Devices and methods for training an artificial neural network and predicting the stress on a vehicle occupant

Abstract

Devices and methods for training an artificial neural network (101) and for predicting the load on an occupant (105) of a vehicle, wherein a time series of an intrusion kinematics (102) during an accident is provided, wherein a load on a dummy equipped with the sensor, detected by a sensor during the accident, is provided, wherein an artificial neural network (101) is provided which is trained to map the time series of the intrusion kinematics (102) to the load on the occupant (105) during the accident, wherein the time series of the intrusion kinematics (102) is mapped to the load on the occupant (105) by the artificial neural network (101), wherein the artificial neural network (101) is trained depending on a deviation (106) of the load on the occupant (105) from the load on the dummy (104).

Description

[0001] The invention relates to devices and methods for training an artificial neural network and for predicting a load on an occupant of a vehicle.

[0002] DE 10 2013 221 282 A1 discloses a determination of an intrusion parameter of an intrusion in order to enable an improved initiation of security measures, in particular a faster initiation of security measures.

[0003] In an accident simulation in the field of passive vehicle safety, dummies equipped with sensors to record the loads generated during an accident are used to prevent injuries to human occupants.

[0004] The devices and methods according to the independent claims provide the possibility of calculating the load of an occupant of a vehicle.

[0005] A method for training an artificial neural network to predict the load of a vehicle occupant provides that a time series of the intrusion kinematics is provided during an accident, wherein a load detected by a sensor during the accident is provided to a dummy equipped with the sensor, wherein an artificial neural network is provided which is designed to map the time series of the intrusion kinematics to the load of the occupant during the accident, wherein the time series of the intrusion kinematics is mapped to the load of the occupant using the artificial neural network, wherein the artificial neural network is trained depending on a deviation of the load of the occupant from the load of the dummy. The time series of the intrusion kinematics include the intrusion acceleration and speed as well as the intrusion path. The artificial neural network is, for example,An artificial neural network, specifically a convolutional neural network (CNN), with fully connected layers, which is coupled with a long short-term memory (LSTM). The artificial neural network enables calculations with less computing time than finite element methods. The input to the artificial neural network consists of a time series, e.g., in ISOMME format. The input is thus independent of interconnection, which poses challenges in the context of algorithmic processing of finite element simulation data.

[0006] It can be provided that the load of the dummy comprises a plurality of parts that are assigned to a respective sensor with which the dummy is equipped, wherein the load of the occupant comprises a plurality of parts, each of which is assigned to one of the plurality of parts of the load of the dummy, wherein the artificial neural network is designed to map the time series of the intrusion kinematics to the parts of the load of the occupant, and wherein the artificial neural network is trained depending on the deviations of the parts of the load of the dummy and the load of the occupant assigned to the same sensor. This provides the possibility of providing the load in a differentiated manner according to different parts of a body of the dummy and thus of the occupant.

[0007] To calculate a temporal sequence, the load on the dummy can be provided as a time series of sensor signals recorded by the sensor, while the load on the occupant can be provided as a prediction for the time series of sensor signals recorded by the sensor. The output consists of a time series, e.g., in ISOMME format. A uniform input and output format, e.g., ISOMME, can be provided.

[0008] In order to enable a differentiation of the temporal sequence of the load, it can be provided that the parts of the load of the dummy each comprise a time series of sensor signals recorded with the respective sensor, wherein the parts of the load of the occupant each comprise a prediction for the time series from the part of the load of the dummy that is assigned to the respective part of the load of the occupant.

[0009] It can be provided that time series of the intrusion kinematics are provided at different locations of the vehicle, in particular a door and a vehicle pillar, wherein the artificial neural network is configured to map the time series of the intrusion kinematics of the different locations to the load on the occupant. This allows the different locations to be taken into account.

[0010] Preferably, a parameter set is provided that characterizes the vehicle or the dummy, or an arrangement of the dummy in the vehicle, wherein the artificial neural network is configured to map the time series of the intrusion kinematics and the parameter set to the occupant's load. This provides the possibility of a structural evaluation of different vehicle designs from the occupant's perspective.

