DE102022205941A1 - Method and device for validating an operation of a vehicle and use of such a method or device - Google Patents

Abstract

The invention relates to a computer-implemented method for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, characterized by the steps: - providing recorded measurement data and / or simulation data from traffic scenarios or vehicle events, - determining anomalies by outliers in the recorded measurement data and / or simulation data are determined as anomalies as part of a statistical evaluation, - applying the extreme value theory to the anomalies and estimating a probability of occurrence of an even more extreme anomaly based on the applied extreme value theory, whereby a more extreme anomaly has a higher extreme value than the previously known anomaly, so that the more extreme one Anomaly corresponds to a previously unknown critical traffic scenario or critical vehicle event. Furthermore, the invention relates to a device and a use.

Description

The invention relates to a computer-implemented method and a device for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components. The invention further relates to the use of such a method.

Fully autonomous vehicles or semi-autonomous vehicles with a high level of autonomy have a number of clear advantages over vehicles with manual operation, such as a reduction in accidents due to human error, optimal driving style, etc.

However, the development costs of such vehicles are enormous. Such vehicles must also meet high safety standards.

The safety standards are set, for example, by SOTIF (Safety of the Intended Functionality), which contains guidelines for autonomous vehicles at levels 0, 1 and 2 (and higher).

All of these factors have a greater impact on the development phase of the autonomous system. The question of the safe operation of an AD prototype that is still in an early development phase in traffic is therefore a very complex task.

In order to make an autonomous vehicle suitable for series use, it is usually tested in real traffic or in simulation. This means that a large number of test kilometers are driven with real test vehicles or in simulation. This leads to a high level of testing effort in terms of time and costs. However, this also poses a number of problems for further development.

For the transition of a technical system from the development phase to series production, the release of this system is necessary. Release only takes place when the previously defined requirements, in particular the requirements for the security of the technical system, are met by this technical system. Release concepts that are currently used in the automotive industry achieve releases for different levels of automation, where controllability by a human driver is always ensured, be it through direct control of the driver or through a fallback level to the driver. A special feature of autonomous driving is the elimination of the fallback level to the human driver from a certain automation level, for example SAE Level 3+.

There are no applicable procedures in the state of the art for the release and calculation of the expected residual risk of a system, although such procedures are required in the SOTIF standard ISO/PAS 21448:2019 “Road vehicles -- Safety of the intended functionality”.

The DE 10 2020 207 909 A1 discloses a computer-implemented method for testing a vehicle system, with the steps: data from a digital road map are read in, zones are determined for the digital road map, depending on the specific zones, possible sequences of trips along a road on the digital road map are determined, in a simulation a behavior of the vehicle system is determined for at least one of the possible processes, depending on a comparison of the determined behavior with at least one predetermined requirement, it is determined whether the vehicle system has an error or a weakness.

It is therefore an object of the invention to provide an improved method and an improved device for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, in particular with regard to the required scope of protection.

This task is solved by a computer-implemented method with the features of claim 1 and a device with the features of claim 10 and a use with the features of claim 14.

The subclaims list further advantageous measures that can be suitably combined with one another to achieve further advantages.

• - Bereitstellen von erfassten Messdaten und/oder Simulationsdaten von Verkehrsszenarien oder Fahrzeugereignissen,
• - Bestimmung von Anomalien, indem Ausreißer in den bereitgestellten Messdaten und/oder Simulationsdaten als Anomalien im Rahmen einer statistischen Auswertung ermittelt werden,
• - Anwenden der Extremwerttheorie auf die Anomalien und Schätzen einer Auftrittswahrscheinlichkeit einer noch extremeren Anomalie anhand der angewendeten Extremwerttheorie, wobei eine extremere Anomalie einen höheren Extremwert als die bislang bekannte Anomalie aufweist, so dass die extremere Anomalie einem bislang unbekannten kritischen Verkehrsszenario oder kritischen Fahrzeugereignis entspricht.

The task is solved by a computer-implemented method for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, comprising the steps:

• - Providing recorded measurement data and/or simulation data from traffic scenarios or vehicle events,
• - Determination of anomalies by identifying outliers in the measurement data and/or simulation data provided as anomalies as part of a statistical evaluation,
• - Applying extreme value theory to the anomalies and estimating a probability of occurrence of an even more extreme anomaly based on the applied extreme value theory, whereby a more extreme anomaly has a higher extreme value than the previously known anomaly, so that the more extreme anomaly corresponds to a previously unknown critical traffic scenario or critical vehicle event.

A vehicle component can be understood to mean, for example, a driver assistance system or a vehicle function or the mode of operation and its evaluation, for example of sensors or the sensors themselves.

A scenario can be described as a describable traffic procedure that takes place in a geographically limited observation area, for example a road segment intersection. Critical scenarios are referred to as scenarios in which a given level of criticality exceeds a certain tolerance. A measure of criticality can be, for example, the distance to other vehicles, the time to impact with another vehicle or something similar.

In relation to traffic scenarios, anomalies include, for example, excessive speed, for example driving through an intersection at a previously unprecedented speed of 120. Other anomalies include, for example, previously unknown errors in the sensor evaluation system/sensor technology or the failure of a transmission component that previously worked reliably, etc.

