attributed to various factors such as material properties or geometric differences. This means that not every dummy of a given type is exactly the same. This variability in the dummy properties leads to differences in crash test results. The certification procedures for crash test dummies take these variations into account by allowing relatively large tolerance limits, known as corridors. This means that a particular result is considered acceptable as long as it falls within these limits, even if a certain degree of variability exists.

[0008] Occupant protection in vehicles is typically developed using computer simulations with a standardized virtual dummy model. Simulations are playing an increasingly important role because they are more cost-effective and deliver results quickly. A virtual dummy model is a digital or computer-generated replica of an average real-world crash test dummy, used in simulation environments to test and analyze vehicle accidents. These dummy models are essential tools in vehicle safety, enabling the virtual simulation of various aspects of crash tests, as they can be easily adapted to different crash test scenarios. For example, different speeds, impact angles, or occupant positions can be varied to examine in detail the forces and stresses acting on the dummy during a crash.Virtual dummy models allow for numerous test runs without additional material costs, thus complementing physical tests by providing additional insights and making the testing process more efficient. This, in turn, enables more efficient and effective vehicle development and safety analysis.

[0009] To ensure that computer-generated crash test simulations accurately reflect reality, the crash test simulation results undergo validation. This validation compares the simulation results with those from real-world crash tests. It covers aspects such as vehicle deformation, acceleration, impact forces, occupant injury criteria, and more. The goal of the validation is to verify that the parameters used in the simulation (material properties, vehicle geometry, occupant positioning, etc.) are correct and accurate. The validation also checks whether all numerical methods used in the simulation are reliable and appropriate, and whether the assumptions and boundary conditions used in the simulation correspond to those in real-world crash tests.

[0010] This standardized virtual dummy model is frequently used to validate simulation results, even though differences exist between the properties of the virtual dummy model and a real hardware dummy. These differences can affect the accuracy and reliability of the simulation results.

[0011] US Patent 2018 / 0285516 A1 discloses a system for generating a user-specific finite element model for a crash test dummy, wherein a plurality of finite element factors and a performance parameter associated with a crash test dummy are selected, wherein a performance parameter value is determined as a function of the plurality of finite element factors, wherein an optimized factor value is determined for each of the finite element factors to minimize a difference between the performance parameter value and a certification test value, and wherein a user-specific finite element model for the crash test dummy is generated using the determined optimized values ​​for each of the finite element factors.

[0012] The invention is based on the objective of providing possibilities to improve a user-specific virtual dummy model for virtual computer simulations of crash tests with a vehicle for certification purposes in such a way that real hardware dummies are better represented by the user-specific dummy model, so that the computer simulations of crash tests carried out with the improved virtual dummy model are more reliable and accurate.

[0013] This problem is solved according to the invention with respect to a method by the features of claim 1, with respect to a system by the features of claim 8, and with respect to a computer program product by the features of claim 15. The further claims relate to preferred embodiments of the invention.

[0014] The creation of a user-specific virtual dummy model according to the invention using an AI regression model enables more accurate prediction of the results of a virtual crash test, since the input parameters of the user-specific dummy model are known. Furthermore, the validation of computer-aided crash test simulation models becomes more accurate and reliable because the uncertainties regarding the virtual dummy model are already reduced. Since the AI ​​regression model is trained with existing knowledge about the varying properties of hardware dummies and experimental boundary conditions, as well as simulated data from stochastically selected virtual dummy models, each with different varying input parameters and properties, these variations can be automatically taken into account. Overall, this leads to improved accuracy and reliability of crash simulations.This allows design parameters such as material selection, component thickness, shape, etc. of vehicle components to be more precisely adjusted and optimized in order to improve the crash safety of vehicle components and thus increase the safety and performance of vehicle bodies.

[0015] According to a first aspect, the invention provides a method for creating a user-specific virtual dummy model for conducting computer-aided crash test simulations for the certification and improvement of design parameters of parts and components of a vehicle. The method comprises the following process steps: - Recording time series of stress values ​​of a real dummy and time series of an impactor using sensors and storing the time series as input data; - Processing the input data with a dimension reduction software application to convert the input data into simpler numerical values ​​(features); - Transferring the converted input data to a trained AI model, where the AI ​​model is at least designed as an AI regression model; - Calculating a user-specific virtual dummy model with at least one fixed value for at least one input parameter, where an input parameter relates to at least one property of the real dummy or at least one experimental boundary condition of a certification test of the dummy; - Outputting the user-specific virtual dummy model as output data.

[0016] In a further training course, it is planned that the AI ​​model will use linear regression, polynomial regression, decision trees, random forest, neural networks, support vector machines (SVM) and / or gradient tree boosting, in particular XGBoost as algorithm(s).

[0017] In an advantageous embodiment, it is provided that sliding means, the determination of a gradient and / or maximum and minimum values, and a principal component analysis (PCA) are performed for data dimension reduction, wherein PCA models for the different load values ​​of the dummy are trained with a large amount of experimental and simulated data in order to identify the principal components in the input data.