[0011] For example, it is provided that a set of time series of intrusion kinematics is provided, comprising time series representing the intrusion kinematics in different accidents. A set of parameter sets is provided, each of which, for the set of time series, comprises a parameter set characterizing the vehicle or the dummy, or an arrangement of the dummy in the vehicle during the accident represented by the respective time series. Associated pairs of time series and parameter set are mapped to a respective load for the occupant using the artificial neural network, and the artificial neural network is trained based on the deviations determined for each pair. This further improves the structural assessment.

[0012] A method for predicting the load of an occupant of a vehicle provides that a time series of intrusion kinematics is provided in an accident, an artificial neural network is provided which is designed to map the time series of the intrusion kinematics to the load of the occupant during the accident, wherein the time series of the intrusion kinematics is mapped to the load of the occupant during the accident using the artificial neural network, or that several time series of different intrusion kinematics are provided in an accident, an artificial neural network is provided which is designed to map the time series of the intrusion kinematics to the load of the occupant during the accident, wherein the time series of the intrusion kinematics are mapped to the load of the occupant during the accident using the artificial neural network. This provides load values, e.g.even before a restraint system was designed for a frontal or side impact.

[0013] It can be provided that, in order to predict the load, a parameter set is provided which characterizes the vehicle or the occupant, or an arrangement of the occupant in the vehicle during the accident, wherein the artificial neural network is designed to map the parameter set and the time series of the intrusion kinematics to the load of the occupant, and wherein the parameter set and the time series of the intrusion kinematics are mapped to the load of the occupant, or wherein the artificial neural network is designed to map the parameter set and the time series of the intrusion kinematics to the load of the occupant, and wherein the parameter set and the time series of the intrusion kinematics are mapped to the load of the occupant. This provides a structural assessment of the vehicle designed according to the parameter set from the occupant's perspective.

[0014] For a detailed assessment of the load, e.g. on different parts of the body, the artificial neural network can be designed to map several time series, each of which is assigned to a part of the load, onto the part of the load assigned to the respective time series.

[0015] A device for training an artificial neural network for predicting a load of an occupant of a vehicle is designed to carry out the method for training the artificial neural network.

[0016] A device for predicting a load of an occupant of a vehicle is designed to carry out the method for predicting the load.

[0017] Further advantageous embodiments can be found in the following description and the drawing. The drawing shows: Fig. 1 is a schematic representation of a device for training an artificial neural network to predict a load on an occupant of a vehicle, Fig. 2 is a schematic representation of a device for predicting a load on an occupant of a vehicle, Fig. 3 a schematic representation of an artificial neural network for predicting the load of a vehicle occupant, Fig. 4 a schematic representation of an arrangement of sensors in the vehicle, Fig. 5 a schematic representation of an arrangement of a doll in the vehicle, Fig. 6 is a flowchart showing steps of a method for training an artificial neural network to predict a load on an occupant of a vehicle, Fig. 7 is a flowchart showing steps of a method for predicting a load on an occupant of a vehicle.

[0018] In Fig. 1 schematically illustrates a device 100 for training an artificial neural network 101. In the example, the device 100 for training the artificial neural network comprises the artificial neural network 101.

[0019] The artificial neural network 101 is trained based on a vehicle accident simulation in the field of passive vehicle safety. In this simulation, intrusion kinematics are measured at various locations on the vehicle using sensors or calculated based on sensor signals. A dummy located inside the vehicle is equipped with sensors to record the loads on the dummy during the accident. The sensors for recording the loads on the dummy are arranged, for example, on different parts of the dummy's body.

[0020] A first input variable of the artificial neural network 101 is a time series of intrusion kinematics 102. The time series is recorded, for example, with a sensor arranged in the vehicle during the accident or calculated based on a sensor signal from a sensor arranged in the vehicle. The first input variable can also comprise multiple time series of intrusion kinematics 102, recorded with multiple sensors arranged in the vehicle during the accident, or calculated based on a sensor signal from each of several sensors arranged in the vehicle.

