System and method for optimizing a machine learning model for predicting crash test results

Abstract

The invention relates to a system (100) for optimizing a machine learning model for predicting crash test results for vehicle development. The system (100) comprises a data preparation module (200) that splits a data set (210) containing crash test and simulation data into a first training data set (220), a validation data set (230), and a test data set (240); a training module (300) with an AI model (350) for predicting crash test results, wherein the AI ​​model (350) is trained with the first training data set (220); and a data value module (400) that calculates data values ​​(φ(i)) for the data points (i) of the crash test and simulation data of the training data set (220) with respect to the validation data set (230).a data cleaning module (500) trained to identify and remove training data points (i) with a low impact on the prediction accuracy of the AI ​​model (350), wherein the data cleaning module (500) uses the computed data values ​​(φ(i)) to determine the least valuable data points (i) and to create a second training data set (520), and wherein the second training data set (520) is subsequently used to train the AI ​​model (350) to create a modified AI model (370).

Description

[0001] The invention relates to a system, a method and a computer program product for optimizing a machine learning model for predicting crash test results for vehicle development.

[0002] Ensuring compliance with passive safety requirements and demonstrating the effectiveness of developed safety measures are essential components of the new vehicle development process. This process encompasses a wide variety of test types, ranging from tests of static individual components to dynamic tests on subsystems and complete systems. The level of integration of these tests steadily increases throughout the development process.

[0003] To ensure the structural integrity of the vehicle body and to realistically test restraint systems such as airbags and seat belts, crash tests are conducted with the complete vehicle. These crash tests simulate various accident scenarios under controlled conditions. In a frontal impact test, for example, a rolling vehicle is accelerated to a defined speed using a cable-operated device and then driven into a fixed barrier. This corresponds to a typical frontal impact, which is one of the most common types of accidents. During the impact, sensors in crash test dummies record the forces that the occupants would be subjected to in a real accident. This data makes it possible to evaluate the effectiveness of protective measures such as crumple zones and restraint systems and to ensure that legal requirements for occupant protection are met.

[0004] Given the increasing complexity of regulations, ever-shorter development cycles, and rising cost pressures in the automotive industry, the vehicle development process is becoming increasingly digitalized. Virtualization is replacing physical testing with computer-aided simulations in many cases. This begins as early as the start of the development process, when initial vehicle designs are available. Using numerical simulation methods, particularly the finite element method (FEM), the behavior of vehicle components and the vehicle as a whole can be virtually calculated under various loads. This method enables engineers to predict the physical effects of a real-world accident as accurately as possible and to identify and address potential weaknesses early on.

[0005] The advantage of virtual validation lies in the fact that time-consuming and costly real-world crash tests can be partially replaced. Simulations can provide valuable insights as early as the design phase, leading to optimized design decisions later on. This not only accelerates the development process but also reduces the number of necessary hardware tests to a minimum.

[0006] In addition to traditional safety methods such as physical crash tests and virtual simulations, machine learning models are increasingly being used to further optimize vehicle safety. These models utilize advanced algorithms, particularly neural networks, to make predictions about the course of accident events based on large datasets. A particularly important application is the prediction of the chest acceleration of a crash test dummy, which is a crucial indicator of the stress and the risk of injury to the occupants.

[0007] The major advantage of machine learning methods lies in their ability to be used in very early stages of the development process, long before physical prototypes are available or detailed virtual simulations have been conducted. This allows engineers to address safety-relevant aspects of the vehicle early on and make appropriate adjustments. The input for the machine learning models consists of design-relevant vehicle properties. These include, among other things, the vehicle structure, the specifications of restraint systems such as airbags and seat belts, and specific details of the test setup, such as impact speed or the type of barrier.

[0008] Training these AI models requires enormous amounts of data. Typically, the training is based on more than 2,000 crash tests and crash simulations, encompassing data from various vehicle manufacturers. This wealth of data enables the algorithms to achieve high predictive accuracy and cover a wide range of vehicle types and impact scenarios. By integrating this data into machine learning models, complex relationships between design parameters and the resulting safety behavior of the vehicle can be determined. This allows potential weaknesses to be identified early and addressed in a targeted manner, without having to wait for the time-consuming execution of real-world crash tests.

[0009] Another advantage of this method lies in the continuous improvement of the AI ​​model. With each new crash test and simulation, the database is expanded, leading to increasingly accurate predictions. These machine learning algorithms enable engineers to gain development knowledge significantly earlier and more efficiently than with conventional methods. This not only shortens development times but also contributes to lower development costs, as fewer physical prototypes are required.

[0010] In the long term, these technologies could even lead to a drastic reduction in the number of necessary real-world crash tests, while simultaneously further increasing vehicle safety. Machine learning thus has enormous potential to transform automotive development and lead it into an even more data-driven and digitized future.

[0011] Machine learning models are used as approximation functions to predict an output based on a variety of input parameters. The training process allows the model to learn an approximation of the underlying function, resulting in a learned approximation function. Training data is often divided into three parts: a training dataset for training the model, a validation dataset for adjusting the hyperparameters, and a test dataset for final evaluation of the model.

[0012] In machine learning practice, a distinction is made between regression, when a continuous feature is to be estimated, and classification, when a category is to be determined. A growing research topic in machine learning is data valuation, which aims to quantify the inherent value of each training instance. In particular, it is assumed that contradictory data points are less useful for model training.

