Method for verifying the plausibility of a crash test for a motor vehicle

Abstract

The invention relates to a method for checking the plausibility of a crash test for a motor vehicle, comprising the following steps: a. Carrying out a crash test with a motor vehicle with previously known parameters relating to the vehicle structure of the motor vehicle, during which measured values ​​are recorded, namely: - by means of numerical simulation on a computer; and / or - by means of a physical test setup; b. using a pre-trained neural network, estimating measured values ​​based on the previously known parameters concerning the vehicle structure; and c. using a computer, overlapping the recorded measured values ​​from step a. with the estimated measured values ​​from step b. and outputting them in graphical form, making any deviation visible. The method proposed here enables the plausibility of a crash test result to be verified quickly and cost-effectively.

Description

The invention relates to a method for plausibility checking a crash test for a motor vehicle.The safety of, for example, vehicle occupants is to be checked by means of crash tests and crash simulations for the permissibility or insurance classification of a motor vehicle. The safeguarding of all requirements for the passive safety or the effectiveness of the developed safety measures is an essential component of the development process of new motor vehicles. Starting from static component tests via dynamic tests in the subsystem and overall system, the protection takes place with different types of tests with increasing degree of integration.For structural protection and for realistic protection of the restraint systems, crash tests are carried out with overall vehicles. For frontal impact examinations, a rollable overall vehicle (usually with the aid of a cable pull device) is accelerated to the predefined test speed and collides against a defined barrier. On the basis of the dummy measured values, the effectiveness of the developed self-protection measures can thus be checked and the fulfilment of the requirements with respect to the occupant load in the overall vehicle combination can be demonstrated.Due to the ever increasing number of new requirements, ever decreasing development cycles, as well as enormous cost pressure, there is an increasing virtualization of the product manufacturing process. Starting from an initial design state or design state, the virtual function protection takes place with the aid of numerical simulation. For virtual safeguarding of the requirements in the field of vehicle safety, finite element methods [FEM] have become established as an essential development tool. As a result of ever better prognosis capability, i.e. the most accurate possible prediction of the physical relationships of reality, more and more time-consuming and cost-intensive hardware attempts can be replaced by virtual protection methods.For example, in DE 10 2020 115 192 A1 and DE 10 2020 132 042 A1 (see also publication in Belaid, Rabus, Kreistel 2021: "CrashNet: an encoder-decoder architecture to predictive crash test exits", https: / / web.archive.org / web / 20220930013940 / https: / / link.springer.com / article / 10.100 7 / s10618-021-00761-9), a further securing method based on machine learning models was introduced. This securing method enables a prediction of the course of the breast acceleration of a dummy with high accuracy even before virtual simulations can be carried out or physical prototypes are available for crash tests. Configuration-relevant properties of the vehicle (for example properties of the vehicle structure, of the restraint system or of the experimental configuration) serve as input for such a machine learning model. Thus, even today important development knowledge can be generated significantly earlier than with conventional protection methods (experiment and simulation).EP 4 293 334 A1 also concerns a method for controlling a crash test device for testing a vehicle chassis / vehicle body on the basis of acceleration pulses. The method includes obtaining information about an adjusted pulse based on a pulse to be experienced by the chassis or body of the vehicle. The method further includes using a neural network to determine configuration parameters for the crash test device based on the adjusted impulse and providing the configuration parameters for the crash test device. Thus, embodiments provide a more efficient concept for controlling a crash test device.When analyzing results and crash tests and crash simulations, it is frequently the question whether the results determined (physically, numerically or by pre-trained neural network) are also plausible. Plausibility checking is enormous and is currently carried out on the basis of the individual experience of the responsible engineers. Despite increasing prediction quality in virtual simulations, the simulation models are still very error-prone today. Whether results of a simulation actually correctly represent the reality is often difficult to assess.Proceeding from this, the object of the present invention is to overcome at least partially the disadvantages known from the prior art. The features according to the invention are evident from the independent claims, for which advantageous embodiments are shown in the dependent claims. The features of the claims can be combined in any technically expedient manner, wherein the explanations from the following description and features from the figures, which comprise supplementary configurations of the invention, can