DE102023117183A1 - Method and system for improving the system behavior of a technical entity - Google Patents

Abstract

The invention relates to a method for improving the system behavior of a technical entity (10) in a scenario, such as in particular a vehicle in a driving situation, comprising:
- generating a simulation model (350) for simulating the system behavior of the technical entity (10) in a simulation scenario and training with training data sets;
- providing input data (240) via a user interface (220) or from a database (290), wherein the input data (240) comprise parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10);
- performing a simulation using the input data (240) and the trained simulation model (450) from a simulation module (400) for the simulation task;
- Generating simulation results (340);
- passing the simulation results (340) to a learning reinforcement module (400);
- Calculating modeled values for the parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10) from the simulation results (340) to achieve an improved system behavior of the entity (10);
- generating and outputting output data (550) by an output module (500), wherein the output data (550) contain optimized results, in particular in the form of the modeled values for the parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10).

Description

The invention relates to a method, a system and a computer program product for improving the system behavior of a technical entity in a scenario, such as in particular a vehicle in a driving situation.

Passive vehicle safety refers to technologies and features of a vehicle that help protect occupants and minimize injuries in the event of an accident. These include, for example, airbags, seat belts, head restraints and reinforced passenger cells. In contrast, active vehicle safety refers to technologies that help prevent accidents or mitigate their consequences, such as anti-lock braking systems (ABS) and electronic stability control (ESC).

Crash simulation is used to predict the behavior of a structure or a technical component or system, e.g. a vehicle with a dummy in the interior, in the event of an impact or crash. This is done by creating a computer model of the structure or component, specifying the properties of the materials used in the construction and simulating the effects of an impact on the model. Crash simulation is therefore an important tool for product development in various industries, including the automotive and aerospace industries. In the automotive industry, crash simulation is used to evaluate and optimize the safety of vehicle designs. In aerospace, crash simulation is used, for example, to evaluate the strength and durability of aircraft components.

The results of the crash simulation make it possible to identify potential weak points and improve the performance of structures or technical systems before they are built and tested under real conditions. This saves costs and time and, above all, increases the safety of occupants and users.

In particular, the finite element method (FEM) is used in crash simulation to model and predict the behavior of components, structures and systems in an impact. FEM software can calculate the behavior of materials under different stress and deformation conditions and analyze the resulting forces and accelerations in the structure. The advantage of applying the finite element method in crash simulation is the ability to simulate the behavior and reactions of components or structures under different crash scenarios. This can help to find errors and weak points in existing designs and improve new developments.

The finite element method (FEM) is a method for the numerical solution of partial differential equations in mechanics. It is used to calculate the stresses and deformations of components under static or dynamic loading. In FEM simulation, the system or component to be analyzed is discretized into a finite number of elements, with each element represented by a set of equations based on its geometry, material properties and boundary conditions.

To simulate the impact of a vehicle with a dummy, the user enters the crash test scenario parameters such as speed and impact angle and starts the simulation. The software generates a digital animation of the impact, showing how the vehicle and dummy behave during the impact. The simulation can also provide detailed data on the forces and loads acting on the dummy, which can be used to evaluate vehicle safety and improve vehicle design.

However, carrying out crash simulations based on FEM simulation requires a lot of experience and know-how. In addition, all relevant parameters - from the vehicle geometry to the type of impact load to the material data - must be taken into account. In addition, the FEM simulation is only as accurate as the assumptions and data that go into the analysis. Simulations for the same technical system or component can therefore lead to different results. In addition, an impact is a complex and non-linear system behavior, so there is no proportional relationship between the forces that occur and the effects. This means that a small change in an impact scenario can have a major impact on the system behavior.

Repeatedly running simulations to determine optimal parameters, e.g. for a vehicle restraint system, is therefore very time-consuming and computationally intensive. Since the parameter changes are typically entered by an engineer, it can take a very long time to find an optimal parameter configuration due to the complex system behavior and the highly nonlinear dependencies of the parameters.

Therefore, it is desirable to perform a structural optimization to determine the structure of a System, especially a vehicle, so that the desired safety properties are achieved more quickly. However, the application of known optimization algorithms such as gradient methods, evolution algorithms, annealing methods or particle swarm optimization often leads to unsuitable or inaccurate results because the number of parameters is too large and the dependency behavior of the parameters cannot be accurately mapped.

The US 6 850 821 B1 describes a method for cascading system-level objectives to component-level design objectives using machine learning and design synthesis techniques. Surrogate models are created from simulation results, test results, or benchmarking data.

The DE 10 202020 128 881 A1 discloses improved component sizing by inputting histograms of operating parameters and environmental data of the test vehicle into a machine learning program that outputs a classification that is in turn used by a predictive component model.

In the EP 4 060 443 A1 A method for producing a composite structure is disclosed in which neural networks are used to classify inconsistencies.

The invention is based on the object of demonstrating possibilities for improving the system behavior of an entity such as a technical system or a component in a scenario, in particular for a restraint system in a vehicle in a driving situation, and thus creating the prerequisite for an improved structural optimization of the entity.

The present invention enables parameters, parameter values and/or properties of an entity, such as a restraint system in a vehicle, to be optimized with regard to the system behavior in a specific scenario, such as an impact situation. By coupling a pre-trained simulation model with a learning reinforcement module, optimized values for parameters, parameter values and/or properties of the entity can be calculated without having to carry out a large number of simulations, each with modified simulation data. Since, for example, the system behavior of a vehicle with a dummy in the interior is complex and highly nonlinear, it is difficult for an engineer to decide which parameters, parameter values and/or properties should be changed and to what extent in order to achieve acceptable system behavior of the entity, e.g. the vehicle.