[0013] A machine learning model can thus be trained with all available crash test data to predict occupant loads (e.g., in the form of chest acceleration). However, a machine learning model trained on a wide variety of vehicle models (e.g., vans, SUVs) from a wide variety of manufacturers may struggle to make accurate predictions for specific vehicle models (e.g., sports cars) from a particular manufacturer. Differences in vehicle structure, material usage, and design can lead to varying occupant loads.

[0014] The significant shift in the data distribution between training and a specific application thus reduces the model performance of the machine learning model.

[0015] This problem of domain shift occurs in machine learning when an AI model is trained on data whose distribution differs from the data encountered during application. In the context of crash test data, this leads to overfitting, as the AI ​​model has learned certain patterns in the training data that are not applicable to all other vehicle types. The machine learning model is then unable to make generalizable predictions for unknown data. This impairs the performance of the machine learning model. For example, estimates of chest acceleration for sports cars can vary significantly due to structural differences.

[0016] Therefore, taking distribution shifts into account is crucial for optimizing the performance of machine learning models in variable and complex application areas such as predicting occupant loads in vehicle crashes.

[0017] US Patent 2024 / 0184272 A1 D1 discloses a method in which a machine learning model is trained with a training dataset to make predictions or decisions in a non-public communications network. The determination of whether the trained model is valid or invalid is based on the evaluation of predictions or decisions made by the trained model from a validation dataset. If the validation is unsuccessful, the training dataset and the model are analyzed to determine which additional training data needs to be added to the training dataset. The model is then retrained with the training dataset augmented with the additional training data.

[0018] German patent DE 10 2023 103 652 A1 describes a computer-implemented method for determining a computational output variable using a trained model. First, a computational input variable is defined. Depending on this, suitable training and validation datasets are selected. The model is trained using the training datasets by learning to determine the corresponding data output variables from the input variables. Subsequently, the trained model is checked against the validation datasets, identifying deviations between the calculated and actual data output variables. Finally, the validated model is used to calculate the computational output variable from the input variable.

[0019] German patent DE 10 2023 114 256 A1 relates to a system and method for generating a predictive index based on data about customers' digital intentions. The method includes generating a multitude of predictive models using a unique machine learning algorithm for each model, determining an optimal predictive model based on an overall model score assigned to each of the multitude of predictive models, and generating a predictive index based on the optimal predictive model. The predictive index can be used to adjust the supply of one or more products based on predicted customer demand.

[0020] German patent DE 10 2017 107 837 A1 describes an adaptable sensor arrangement with at least one sensor element for acquiring sensor data. It also includes a control and evaluation unit with a classifier and an interface to a higher-level computer network. The control and evaluation unit is designed to select additional sensor data, which will also be classified in the future, as part of an extension function. This data is transmitted to the computer network, evaluated there, and returned in the form of derived classifier data. Based on this, the classifier is adjusted so that the sensor arrangement can subsequently classify the newly selected sensor data as well.

[0021] German patent DE 10 2019 213 061 A1 relates to a method, a computer program, and a device for providing a classifier for an AI module, as well as the provided classifier itself. It also relates to a method, a computer program, and a device for configuring a control system with a library of AI modules, and a motor vehicle equipped with such a system. First, an AI module to be classified and a suitable test dataset are selected. Then, the AI ​​module is applied to the data points of the test dataset, for which basic truths and contextual parameters are known. Based on the outputs of the AI ​​module, a functional rating is determined for each data point. Finally, a classifier is created that outputs the functional rating of the AI ​​module for given contextual parameters.

[0022] German patent application DE 20 2022 105 865 U1 describes a system based on IoT and machine learning. The system includes data preprocessing to convert a dataset into a suitable format, feature selection to remove irrelevant features and select relevant ones, and a classifier to divide the data into different classes. Furthermore, the data is divided into training and validation datasets, the classifier's performance is evaluated, and the results are stored.

[0023] The object of the present invention is to provide possibilities for optimizing a machine learning model for predicting crash test results for vehicles of a specific vehicle manufacturer, in order to increase the predictive accuracy of the learning model, in particular for manufacturer-specific restraint systems, and thus to improve their effectiveness in the vehicle.

[0024] This problem is solved according to the invention with respect to a system by the features of claim 1, with respect to a method 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.

[0025] The system and method according to the invention enable the optimization of occupant load prediction for the vehicles of a specific vehicle manufacturer A. Targeted data cleansing and the inclusion of suitable data from other vehicle manufacturers reduce the risk of overfitting to the specific data of vehicle manufacturer A. A broad training dataset, which also includes data from other manufacturers, ensures that the AI ​​model is not exclusively specialized for the characteristics of a single manufacturer, but adequately considers them. In this way, the modified AI model contributes to the development of overall more efficient and specific restraint systems for the vehicles of a particular vehicle manufacturer.