[0018] In a further embodiment, it is provided that training data is used to train the AI ​​model, wherein the training data comprises first data and second data, wherein the first data comprise time series of load values ​​for a plurality of real dummies and time series of at least one impactor from a plurality of certification tests of the plurality of real dummies, wherein the plurality of real dummies represent different individual versions of a particular dummy type and each differ from one another in terms of input parameters and properties such as the material thickness of individual components, and wherein the second data comprise simulation-generated time series of load values ​​for several variants of a virtual dummy model, each with different varying input parameters and properties.wherein the simulation-generated time series are determined in a large number of simulation-generated certification tests of the variants of the virtual dummy model, and wherein the variants of the dummy model are generated by a stochastic simulation within the framework of a sampling from probability distributions of the scattering input parameters and properties of the dummy.

[0019] In particular, at least one load value represents an acceleration of the dummy or one of its parts, a direct force acting on the dummy, a pressure exerted on a specific area of ​​the dummy, a deformation of body parts of the dummy, a torque acting on the dummy, a speed at which the dummy or a part of it moves.

[0020] Advantageously, a time series of the impactor represents the force exerted by the impactor on the dummy, the speed and acceleration of the impactor upon impact with the dummy, and the change in the spatial position of the impactor.

[0021] In particular, at least one design parameter is a dimension, a material, a shape, a manufacturing process, a surface finish, a density, a temperature resistance, a tensile strength, a compressive strength, a flexural strength, a fracture behavior, a plastic deformation, a characteristic value, a measured value, a safety factor and / or a material property.

[0022] According to a second aspect, the invention provides a system for creating a user-specific virtual dummy model for conducting computer-aided crash test simulations for the certification and improvement of design parameters of vehicle parts and components. The system comprises an input module, a database, a dummy model creation module with an AI model, and an output module. The input module is configured to acquire time series of load values ​​from a real dummy and time series from an impactor using sensors. The database is configured to store the time series as input data. A dimensionality reduction software application is configured to convert the input data into simpler numerical values ​​(features) and to transfer the converted input data to a trained AI model.The AI ​​model is configured, at least as an AI regression model, to compute a user-defined virtual dummy model with at least one fixed value for at least one input parameter, where an input parameter relates to at least one property of the real dummy or at least one experimental boundary condition of a certification test of the dummy. The output module is configured to output the user-defined virtual dummy model as output data.

[0023] In a further training course, it is planned that the AI ​​model will use linear regression, polynomial regression, decision trees, random forest, neural networks, support vector machines (SVM) and / or gradient tree boosting, in particular XGBoost as algorithm(s).

[0024] In an advantageous embodiment, it is provided that sliding means, the determination of a gradient and / or maximum and minimum values, and a principal component analysis (PCA) are performed for data dimension reduction, wherein PCA models for the different load values ​​of the dummy are trained with a large amount of experimental and simulated data in order to identify the principal components in the input data.

[0025] In a further embodiment, it is provided that training data is used to train the AI ​​model, wherein the training data comprises first data and second data, wherein the first data comprise time series of load values ​​for a plurality of real dummies and time series of at least one impactor from a plurality of real certification tests of the plurality of real dummies, wherein the plurality of real dummies represent different individual versions of a particular dummy type and each differ from one another in terms of input parameters and properties such as the material thickness of individual components, and wherein the second data comprise simulation-generated time series of load values ​​for several variants of a virtual dummy model, each with different varying input parameters and properties.wherein the simulation-generated time series are determined in a large number of simulation-generated certification tests of the variants of the virtual dummy model, and wherein the variants of the dummy model are generated by a stochastic simulation within the framework of a sampling from probability distributions of the scattering input parameters and properties of the real dummy.

[0026] In particular, at least one load value represents an acceleration of the dummy or one of its parts, a direct force acting on the dummy, a pressure exerted on a specific area of ​​the dummy, a deformation of body parts of the dummy, a torque acting on the dummy, a speed at which the dummy or a part of it moves.

[0027] Advantageously, a time series of the impactor represents the force exerted by the impactor on the dummy, the speed and acceleration of the impactor upon impact with the dummy, and the change in the spatial position of the impactor.

[0028] In particular, at least one design parameter is a dimension, a material, a shape, a manufacturing process, a surface finish, a density, a temperature resistance, a tensile strength, a compressive strength, a flexural strength, a fracture behavior, a plastic deformation, a characteristic value, a measured value, a safety factor and / or a material property.

[0029] According to a third aspect, the invention provides a computer program product comprising an executable program code configured to perform the method according to the first aspect when executed.

[0030] The invention will now be explained in more detail with reference to exemplary embodiments shown in the drawing.

[0031] This shows: Fig. 1. A perspective view of a thorax certification test for a dummy in a vehicle; Fig. 2a a representation of a force-time diagram of the force applied to the dummy Fig. 1 acting force with different curve profiles in various crash tests; Fig. 2b a representation of a displacement-time diagram for a first rib of the dummy made of Fig. 1 with different curve profiles in different crash tests; Fig. 2c a representation of a displacement-time diagram for a second rib of the dummy made of Fig. 1 with different curve profiles in different crash tests; Fig. 2d a representation of a displacement-time diagram for a third rib of the dummy made of Fig. 1 with different curve profiles in different crash tests; Fig. 3 a perspective view of part of the first rib in the arm area of ​​a dummy; Fig. 4a a histogram showing the variation in the material thickness of a rib in a dummy; Fig. 4b a diagram showing the normal distribution of wall thickness for different materials; Fig. 5 a block diagram with a schematic representation of a system according to the invention; Fig. 6 a flowchart to explain the individual process steps of a process according to the invention; Fig. 7 a block diagram of a computer program product according to an embodiment of the third aspect of the invention.