The invention can therefore determine an optimal configuration of parameters, parameter values and/or properties of an entity for the system behavior in a specific scenario, so that in subsequent simulations only a fine-tuning of the parameters, parameter values and/or properties is required. This can reduce the execution of a large number of simulations, which can save computing time and thus costs and energy. In particular, a pre-trained simulation model can be used to use finite element methods directly in a learning reinforcement module. The invention can thus help to increase and improve the safety and performance of an entity.

This object is achieved according to the invention with regard to a method by the features of patent claim 1, with regard to a system by the features of patent claim 11, and with regard to a computer program product by the features of patent claim 15. The further claims relate to preferred embodiments of the invention.

• - Erzeugen eines Simulationsmodells zur Simulation des Systemverhaltens der technischen Entität in einem Simulationsszenario und Trainieren des Simulationsmodells mit Trainingsdatensätzen, wobei die Trainingsdatensätze Messignale von Sensoren enthalten;
• - Bereitstellen von Eingabedaten über eine Benutzerschnittstelle oder aus einer Datenbank, wobei die Eingabedaten insbesondere Parameter, Parameterwerte und/oder Eigenschaften der Entität für eine bestimmte Simulationsaufgabe hinsichtlich des Systemverhaltens in einem bestimmten Szenario umfassen;
• - Durchführen einer Simulation unter Verwendung der Eingabedaten und des trainierten Simulationsmodells von einem Simulationsmodul für die Simulationsaufgabe;
• - Erzeugen von Simulationsergebnissen, wobei die Simulationsergebnisse das Systemverhalten der Entität in dem Simulationsszenario wiedergeben;
• - Weitergeben der Simulationsergebnisse an ein Lernverstärkungsmodul;
• - Berechnen von modellierten Werten für die Parameter, Parameterwerte und/oder Eigenschaften der Entität aus den Simulationsergebnissen zur Erzielung eines verbesserten Systemverhaltens der Entität in dem Simulationsszenario;
• - Erzeugen und Ausgeben von Ausgabedaten durch ein Ausgabemodul, wobei die Ausgabedaten optimierte Ergebnisse insbesondere in Form der modellierten Werte für die Parameter, Parameterwerte und/oder Eigenschaften der Entität enthalten.

According to a first aspect, the invention provides a method for improving the system behavior of a technical entity in a scenario, such as in particular a vehicle in a driving situation. The method comprises the method steps:

• - Generating a simulation model for simulating the system behavior of the technical entity in a simulation scenario and training the simulation model with training data sets, wherein the training data sets contain measurement signals from sensors;
• - Providing input data via a user interface or from a database, wherein the input data comprises in particular parameters, parameter values and/or properties of the entity for a specific simulation task with regard to the system behaviour in a specific scenario;
• - Performing a simulation using the input data and the trained simulation model from a simulation module for the simulation task;
• - generating simulation results, the simulation results representing the system behaviour of the entity in the simulation scenario;
• - Passing the simulation results to a learning reinforcement module;
• - Calculate modeled values for the parameters, parameter values and/or eigen properties of the entity from the simulation results to achieve improved system behavior of the entity in the simulation scenario;
• - Generating and outputting output data by an output module, wherein the output data contains optimized results, in particular in the form of the modeled values for the parameters, parameter values and/or properties of the entity.

In an advantageous embodiment, it is provided that the training data for the simulation model represent input and/or output data of a FEM simulation such as in particular parameters, time series data and finite element attributes, wherein finite element attributes are node positions of finite elements in 3-dimensional space.

In a further development, it is planned that in the case of inconsistent data formats of the input and output data of the FEM simulation, an adjustment is carried out by a transformation into a uniform voxel grid.

In a further development, it is provided that the learning reinforcement module comprises at least one learning reinforcement agent, an action module, an environment module, a state module and a reward module.

In an advantageous embodiment, it is provided that a state of the entity is determined by the state module, wherein a state is determined by parameters and/or parameter values and/or a property of the entity.

In a further embodiment, it is provided that the simulation model uses at least one neural network, in particular recurrent neural networks or convolutional neural networks, or gradient tree boosting as algorithm.

In particular, the reinforcement learning algorithm of the learning reinforcement module is designed as a Markov decision process or as Temporal Difference Learning (TD-Learning) or as Q-Learning or as SARSA or as Monte Carlo simulation or as Actor-Critic or as Proximal Policy Optimization.

Advantageously, at least one parameter and/or one parameter value and/or one property entity represents a position, a dimension, a material, a shape, a characteristic value, a measured value, a safety factor, a speed, an acceleration, a force, a torque, a rotational speed, a state of charge, a specific fuel consumption, a thermal resistance, an electrical resistance, a conductivity, a strength, a fracture behavior, a density, a plastic deformation, a stress, pressure, bending resistance, a material property or a surface property.

In a further embodiment, it is provided that a calculation function and/or an action is selected based on a policy for the state for modifying at least one parameter and/or a parameter value and/or a property of the entity by the learning reinforcement agent; wherein a new state is calculated by the environment module based on the modified value for the parameter and/or the parameter value and/or the property; wherein the new state is compared with a target state and a deviation for a comparison result is calculated in the state module; wherein a reward is determined by the reward module for the comparison result; and wherein the states and/or actions are changed until the target state is reached.

In particular, it is intended that the output data contain design plans, images and graphics, video sequences, acoustic sequences, holograms, key figures, diagrams, mathematical formulas and/or Power Point presentations, etc.

In a further development, it is planned that the environment module includes a software application for simulating the system behavior of the entity using a finite element method and that the results of the FEM simulation are forwarded directly to the state module.

In a further embodiment, the learning reinforcement module performs a multifunctional geometry optimization through a plurality of iterations.