[0026] According to a first aspect, the invention provides a system for optimizing a machine learning model for predicting crash test results for vehicle development. The system comprises a data preparation module that splits a dataset containing crash test and simulation data into a first training dataset, a validation dataset, and a test dataset, wherein the training dataset contains data from crash tests and simulations of a large number of vehicles from a large number of vehicle manufacturers, and the validation dataset contains data from crash tests and simulations of a large number of vehicles from a specific vehicle manufacturer; a training module with an AI model for predicting crash test results, wherein the AI ​​model is trained using the first training dataset;a data value module that calculates a data value for each training data point of the crash test and simulation data of the training data set in relation to the validation data set, where this data value represents the contribution of the respective training data point to the predictive performance of the AI ​​model with respect to the manufacturer-specific validation data set;A data cleaning module trained to identify and remove training data points with a low impact on the predictive accuracy of the AI ​​model, wherein the data cleaning module uses the calculated data values ​​to determine the least valuable training data points with respect to the manufacturer-specific validation dataset and thereby create a second training dataset that improves the quality of the training data and is adapted to the specific technical requirements of the specific vehicle manufacturer, and wherein the second training dataset is subsequently used to train the AI ​​model to create a modified AI model that generates optimized predictions of crash test results for vehicles of the specific vehicle manufacturer.

[0027] In a further development, it is planned that the AI ​​model will include a neural network with an encoder-decoder architecture that learns the relationship between crash test parameters and the results of physical crash tests as well as virtual simulations.

[0028] In an advantageous embodiment, the data value module calculates the data values ​​using algorithms such as Data Shapley or Data-OOB to determine the exact contribution of each data point to the prediction performance of the AI ​​model.

[0029] In another embodiment, the data preparation module includes preprocessing of the crash test and simulation data, which removes data anomalies from the data set in order to improve the training quality of the AI ​​model.

[0030] Advantageously, the training module uses supervised learning techniques to continuously improve the predictive accuracy of the modified AI model based on feedback from the validation dataset.

[0031] In particular, the data cleansing module uses an iterative data cleansing procedure, where the performance of the modified AI model is monitored after each adjustment of the training data to ensure that the prediction quality remains optimized at all times.

[0032] Advantageously, the modified Kl model is optimized for different vehicle categories (e.g., cars, SUVs, trucks) to develop tailored restraint systems that meet the specific safety requirements of each category.

[0033] According to a second aspect, the invention provides a method for optimizing a machine learning model for predicting crash test results for vehicle development. The method comprises the following steps: - Providing a dataset that includes crash test and simulation data, and splitting the dataset into an initial training dataset, a validation dataset and a test dataset, wherein the initial training dataset contains data from crash tests and simulations of a large number of vehicles from various vehicle manufacturers and the validation dataset contains data from crash tests and simulations of a specific vehicle manufacturer; - Training a AI model using the initial training dataset to predict crash test results; - Calculating a data value for each training data point of the crash test and simulation data of the first training data set with respect to the validation data set using a data value module, where this data value represents the contribution of the respective training data point to the prediction performance of the AI ​​model with respect to the manufacturer-specific validation data set; - Transmitting the calculated data values ​​to a data cleansing module; - Identifying and removing training data points with a low impact on the prediction accuracy of the AI ​​model using the data cleaning module, with the calculated data values ​​being used to determine the least valuable training data points in relation to the manufacturer-specific validation data set in order to improve the quality of the training data and adapt it to the specific technical requirements of the vehicle manufacturer; - Creating a second training dataset after removing the identified training data points, where the second training dataset contains data with a higher impact on prediction accuracy; - Training the AI ​​model using the second training dataset to create a modified AI model that generates optimized predictions of crash test results for vehicles of the specific vehicle manufacturer.

[0034] In a further training course, it is planned that the AI ​​model will include a neural network with an encoder-decoder architecture that learns the relationship between crash test parameters and the results of physical crash tests as well as virtual simulations.

[0035] In an advantageous embodiment, the data value module calculates the data values ​​using algorithms such as Data Shapley or Data-OOB to determine the exact contribution of each data point to the predictive performance of the AI ​​model.

[0036] In another embodiment, the data preparation module includes preprocessing of the crash test and simulation data, which removes data anomalies from the data set in order to improve the training quality of the AI ​​model.

[0037] Advantageously, the training module uses supervised learning techniques to continuously improve the predictive accuracy of the modified AI model based on feedback from the validation dataset.

[0038] In particular, the data cleansing module uses an iterative data cleansing procedure, where the performance of the modified AI model is monitored after each adjustment of the training data to ensure that the prediction quality remains optimized at all times.

[0039] Advantageously, the modified Kl model is optimized for different vehicle categories (e.g., cars, SUVs, trucks) to develop tailored restraint systems that meet the specific safety requirements of each category.

[0040] According to a third aspect, the invention provides a computer program product with an executable program code that is configured to perform the method according to the second aspect when executed.

[0041] The invention will now be explained in more detail with reference to an embodiment shown in the drawing.

[0042] It shows: Fig. 1 a block diagram to illustrate an embodiment of a system according to the invention; Fig. 2 a flowchart to explain the individual process steps of a process according to the invention; Fig. 3 a block diagram of a computer program product according to an embodiment of the third aspect of the invention.

[0043] Additional features, aspects and advantages of the invention or its embodiments become apparent from the detailed description in conjunction with the claims.

[0044] Fig. Figure 1 shows a system 100 according to the invention for optimizing a machine learning model for predicting crash test results for vehicle development. The system 100 comprises a data preparation module 200, a training module 300 with an AI model 350 (AI, Artificial Intelligence), a data value module 400, and a data cleansing module 500. The AI ​​model 350 is designed as a machine learning model.

