Computer-implemented method for simulating a motor vehicle accident using an artificial neural network

Abstract

The invention relates to a computer-implemented method for simulating a motor vehicle accident using an artificial neural network (1), comprising the following steps: - Receipt of input data (14) related to the accident; - Transformation of the input data (14) into transformation data (12); and - Simulation of the accident by the artificial neural network (1) using the transformation data (12).

Description

[0001] The present invention relates to a computer-implemented method for simulating a motor vehicle accident using an artificial neural network.

[0002] German patent application DE 10 2020 115 192 A1 discloses a method for simulating a motor vehicle accident using an artificial neural network. In a hidden layer of the network, input scalars related to the accident are transformed into a transformed signal. This transformation involves a mathematical convolution of the input scalars. The accident is then simulated using the transformed signal.

[0003] In contrast, the present invention is based on the objective of creating an improved method for simulating a motor vehicle accident.

[0004] This problem is solved by a method according to claim 1. Embodiments of the invention are specified in the dependent claims.

[0005] The procedure involves receiving input data related to the accident. This input data may include, for example, information describing the characteristics of the vehicle and / or persons inside it. The input data may describe, for instance, the positions, lengths, widths, and / or heights of structural and / or body parts.

[0006] The input data is transformed into transformation data. This can be done, for example, in a hidden layer of the artificial neural network. The accident is then simulated by the artificial neural network using the transformation data.

[0007] The artificial neural network can, for example, have an encoder-decoder architecture. For instance, convolutions in different dimensions can be used. Preferably, a one-dimensional convolution can be used for the encoder and a two-dimensional convolution for the decoder.

[0008] This is a computer-implemented method. This can mean, for example, that the method is carried out using a digital processing device and a digital data storage device. Instructions can be stored in the digital data storage device, which can be read and executed by the processing device. The instructions can be configured to cause the processing device, upon execution of the instructions, to carry out a method according to an embodiment of the invention.

[0009] According to one embodiment of the invention, the input data can be configured as an input signal. The transformation data can comprise scalars. The scalars can be converted into a transformation signal before the simulation. The simulation can then be performed using this transformation signal.

[0010] This embodiment is advantageous because the scalars transformed from the input signal make it possible to explain how the artificial neural network processes the input signal during the simulation. For example, the scalars can be generated in a hidden layer of the neural network. They can then be analyzed by a data specialist to determine the influence of the input signal on the simulation. Furthermore, it can also be verified whether an appropriate portion of the input signal is being used.

[0011] The scalars can, for example, contain information about the input signal. The transformation signal can, for instance, include several values ​​for the acceleration of a body part of a person in a motor vehicle during the simulated accident.

[0012] According to one embodiment of the invention, the input data can be configured as input scalars.

[0013] According to one embodiment of the invention, the transformation data can be configured as a single transformation scalar. For example, the transformation scalar can include an indication of the severity of the accident. The transformation scalar can, for example, include an indication of the maximum acceleration of a body, a body part, a component of the motor vehicle, or the motor vehicle itself.

[0014] This embodiment is particularly advantageous because the single transformation scalar contains less information than the input data, making it easier for a data specialist to assess whether the correct portion of the input data is being used and how the simulation is affected by the input data. This improves both the precision and the understanding of how the input data influences the transformation scalar.

[0015] According to one embodiment of the invention, the transformation can be carried out using a transformer architecture or an attention mechanism. For the purposes of this description, a transformer architecture is understood to be, in particular, an architecture with encoders and decoders connected in series. The encoders and decoders may have different dimensions. In this embodiment, folding can be avoided. Practical tests have shown that this embodiment produces particularly realistic results.

[0016] According to one embodiment of the invention, the input data can be configured as input scalars. The transformation data can be configured as a first embedded signal and a second embedded signal. The first embedded signal can, for example, provide information about the position and movement of a body or body parts of a person, and the second embedded signal can provide information about the position and movement of components of the motor vehicle. For the purposes of this description, an embedded signal is understood to be, in particular, a signal that does not appear outside the artificial neural network. Practical tests have shown that this embodiment produces particularly realistic results. For example, the first and second signals can be used to generate...

[0017] According to one embodiment of the invention, the first embedded signal and / or the second embedded signal can each be multidimensional. For example, the signals can each have multiple values ​​in multiple dimensions, such as 32 or more dimensions.

[0018] According to one embodiment of the invention, it is possible for the transformation data in the simulation to influence only a present or a future state of the simulation. This can be achieved, for example, through the structure of the artificial neural network. For instance, the neurons of the artificial neural network can receive information and / or data exclusively from the past or the present.

[0019] According to one embodiment of the invention, the transformation data can be inputted into an input layer of the artificial neural network. This is particularly advantageous if the artificial neural network has an architecture as described in DE 10 2020 132 042 A1. Practical tests have shown that this embodiment produces particularly realistic results.