[0032] Additional features, aspects and advantages of the invention or its embodiments become apparent from the detailed description in conjunction with the claims.

[0033] The use of crash simulation results in the design of body components is a crucial step in the development process, ensuring, for example, that a vehicle meets the required safety standards and that occupants are optimally protected in the event of an accident. Crash simulation results are used to evaluate and validate various design concepts for body components. By simulating crash scenarios, it is possible to determine which design concepts offer the best performance in terms of structural integrity and safety. Based on the simulation results, the effects of different materials and structures on the crash behavior of body components can be assessed. This allows for the selection of suitable materials and structures to, for example, maximize the strength and stiffness of the components while simultaneously minimizing vehicle weight.This includes creating and optimizing design parameters such as a component's wall thickness to improve structural integrity while simultaneously optimizing vehicle weight. Simulation results are also used to further optimize and improve existing body components. Various design parameters, such as shape, thickness, and joining methods, can be precisely calibrated to optimize component performance and maximize safety in the event of an accident. The use of crash simulations can significantly reduce development time and costs. Instead of building physical prototypes and conducting costly crash tests, virtual crash tests can be performed, and various scenarios can be evaluated quickly and cost-effectively.

[0034] The term "design" in connection with the present invention refers, on the one hand, to the process of planning, designing, and shaping a component, such as a connecting element for a crossmember of a vehicle body. The design process involves the application of specific knowledge and experience to solve problems and develop solutions that meet the requirements of the component. On the other hand, the result of this process, the finished component itself, is also referred to as a design.

[0035] The term "design design" refers to the specific plan or detailed representation of a design created during the development process. A design design is essentially a visual or written representation of how the final component will look and function. It can include technical drawings, schematic diagrams, models, prototypes, or other forms of representation that describe the various aspects of the design in detail, including design parameters such as dimensions, materials, functionalities, and potentially manufacturing processes. The design design enables a detailed understanding of the design, the anticipation of potential problems, the implementation of modifications, and the efficient and precise execution of manufacturing.

[0036] A "design parameter" can be a dimension, a material, a shape, a manufacturing process, a surface finish, a density, a temperature resistance, a tensile strength, a compressive strength, a flexural strength, a fracture behavior, a plastic deformation, a characteristic value, a measured value, a safety factor and / or a material property.

[0037] Fig. Figure 1 shows a perspective view of a setup for a thorax certification test with a dummy 10 in a vehicle. A thorax certification test is a specific test method used in crash tests to assess the strength and safety of the chest area (thorax) of dummies 10. This type of test is used to ensure that the dummies 10 are properly calibrated in crash tests and deliver realistic results by measuring the mechanical stress on the thorax (chest area) under simulated force.

[0038] In this test, the dummy 10 is subjected to specific forces using an impactor 20, usually by striking or pressing on the chest area. The dummy 10's reaction to these forces is then analyzed to ensure that it corresponds to the expected values.

[0039] The Impactor 20 is designed to apply a specific force to perform the desired test. The shape and size of the Impactor 20 can be varied to simulate different types of force application. Fig. The Impaktor 20 has a flat surface, but can also have a round or specially shaped surface. The Impaktor 20 can be precisely controlled to achieve a specific force, speed, or position. This enables consistent and reproducible testing.

[0040] Fig. Figure 2a shows a force-time diagram of the force exerted on dummy 10 by impactor 20, with different curve shapes for various crash tests. As can be clearly seen, the magnitude of the force acting on dummy 10 varies within a range around a mean value µ.

[0041] Fig. Figure 2b shows a representation of a displacement-time diagram for a first rib of dummy 10 with different curve profiles in different crash tests. Fig. Figure 2c shows a representation of a displacement-time diagram for a second rib of dummy 10 with different curve profiles in different crash tests, and Fig. Figure 2d shows a displacement-time diagram for the third rib of dummy 10, with different curve shapes for various crash tests. In the diagrams, the maximum displacement of the respective rib varies within a range around a mean value µ.

[0042] Thorax certification is part of a broader certification process for crash test dummies. In addition to the thorax certification test, which assesses the function and accuracy of the chest area, there are other certification tests that vary depending on the dummy type and the focus of the investigation. For example, the head certification test examines how the dummy's head reacts to forces, measuring parameters such as acceleration and force transmission. The neck certification test assesses the mobility and resilience of the dummy's neck, measuring how the neck responds to various forces and moments. The pelvic certification test verifies whether the dummy's pelvis reacts appropriately under different loads. Finally, the leg and knee certification test examines how the dummy's legs and knees respond to various forces and deformations.The shoulder certification test assesses the mechanical stability of the shoulder area and how it reacts to stresses that can occur during various types of impacts.