[0045] System 100 thus consists of various modules, each responsible for a specific process step. Each module performs specialized tasks and interacts with the other modules to enable efficient and optimized crash test and simulation planning.

[0046] The data preparation module 200 prepares the data 210 and divides it into an initial training data set 220, a validation data set 230, and a test data set 240. This enables targeted adaptation of the AI ​​model 350 to the specific characteristics of a particular vehicle manufacturer A.

[0047] The training module 300 contains the machine learning model (referred to here as KL model 350) which learns from the training data of the training data set 220 to predict crash test results.

[0048] The data value module 400 evaluates the training data of the training dataset 220 with regard to its contribution to the model performance of the AI ​​model 350 in relation to the manufacturer-specific validation dataset 230. This helps to identify relevant data points and to exclude less useful or counterproductive data.

[0049] The data cleansing module 500 performs targeted data cleansing by removing potentially disruptive or irrelevant training data based on the data values ​​calculated in the data value module 400. This allows the AI ​​model 350 to be specifically adapted to the characteristics of vehicles from the specific manufacturer A.

[0050] The data preparation module 200, the training module 300, the data value module 400, and the data cleansing module 500 can each be equipped with a storage unit and / or a processor. In particular, they are combined in an integrated computing unit, such as a cloud environment.

[0051] A "module" is defined as a specialized unit of software and / or hardware components that performs a specific function. It is independent, receives input, processes it, and provides output, and can communicate with other modules via interfaces.

[0052] A "processor" can be a main processor, a microcontroller, or a programmable processor, even in virtualized form, and can include complex computing units as well as graphics modules. It performs the configuration steps of the method according to the invention.

[0053] A "storage unit" can be volatile or permanent storage, e.g., RAM, a hard drive, or cloud storage.

[0054] The modules can be integrated, in particular, into a cloud computing infrastructure. Cloud computing offers scalable computing power as well as storage and network resources, thus ensuring high flexibility, cost optimization, and fast access. Cryptographic encryption methods can be used to secure the connection.

[0055] The data preparation module 200 is responsible for dividing and managing a data set 210 containing existing crash test and simulation data from a large number of vehicles of at least one specific vehicle manufacturer A and a large number of vehicles from a large number of vehicle manufacturers B1, B2, ..., Bn into the first training data set 220, the validation data set 230 and the test data set 240.

[0056] Dataset 210, containing existing crash test and simulation data, primarily includes NHTSA data based on information and statistics collected and published by the National Highway Traffic Safety Administration (NHTSA) in the USA. The NHTSA is an agency of the US Department of Transportation whose goal is to improve road safety. The agency collects data on various aspects of road safety, with a particular focus on accidents, vehicle recalls, crash tests, and vehicle performance. The NHTSA also conducts crash tests on vehicles to evaluate their safety levels. The results of these tests, such as safety ratings (5-star rating), are publicly available.The NHTSA data is highly relevant to a wide range of people, as it helps to optimize vehicle safety, identify accident causes, and monitor the overall state of road safety.

[0057] The first training dataset 220 comprises data from physical crash tests and virtual simulations of a large number of vehicles from various vehicle manufacturers B1, B2, ..., Bn, which were integrated into the training dataset 220. In contrast, the validation dataset 230 and the test dataset 240 contain only the data from the specific vehicle manufacturer A, which are randomly distributed between the two datasets 230 and 240.

[0058] The data preparation module 200 comprises a software application designed to split the existing crash test and simulation data of dataset 210 into the initial training dataset 220, the validation dataset 230, and the test dataset 240, and to store and manage these datasets 210, 220, 230, and 240. In particular, the data preparation module 200 ensures data integrity and the traceability of new data points from new crash test and simulation data.

[0059] Validation dataset 230 contains exclusively the data of the specific vehicle manufacturer A. This ensures that AI model 350 is optimally adapted to the specific characteristics and requirements of manufacturer A's vehicles. By focusing on a specific validation dataset 230, AI model 350 is guaranteed to consider the realistic scenarios, behaviors, and vehicle models relevant to manufacturer A's vehicles. This approach minimizes the risk of over-fitting, as AI model 350 is tested with manufacturer-relevant validation data that is not integrated into the broader initial training dataset 220. This is of great importance for manufacturer A's vehicle development, as safety standards and requirements can vary from manufacturer to manufacturer.

[0060] Training module 300 contains AI model 350 for predicting crash test results, such as the behavior of restraint systems in a vehicle. AI model 350 within training module 300 uses machine learning (ML) and artificial intelligence (AI) algorithms. Specifically, AI model 350 employs a neural network.

[0061] A neural network consists of neurons arranged in multiple layers and interconnected in various ways. A neuron is capable of receiving information at its input from the outside or from another neuron, evaluating it in a specific way, and then passing it on in a modified form to another neuron at its output, or outputting 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 result in different types of neural networks, such as...Feedforward networks (FFN), recurrent networks (RNN), or convolutional neural networks (CNN) are all examples of neural networks. These networks can be trained using unsupervised or supervised learning.

[0062] The Convolutional Neural Network (CNN) is a specific type of artificial neural network. It has multiple convolutional layers and is well-suited for machine learning and artificial intelligence (AI) applications in the field of pattern and image recognition. The individual layers of the CNN are the convolutional layer, the pooling layer, and the fully connected layer. The convolutional layer is capable of recognizing and extracting individual features from the input data. In pattern and image recognition, these can be features such as lines, edges, or specific shapes. The input data is processed in the form of tensors, such as a matrix or vectors. The pooling layer, also called the subsampling layer, condenses and reduces the resolution of the recognized features using appropriate filtering functions. The reduced data volume increases the processing speed.