[0021] An optional second input variable of the artificial neural network 101 is a parameter set 103. The parameter set 103 characterizes the vehicle or the doll, or an arrangement of the doll in the vehicle.

[0022] A reference value for training is a load on the dummy 104 recorded with a single sensor during the accident. The reference value can also include multiple portions of the load on the dummy 104 recorded with multiple sensors during the accident. For example, each sensor provides a portion of the load on the dummy 104.

[0023] The artificial neural network 101 is designed to map the time series of the intrusion kinematics 102 to a load of an occupant 105 during the accident. The artificial neural network 101 can be designed to map multiple time series of the intrusion kinematics 102 to the load of the occupant 105 during the accident. The artificial neural network 101 can be designed to map multiple time series of the intrusion kinematics 102 to parts of the load of the occupant 105 during the accident. The parts of the load of the occupant 105 are assigned, for example, to the parts of the load of the dummy 104. In the example, the parts of the load of the dummy 104 are assigned to a respective position of the sensor on the dummy's body. In the example, the parts of the load of the occupant 105 are assigned to a position on the occupant's body corresponding to the respective position of the sensor on the dummy.

[0024] The load on the dummy 104 and the load on the occupant 105 may be time series. The parts of the load on the dummy 104 and the parts of the load on the occupant 105 may be time series.

[0025] The device 100 for training the artificial neural network 101 is configured to train the artificial neural network 101 depending on a deviation 106 of the load of the occupant 105 from the load of the dummy 104. The device 100 for training the artificial neural network 101 can be configured to train the artificial neural network 101 depending on a respective deviation 106 of the mutually associated parts of the load of the occupant 105 and parts of the load of the dummy 104 from each other.

[0026] The device 100 for training the artificial neural network 101 is designed to carry out a method for training the artificial neural network 101 to predict the load of the occupant of the vehicle.

[0027] In Fig. 2 schematically illustrates a device 200 for predicting the load on the occupant 105 depending on a predefined time series of the intrusion kinematics 102 or a plurality of predefined time series of the intrusion kinematics 102. Optionally, the device 200 is configured to predict the load on the occupant 105 depending on a predefined parameter set 103 and a predefined time series of the intrusion kinematics 102. Optionally, the device 200 is configured to predict the load on the occupant 105 depending on a predefined parameter set 103 and a plurality of predefined time series of the intrusion kinematics 102. In the example, the device 200 for predicting the load comprises the artificial neural network 101.

[0028] The load prediction device 200 is configured to execute a load prediction method.

[0029] In Fig. 3 shows a schematic representation of the artificial neural network 101. In the example, the artificial neural network 101 comprises a deep neural network 107, in particular a convolutional neural network (CNN) with fully connected layers, dense layers, which is fed back to a long short-term memory (LSTM) 108.

[0030] The training is based on n simulations of an accident with the dummy, ie for each simulation on m time series s1, s2, s3,..., s m the parts of the load on the dummy 102 that were measured with m sensors in the respective simulation or calculated from the sensor signals of m sensors.

[0031] Optionally, the training is based on n parameter sets 103, each containing z parameters p1, p2, ..., p z The n parameter sets 103 are assigned to the respective time series that were measured or calculated during the simulation with the respective parameter set 103.

[0032] In the example, an output 109 of the deep neural network 107 per sensor comprises n predictions for the load of the occupant s' i .

[0033] In the example, the load of the occupant 105 comprises m time series s'1, s'2, s'3,..., s' m the parts of the load of the occupant 105.

[0034] The training is based on the m time series s'1, s'2, s'3,..., s' m the parts of the load of the occupant 105.

[0035] Fig. 4 shows a schematic representation of a door 401 and a vehicle pillar 402, e.g. the B-pillar, of an exemplary vehicle 403 and an arrangement of sensors 404 on the door 401 and the B-pillar 402. The sensors 404 are, for example, acceleration sensors that detect an acceleration differentiated according to three orthogonal directions x, y, z.