[0043] There are specific norms and standards that define how these tests are to be carried out and which results are considered acceptable. Certification tests serve to verify the mechanical and sensory integrity of crash test dummies before they are used in crash tests.

[0044] Fig. Figure 3 shows a perspective view of part of the first rib in the arm area of ​​dummy 10. The first rib in a person's rib cage, located near the arm, is the uppermost of the twelve pairs of ribs and connects the sternum to the spine. It lies directly below the clavicle and, together with the clavicle and other structures, forms the upper boundary of the rib cage. Due to its position, the first rib plays a crucial role in the stability of the upper rib cage and is part of the anatomical structures that connect the arm to the torso.

[0045] A force transducer 12 is arranged between the arm 11 of the dummy 10, which transmits the force acting on the arm 11 to an outer rib 14. The outer rib 14 includes a rib reinforcement 15. An inner rib 17 and damping material 18 are arranged in the inner region of the outer rib.

[0046] The variation in the stiffness of the ribs in the fuselage of dummy 10 is due to the large variability of the geometry resulting from the different wall thicknesses of the materials used for the ribs, such as metals, plastics and elastomers used for structural elements, joints and coverings, which leads to a variation in the load values ​​of dummy 10.

[0047] Fig. Figure 4a shows a histogram 30 with variations in the material thickness of the damping material 18. The mean µ of the material thickness in this example is µ = 1.62 mm and the standard deviation σ is σ = 0.05. The distribution of the wall thicknesses is essentially a normal distribution 32. As in Fig. As shown in Figure 4b, the normal distribution 32 can be different for different materials with respect to material thickness, i.e. some materials have a lower dispersion than others.

[0048] By drawing samples from the normal distributions 32, for example for the different materials used, the scattering numerical properties and parameters of the dummy 10 can be integrated into a stochastic simulation.

[0049] In order to take these distributions 32 into account for the various properties and parameters of a real hardware dummy 10, such as in particular the wall thicknesses of the different materials for the different components of the hardware dummy 10, in a user-specific virtual dummy model 50, a trained AI model 430 is used according to the invention to create a user-specific virtual dummy model 50 with defined input parameters 55.

[0050] Fig. Figure 5 shows a system 100 according to the invention for creating a user-specific virtual dummy model 50 with an input module 200, a database 300, a dummy model creation module 400 with an AI model 430 and an output module 500.

[0051] The input module 200, the database 300, the dummy model creation module 400 and the output module 500 can each be equipped with a processor and / or a memory unit.

[0052] In the context of the invention, a "module" is defined as a self-contained, specialized unit of software and / or hardware components. A module is designed to perform a specific function or task and is independent and self-contained; that is, it accepts specific inputs, performs internal processing, and then delivers specific outputs or results. A module can communicate with other modules or components via interfaces. These interfaces determine how data or commands are input into the module and how results or information are output.

[0053] In the context of the invention, a "processor" can be, for example, a machine or an electronic circuit. A processor can, in particular, be a central processing unit (CPU), a microprocessor, or a microcontroller, such as an application-specific integrated circuit or a digital signal processor, optionally in combination with a memory unit for storing program instructions. A processor can also be a virtualized processor, a virtual machine, or a soft CPU. It can, for example, also be a programmable processor equipped with configuration steps for executing the method according to the invention, or configured with configuration steps such that the programmable processor implements the features of the method, the modules, or other aspects and / or partial aspects of the invention.In particular, the processor can contain highly parallel computing units and powerful graphics modules.

[0054] In the context of the invention, a "storage unit" or "storage module" and the like can refer, for example, to volatile memory in the form of random access memory (RAM), permanent storage such as a hard drive or data carrier, or, for example, a replaceable storage module. The storage module can also be a cloud storage solution.

[0055] The term "database" refers to both a storage algorithm and the hardware in the form of a storage unit.

[0056] Furthermore, communication links are provided for the exchange and transmission of data between the individual modules, in particular as wireless communication links, e.g. as mobile communication links (e.g. 4G LTE, 5G, 6G) and / or as near-field communication links, e.g. Bluetooth. ® , Ethernet, NFC (Near Field Communication) or Wi-Fi ® are trained.

[0057] In particular, Database 300 and / or the Dummy Model Creation Module 400 can be integrated into a cloud computing infrastructure. A cloud computing infrastructure offers the ability to scale resources up or down as needed, allowing computing power, storage space, or network resources to be easily adapted to changing requirements. This scalability enables cost optimization and efficient resource allocation without large investments in hardware. Furthermore, users can access applications and data from anywhere with internet access. Cloud computing also offers flexibility in software deployment, enabling applications to be deployed and updated quickly and without interruption. Cryptographic encryption methods can also be used to protect the connection to the cloud computing infrastructure via a mobile network.Since a fast calculation result is desired, the processing speed is crucial for the numerous calculations in the dummy model creation module 400. To enable real-time execution of these calculations, the use of cloud computing infrastructure is therefore advantageous, as this ensures rapid computation.

[0058] The input module 200 includes a variety of sensors 250 such as accelerometers, strain gauges, and pressure sensors, as well as high-speed cameras for capturing crash test data. It also includes a user interface that can utilize various platforms such as desktop computers, laptops, tablet PCs, or smartphones for interactive communication with the dummy model creation module 400.