[0063] The architecture of the AI ​​Model 350 comprises several components to manage the complexity of crash test predictions. Specifically, the AI ​​Model 350 employs an encoder-decoder architecture. The encoder processes the input data (e.g., vehicle type, crash test dummy configuration, restraint system), extracts relevant features, and creates a latent representation of the input data. The decoder uses this latent representation to generate the crash test predictions. This separation allows the encoder and decoder to efficiently handle specific tasks while working together to manage the complexity of crash test predictions.

[0064] Within the encoder-decoder architecture, transformers are used to efficiently process sequences of features. These transformers utilize mechanisms such as self-attention to consider interactions between different input parameters and model their influence on the crash test result. This technology is particularly useful for capturing and understanding dependencies between various test parameters.

[0065] The AI ​​model 350 can also use a language model to process data in natural language. This language model is used to convert descriptions of test conditions, technical specifications, or expert notes into machine-readable parameters. The language model can analyze both technical documentation and process user input in natural language.

[0066] The AI ​​model 350 is trained using the initial training dataset 220, which is stored in the data preparation module 200. This initial training dataset 220 includes both real crash test data and simulation results. During the training process, the neural network of the AI ​​model 350 learns to generate predictions about crash test results. Furthermore, the AI ​​model 350 learns to understand the interactions between different configuration parameters (e.g., how the seat position, the restraint system, or the airbag deployment time affect the test result). It can therefore recognize patterns and relationships in the data that might be difficult for a human expert to discern, especially with complex parameters.

[0067] After the training process, the AI ​​Model 350 is able to accurately predict crash test results. By selectively choosing test parameters that optimally cover the search space, the AI ​​Model 350 minimizes the number of physical crash tests and simulations required to improve its predictive performance. The AI ​​Model 350 can be applied to new vehicle types or innovative restraint systems and continuously adapts its predictions based on new data.

[0068] The AI ​​model 350 is tested in data value module 400 using the validation dataset 230. By splitting the dataset 210 into an initial training dataset 220 and a manufacturer-specific validation dataset 230, it is ensured that the AI ​​model 350 is trained with a broad range of training data for vehicles from various manufacturers B1, B2, ..., Bn, which are not specifically related to vehicles from manufacturer A. This ensures that the AI ​​model 350 develops a general suitability for different vehicle types, while the validation is focused on the specific characteristics and safety requirements of manufacturer A. This maximizes the validity of the validation for a particular vehicle manufacturer.

[0069] The data value module 400 now calculates data values ​​φ(i) for the available crash test data based on the validation dataset 230, i.e., related to the vehicles of a specific vehicle manufacturer A. It evaluates the relevance and contribution of each individual crash test or simulation to improving prediction accuracy. The determined data values ​​φ(i) are then used for data cleaning of the training dataset 220. Methods such as Data Shapley or Data-OOB can be used to calculate the data values ​​φ(i) and assess the influence of each crash test.

[0070] The data value module 400 is an essential component of the system 100 according to the invention, as it evaluates the value and influence of individual data points i (e.g., crash tests and simulations) on the training and prediction accuracy of the AI ​​model 350. Specific methods are used for this purpose, which quantitatively determine the contribution of each data point i.

[0071] A data point i within the scope of the invention is a specific composition or assignment of crash test or simulation data. A data point i can, for example, comprise the following: - A specific vehicle configuration (e.g., model, size, safety features) - Specific crash parameters (e.g. impact speed, impact angle) - A dummy configuration (e.g., position and sensor values) - Specific results of a crash simulation or a physical crash test (e.g. injury values, vehicle damage)

[0072] The data points i thus contain information about the behavior of the vehicles in crash tests or simulations. The influence of a data point i in the first training data set 220 on the AI ​​model 350 varies depending on its relevance and the extent to which the data point i improves or worsens the AI ​​model 350.

[0073] A data point i is thus assigned a data value φ(i) that evaluates the influence of this specific composition on the prediction accuracy of the Kl model 350, for example by methods such as Data Shapley or Data-OOB.

[0074] Data Shapley is a method based on the concept of the Shapley value from game theory. It aims to calculate the contribution of each data point i to the overall performance of AI Model 350. Data Shapley evaluates each data point i in terms of how much it contributes to improving prediction accuracy by considering all possible combinations of data points i. A data point i receives a high value if it significantly improves AI Model 350, and a low or even negative value if it contributes little or worsens AI Model 350. The method is computationally expensive because it iterates through many combinations of data points i, but it provides precise results regarding the importance of individual data points i.

[0075] Data-OOB (Out-of-Bag) is a statistical method used to evaluate the importance of individual data points i or subsets of data. It is based on the concept of out-of-bag error, which originates from the random forest method. In Data-OOB, models are trained on a portion of the data, while the remaining data is used for validation. A data point i has a higher value if the model trained without that data point i exhibits poorer predictive accuracy, and a lower value if its absence does not result in any deterioration. This method is less computationally intensive than Data Shapley and can be executed more quickly, making it attractive in many scenarios.