[0036] Fig. 5 shows a schematic representation of an arrangement of an exemplary doll 501 in the vehicle 403.

[0037] The parameter set 103 includes, for example, a particularly minimum Y-distance 502 of a head of the doll 501 from a window 503 of the vehicle 403.

[0038] The parameter set 103 includes, for example, a particularly minimum Y-distance 504 of a shoulder of the doll 501 from a side panel 505 of the vehicle 403.

[0039] The parameter set 103 includes, for example, a particularly minimum Y-distance 506 of a hip of the doll 501 from the side panel 505 of the vehicle 403.

[0040] The parameter set 103 may include a type of seat of the vehicle 403, e.g., sport or comfort.

[0041] The parameter set 103 may include a type of door panel of the vehicle 403, e.g. slush or leather.

[0042] The parameter set 103 may include a type of doll or a gender of the doll, e.g., male, female, child.

[0043] The parameter set 103 may include relevant parameters of a restraint system or an occupant protection system, in particular an airbag, preferably an ignition time of the airbag.

[0044] The parameter set 103 may include relevant parameters of the vehicle, e.g. a vehicle mass.

[0045] Fig. 6 shows a flowchart with steps of the method for training the artificial neural network 101 to predict the load of the occupant of the vehicle.

[0046] The method for training the artificial neural network 101 to predict the load of the occupant 105 of the vehicle includes a step 601.

[0047] In step 601, the artificial neural network 101 is provided.

[0048] The artificial neural network 101 is designed to map at least one time series of the intrusion kinematics 102 to the load of the occupant 105 during the accident.

[0049] The artificial neural network 101 can be designed to map the time series of the intrusion kinematics 102 to the parts of the load of the occupant 105

[0050] Subsequently, a step 602 is executed.

[0051] In step 602, the at least one time series of the intrusion kinematics 102 is provided in the event of an accident.

[0052] It can be provided that time series of the intrusion kinematics 102 are provided at different locations, in particular the door 401 and the vehicle pillar 402, of the vehicle.

[0053] It may be provided that the parameter set 103 is provided. It may be provided that the time series or time series for each simulation are provided together with the parameter set 103 used in the simulation.

[0054] For example, a set of intrusion kinematics time series 102 is provided, which includes time series representing the intrusion kinematics in different accidents. For example, a set of parameter sets 103 is provided, each of which includes a parameter set 103 for each time series of the set of time series, which characterizes the vehicle or the dummy, or an arrangement of the dummy in the vehicle, during the accident represented by the respective intrusion kinematics time series 102.

[0055] Subsequently, a step 603 is executed.

[0056] In step 603, the load of the doll 104 detected by at least one sensor during the accident is provided.

[0057] The load of the doll 104 may include the plurality of parts associated with a respective sensor with which the doll is equipped.

[0058] The load on the doll 104 may include a time series of sensor signals acquired by the sensor.

[0059] The parts of the load on the doll 104 can each comprise a time series of sensor signals acquired with the respective sensor.

[0060] A step 604 is then executed.

[0061] In step 604, the at least one time series of the intrusion kinematics 102 is mapped to the load of the occupant 105 using the artificial neural network 101.

[0062] The load of the occupant 105 may include the plurality of parts each associated with one of the plurality of parts of the load of the dummy 104.

[0063] The occupant load 105 may include a prediction for the time series of sensor signals acquired by the sensor.

[0064] The parts of the load of the occupant 105 may each comprise a prediction for the time series from the part of the load of the dummy 104 that is associated with the respective part of the load of the occupant 105.

[0065] It can be provided that the artificial neural network 101 maps the time series of the intrusion kinematics 102 of the different locations to the load of the occupant 105.

[0066] It can be provided that the time series of the intrusion kinematics 102 and the parameter set 103 are mapped to the load of the occupant 105 using the artificial neural network 101.