[0059] Database 300 contains test data sets 320 for training the AI ​​model 430. Additionally, the database stores the results of current crash tests as input data 350.

[0060] The AI ​​model 430 is designed as an AI regression model. An AI regression model is a machine learning model that aims to predict quantitative values ​​and / or analyze the relationship between different variables. In this context, the term "regression" refers to the fact that AI model 430 attempts to find a mathematical function or equation that best describes the relationship between input and output data. AI regression models can be linear or nonlinear. A linear AI regression model describes a direct, proportional relationship between input and output data, while nonlinear AI regression models consider more complex relationships.

[0061] The AI ​​model 430 can be based on various machine learning algorithms. In particular, linear regression, polynomial regression, decision trees, random forests, neural networks, and support vector machines (SVMs) are used, each of which has its specific strengths.

[0062] Linear regression is a statistical method that uses a straight line to model the relationship between a dependent variable (target variable) and one or more independent variables (predictors). Linear regression is suitable for relationships that are linear or approximately linear.

[0063] Polynomial regression is an extension of linear regression that uses a higher-degree polynomial to model more complex relationships. Polynomial regression is suitable for data that are not linear, but curved or otherwise complex. Decision trees:

[0064] Decision trees are models that make data-driven decisions using a tree-like structure. A decision tree consists of nodes and edges. Each internal node represents a decision or condition, and each branch or edge indicates which decision will be made in the next step. The leaves represent the final results or categories.

[0065] Random Forest is an ensemble model that combines multiple decision trees to increase prediction accuracy. The Random Forest algorithm generates many decision trees, each trained on randomly selected subsets of the data. The final prediction is made by averaging (in regression) or by majority vote (in classification) the results of all trees.

[0066] A neural network consists of neurons arranged in multiple layers and interconnected in various ways. A neuron can receive information at its input from the outside or from another neuron, process it in a specific manner, and then pass it on in a modified form to another neuron at its output, or output it as a final result. Between the input and output neurons are the so-called hidden neurons. Depending on the type of network, there can be several layers of hidden neurons. They are responsible for the transmission and processing of information. The output neurons ultimately deliver a result and transmit it to the outside world. The arrangement and interconnection of the neurons give rise to different types of neural networks, such as feed-forward networks, recurrent networks, or convolutional neural networks.The networks can be trained through unsupervised or supervised learning.

[0067] Support Vector Machines (SVMs) are models primarily used for classification, but they are also suitable for regression. The basic idea of ​​an SVM is to find an optimal dividing line (or hyperplane) between different classes, maximizing the distance between the line and the nearest data points of each class. These data points are called support vectors. SVMs can also handle nonlinear relationships by using kernels that project the data into higher-dimensional spaces to enable linear separation.

[0068] In particular, the invention provides for the use of gradient tree boosting as the algorithm for the AI ​​regression model in order to generate a particularly powerful predictive model. XGBoost is a specific implementation of this method, characterized by speed, efficiency, and accuracy.

[0069] Gradient tree boosting is a form of "boosting," a machine learning technique that combines multiple weak models into a strong one. A "weak model" is one with low predictive power, such as a single decision tree with limited depth. In the boosting process, these weak models are added incrementally, with each new model aiming to correct the errors of the previous one. The basic idea behind gradient tree boosting is to iteratively train decision trees, with each tree attempting to minimize the residual errors of the preceding tree. The term "gradient" refers to the fact that the corrections are based on the gradient, i.e., the derivative of a loss function that measures how far the current predictions are from the actual values. This approach is used to steer the models in a direction that minimizes errors.In each iteration, a new decision tree is added, and the predictions of the boosting model are a combination of the predictions from all trees, with the predictions of the different trees being weighted. The boosting process continues until a predetermined number of trees is reached or another criterion is met.

[0070] XGBoost, short for "eXtreme Gradient Boosting", is a special implementation of Gradient Tree Boosting and is characterized by its high performance, speed and adaptability.

[0071] The AI ​​model 430 is trained using suitable training datasets 320. The training datasets 320 comprise initial data 325 relating to real crash test dummies 10, and secondary data 327 relating to virtual dummy models. A virtual dummy model is a digital or computer-generated replica of an average real crash test dummy 10. The initial data 325 were obtained in a variety of real certification tests for at least one specific body region of the real dummy 10. The real certification tests were conducted using a variety of real dummies 10, each representing an individual version of a specific dummy type, such as an average adult or an average child, and which may differ from one another in terms of variable properties such as the material thickness of individual components.The first data 325 are time series of various load values ​​for this multitude of real dummies 10 of a specific dummy type, recorded during the impact by at least one impactor 20. The time series provide detailed information about the temporal progression of an impact and make it possible to identify safety risks, optimize test procedures, and develop safer products and vehicles.