According to a second aspect, the invention provides a system for improving the system behavior of a technical entity in a scenario, such as in particular a vehicle in a driving situation. The system comprises an input module, a simulation module, a learning reinforcement module and an output module. The simulation model is designed to generate a simulation model for simulating the system behavior of the technical entity in a simulation scenario by training with training data sets, wherein the data of the training data sets contain measurement signals from sensors. The input module is designed to provide input data via a user interface or from a database, wherein the input data in particular parameters, parameter values and/or properties of the entity for a specific simulation task such as the system behavior in a specific scenario. The simulation module is further designed to carry out a simulation using the input data and the trained simulation model for the simulation task and to generate simulation results, wherein the simulation results reflect the system behavior of the entity in the simulation scenario. The learning reinforcement module is designed to calculate modeled values for the parameters, parameter values and/or properties of the entity from the simulation results in order to achieve improved system behavior of the entity in the simulation scenario. The output module is designed to generate and output output data, wherein the output data contains optimized results, in particular in the form of the modeled values for the parameters, parameter values and/or properties of the entity.

In a further development, it is provided that the learning reinforcement module comprises at least one learning reinforcement agent, an action module, an environment module, a state module and a reward module.

In a further embodiment, it is provided that the environment module comprises a software application for simulating the system behavior of the entity by means of a finite element method and the results of the FEM simulation are forwarded directly to the state module.

Advantageously, the simulation model uses at least one neural network as an algorithm, in particular recurrent neural networks or convolutional neural networks, or gradient tree boosting.

In particular, at least one parameter and/or a parameter value and/or a property of the entity represents a position, a dimension, a material, a shape, a characteristic value, a measured value, a safety factor, a speed, an acceleration, a force, a torque, a rotational speed, a state of charge, a specific fuel consumption, a thermal resistance, an electrical resistance, a conductivity, a strength, a fracture behavior, a density, a plastic deformation, a stress, pressure, bending resistance, a material property or a surface property.

According to a third aspect, the invention provides a computer program product comprising executable program code configured to carry out the method according to the first aspect when executed.

The invention is explained in more detail below with reference to embodiments shown in the drawing.

• 1 ein Blockdiagramm zur Erläuterung eines Ausführungsbeispiels eines erfindungsgemäßen Systems;
• 2 eine Darstellung eines Lernverstärkungsmoduls;
• 3 ein Flussdiagramm zur Erläuterung der einzelnen Verfahrensschritte eines erfindungsgemäßen Verfahrens;
• 4 ein Blockdiagramm eines Computerprogrammprodukt gemäß einer Ausführungsform des dritten Aspekts der Erfindung.

It shows:

• 1 a block diagram for explaining an embodiment of a system according to the invention;
• 2 a representation of a learning reinforcement module;
• 3 a flow chart explaining the individual process steps of a process according to the invention;
• 4 a block diagram of a computer program product according to an embodiment of the third aspect of the invention.

Additional features, aspects and advantages of the invention or embodiments thereof will become apparent from the detailed description taken in conjunction with the claims.

A restraint system in automotive engineering is a safety mechanism designed to protect occupants in the event of an accident or heavy braking. It typically includes seat belts, airbags, and head restraints that help keep occupants in the vehicle and minimize injuries. A capable restraint system is an important part of the safety equipment of any modern vehicle. To simulate the behavior of the human body in an accident or other collision, dummies are usually used when testing restraint systems in vehicles. They are specifically designed to measure the forces and loads the body is subjected to in an accident and provide valuable data. The dummies are equipped with sensors that are attached all over the body and measure the forces and loads that occur in an accident. The sensors in a crash test dummy collect data about the movements of the dummy and the forces acting on it during an accident. This data is then evaluated to assess the safety of the vehicle being tested. The data is used to improve the vehicle design or safety equipment.

A computer simulation with a crash test dummy in a vehicle is a virtual model of a crash test created using computer software, often based on the finite element method. The simulation includes usually a digital representation of a vehicle, a crash test dummy, and other objects and environmental factors relevant to the test scenario. The simulation makes it possible to analyze the behavior of the vehicle and dummy during an impact without having to conduct a physical test. The simulation software is usually designed to replicate the physics of a real impact as closely as possible, taking into account factors such as the speed, weight and dimensions of the vehicle, as well as the characteristics of the crash test dummy and any other objects in the simulation. The software contains algorithms that simulate the forces and loads the dummy is subjected to during the impact.

To perform the simulation, the user enters the parameters of the crash test scenario, such as speed and impact angle, and starts the simulation. The software generates a digital animation of the impact, showing how the vehicle and dummy behave during the impact. However, repeatedly running simulations to determine optimal parameters, e.g. for a restraint system in the vehicle, is very time- and computationally intensive. Since the parameter changes are typically entered by an engineer, it can take a very long time to find an optimal parameter configuration due to the complex system behavior and the highly nonlinear dependencies of the parameters.

The system 100 according to the invention is based on reinforcement learning methods and is in 1 The system 100 comprises an input module 200, a simulation module 300, a learning reinforcement module 400 and an output module 500. The simulation module 300 comprises a simulation model 350 with software for carrying out a simulation of the system behavior of a technical entity 10. The entity 10 can be a part, a component or a technical system. In particular, the entity 10 can be a vehicle with a dummy in the interior, so that, for example, to analyze the restraint system, a simulated crash test is carried out with the vehicle and the crash test dummy. The technical entity 10 can also be an aircraft, a drone, a motorcycle, a bicycle, a household appliance, an electric motor, a solar system, a wind turbine, a production plant, etc.

The entity 10 can be described by parameters p _i and/or parameter values and/or properties e _i . A parameter p _i and/or a parameter value and/or a property e _i can be a position, a dimension, a material, a shape, a characteristic value, a measured value, a safety factor, a speed, an acceleration, a force, a torque, a rotational speed, a state of charge, a specific fuel consumption, a thermal resistance, an electrical resistance, a conductivity, a strength, a fracture behavior, a density, a plastic deformation, a stress, pressure, bending resistance, a material property or a surface property.