[0076] The data value φ(i) is typically a numerical value that quantifies the relevance or importance of the specific composition for the prediction accuracy of the Kl model 350. It can take various forms depending on the calculation method used. Some typical forms of the data value φ(i) are listed below: - Numerical value: The data value φ(i) can simply be a scalar value representing the influence of data point i on the model prediction. Higher values ​​could indicate greater relevance or importance for model performance. - Probabilities: In some approaches, the data value φ(i) can also represent a probability that indicates how likely it is that a certain parameter configuration will lead to a certain prediction of the AI ​​model 350. - Gradient values: In methods such as Data Shapley or Data-OOB, the data value can also be derived from gradient information, which indicates how much the model prediction would change if the data point i were included in or removed from the first training data set 220. - Importance of a feature: The data value φ(i) can also be interpreted as a measure of the importance of a feature associated with the composition.

[0077] The data value φ(i) is thus a numerical measure that indicates the significance or influence of a particular data point i for the Kl model 350.

[0078] The data value module 400 uses the validation data set 230 to perform the calculations of the data values ​​φ(i). The primary goal is to determine the data value φ(i) of each individual crash test or simulation, which is relevant for improving the predictive performance of the Kl model 350.

[0079] The data preparation module 200 transmits all data points i, including all crash test and simulation results from the training dataset 220 and the validation dataset 230, to the data value module 400. The data value module 400 then uses a software application 450 with algorithms such as Data Shapley or Data-OOB to calculate the value of each individual data point i. Each data point i thus receives a data value φ(i), which quantifies its influence on the prediction accuracy of the Kl model 350 with respect to the validation dataset 230.

[0080] The calculated data values ​​φ(i) for the data points i are now transmitted to the data cleaning module 500. The data cleaning module 500 comprises a software application 550 which uses the calculated data values ​​φ(i) for the data points i of the first training data set 220 to calculate a second, optimized training data set 520 by targeted cleaning according to two selectable approaches. - Approach A: This approach is used particularly when employing Data Shapley. Since the Data Shapley value quantifies the marginal contribution of each data point i to the model performance of the Kl-Model 350, all training data points i with a negative Shapley value are removed, as these could adversely affect the model accuracy. While this approach is mathematically advantageous, it requires significant computational resources because the Kl-Model 350 must be trained multiple times to calculate the marginal contribution for each data point i. - Approach B: In an iterative cleanup procedure, the data points i with the lowest data values ​​φ(i) are successively removed from the training dataset 220. After each step, the AI ​​model 350 is trained with the updated training dataset 520 and its performance is checked using the validation dataset 230. This process is repeated until the predictive performance of the AI ​​model 350 converges, i.e., no further performance improvement is achieved.

[0081] The AI ​​model 350 is then retrained with the optimized training dataset 520, resulting in a modified AI model 370 specifically tailored to the relevant and high-quality data in the second training dataset 520. This modified AI model 370 is first validated with the validation dataset 230 to ensure it meets the requirements and specific characteristics of vehicle manufacturer A. Finally, a test is conducted with the test dataset 240 to evaluate the final performance of the modified AI model 370 and confirm its suitability for real-world application scenarios.

[0082] The data cleaning module 500 employs an iterative data cleaning process, where the performance of the modified AI model 370 is verified after each adjustment of the training data. This ensures that the cleaning steps, such as removing low-value data points i (i.e., those with little impact on prediction accuracy), actually contribute to optimizing prediction quality. The modified AI model 370 can be revalidated after each cleaning loop to ensure that its performance remains stable or even improves.

[0083] With the modified Kl model 370, predictions for vehicles of the specific vehicle manufacturer A are more reliable than before, while the prediction accuracy for vehicles of other manufacturers may be lower.

[0084] The data from a specific vehicle manufacturer A can contain a variety of special features based on that manufacturer's specific design and safety requirements and target groups. Some examples of such special features are listed below: - Specific vehicle architecture: Differences in body shape, materials, frame thickness and geometry that play a role in crash tests. These features influence how the forces act on the passenger cell in an accident. - Safety equipment and restraint systems: Manufacturer A may use specially developed airbag systems, seat belts and deformation zones that influence the vehicle's crash behavior and the strain on the occupants. - Vehicle weight and center of gravity: These characteristics have a major influence on crash behavior, especially in rollover scenarios and with regard to side impact protection. - Typical seating positions and configurations: Manufacturer A may prefer certain seating configurations, such as specific seat heights, seat distances, or tilt angles. These configurations influence the loads to which occupants are exposed in various accident scenarios. - Material properties and innovative designs: The use of lightweight materials or special composite materials that affect occupant protection and structural integrity is another possible special feature. - Target group-specific variables: For example, if manufacturer A focuses more on sports cars or family vehicles, the crash tests and simulations may include data for more common occupants such as children or elderly people.

[0085] One use case is, for example, the adaptation of the performance of the modified Kl model 370 with regard to occupant prediction for crash test dummies with special body shapes, e.g. for women or children, when the crash test and simulation data of vehicle manufacturer A contains more data for these crash test dummy types compared to the training data set.

[0086] By specifically evaluating and adapting the data of the first training data set 220, the second training data set 520 is optimized accordingly, so that the KL model 370 is better adapted to the requirements of the vehicle manufacturer A.