[0020] According to one embodiment of the invention, the transformation data can be input into different, particularly hidden, layers of the artificial neural network and / or used at different times during the simulation. For example, a first subset of the transformation data can be used in a first layer and / or at a first time during the simulation, while a second subset of the transformation data can be used in a second layer and / or at a second time during the simulation. This is also particularly advantageous if the artificial neural network has an architecture as described in DE 10 2020 132 042 A1. This embodiment also allows for the generation of particularly realistic results.

[0021] During the simulation, information and / or data are generated that can be used, for example, to improve the crash safety of a motor vehicle by modifying one or more components.

[0022] Further features and advantages of the present invention will become clear with reference to the following description of preferred embodiments and the accompanying figures. The same reference numerals are used for identical or similar features and for features with identical or similar functions. Fig. 1 a schematic representation of an artificial neural network according to an embodiment of the invention; Fig. 2 a schematic representation of an artificial neural network according to an embodiment of the invention; and Fig. 3 a schematic representation of an artificial neural network according to an embodiment of the invention.

[0023] The in Fig. The artificial neural network shown in Figure 1 comprises an input layer 3, several hidden layers 4, 5 and 6, and an output layer 2. Each of the layers 2-6 has several artificial neurons represented as circles.

[0024] In Fig. Figure 1 represents the time elapsed during the simulation. Data is transferred from the neurons in the direction from input layer 3, through the hidden layers 4-6, to output layer 2, exclusively to neurons in the next layer simulating the same point in the simulation or a later point in the simulation. This ensures that a later point in the simulation cannot influence an earlier point in the simulation.

[0025] For example, if input data relates to the movement and / or location of a motor vehicle or its components, and output layer 2 contains information about the movement and / or location of a human body or body parts, subsequent states of the motor vehicle components or the motor vehicle itself cannot influence earlier states of the body parts or the body. This results in a more physically realistic simulation.

[0026] The in Fig. The artificial neural network shown in Figure 2 comprises an input layer 7, several hidden layers 8, 9, and 10, and an output layer 11. The artificial neural network can be structured, for example, as described in DE 10 2020 132 042 A1. In a dense layer, input data 14 is transformed into transformation data 12. The transformation data 12, together with a one-dimensional vector 13, is fed into the input layer 7 of the artificial neural network. The one-dimensional vector 13 can, for example, represent the location and / or movement of a motor vehicle or parts thereof. The processing of the one-dimensional vector 13 and the transformation data 12 can be carried out as described in DE 10 2020 132 042 A1. By inputting the transformation data 12 into the input layer 7, a particularly realistic simulation is achieved.

[0027] The output data provided by the artificial neural network 11 can, for example, include an indication of the severity of the simulated accident. This could, for instance, be a one-dimensional vector. The severity of the accident could be indicated, for example, by the severity of an injury to a body part.

[0028] The in Fig. The artificial neural network shown in Figure 3 resembles the one in Figure 3. Fig. The artificial neural network shown in Figure 2. However, the transformation data 12 are not entered into input layer 7, but into different hidden layers 9 and 15. Under certain circumstances, depending on the type of simulated accident, this can enable particularly realistic simulations, since the transformation data 12 can be entered where they have a particularly high impact on realism. QUOTES INCLUDED IN THE DESCRIPTION

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

[0000] DE 10 2020 115 192 A1

[0002] DE 10 2020 132 042 A1 [0019, 0020, 0026]

Claims

[1] Computer-implemented method for simulating a motor vehicle accident using an artificial neural network (1), comprising the following steps: - receipt of input data (14) related to the accident; - transforming the input data (14) into transformation data (12); and - Simulation of the accident by the artificial neural network (1) using the transformation data (12). [2] Method according to claim 1, characterized by that the input data (14) are designed as an input signal, wherein the transformation data (12) comprise scalars, wherein the scalars are converted into a transformation signal before the simulation, and wherein the simulation is carried out using the transformation signal. [3] Method according to claim 1, characterized by that the input data (14) are designed as input scalars [4] Method according to the preceding claim, characterized by that the transformation data (12) are formed as a single transformation scalar. [5] Method according to claim 1, characterized by that the transformation is performed using a transformer architecture or an attention mechanism. [6] Method according to claim 1, characterized by that the input data (14) are designed as input scalars, wherein the transformation data (12) are designed as a first embedded signal and a second embedded signal. [7] Method according to the preceding claim, characterized by that the first embedded signal and / or the second embedded signal are each multidimensional. [8] Method according to one of the preceding claims, characterized by that during the simulation the transformation data (12) only influence a current or a future state of the simulation. [9] Method according to one of the preceding claims, characterized by that the transformation data (12) are input into an input layer (3; 7) of the artificial neural network (1). [10] Method according to one of claims 1 to 8, characterized by that the transformation data (12) are entered into different layers (9; 15) of the artificial neural network (1) and / or used at different times when carrying out the simulation.

Images