The input module 200, the simulation module 300, the learning reinforcement module 400 and the output module 500 can be designed as independent computer units or as a cloud-based solution. In particular, they can each be provided with a processor and/or a memory unit.

In the context of the invention, a “processor” can be understood to mean, for example, a machine or an electronic circuit. A processor can in particular be a main processor (Central Processing Unit, CPU), a microprocessor or a microcontroller, e.g. an application-specific integrated circuit or a digital signal processor, possibly in combination with a memory unit for storing program instructions, etc. A processor can also be a virtualized processor, a virtual machine or a soft CPU. It can also be, for example, a programmable processor that is equipped with configuration steps for carrying out the aforementioned method according to the invention or is configured with configuration steps such that the programmable processor implements the inventive features of the method, the entity, 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.

In the context of the invention, a "memory unit" or a "memory module" and the like can be understood to mean, for example, a volatile memory in the form of a random access memory (RAM) or a permanent memory such as a hard disk or a data carrier or, for example, a removable memory module. The memory module can also be a cloud-based storage solution.

In the context of the invention, a “module” can be understood as, for example, a processor and/or a memory unit for storing program instructions. For example, the processor is specially designed to execute the program instructions in such a way that the processor and/or the control unit executes functions in order to implement the method or to carry out or implement a step of the method according to the invention.

The simulation model 350 is designed to simulate the system behavior of the entity 10 in a scenario based on input data. A scenario is a specific situation or a temporal development in which forces act on the entity 10 that trigger a specific system behavior of the entity 10, such as an impact situation in a vehicle. For example, the behavior of a crash test dummy in a vehicle in an impact can be predicted based on the parameters of a restraint system, such as an airbag. For an airbag, the threshold values for the triggering mechanism, the number of airbags that trigger, and the inflation speed of the airbags can be varied. In addition, sensors and/or algorithms can be used to determine the need for and the exact time of triggering and to enable individual triggering behavior. These sensors and algorithms are then also used as digital models in the simulation.

The simulation model 350 includes in particular neural networks that have been trained with suitable training data sets. However, other algorithms such as gradient tree boosting can also be used to make predictions about the behavior of a system.

A neural network consists of neurons that are arranged in several layers and connected to one another in different ways. A neuron is able to receive information from outside or from another neuron at its input, evaluate it in a certain way and pass it on in a modified form to another neuron at the neuron's output or output it as the final result. Hidden neurons are located between the input neurons and the output neurons. Depending on the type of network, there can be several layers of hidden neurons. They ensure that the information is forwarded and processed. Output neurons finally provide a result and pass it on to the outside world. The arrangement and interconnection of the neurons creates different types of neural networks, such as feed-forward networks, recurrent networks or convolutional neural networks. The networks can be trained using unsupervised or supervised learning.

The convolutional neural network (CNN) is a special form of an artificial neural network. It has several convolutional layers and is very suitable for machine learning and applications of artificial intelligence (AI) in the field of pattern recognition. The individual layers of the CNN are the convolutional layer, the pooling layer and the fully linked layer. The convolutional layer is the actual convolutional layer and is able to recognize and extract individual features in the input data. In pattern and image recognition, these can be features such as lines, edges or certain 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 suitable filter functions. The reduced amount of data increases the computing speed. Since the CNN is divided into several local, partially networked layers, it requires significantly less storage space than fully networked neural networks. The training time of the convolutional neural network is also significantly shorter. By using modern graphics processors, CNNs can be trained very efficiently.

Gradient Tree Boosting (GBT) is a machine learning algorithm that combines the concepts of gradient boosting and decision trees. It is a powerful technique for solving regression and classification tasks. In gradient tree boosting, the algorithm iteratively generates a set of decision trees. Each tree is constructed to correct the errors of the previous trees in the ensemble. In this way, a loss function is minimized by adjusting the model's predictions based on the gradient of the loss function.

In particular, for reliable prediction, regression analysis is an important aspect of simulation model 350. It is a procedure of estimating an unknown function from data in order to make predictions or identify patterns in the data. Regression analysis is a statistical method to find a function that describes the relationship between a dependent variable and one or more independent variables. It is mainly used to understand and predict the value of a continuous target variable based on the values of other variables. In regression analysis, the dependent variable, also called the response variable or target variable, is the variable of interest whose behavior is to be predicted. The independent variables, also called predictor variables or features, are the variables that are believed to be related to the dependent variable and can help in prediction.

The goal of regression analysis is to estimate the parameters of the regression model that best describe the relationship between the dependent variable and the independent variables. The parameters are the coefficients or weights assigned to each independent variable that indicate the strength and direction of its influence on the dependent variable. There are several types of regression analysis, notably simple linear regression with a single independent variable, multiple linear regression with several independent variables, and polynomial regression, which takes into account nonlinear relationships.

In regression analysis, the parameters of the regression model are estimated using various statistical techniques, such as ordinary least squares (OLS) or maximum likelihood estimation. Once the model is trained, it can be used to predict new input data.

The input module 200 can be integrated into a computer, a tablet, a smartphone, etc. and has a user interface 220 for entering input data 240, such as parameters of an airbag. The user interface 220 is designed in particular as a touchscreen, via which numerical values, a drawing or a diagram and other inputs can be entered. The input module 200 can additionally have a keyboard, a microphone and/or a camera. The touchscreen is designed in particular as a resistive, capacitive or inductive system. However, other technologies can also be used. In particular, the input module 200 has an input software application 250 for entering the input data 240 to carry out a simulation with the simulation model 350.