[0067] It can be provided that the artificial neural network 101 maps the time series of the intrusion kinematics 102 and the parameter set 103 to the load of the occupant 105.

[0068] It can be provided that mutually associated pairs of time series of the intrusion kinematics 102 and parameter set 103 are mapped to a respective load for the occupant 105 using the artificial neural network 101.

[0069] A step 605 is then executed.

[0070] In step 605, the artificial neural network 101 is trained depending on the deviation 106 of the load of the occupant 105 from the load of the dummy 104.

[0071] It can be provided that the artificial neural network 101 is trained depending on the deviations 106 of the parts of the load of the dummy 104 and the load of the occupant 105 assigned to the same sensor.

[0072] It can be provided that the artificial neural network 101 is trained depending on the deviations 106 determined for each pair of time series and parameter set.

[0073] It can be provided that step 604 is carried out repeatedly in order to map the time series of the intrusion kinematics or the intrusion kinematics 102 of the n simulations, optionally together with the respective parameter set 103, to the load of the occupant 105 or the parts of the loads of the occupant 105.

[0074] The training may provide for using a gradient descent method to determine weights of the artificial neural network 101 that minimize the deviations 106, depending on the deviations 106 that are determined for several of the time series.

[0075] Fig. Figure 7 shows a flowchart showing steps of the method for predicting the load of the vehicle occupant.

[0076] The method for predicting the load of the occupant 105 includes a step 701.

[0077] In step 701, the artificial neural network 101 is provided. For example, the artificial neural network 101 is provided with the method for training.

[0078] Subsequently, a step 702 is executed.

[0079] In step 702, a time series of intrusion kinematics 102 is provided during an accident.

[0080] It may be provided that several time series of different intrusion kinematics 102 are provided in the event of an accident.

[0081] It may be provided that a parameter set 103 is provided which characterizes the vehicle or the occupant, or an arrangement of the occupant in the vehicle at the time of the accident.

[0082] Subsequently, a step 703 is executed.

[0083] In step 703, the time series of the intrusion kinematics 102 is mapped to the load of the occupant 105 during the accident using the artificial neural network 101.

[0084] It can be provided that several time series of different intrusion kinematics 102 are mapped to the load of the occupant 105 during the accident using the artificial neural network 101.

[0085] It can be provided that the artificial neural network 101 maps several time series, each associated with a part of the load, to the part of the load associated with the respective time series.

[0086] It can be provided that the parameter set 103 and the time series of the intrusion kinematics 102 are mapped to the load of the occupant 105 using the artificial neural network 101.

[0087] It can be provided that the parameter set 103 and the time series of the intrusion kinematics 102 are mapped to the load of the occupant 105 using the artificial neural network 101. QUOTES CONTAINED IN THE DESCRIPTION

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

[0000] DE 10 2013 221 282 A1

[0002]