[0072] The second set of data, 327, consists of simulation-generated time series of load values ​​for several variants of a virtual dummy model. These variants of the virtual dummy model relate to different values ​​of the varying parameters and properties of the real dummies 10. These simulation-generated time series were obtained in a large number of simulation-generated certification tests of the variants of the virtual dummy model. The variants of the virtual dummy model represent the varying input parameters and properties of the real dummies 10.To reduce the data volume from this large number of possible virtual dummy model variants, the variants provided to AI model 430 are determined by a stochastic simulation using sampling from probability distributions of the variable input parameters and properties of the real dummies 10, as well as the various variable experimental boundary conditions. A Monte Carlo simulation can be used for sampling to simulate a wide range of possible outcomes.

[0073] Dummy load values ​​are measurements that quantify the physical forces, moments, accelerations, deformations, or other parameters to which the dummy 10 is subjected during a crash test or similar simulation. These values ​​are used to analyze the effects of the forces acting on the dummy 10 in order to assess the potential risk of injury to vehicle occupants. Typical load values ​​for a dummy 10 include the acceleration of the dummy 10 or one of its parts, the direct forces acting on the dummy 10, the pressure exerted on a specific area of ​​the dummy 10, the deformations of body parts of the dummy 10, the torques acting on the dummy 10, and the velocity at which a dummy 10 or a part thereof is moving.

[0074] A time series of the impactor 20 can contain a variety of measured values, including the force exerted by the impactor 20 on a dummy 10, the velocity and acceleration of the impactor 20 upon impact with the dummy 10, and the spatial position of the impactor 20, which provides information about the location of the impact. The time series provides information about the dynamics of the impact, such as the peak force, the impact duration, the impact velocity, and the force profile.

[0075] The first data points 325 and the second data points 327 are converted into simpler numerical values ​​(features) by a software application 450 so that they can be processed by the AI ​​model 430. Extracting these features is necessary because AI regression models struggle to work efficiently and accurately with excessively large or high-dimensional data. Dimension reduction allows the AI ​​model 430 to process high-dimensional data without loss.

[0076] Dimensional reduction smooths or averages a moving time window using sliding means. For a specific segment of the curve (time window), the mean value for that segment is calculated. The time window is then shifted slightly, and the mean is calculated again. This results in a reduced and smoothed representation of the curve, which minimizes fluctuations and highlights general trends. Next, the slope or rate of change (gradient) of the curve is determined. The gradient reveals how quickly a measured quantity or value changes over time. It is an important characteristic for identifying dynamics and trends in time series. The maximum or minimum value can then be determined as the highest or lowest point of a curve. This can be of interest, for example, when studying peak loads or peak values ​​in crash tests.

[0077] Additionally, a principal component analysis (PCA) is performed. PCA is a technique for reducing dimensionality. It identifies the principal axes or "principal components" in a multidimensional space and allows high-dimensional data to be projected into a smaller space. In this way, the most important patterns or structures in the data can be extracted. To this end, PCA models for the various load values ​​of Dummy 10 are trained using a large amount of experimental and simulated data to identify the principal components. These trained PCA models then serve as reference points for projecting new data into a reduced space. Once the PCA model is trained for a load value, it can be used to decompose new time series or curves into their principal components.These extracted principal components then represent the numerical values ​​that can be used as input for the AI ​​model 430. These numerical values ​​are now much easier to process, while preserving the essential properties of the original curves or time series. The AI ​​model 430 is trained on the extracted features from the first data 325 and the second data 327 of the training datasets 320 to generate predictions for a user-specific virtual dummy model 50.

[0078] The input parameters 55 of the user-specific virtual dummy model 50, which can influence the outcome of a virtual crash test performed with the virtual dummy model 50, include material properties such as the material thickness, elasticity, or density of certain components of the real dummy 10. Other input parameters 55 may relate to the positioning of the dummy 10, as the precise position of the head, arms, or legs can vary. Furthermore, the flexibility or stiffness of the joints can vary. These input parameters 55 can exhibit variance; that is, they do not always have exactly the same value for all real dummies 10 of a dummy type. The probability distribution for these input parameters 55 describes the expected values ​​and the range within which they can vary.

[0079] In addition to the properties of the dummy 10, the experimental boundary conditions of a crash test can also vary; these are also important as input parameters 55 of the user-specific virtual dummy model 50. These include the impact velocity, the impact angle, the speed, the force, and environmental factors such as temperature or humidity, which can influence the test behavior. These conditions are also described by probability distributions to represent the variability in the test conditions.

[0080] The second set of data, 327, containing load values ​​and time series for different variants of a virtual dummy model, where the variants are generated by sampling probability distributions of the properties of real dummies 10 and the experimental boundary conditions of a conducted crash test, thus considers a broader spectrum of possible real dummy configurations 10 that can be used in a standardized crash test. A standardized crash test is a test procedure that uses specific, clearly defined conditions and parameters, such as impact velocity, impact angle, dummy positioning, etc., to evaluate the safety behavior of vehicles or components. The different variants of the virtual dummy model take into account these variations that can occur within a standardized crash test.

[0081] The AI ​​model 430 is trained with the first data 325 and the second data 327 of the training data sets 320 to determine the values ​​of the scattering input parameters 55 and on this basis to generate a user-specific virtual dummy model 50 with fixed input parameters 55.