Furthermore, the input module 200 can be connected to sensors 280, which transmit measurement signals as input data 240 from a system, such as a vehicle with a dummy, to the input module 200. The sensors 280 can be designed in particular as pressure sensors, piezo sensors, torque sensors, rotational speed sensors, acceleration sensors, speed sensors, capacitive sensors, inductive sensors, temperature sensors, image-recording cameras and/or optical distance and speed measuring systems. The measurement data recorded by the sensors are images and/or data about physical or chemical processes.

Furthermore, the input module 200 and/or the simulation module 300 can be connected to at least one database 290. The term "database" refers both to a storage algorithm and to the hardware in the form of a storage unit. In particular, the database 290 can be designed as a cloud computing infrastructure.

In particular, historical data in the form of images, graphics, time series, parameters, etc. are stored in the database 290. For example, target and target values that define a safety standard can be stored in the database 290. In addition, a digital twin of a system component of the entity 10, e.g. a pillar in a vehicle or an airbag, can be stored in the database 290. A digital twin represents a digital model in the virtual world and contains simulations and algorithms that describe the properties and behavior of an object. In addition, simulation results from component or system simulations can be stored in the database 290. The simulations are in particular crash simulations with a vehicle and a dummy in the interior, whereby simulation parameters such as the vehicle speed, the properties of the obstacle, the seating position of the dummy, the seat construction, fittings in the vehicle interior, safety devices such as airbags and seat belts can vary.

The input data 240 can be further processed by the input module 200 or passed directly to the simulation module 300. In particular, coding of the input data can be provided to enable better processing of the input data 240 by the simulation module 300. For example, the input module 200 can have a transformer module with an encoder and a decoder to translate the input data into another character string. The encoder and the decoder use artificial intelligence algorithms such as deep learning and machine image analysis to enable, for example, extraction and classification of features relevant to the simulation. Deep neural networks and/or recurrent neural networks and/or convolutional neural networks are used as algorithms. In particular, attention algorithms are used that mimic cognitive attention and use convolutional neural networks. The transformer module is trained accordingly with training data.

If the input module 200 and the simulation module 300 are not integrated in the same computing unit, communication connections such as a CAN bus system (controller Area Network) or wireless connections can be provided. A wireless communication connection is designed in particular as a mobile radio connection and/or near field communication connection such as Bluetooth®, Ethernet, NFC (near field communication) or Wi-Fi®.

It can be planned to equip the input module 200 with mobile radio modules of the 5G or 6G standard. 5G is the fifth generation mobile radio standard and is characterized by higher data rates of up to 10 Gbit/sec compared to the 4G mobile radio standard, the use of higher frequency ranges such as 2100, 2600 or 3600 megahertz, a higher frequency capacity and thus a higher data throughput. The latency times are very short at a few milliseconds to less than 1 ms, so that real-time transmission of data and calculation results is possible.

The high speed of data transmission is advantageous when cloud-based solutions are used for data processing. Cloud-based solutions offer the advantage of high and therefore fast computing power. If the simulation module 300 and/or the learning reinforcement module 400 are integrated into the cloud computing infrastructure 700, the input data 240 transmitted by the input module 200 can be sent to the simulation module 300 in real time. The simulation can then be carried out in the cloud computing infrastructure using the simulation model 350.

Cryptographic encryption methods can be provided to protect communication with the cloud computing infrastructure via a mobile radio connection. In particular, cryptographic encryption methods are provided for the connection between the input module 200 and the simulation module 300 and/or for the connection between the learning reinforcement module 400 and the output module 500.

The input data 240 may at least partially include voxel grid data formats. A voxel grid is a three-dimensional grid composed of voxels (volume elements). Each voxel represents a small portion of 3D space and contains information such as color, material, or texture of the object located at that grid cell in space. Voxel grids are often used in computer graphics to calculate or simulate complex 3D scenes.

While in the finite element method (FEM) the geometry is decomposed into a network of finite elements and these elements are typically represented as triangles or tetrahedra in 2D or 3D analyses, voxel grids are commonly used in voxel-based modelling and simulation techniques such as voxel-based finite difference methods or voxel-based finite volume methods. In these methods, the area of interest is discretised into a regular grid of cubic cells, the so-called voxels. Voxel-based methods are particularly suitable for modelling complex geometries. A voxel-based approach is therefore particularly interesting for the topology optimization of the entire vehicle, since the entire geometry can be mapped very homogeneously.

Although voxel-based modeling and FEM share some similarities, such as discretization and numerical analysis, they are different methods with different applications. FEM is more associated with traditional network-based approaches, while voxel-based techniques are often used in areas where dynamic processes are studied, such as the analysis of flows in porous media. However, the voxel grid can be converted to a finite element mesh by various techniques to enable finite element analysis.

The simulation module 300 uses the simulation model 350 to calculate simulation results 340 that include a prediction of the system behavior of the entity 10 in a scenario, such as the movement of the dummy during the impact and the deformation of the vehicle in an impact situation.

According to the invention, the input data 240 and the calculated simulation results 340 are transferred to the learning reinforcement module 400. As in 2 As shown, the reinforcement learning module 400 includes a reinforcement learning agent (RV) 410, an action module 420, an environment module 430, a state module 440, and a reward module 450.

In the learning reinforcement module 400, the learning reinforcement agent 410 selects a specific state s _i ∈ S from a set of available states for at least one action a _i ∈ 4 from a set of available actions. According to the invention, the actions a _i are the parameters p _i of the input data 240. For the selected action a _i, the LV agent 410 receives a reward that can be positive, neutral or negative. The states s _i are obtained by the LV agent 410 from the state module 440, which derives the states s _i from the simulation results 340. A state s _i ∈ S is determined by the selection of specific parameters p _i , parameter values and/or properties e _i of the entity 10. In addition, further data can be obtained from the various various sensors 280 and/or the database 290. These processed data are assigned states s _i which the LV agent 410 can access.