also be used for this purpose.The invention relates to a method for plausibility checking a crash test for a motor vehicle, having the following steps: a. carrying out a crash test using a motor vehicle having previously known parameters relating to the vehicle structure of the motor vehicle, measured values being recorded during the crash test, namely:by means of numerical simulation on a computer; and / orby means of a physical test setup;b. by means of a pre-trained neural network, estimating measured values on the basis of the previously known parameters relating to the vehicle structure; and c. by means of a computer, overlapping the detected measured values from step a. with the estimated measured values from step b. and outputting in graphically prepared form, wherein any deviation present is made visible.Ordinal numbers used in the preceding and following description, unless expressly stated to the contrary, serve only for the clear distinguishability and do not represent an order or ranking of the designated components. An ordinal number greater than one does not require that a further such component must necessarily be present.Whereas generally little confidence is provided to the results of a machine learning algorithm or pre-trained neural network, because the decision-making process is usually not (reliably) comprehensible, a departure from this idea is proposed here. The weakness of the respective systems is compensated for with good certainty in a meaningful concatenation of the instruments. While a numerical simulation is therefore to be regarded as highly reliable on the basis of the mathematical algorithm, the numerical simulation is completely dependent on how it is fed from the human side, that is to say on the basis of which information the numerical simulation is carried out. The same applies to a hardware test setup in which incorrect or incomplete specifications may have been implemented or errors may generally occur during implementation.A pre-trained neural network, in contrast, has the strength that it relates its prior knowledge (the training data) from already plausibility checked tests. The error rate in this database is almost zero. Therefore, if one of the result of estimation by the pre-trained neural network occurs in a physical experiment or a numerical simulation (i.e., the crash test), then the probability that a fault is present is high. Whether this has now been made in the estimation or in the crash test, it is important to know a technical engineer and possibly to find out attempts at counteraction. However, in the case that the crash test and the estimation of the pre-trained neural network result in a good-agreement result, it can be highly likely that the results are plausible. The latter is the rule.A crash test is therefore to be carried out, wherein depending on the task or level of development, it is carried out as a numerical simulation (for example of a subsystem of the motor vehicle) and / or as physical tests (for example of the overall vehicle). It should be noted that this can be carried out before, simultaneously or after step b. This is only the step which is to be executed first, because its result is to be checked for plausibility.The numerical simulation is, for example, an FEM simulation, which is frequently preferred on account of the detailed insight into the deformation behavior and stress behavior of structures. In some cases, however, other methods are also used, such as, for example, the so-called multi-body simulation [MBS], the discrete element method [DEM] and the boundary element method [BEM]. The advantage of the MBS simulation is the possibility of simulating dynamic systems with many bodies, which is at the same time efficient in the calculation of the mechanical movement states. The DEM is well suited for simulating granular materials such as bulk material or for modeling breaking processes, but not ideal for simulating continuous solid bodies such as body panels. BEM solves structural mechanics problems in which the regions of interest are defined by surfaces, however, BEM is less efficient in volume elements and not suitable for complex fracture mechanics common in crash tests.The physical test setup is, for example, the classic crash test with an overall vehicle, as carried out, for example, for frontal impact examinations. In another application, subsystems are examined, such as the triggering behavior of a pyrotechnic system (for example, an airbag or restraint system in a seat belt system). This also includes abstract experiments, such as, for example, investigations made from material science, the present method being advantageous above all for a complex requirement situation.Furthermore, an estimation is to be carried out by means of a pre-trained neural network, specifically for plausibility checking the result, i.e. the captured measured values from step a., with the aid of the result of step b., i.e. the estimated measured values from this step b.. The pre-trained neural network is pre-trained, for example as in the previously mentioned publications. In this case, plausibility-checked or verified test results, and also their parameters of (simulated and / or physically carried out) crash tests, are used as the basis for the training.A suitable pre-trained neural network is based, for example, on one of the following currently available products: ChatGPT® from OpenAI Incorporation, PaLM® from Alphabet Incorporation, GPT-JT from CloudPlat together.ai, or LLaMA® from Meta Plat Incorporation. It should be noted that