[0082] Furthermore, the performance of the AI ​​model 430 is verified using a validation dataset, particularly based on error measures such as the mean squared error, the mean absolute error, and the coefficient of determination. These measures indicate how well the AI ​​model 430 is able to approximate or predict the actual values ​​of the variable input parameters 55 of the virtual dummy 50.

[0083] The trained AI model 320 is now used for new input data 350, acquired by sensors 250. The input data 350 consists of data from certification tests of a dummy 10, specifically a certification test of a particular body region of the dummy 10, such as the thorax, and includes time series data on various stress levels of a dummy 10 and an impactor 20. This input data 350 is then converted by the software application 450 into lower-dimensional data so that it can be processed by the AI ​​model 430. The trained PCA models for the various stress levels are used to decompose the new data into its principal components. The extracted principal components of the new input data 350 represent the numerical features used as input for the trained AI model 430.

[0084] The trained AI model 430 can now generate a user-specific virtual dummy model 50 from the input data 350 for a specific certification test by adjusting the input parameters 55 of a standardized virtual dummy model. The specific properties of a particular dummy 10 and the experimental boundary conditions present in a specific certification test are incorporated into the user-specific virtual dummy model 50 as specific values ​​of the input parameters 55. The specific dummy 10 and the specific experimental boundary conditions are now accurately and correctly represented by the user-specific virtual dummy model 50. The user-specific virtual dummy model 50 with the calculated input parameters 55 is output as output data 550 by the output module.

[0085] This user-specific virtual dummy model 50 can now be used for virtual simulations of crash tests for the certification and verification of design aspects of the vehicle, such as the design of certain body parts or the design of safety devices, such as an airbag.

[0086] In Fig. Section 6 describes the procedural steps for creating a user-specific virtual dummy model 50 for carrying out computer-aided crash test simulations for the certification and improvement of design parameters of parts and components of a vehicle.

[0087] In step S10, time series of load values ​​of a real dummy 10 and of time series of an impactor 20 are recorded using sensors 250 and stored as input data 350.

[0088] In step S20, the input data 350 is processed with a software application 450 for dimension reduction in order to convert the input data 350 into simpler numerical values ​​(features).

[0089] In step S30, the converted input data 350 are transferred to a trained AI model 430, where the AI ​​model 430 is at least designed as an AI regression model.

[0090] In step S40, a user-specific virtual dummy model 50 is calculated with at least one defined value for at least one input parameter 55, where an input parameter 55 refers to at least one property of the real dummy 10 or at least one experimental boundary condition of a certification test of the dummy 10.

[0091] In step S50, the user-specific virtual dummy model 50 is output as output data 550.

[0092] Fig.Figure 7 schematically represents a computer program product 900 comprising an executable program code 950 configured to execute the method according to the first aspect of the present invention.

[0093] The creation of a user-specific virtual dummy model according to the invention using an AI regression model enables more accurate prediction of the results of a virtual crash test, since the input parameters of the user-specific dummy model are known. Furthermore, the validation of computer-aided crash test simulation models becomes more accurate and reliable because the uncertainties regarding the virtual dummy model are already reduced. Since the AI ​​regression model is trained with existing knowledge about the varying properties of hardware dummies and experimental boundary conditions, as well as simulated data from stochastically selected virtual dummy models, each with different varying input parameters and properties, these variations can be automatically taken into account. Overall, this leads to improved accuracy and reliability of crash simulations.This allows design parameters such as material selection, component thickness, shape, etc. of vehicle components to be more precisely adjusted and optimized in order to improve the crash safety of vehicle components and thus increase the safety and performance of vehicle bodies. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 2018 / 0285516 A1

[0011]