The choice of the selected action a _i is based on a strategy or policy. For the selected action a _i, the agent 410 receives a reward r _i ∈ ℝ from the reward module 450. The strategy in turn changes because it is adapted on the basis of the rewards r _i received. The strategy is used to determine which action a _i ∈ A is to be selected from the set of available actions for a specific state s _i ∈ S from the set of available states. This creates a new state s _i+1 for which the LV agent 410 receives a reward r _i . A strategy thus determines the assignment between a state s _i and an action a _i and thus specifies which action a _i to be carried out is to be selected for a state s _i . The goal of the LV agent 410 is to maximize the rewards r _i achieved.

In the action module 420, the actions a _i selected by the LV agent 410 are executed. An action a _i adjusts a parameter p _i and/or a parameter value and/or a property e _i of the entity 10. Preferably, an action a _i is one of the three actions A(+), A(0) and A(-). A positive action A(+) refers to an action that increases the value for a parameter p _i , a neutral action A(0) is an action in which the value of the parameter p _i remains the same, and a negative action A(-) reduces the value of the parameter p _i .

The environment module 430 calculates the states s _i based on the selected action a _i and taking into account the requirements and constraints determined by the simulation results 340. In addition, it can be provided that further constraints are taken into account, for example with regard to the cost structure, energy costs, environmental balance, availability or the delivery situation. These constraints can be transmitted separately to the learning reinforcement module 400.

In the reward module 450, a reward r _i is assigned to the degree of deviation Δ between the calculated value for the state s _i and the target value of a target state s _t . Since the degree of deviation Δ depends on the selection of the respective action a _i , the reward r _i for the respectively selected action a _i is preferably stored in a matrix or a database. A reward r _i preferably has the values +1 and -1, whereby a small or positive deviation Δ between the calculated state s _i and the target state s _t is rewarded with +1 and is thus reinforced, while a significant negative deviation Δ is rewarded with -1 and is thus negatively evaluated. However, it is also conceivable that values > 1 and values < 1 are used. The result state is reached when the calculated states s _i are equal to or greater than the target state s _t .

Preferably, a Markov decision process is used as the algorithm for the LV agent 410. However, it can also be provided to use a temporal difference learning (TD learning) algorithm. The LV agent 410 with a TD learning algorithm does not adjust the actions a _i only when it receives the reward, but after each action a _i on the basis of an estimated expected reward. Furthermore, algorithms such as Q-learning and SARSA are also conceivable, or Actor-Critic or Monte Carlo simulations, with which dynamic programming and adjustment of the strategy by iteration processes is possible.

In addition, the LV agent 410 and/or the action module 420 and/or the environment module 430 and/or the state module 440 and/or the reward module 450 contain calculation methods and algorithms f _i for mathematical regression methods or physical model calculations that describe a correlation between selected parameters p _i and a target state s _t . The mathematical functions f _i can be statistical methods such as mean values, minimum and maximum values, lookup tables, models for expected values, linear regression methods or Gaussian processes, fast Fourier transformations, integral and differential calculus, Markov methods, probability methods such as Monte Carlo methods, temporal difference learning, but also extended Kalman filters, radial basis functions, data fields, convergent neural networks, deep neural networks, and/or feedback neural networks. Based on the actions a _i and the rewards r _i , the LV agent 410 and/or the action module 420 and/or the environment module 430 and/or the state module 440 and/or the bypass module 430 selects one or more of these computation functions f _i for a state s _i .

Now a second cycle begins to optimize the parameters p _i and/or parameter values and/or properties e _i of the entity 10. Here, the LV agent 410 can select another action a _i+1 and/or another calculation function f _i+1 and/or another parameter p _i+1 according to the defined strategy or guideline. The result is again fed to the state model dul 440 and the result of the comparison is evaluated in the reward module 450. The LV agent 410 repeats the optimization process for all intended actions a _i , calculation functions f _i and parameters p _i until the greatest possible agreement between a calculated state s _i and a target state s _t is achieved. Preferably, the result state of the optimization of the entity 10 is reached when the deviation Δ is in the range of +/- 5%. The LV agent 410 thus optimizes its behavior and thus the strategy with which an action a _i is selected until the calculated states s _i converge. The optimized result of the design of the entity 10 by modified parameters p _i and/or parameter values and/or properties e _i is output in the form of output data 550 by the output module 500.

In the optimized result state, the optimized values for the parameters p _i and/or parameter values and/or properties e _i of the entity 10 are defined, with which a specific behavior of the entity 10, such as a specific safety criterion in the event of a collision between a vehicle and a dummy, can be achieved. If the simulation of the simulation model 350 has shown, for example, that the acceleration acting on a dummy head in the event of an impact is too great, the parameters p _i and/or parameter values and/or properties e _i of the entity 10 are modified and adapted by the learning reinforcement module 400 until an acceptable acceleration value is reached.

In a further development, it can be provided that the simulation module 300 is only used for pre-training the learning reinforcement module 400. In this case, the environment module 430 of the learning reinforcement module 400 has a simulation software 435 that is designed in particular for carrying out finite element simulations. The results of the finite element simulation are then fed directly to the state module 440, which derives the corresponding states s _i from them, which the learning reinforcement agent 410 can access.

The pre-trained learning reinforcement module 400 is thus used according to the invention to solve similar optimization tasks. For example, the weight of a dummy can be varied. If the learning reinforcement module 400 was trained with a dummy with a body weight of 90 kg, the restraint system can now be optimized for a dummy with a body weight of 55 kg without the simulation results 340 having to be calculated by the simulation module 300. In this case, the suggestions of the RV agent 410 are recalculated directly by the simulation software 435 of the environment module 430 to determine new states s _i .