the algorithm can be used in a self-learning and / or self-amplifying manner in order to further improve future estimates of measured values, if appropriate. Results are used for continued training or the estimation is refined by suitably implemented weighting of deviations from an expected result.Using the measured values from step a. and step b., a graphically prepared output is now generated, which enables a human (or computer) to ascertain deviations between the results. In one embodiment, a diagram is created via acceleration values (i.e. the ordinate with the acceleration) over a time profile (i.e. the abscissa with time). For example, the acquired measurement values are displayed in a first color and the estimated measurement values are displayed in a second color. In one embodiment, a previously known corridor is additionally predefined as an envelope, so that a rough plausibility check is already made possible by it, for example if at least one of the measured values lies outside such a corridor. For a technical engineer, it is easy to judge on the basis of two curves whether these sufficiently coincide with one another or whether an error exists or could exist here because a deviation is too strong according to his technical assessment.It is furthermore proposed in an advantageous embodiment of the method that in step b. stepped standard deviations from the measured values are estimated and in step c. are reproduced graphically prepared in the overlap.By means of stepped standard deviations, it is possible to simplify the estimateability of the deviation. For example, a first standard deviation is designed such that it is a normal, i.e. plausible, deviation. For example, in the case of such a slight deviation, an assessment by a technical engineer becomes redundant, that is to say is required in an unsolicited or involuntary manner or merely in a random sample manner. A second standard deviation is, for example, one which necessarily requires a separate expert check. A third standard deviation is, for example, an error and at least the experiment according to step a. or according to step b. must be adapted and repeated. With these standard deviations, it is possible to make the plausibility check more efficient and, if appropriate, to automate it further, provided that the deviations are sufficiently small.It should be noted that the standard deviations are percentiles. The standard deviations are preferably produced separately or as independent task with an independent result by means of the pre-trained neural network. For example, the pre-trained neural network is given the information that specific sensors can deliver specific measurement errors, such as an acceleration sensor used in a test setup having its own inertial mass such that a delayed reaction and / or a (self) oscillation occurs, which can superimpose the measurement. This could therefore correspond to a first standard deviation. In contrast, a measurement of an acceleration sensor, which has become detached from its intended measurement location for example because of a technical defect, would output results within or even beyond the third standard deviation.Within a second standard deviation, for example, a transverse influence which is not taken into account in the simulation compared to the learned (therefore ascertained in the estimation) or physical (therefore detected in the crash test) could be, for example, on the basis of a transverse acceleration in the event of a frontal impact as a result of a yielding of a belt and / or a vibration in the belt fastening location (for example the B-pillar). Such non-consideration is intentionally carried out, for example, in order to make the simulation more cost-effective and / or faster. If the deviation during plausibility checking is still sufficiently small, and this could be the reason according to expert assessment, the result would be plausible despite the deviation.It is furthermore proposed in an advantageous embodiment of the method that in a step d. a possibly present deviation between the measured values is analyzed with computer assistance, wherein known possible causes are determined and output, wherein the analysis is preferably carried out by means of a pre-trained neural network.Here, a further step d. is proposed following step c., wherein a possible cause of a deviation between the captured measured values and the estimated measured values is determined, specifically with computer assistance.In one embodiment, a similar constellation (for example, a comparable test arrangement for another motor vehicle) is found by comparing previous tests, wherein there was preferably a similar deviation. This can be easily found with computer assistance, for example, by means of a corresponding key interpretation of trial data stored (for example, in an ERP system [Enterprise Resource Planning]). If comparable errors were present in the test data found, these are listed here as possible causes from the accompanying documentation. The experience from previous experiments is thus well-prepared and not just the experience of the technical engineer or the reminder thereof is necessary.In one embodiment, such effects (for example intrinsic, as so-called hidden information) are already known during the training (training) of the neural network for carrying out the estimation of measured values in step b, and the pre-trained neural network can immediately provide possible reasons for this with the generated standard deviations or, in a further step, at least one possible cause for