[0087] By taking into account the specific characteristics of vehicle manufacturer A during model optimization, the modified Kl-Model 370 can improve the safety standards of manufacturer A's vehicles even more effectively and better reflect the real-world scenarios of this manufacturer A.

[0088] In a further development, it may be envisaged that the data preparation module 200 additionally uses anomaly detection algorithms to identify erroneous, inconsistent, or implausible data points i that could negatively affect the training of the AI ​​model 350. These could be, for example, measured values ​​that lie outside a defined range, or data points i that are unreliable due to measurement errors or inadequate test conditions.

[0089] Additionally, the data preparation module can apply 200 data enrichment techniques to increase the diversity of the initial training dataset. This can be achieved by generating new data points from existing data through transformations (e.g., rotation, scaling) or through synthetic data generation. This extension makes the AI ​​model more stable, as it is trained on a broader data set.

[0090] The data preparation module 200 can also extract or generate relevant features from the crash test and simulation data that are particularly important for the predictive performance of the Kl model 350. This includes the identification of key variables that have a significant influence on the crash test results and the generation of additional features that better capture complex relationships.

[0091] Furthermore, it may be provided that the training module 300 uses supervised learning methods in which the modified AI model 370 is iteratively trained on the data from the second training dataset 520. In this process, the training module 300 uses known target values ​​(e.g., measured load values ​​and injury risks) as a reference to minimize the errors between the predictions of the modified AI model 370 and the actual results.

[0092] Furthermore, the training module 300 can also use methods for optimizing hyperparameters, such as the number of layers or the learning rate, to adapt the structure of the modified AI model 370 to the complexity of the crash test data. The hyperparameters are also adjusted based on the validation data to avoid overfitting and to improve the generalizability of the modified AI model 370.

[0093] The learning rate is an important hyperparameter that determines how much the weights of AI Model 370 are adjusted after each training iteration. A learning rate that is too high can cause the modified AI Model 370 to fluctuate excessively and fail to recognize important patterns, while a learning rate that is too low slows down the training process and may cause the modified AI Model 370 to get stuck in a local minimum. An optimally set learning rate allows AI Model 370 to converge effectively and efficiently, i.e., to minimize the errors between predictions and actual target values ​​and to develop high-performance predictions.

[0094] In Fig. Figure 2 shows the process steps for optimizing a machine learning model to predict crash test results for vehicle development.

[0095] In step S10, a data set 210 is provided, which includes crash test and simulation data, and the data set 210 is split into a first training data set 220, a validation data set 230 and a test data set 240, wherein the first training data set 220 contains data from crash tests and simulations of a variety of vehicles from different vehicle manufacturers B1, B2, ..., Bn and the validation data set 230 contains data from crash tests and simulations of a specific vehicle manufacturer A.

[0096] In step S20, a Kl model 350 is trained using the first training data set 220 to predict crash test results.

[0097] In step S30, data values ​​φ(i) for the crash test and simulation data of the first training data set 220 are calculated based on the validation data set 230 using a data value module 400, where a data value φ(i) represents the contribution of a data point i to the prediction performance of the AI ​​model 350.

[0098] In step S40, the calculated data values ​​φ(i) are transmitted to a data cleansing module 500.

[0099] In step S50, training data points i with a low influence on the prediction accuracy of the Kl model 350 are identified and removed by the data cleaning module 500, whereby the calculated data values ​​φ(i) are used to determine the least valuable data points i in order to improve the quality of the training data and adapt it to the specific technical requirements of the vehicle manufacturer A.

[0100] In step S60, a second training data set 520 is created after the removal of the identified data points i, where the second training data set 520 contains data with a higher influence on the prediction accuracy.

[0101] In step S70, the AI ​​model 350 is trained using the second training data set 520 to create a modified AI model 370 that generates optimized predictions of crash test results for vehicles of the specific vehicle manufacturer A.

[0102] Fig. Figure 3 schematically represents a computer program product 900 comprising an executable program code 950 configured to perform the method according to the second aspect of the present invention.

[0103] The system and method according to the invention enable the optimization of occupant impact prediction for the vehicles of a specific vehicle manufacturer A. Targeted data cleansing and the inclusion of suitable data from other vehicle manufacturers reduce the risk of overfitting to the specific data of vehicle manufacturer A. A broad training dataset, which also includes data from other manufacturers, ensures that the modified AI model 370 is not exclusively specialized for the characteristics of a single manufacturer, but adequately considers them. In this way, the modified AI model 350 contributes to the development of overall more efficient and specific restraint systems for the vehicles of a particular vehicle manufacturer, thereby improving occupant protection. Reference sign 200 Data preparation module 210 record 220 first training data set 230 Validation data set 240 test data set 300 training module 350 AI model 400 Data Value Module 450 software applications 500 Data Cleansing Module 520 second training data set 550 Software application 900 computer program product 950 program code φ(i) data value i data point