The output module 500 is designed to output the optimized results in the form of output data 550. The output module 500 can also be integrated in a hardware device such as a computer, a tablet, a smartphone, etc., wherein in particular the hardware device for the input module 200 and the hardware device for the output module 500 can be identical.

The optimized results in the form of output data 550 can be output directly by the output module 500. However, it can also be provided that the output module 500 has an output transformer module which translates the optimized results calculated by the learning reinforcement agent 400 into a data format accessible to a user. The data formats can be, for example, design plans, images and graphics, video sequences, acoustic sequences, holograms, key figures, diagrams, mathematical formulas, Power Point presentations, etc. For this purpose, the output transformer module can also have at least one encoder and one decoder or encoders and decoders connected in series.

In particular, it can be provided that the calculation results are stored in the form of states, actions, rewards and strategies in the cloud computing infrastructure and are available via the Internet. The LV agent 410, the action module 420, the environment module 430, the state module 440 and the reward module 450 have the necessary technical interfaces and protocols for accessing the cloud computing infrastructure. This can increase computing efficiency because the access options and access speeds to already calculated states, actions, rewards and strategies are simplified.

Since a fast calculation result is often desired, the computing speed is crucial for the large number of computing operations in the simulation module 300 and the learning reinforcement module 400. In order to enable the computing operations to be carried out in real time, the use of the cloud computing infrastructure is therefore advantageous, as this can ensure fast calculation.

In a further development, it can also be provided to use the system 100 according to the invention for a multifunctional geometry optimization, in particular of a vehicle. Example For example, different geometric requirements for stiffness arise for a side impact than for a front impact or a rear impact. Through mutual optimization, an optimal solution can be found with the help of the learning reinforcement algorithm 410 that covers a large number of scenarios in such a way that an optimal holistic result is found. It is also possible to weight the probabilities of the various impact scenarios, e.g. using key performance indices (KPIs), and to take this weighting into account when calculating the optimization parameters by the learning reinforcement agent 410.

For the multifunctional geometry optimization, standardized data formats based on finite elements, in particular voxel grids or parameters, are used as input data 240 for the environment module 430. In addition, the environment module 430 comprises at least one neural network for processing the standardized data formats based on finite elements, in particular voxel grids or parameters, since this is particularly suitable for learning relationships in high-dimensional data.

In 3 the method steps for improving the system behavior of a technical entity 10 in a scenario, such as in particular a vehicle in a driving situation, are shown.

In a step S10, a simulation model 350 is used to simulate the system behavior of the technical entity 10 in a simulation scenario and to train the simulation model 350 with training data sets, wherein the data of the training data sets contain measurement signals from sensors 280.

In a step S20, input data 240 is provided via a user interface 220 or from a database 290, wherein the input data 240 comprises parameters p _i , parameter values and/or properties e _i of the entity 10 for a specific simulation task such as the system behavior in a specific scenario.

In a step S30, a simulation is performed using the input data 240 and the trained simulation model 450 by a simulation module 400 for the simulation task.

In a step S40, simulation results 340 are generated, wherein the simulation results 340 reflect the system behavior of the entity 10 in the simulation scenario.

In a step S50, the simulation results 340 are passed on to a learning reinforcement module 400.

In a step S60, modeled values for the parameters p _i , parameter values and/or properties e _i of the entity are calculated from the simulation results in order to achieve an improved system behavior of the entity 10 in the simulation scenario.

In a step S70, output data 550 are generated and output by an output module 500, wherein the output data 550 contains optimized results, in particular in the form of the modeled values for the parameters p _i , parameter values and/or properties e _i of the entity 10.

4 schematically illustrates a computer program product 900 comprising an executable program code 950 configured to carry out the method according to the first aspect of the present invention.

The invention enables parameters, parameter values and/or properties of an entity, such as a restraint system in a vehicle, to be optimized with regard to the system behavior in a specific scenario, such as an impact situation. By coupling a pre-trained simulation model with a learning reinforcement module, optimized values for parameters, parameter values and/or properties of the entity can be calculated without having to carry out a large number of simulations, each with modified simulation data. Since, for example, the system behavior of a vehicle with a dummy in the interior is complex and highly nonlinear, it is difficult for an engineer to decide which parameters, parameter values and/or properties should be changed and to what extent in order to achieve acceptable system behavior of the entity, such as the vehicle.

The invention can therefore determine an optimal configuration of parameters, parameter values and/or properties of an entity for the system behavior in a specific scenario, so that in subsequent simulations only a fine-tuning of the parameters, parameter values and/or properties is required. This can reduce the execution of a large number of simulations, which can save computing time and thus costs and energy. In particular, a pre-trained simulation model can be used to use finite element methods directly in a learning reinforcement module. The invention can thus help to increase and improve the safety and performance of an entity.

reference sign

10

entity

100

system

200

input module

220

user interface

240

input data

250

input software application

280

sensors

290

database

300

simulation module

340

simulation results

350

simulation model

400

learning reinforcement module

410

learning reinforcement agent

420

action module

430

environment module

435

software application

440

state module

450

reward module

500

output module

550

output data

900

computer program product

950

program code

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. 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

• US 6 850 821 B1 [0011]
• DE 10 202020 128 881 A1 [0012]
• EP 4 060 443 A1 [0013]

Claims (15)