the specific deviation can be determined with the aid of a comparing pre-trained neural network. Alternatively, a separately pre-trained neural network is used for estimating a cause of the fault.It is furthermore proposed in an advantageous embodiment of the method that in a step e. on the basis of a found cause for a deviation:the pre-trained neural network is further trained;the crash test, preferably the associated numerical simulation, is adapted and repeated accordingly; and / ora crash test carried out in the future is informed and / or considered automatically to this cause in a computer-assisted manner,wherein a measure for remedying the found cause is preferably found with the aid of a pre-trained neural network.In an advantageous embodiment, each experiment or at least the experiments running satisfactorily (but not, or not precisely, the error-free) is used for a further improvement of the method proposed here.As already mentioned above, it is possible and particularly expedient in this respect to train the pre-trained neural network further, i.e. to enlarge or improve the database, starting from which the pre-trained neural network operates its estimate. In a preferred embodiment, a weighting is used in order to achieve a learning effect and to avoid a pure replacement of available knowledge being achieved. In one embodiment, this is an automated process, namely within the scope of self-learning and / or self-reinforcement.In one specific embodiment, the crash test is highly likely to be defective. It is precisely in the case of a numerical simulation that it is possible to repeat with reasonable time and cost, a change then being carried out which corresponds to the ascertained cause. For example, a determined transverse influence is taken into account as a (possible) cause. It should be noted that in some applications, repetition is carried out within the scope of a larger test arrangement, for example first a physical test with a subsystem and then the (adapted) subsystem integrated into a numerically simulated overall system.In one embodiment, an indication is generated that will be considered in the future. In one embodiment, this indication is explicitly output in human-readable form, for example as a type of check list. In another embodiment, such a pointer is observed automatically, i.e. added automatically, for example in a numerical simulation, for example a FEM network is created fine-grained at a point detected as critical. In a physical test setup, for example, a specification for fastening a sensor is automatically issued, which ensures a sufficiently stable fastening.In one embodiment, measures (for example intrinsic, as so-called hidden information) are already known during training of the pre-trained neural network for carrying out the estimation of measured values in step b, and the pre-trained neural network can immediately provide possible measures with the generated standard deviations or, in a further step, at least one possible measure for excluding an influence of the (suspected) cause for the specific deviation is proposed with the aid of a comparative pre-trained neural network. Alternatively, a separately pre-trained neural network is used for ascertaining a measure.The invention described above is explained in detail below in view of the relevant technical background with reference to the associated drawings, which show preferred embodiments. The invention is in no way limited by the purely schematic drawings, it being noted that the drawings are not dimensionally accurate and are not suitable for defining size ratios. It is shown in FIG. 1 : Diagram of plausible breast acceleration of a crash test dummy during a frontal impact; FIG. 2 : Diagram of an implausible breast acceleration of a crash test dummy in the event of a frontal impact; FIG. 3 : Diagram of an implausible breast acceleration of a crash test dummy in the event of a frontal impact with standard deviations; and FIG. 4 : shows a flow diagram of a method for plausibility checking a crash test for a motor vehicle.FIG. 1 shows a diagram of a (plausible) breast acceleration of a crash test dummy during a frontal impact, with the time axis 7 as the abscissa and the acceleration axis 8 as the ordinate (decisively as negative acceleration, i.e. deceleration). The course of the vehicle acceleration 9 is shown in dashed lines. In a slight delay, this is followed by the chest of the crash test dummy, namely in a dash-dotted line of the crash test, i.e. on the basis of the recorded measured values 1, and in a thick line of the estimation by the pre-trained neural network, i.e. on the basis of the estimated measured values 2. the deviations 3 are small here or in a less critical range. The background here is that a belt restraint system is activated pyrotechnically, so that the deceleration of the crash test dummy only occurs with a delay with the activation, and a further delay occurs due to a yielding of the belt and the mechanics connected in series. However, a maximum of the load for the crash test dummy occurs approximately at the same (measurement) value and over the same duration. Overall, a very high quality of the agreement of these two curves of the measured values 1, 2 is given. The crash test is thus plausibilized, namely on the basis of the graphical processing (at least for a human being), quickly and easily comprehensible.As in FIG. 1, FIG. 2 shows a diagram of a (implausible) breast acceleration of a crash test dummy during a frontal impact. For understanding the representation, reference is