Claims

System (100) for optimizing a machine learning model for predicting crash test results for vehicle development, comprising a data preparation module (200) that splits a data set (210) containing crash test and simulation data into an initial training data set (220), a validation data set (230), and a test data set (240), wherein the training data set (220) contains crash test and simulation data from a variety of vehicles from a variety of vehicle manufacturers (B1, B2, ..., Bn), and the validation data set (230) comprises crash test and simulation data from a variety of vehicles from a specific vehicle manufacturer (A); a training module (300) with an AI model (350) for predicting crash test results, wherein the AI ​​model (350) is trained using the initial training data set (220);a data value module (400) that calculates a data value (φ(i)) for each training data point (i) of the crash test and simulation data of the training data set (220) with respect to the validation data set (230), wherein this data value (φ(i)) represents the contribution of the respective training data point (i) to the predictive performance of the AI ​​model (350) with respect to the manufacturer-specific validation data set (230);a data cleaning module (500) trained to identify and remove training data points (i) with a low impact on the prediction accuracy of the AI ​​model (350), wherein the data cleaning module (500) uses the calculated data values ​​(φ(i)) to determine the least valuable training data points (i) with respect to the manufacturer-specific validation data set (230) and to create a second training data set (520) that improves the quality of the training data and is adapted to the specific technical requirements of the vehicle manufacturer (A), and wherein the second training data set (520) is subsequently used to train the AI ​​model (350) to create a modified AI model (370) that generates optimized predictions of crash test results for vehicles of the specific vehicle manufacturer. System (100) according to claim 1, wherein the Kl model (350) comprises a neural network with an encoder-decoder architecture that learns the relationship between crash test parameters and the results of physical crash tests as well as virtual simulations. System (100) according to claim 1 or 2, wherein the data value module (400) calculates the data values ​​(φ(i)) using algorithms such as Data Shapley or Data-OOB to determine the exact contribution of each data point (i) to the prediction performance of the Kl model (350). System (100) according to one of claims 1 to 3, wherein the data preparation module (200) comprises preprocessing the crash test and simulation data, which removes data anomalies from the data set (210) to improve the training quality of the AI ​​model (350). System (100) according to one of claims 1 to 4, wherein the training module (300) uses supervised learning techniques to continuously improve the prediction accuracy of the modified AI model (370) based on the feedback from the validation data set (230). System (100) according to any one of claims 1 to 5, wherein the data cleansing module (500) uses an iterative data cleansing method in which the performance of the modified AI model (370) is monitored after each adjustment of the training data to ensure that the prediction quality remains optimized at all times. System (100) according to any one of claims 1 to 6, wherein the modified Kl model (370) is optimized for different vehicle categories (e.g. passenger car, SUV, truck) to develop customized restraint systems that meet the specific safety requirements of each category. Method for optimizing a machine learning model for predicting crash test results for vehicle development, comprising the following steps: - Providing (S10) a dataset (210) comprising crash test and simulation data, and splitting the dataset (210) into an initial training dataset (220), a validation dataset (230), and a test dataset (240), wherein the initial training dataset (220) contains crash test and simulation data from a variety of vehicles from different vehicle manufacturers (B1, B2, ..., Bn), and the validation dataset (230) comprises crash test and simulation data from a specific vehicle manufacturer (A); - Training (S20) a machine learning model (350) using the initial training dataset (220) to predict crash test results;- Calculating (S30) a data value (φ(i)) for each training data point (i) of the crash test and simulation data of the first training data set (220) with respect to the validation data set (230) using a data value module (400), where this data value (φ(i)) represents the contribution of the respective training data point (i) to the predictive performance of the AI ​​model (350) with respect to the manufacturer-specific validation data set (230); - Transmitting (S40) the calculated data values ​​(φ(i)) to a data cleansing module (500);- Identifying and removing (S50) the training data points (i) with a low impact on the prediction accuracy of the AI ​​model (350) by the data cleaning module (500), wherein the calculated data values ​​(φ(i)) are used to determine the least valuable training data points (i) with respect to the manufacturer-specific validation data set (230) in order to improve the quality of the training data and adapt it to the specific technical requirements of the vehicle manufacturer (A); - Creating (S60) a second training data set (520) after removing the identified training data points (i), wherein the second training data set (520) contains data with a higher impact on the prediction accuracy;- Training (S70) the Kl model (350) using the second training data set (520) to create a modified Kl model (370) that generates optimized predictions of crash test results for vehicles of the specific vehicle manufacturer (A). Method according to claim 8, wherein the AI ​​model (350) comprises a neural network with an encoder-decoder architecture that learns the relationship between crash test parameters and the results of physical crash tests as well as virtual simulations. Method according to claim 8 or 9, wherein the data value module (400) calculates the data values ​​(φ(i)) using algorithms such as Data Shapley or Data-OOB to determine the exact contribution of each data point (i) to the prediction performance of the Kl model (350). Method according to one of claims 8 to 10, wherein the data preparation module (200) comprises preprocessing the crash test and simulation data, which removes data anomalies from the data set (210) to improve the training quality of the AI ​​model (350). Method according to one of claims 8 to 11, wherein the training module (300) uses supervised learning techniques to continuously improve the prediction accuracy of the modified AI model (370) based on the feedback from the validation data set (230). Method according to any one of claims 8 to 12, wherein the data cleansing module (500) uses an iterative data cleansing method in which the performance of the modified AI model (370) is monitored after each adjustment of the training data to ensure that the prediction quality remains optimized at all times. Method according to any one of claims 8 to 13, wherein the modified Kl model (370) is optimized for different vehicle categories (e.g. passenger cars, SUVs, trucks) to develop tailored restraint systems that meet the specific safety requirements of each category. Computer program product (900) comprising an executable program code (950) configured to execute the method according to any one of claims 8 to 14 when executed.

Images