Method for improving the system behavior of a technical entity (10) in a scenario, such as in particular a vehicle in a driving situation, comprising: - generating (S10) a simulation model (350) for simulating the system behavior of the technical entity (10) in a simulation scenario and training the simulation model (350) with training data sets, wherein the training data sets contain measurement signals from sensors (280); - providing (S20) input data (240) via a user interface (220) or from a database (290), wherein the input data (240) comprise parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10) for a specific simulation task such as the system behavior in a specific scenario; - carrying out (S30) a simulation using the input data (240) and the trained simulation model (450) from a simulation module (400) for the simulation task; - generating (S40) simulation results (340), wherein the simulation results (340) reflect the system behavior of the entity (10) in the simulation scenario; - forwarding (S50) the simulation results (340) to a learning reinforcement module (400); - calculating (S60) modeled values for the parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10) from the simulation results (340) in order to achieve improved system behavior of the entity (10) in the simulation scenario; - generating and outputting (S70) output data (550) by an output module (500), wherein the output data (550) contain optimized results, in particular in the form of the modeled values for the parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10). procedure according to claim 1 , wherein the learning reinforcement module (400) comprises at least one learning reinforcement agent (410), an action module (420), an environment module (430), a state module (440) and a reward module (450). procedure according to claim 1 or 2 , wherein a state (s _i ) of the entity (10) is determined by the state module (440), wherein a state (s _i ) is determined by parameters (p _i ) and/or parameter values and/or a property (e _i ) of the entity (10). Method according to one of the Claims 1 until 3 , wherein the simulation model (350) uses at least one neural network as algorithm, in particular recurrent neural networks or convolutional neural networks (convolutional neural networks) or gradient tree boosting. Method according to one of the Claims 1 until 4 , wherein the reinforcement learning algorithm of the learning reinforcement module (400) is designed as a Markov decision process or as temporal difference learning (TD-learning) or as Q-learning or as SARSA or as Monte Carlo simulation or as actor-critic or as proximal policy optimization. Method according to one of the Claims 1 until 5 , wherein at least one parameter (p _i ) and/or a parameter value and/or a property (e _i ) of the entity (10) represents a position, a dimension, a material, a shape, a characteristic value, a measured value, a safety factor, a speed, an acceleration, a force, a torque, a rotational speed, a state of charge, a specific fuel consumption, a thermal resistance, an electrical resistance, a conductivity, a strength, a fracture behavior, a density, a plastic deformation, a stress, pressure, bending resistance, a material property or a surface property. Method according to one of the Claims 2 until 6 , wherein a calculation function (f _i ) and/or an action (̇a _i ) is selected based on a policy for the state (s _i ) for modifying at least one parameter (p _i ) and/or a parameter value and/or a property (e _i ) of the entity (10) by the learning reinforcement agent (410); wherein a new state (s _i+1 ) is calculated by the environment module (430) based on the modified value for the parameter (p _i ) and/or the parameter value and/or the property (e _i ); wherein the new state (s _i+1 ) is compared with a target state (s _t ) and a deviation (Δ) for a comparison result is calculated in the state module (440); wherein a reward (r _i ) is determined by the reward module (450) for the comparison result; and wherein the states (s _i ) and/or actions (a _i ) are changed until the target state (s _t ) is reached. Method according to one of the Claims 1 until 7 , wherein the output data (550) contains design plans, images and graphics, video sequences, acoustic sequences, holograms, key figures, diagrams, mathematical formulas and/or Power Point presentations, etc. Method according to one of the Claims 2 until 8 , wherein the environment module (430) comprises a software application (435) for simulating the system behavior of the entity (10) by means of a finite element method and the results of the FEM simulation can be forwarded directly to the state module (440). Method according to one of the Claims 1 until 9 , wherein the learning reinforcement module (400) performs a multifunctional geometry optimization through a plurality of iterations. System (100) for improving the system behavior of a technical entity (10) in a scenario, such as in particular a vehicle in a driving situation, comprising an input module (200), a simulation module (300), a learning reinforcement module (400) and an output module (500), wherein the simulation module (300) is designed to generate a simulation model (350) for simulating the system behavior of the technical entity (10) in a simulation scenario by training with training data sets, wherein the data of the training data sets contain measurement signals from sensors (280); wherein the input module (200) is designed to provide input data (240) via a user interface (220) or from a database (290), wherein the input data (240) in particular comprise parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10) for a specific simulation task such as the system behavior in a specific scenario; wherein the simulation module (300) is designed to carry out a simulation using the input data (240) and the trained simulation model (450) for the simulation task and to generate simulation results (340), wherein the simulation results (340) reflect the system behavior of the entity (10) in the simulation scenario; wherein the learning reinforcement module (400) is designed to calculate modeled values for the parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10) from the simulation results (340) in order to achieve improved system behavior of the entity (10) in the simulation scenario; wherein the output module (500) is designed to generate and output output data (550), wherein the output data (550) contain optimized results, in particular in the form of the modeled values for the parameters (p _i ), parameter values and/or properties (e _i ) of the entity (10). System (100) after claim 11 , wherein the learning reinforcement module (400) comprises at least one learning reinforcement agent (410), an action module (420), an environment module (430), a state module (440) and a reward module (450), and wherein the environment module (430) comprises a software application (435) for simulating the system behavior of the entity (10) by means of a finite element method and the results of the FEM simulation are forwarded directly to the state module (440). System (100) after claim 11 or 12 , wherein the simulation model (350) uses at least one neural network, in particular recurrent neural networks or convolutional neural networks, or gradient tree boosting as algorithm. System (100) according to one of the Claims 11 until 13 , wherein at least one parameter (p _i ) and/or a parameter value and/or a property (e _i ) of the entity (10) represents a position, a dimension, a material, a shape, a characteristic value, a measured value, a safety factor, a speed, an acceleration, a force, a torque, a rotational speed, a state of charge, a specific fuel consumption, a thermal resistance, an electrical resistance, a conductivity, a strength, a fracture behavior, a density, a plastic deformation, a stress, pressure, bending resistance, a material property or a surface property. Computer program product (900) comprising an executable program code (950) which is configured such that, when executed, it carries out the method according to one of the Claims 1 until 10 executes.

Images