Claims

[1] Method for creating a user-specific virtual dummy model (50) for carrying out computer-aided crash test simulations for the certification and improvement of design parameters of parts and components of a vehicle, comprising the following method steps: - Recording (S10) time series of stress values ​​of a real dummy (10) and of time series of an impactor (20) using sensors (250) and storing the time series as input data (350); - Processing (S20) the input data (350) with a software application (450) for dimension reduction in order to convert the input data (350) into simpler numerical values ​​(features); - Transferring (S30) the converted input data (350) to a trained AI model (430), wherein the AI ​​model (430) is at least designed as an AI regression model; - Calculating (S40) a user-specific virtual dummy model (50) with at least one fixed value for at least one input parameter (55), wherein an input parameter (55) relates to at least one property of the real dummy (10) or at least one experimental boundary condition of a certification test of the dummy (10); - Output (S50) of the user-specific virtual dummy model (50) as output data (550). [2] Method according to claim 1, wherein the AI ​​model (430) uses linear regression, polynomial regression, decision trees, random forest, neural networks, support vector machines (SVM) and / or gradient tree boosting, in particular XGBoost as algorithm or algorithms. [3] Method according to claim 1 or 2, wherein sliding means, the determination of a gradient and / or maximum and minimum values, and a principal component analysis (PCA) are performed for data dimension reduction, wherein PCA models for the different load values ​​of the dummy (10) are trained with a large amount of experimental and simulated data to identify the principal components in the input data (350). [4] A method according to any one of claims 1 to 3, wherein training data (320) are used to train the AI ​​model (430), the training data (320) comprising first data (325) and second data (327), the first data (325) comprising time series of load values ​​for a plurality of real dummies (10) and time series of at least one impactor (20) from a plurality of real certification tests of the plurality of real dummies (10), the plurality of real dummies (10) representing different individual versions of a particular dummy type and each differing from the others in terms of input parameters and properties such as the material thickness of individual components, and the second data (327) comprising simulation-generated time series of load values ​​for several variants of a virtual dummy model, each with different varying input parameters and properties.wherein the simulation-generated time series are determined in a multitude of simulation-generated certification tests of the variants of the virtual dummy model, and wherein the variants of the dummy model are generated by a stochastic simulation within the framework of a sampling from probability distributions of the scattering input parameters and properties of the real dummy (10). [5] Method according to any one of claims 1 to 4, wherein at least one load value represents an acceleration of the dummy (10) or one of its parts, a direct force acting on the dummy (10), a pressure exerted on a certain area of ​​the dummy (10), a deformation of body parts of the dummy (10), a torque acting on the dummy (10), a speed at which the dummy (10) or a part thereof moves. [6] Method according to any one of claims 1 to 5, wherein a time series of the impactor (20) represents the force exerted by the impactor (20) on the dummy (10), the velocity and acceleration of the impactor (20) upon impact with the dummy (10), and the change in the spatial position of the impactor (20). [7] Method according to any one of claims 1 to 6, wherein at least one design parameter is a dimension, a material, a shape, a manufacturing process, a surface finish, a density, a temperature resistance, a tensile strength, a compressive strength, a flexural strength, a fracture behavior, a plastic deformation, a characteristic value, a measured value, a safety factor and / or a material property. [8] System (100) for creating a user-specific virtual dummy model (50) for performing computer-aided crash test simulations for the certification and improvement of design parameters of parts and components of a vehicle, comprising an input module (200), a database (300), a dummy model creation module (400) with an AI model (430), and an output module (500), wherein the input module (200) is configured to acquire time series of load values ​​of a real dummy (10) and time series of an impactor (20) using sensors (250); wherein the database (300) is configured to store the time series as input data (350); wherein one software application (450) is designed for dimension reduction to convert the input data (350) into simpler numerical values ​​(features) and to transfer the converted input data (350) to a trained AI model (430);wherein the AI ​​model (430) is configured at least as an AI regression model for calculating a user-specific virtual dummy model (50) with at least one fixed value for at least one input parameter (55), wherein an input parameter (55) relates to at least one property of the real dummy (10) or at least one experimental boundary condition of a certification test of the dummy (10), and wherein the output module is configured to output the user-specific virtual dummy model (50) as output data (550). [9] System (100) according to claim 8, wherein the AI ​​model (430) uses linear regression, polynomial regression, decision trees, random forest, neural networks, support vector machines (SVM) and / or gradient tree boosting, in particular XGBoost as algorithm or algorithms. [10] System (100) according to claim 8 or 9, wherein sliding means, the determination of a gradient and / or maximum and minimum values, and a principal component analysis (PCA) are performed for data dimension reduction, wherein PCA models for the different load values ​​of the dummy (10) are trained with a large amount of experimental and simulated data to identify the principal components in the input data (350). [11] System (100) according to any one of claims 8 to 10, wherein training data (320) are used to train the AI ​​model (430), the training data (320) comprising first data (325) and second data (327), wherein the first data (325) comprise time series of load values ​​for a plurality of real dummies (10) and time series of at least one impactor (20) from a plurality of real certification tests of the plurality of real dummies (10), wherein the plurality of real dummies (10) represent different individual versions of a particular dummy type and each differ from each other by input parameters and properties such as the material thickness of individual components, and wherein the second data (327) comprise simulation-generated time series of load values ​​for several variants of a virtual dummy model, each with different varying input parameters and properties.wherein the simulation-generated time series are determined in a multitude of simulation-generated certification tests of the variants of the virtual dummy model, and wherein the variants of the dummy model are generated by a stochastic simulation within the framework of a sampling from probability distributions of the scattering input parameters and properties of the real dummy (10). [12] System (100) according to one of claims 8 to 11, wherein at least one load value represents an acceleration of the dummy (10) or one of its parts, a direct force acting on the dummy (10), a pressure exerted on a certain area of ​​the dummy (10), a deformation of body parts of the dummy (10), a torque acting on the dummy (10), a speed at which the dummy (10) or a part thereof moves. [13] System (100) according to one of claims 8 to 12, wherein a time series of the impactor (20) represents the force exerted by the impactor (20) on the dummy (10), the velocity and acceleration of the impactor (20) upon impact on the dummy (10), and the change in the spatial position of the impactor (20). [14] System (100) according to any one of claims 8 to 13, wherein at least one design parameter is a dimension, a material, a shape, a manufacturing process, a surface finish, a density, a temperature resistance, a tensile strength, a compressive strength, a flexural strength, a fracture behavior, a plastic deformation, a characteristic value, a measured value, a safety factor and / or a material property. [15] Computer program product (900) comprising an executable program code (950) configured to perform the method according to any one of claims 1 to 7 when executed.

Images