made to the preceding description relating to FIG. 1. In contrast to the profile in FIG. 1, the crash test and the estimation by pre-trained neural network here have a deviation 3 of the recorded measured values 1 from the estimated measured values 2 (additionally in a critical range), namely that of the maximum acceleration. Here, for example, the reason could be that an acceleration sensor has become detached from the crash test dummy or, in a numerical simulation, a strength of the holding mechanics of a belt has been assumed to be too soft or, conversely, a strength has been estimated to be too positive in the estimation by the pre-trained neural network.FIG. 3 shows a diagram, as in FIGS. 1 and 2, of implausible chest acceleration of a crash test dummy during a frontal impact with standard deviations. Here, the course as shown in FIG. 2 is purely exemplary and reference is made to the preceding description in this respect. As additional information, which facilitates the reading out or evaluation of the deviation 3 (compare FIG. 2 ) of the measured values 1, 2 from one another at least for a human being, standard deviations 4, 5, 6 are shown here in three stages. The profile of the recorded measured values 1 projects into the last (third) standard deviation 6 and is thus, for example, a signal that an error is present. Alternatively or additionally, regions are identified which make critical regions of less critical regions of the entire course quickly and easily distinguishable.FIG. 4 shows a flow chart of a method for plausibility checking a crash test for a motor vehicle. In step a., a crash test is carried out, wherein measured values 1 are acquired here physically or by means of numerical simulation. In step b., measured values 2 are estimated by means of a pre-trained neural network. In one embodiment, step a. and step b. are performed in a different time order or (approximately) simultaneously. It should be noted that the operations in step a. and step b. are completely independent of each other in order to increase redundancy security. Alternatively, the necessary parameters are only input once for both steps via a common input mask (for example in an ERP system), which provides a high level of certainty that the same data base relating to the test setup is assumed.In step c., the measured values 1, 2 determined are compared with one another, for example as shown in FIGS. 1 to 3. Alternatively or additionally, in computer-readable form, in order to use this comparison for self-learning and / or self-amplification, for example in a pre-trained neural network.Purely optionally, a step d. is provided, in which any deviation 3 between the measured values that may be present is analyzed with computer assistance, wherein known possible causes for the deviation 3 are determined and output. This is carried out, for example, by the same or a further (specialized) pre-trained neural network. It should be noted that the result can relate both to the crash test and to the estimation, i.e. can detect located errors there in each case and enable the correction thereof.Purely optionally, a step e. is also provided, in which a measure is found, by means of which a fault can be corrected. As soon as this measure is found, step a. and / or step b. is repeated, if appropriate, with step c. following. If a deviation 3 again occurs, step d., and optionally then step e., are also carried out, if appropriate. In one embodiment, this is repeated in the manner of a loop until a sufficient agreement of the measured values 1, 2 is achieved.The method proposed here makes it possible for a result of a crash test to be checked for plausibility quickly and cost-effectively.

Claims

Method for plausibility checking a crash test for a motor vehicle, having the following steps: a. carrying out a crash test with a motor vehicle having previously known parameters relating to the vehicle structure of the motor vehicle, wherein measured values (1) are detected during the crash test, namely: - by means of numerical simulation on a computer; and / or - by means of a physical test setup; b. by means of a pre-trained neural network, estimating measured values (2) on the basis of the previously known parameters relating to the vehicle structure; and c. by means of a computer, overlapping the detected measured values from step a. with the estimated measured values from step b. and outputting in graphically prepared form, wherein a deviation (3) which may be present is made visible.Method according to Claim 1, wherein stepped standard deviations (4, 5, 6) from the measurement values are estimated in step b. and are reproduced graphically prepared in the overlap in step c.Method according to Claim 1 or Claim 2, wherein in a step d. a deviation (3) which may be present between the measured values (1, 2) is analyzed with computer assistance, wherein known possible causes are determined and output, wherein the analysis is preferably carried out by means of a pre-trained neural network.Method according to one of the preceding claims, wherein in a step e. on the basis of a found cause for a deviation (3): - the pre-trained neural network is further trained; - the crash test, preferably the associated numerical simulation, is correspondingly adapted and repeated; and / or - in the case of a crash test carried out in the future, this cause is pointed out and / or observed in an automated manner with computer assistance, wherein a measure for remedying the found cause is preferably found with the aid of a pre-trained neural network.

Images