Claims

[1] Method for training an artificial neural network (101) for predicting a load of an occupant (105) of a vehicle, characterized by in that a time series of intrusion kinematics (102) is provided during an accident (602), wherein a load of a dummy (104) equipped with the sensor, which load is detected by a sensor during the accident, is provided (603), wherein an artificial neural network (101) is provided (601) which is designed to map the time series of the intrusion kinematics (102) to the load of the occupant (105) during the accident, wherein the time series of the intrusion kinematics (102) is mapped (604) to the load of the occupant (105) using the artificial neural network (101), wherein the artificial neural network (101) is trained (605) depending on a deviation (106) of the load of the occupant (105) from the load of the dummy (104). [2] Method according to claim 1, characterized byin that the load of the dummy (104) comprises a plurality of parts (603) which are assigned to a respective sensor with which the dummy is equipped, wherein the load of the occupant (105) comprises a plurality of parts (602), each of which is assigned to one of the plurality of parts of the load of the dummy (104), wherein the artificial neural network (101) is designed (601) to map the time series of the intrusion kinematics (102) onto the parts of the load of the occupant (105), and wherein the artificial neural network (101) is trained (605) depending on the deviations (106) of the parts of the load of the dummy (104) and the load of the occupant (105) assigned to the same sensor. [3] Method according to one of the preceding claims, characterized by that the load of the dummy (104) comprises a time series of sensor signals detected by the sensor (603), wherein the load of the occupant (105) comprises a prediction for the time series of sensor signals detected by the sensor. [4] Method according to one of the preceding claims, characterized by that the parts of the load of the doll (104) each comprise a time series of sensor signals detected with the respective sensor (603), wherein the parts of the load of the occupant (105) each comprise a prediction for the time series from the part of the load of the doll (104) which is assigned (602) to the respective part of the load of the occupant (105). [5] Method according to one of the preceding claims, characterized by that time series of the intrusion kinematics (102) are provided at different locations, in particular a door and a vehicle pillar, of the vehicle (602), wherein the artificial neural network (101) is designed to map (601) the time series of the intrusion kinematics (102) of the different locations to the load of the occupant (105). [6] Method according to one of the preceding claims, characterized bythat a parameter set (103) is provided (602) which characterizes the vehicle or the dummy, or an arrangement of the dummy in the vehicle, wherein the artificial neural network (101) is designed to map the time series of the intrusion kinematics (102) and the parameter set (103) to the load of the occupant (105). [7] Method according to claim 6, characterized bythat a set of time series of the intrusion kinematics (102) is provided (602), which comprises time series that represent the intrusion kinematics in different accidents, wherein a set of parameter sets (103) is provided (602), which for each time series of the set of time series, comprises a parameter set (103) that characterizes the vehicle or the dummy, or an arrangement of the dummy in the vehicle in the accident that the respective time series of the intrusion kinematics (102) represents, wherein mutually associated pairs of time series of the intrusion kinematics (102) and parameter set (103) are mapped (604) to a respective load for the occupant (105) using the artificial neural network (101), and the artificial neural network (101) is trained (605) depending on the deviations (106) determined for each pair. [8] Method for predicting a load of an occupant (105) of a vehicle, characterized bythat a time series of intrusion kinematics (102) is provided in the event of an accident (702), an artificial neural network (101) is provided (701), which is designed to map the time series of the intrusion kinematics (102) to the load of the occupant (105) during the accident, wherein the time series of the intrusion kinematics (102) is mapped with the artificial neural network (101) to the load of the occupant (105) during the accident (703), or that several time series of different intrusion kinematics (102) are provided in the event of an accident (702), an artificial neural network (101) is provided, which is designed to map the time series of the intrusion kinematics (102) to the load of the occupant (105) during the accident (701), wherein the time series of the intrusion kinematics (102) are mapped with the artificial neural network (101) to the load of the Occupants (105) are depicted at the accident (703). [9] Method according to claim 8, characterized by that a parameter set (103) is provided (702) which characterizes the vehicle or the occupant, or an arrangement of the occupant in the vehicle during the accident, wherein the artificial neural network (101) is designed to map the parameter set (103) and the time series of the intrusion kinematics (102) onto the load of the occupant (105) (701), and wherein the parameter set (103) and the time series of the intrusion kinematics (102) are mapped onto the load of the occupant (105) (703), or wherein the artificial neural network (101) is designed to map the parameter set (103) and the time series of the intrusion kinematics (102) onto the load of the occupant (105) (701), and wherein the parameter set (103) and the time series of the intrusion kinematics (102) are mapped onto the load of the occupant (105) (703). [10] Method according to claim 8 or 9, characterized bythat the artificial neural network (101) is designed (701) to map a plurality of time series, each associated with a part of the load, onto the part of the load associated with the respective time series. [11] Device (100) for training an artificial neural network (101) for predicting a load of an occupant of a vehicle, characterized by that the device (100) is designed to carry out the method according to one of claims 1 to 7. [12] Device (200) for predicting a load on an occupant of a vehicle, characterized by that the device (200) is designed to carry out the method according to one of claims 8 to 10.

Images