According to the invention, it was recognized that, for example, in order to recognize how often a critical situation occurs in a specific scenario, the parameters of the scenario must be varied and examined using stochastic tests. However, this scenario-based testing limits the underlying parameters and also takes a very long time.

According to the invention, it was further recognized that previous validation of, for example, a driver assistance system/driving function was only based on a scenario-based validation. To do this, the scenarios must be clustered, although it cannot be predicted whether all scenarios that occur are covered.

According to the invention, it was recognized that by determining the probability of occurrence of an unknown anomaly based on the extreme value theory and therefore a previously unknown traffic scenario or critical vehicle event, it can be recognized, for example, when a scenario space is covered precisely enough for autonomous testing. The method according to the invention thus allows scenario-independent validation and testing by estimating the probability of supposedly assumed rare events, on the basis of which further decisions and analyzes can be made or even a release of the driver assistance system/driving function being tested, for example, can be obtained.

The method according to the invention can also be carried out independently of clustering.

The method according to the invention makes it possible to secure/estimate the scenario space coverage and the estimation of the probability of the occurrence of new, unknown anomalies, independently of assumptions, e.g. a confidence interval, and without restricting the underlying parameters.

In a further training, the maxima and/or the minima in the recorded measurement data and/or simulation data are determined as anomalies. This means that the maximum deflections or maximum outliers or the minimum deflections or minimum outliers are used for the extreme value theory and the extreme value theory is used to determine the probability of exceeding the maximum deflection/minimum deflection.

Furthermore, in a further training, parameters to be examined can be defined in advance and the maxima and/or minima in the recorded measurement data and/or simulation data in the parameter to be examined can be determined as anomalies. This allows the probability of anomalies to occur in relation to critical traffic scenarios or vehicle events to be recorded with regard to particularly dangerous parameters.

In a further training, in order to estimate the probability of occurrence of an even more extreme anomaly, a limit value distribution is generated using extreme value theory based on the detected anomalies, with the probability of occurrence of an even more extreme anomaly being estimated based on the generated limit value distribution. In particular, the extreme value theory can be designed as a generalized (generalized) extreme value theory and/or Pareto distribution, Gumbel distribution or Frechet distribution.

An anomaly is Gumbel distributed if: $$\wedge \left(\text{t}\right)=e\left(〚-\text{e}〛\left(-\text{t}\right)\right)\text{with t}\in ℝ$$

An anomaly is Frechet-distributed with parameter α >0 if: $$\begin{array}{l}\mathrm{\Phi \alpha }\left(\text{t}\right)=\text{e}\left(〚-\text{t}〛\left(-\mathrm{\alpha }\right)\right)\text{with t}>\text{0 and}\\ \mathrm{\Phi \alpha }\left(\text{t}\right)=0\text{with t}\le \text{0}\end{array}$$

An anomaly is Weibull distributed with parameter α > 0 if: $$\begin{array}{l}\mathrm{\Psi \alpha }\left(\text{t}\right)=\text{e}\left(〚-\left(-\text{t}\right)〛\left(-\mathrm{\alpha }\right)\right)\text{with t}\le \text{0}\\ \mathrm{\Psi \alpha }\left(\text{t}\right)=0\text{with t}>\text{0}\end{array}$$

This means that a Frechet distributed anomaly only takes positive values; a Weibull distributed anomaly only takes negative values and a Gumbel distributed anomaly can take both positive and negative values.

The generalized extreme value distribution combines all three distributions into one.

In further training, the outliers are extracted using the Peak Over Threshold method (POT) or using block maxima/block minima, depending on the anomalies to be determined.

In the block maxima/block minima method, the outlier analysis is based on the determination of block maxima (minima) in a given block length. Especially for time periods, such as recording an intersection over several months, these time periods can be divided into equidistant observation windows or blocks. These blocks are examined for anomalies, for example maximum deflections or maxima and/or minimum deflections or minima. For example, anomalies related to speed can be investigated. For example, the annual maxima (minima) can be extracted and an “Annual Maxima Series” (AMS) can be generated. Such a method can be used to accurately determine enough, but not too many, outliers. No threshold values need to be defined so that sufficient data can be used for evaluation.

Alternatively, depending on the anomalies to be determined, the outliers can be extracted using the Peak Over Threshold (POT) method. Peak Over Threshold extracts from a continuous recording the peak values that have been reached for a period in which values exceed or fall below a certain threshold.

In a further embodiment, anomalies in relation to a vehicle event and/or traffic scenario are considered to be the malfunction and/or failure of an internal vehicle component and/or vehicle software. In particular, the vehicle component can be a sensor or sensor system and/or its evaluation system and/or the vehicle software can be designed as AI (artificial intelligence) software. This allows the validation of failures of hardware such as cameras, radar, lidar sensors, or AI (artificial intelligence) software, which are increasingly being used for automated driving and which are necessary, to be accelerated by reducing the likely occurrence of anomalies can be estimated in the measurements of, for example, radar/lidar sensors, anomaly detection in AI (artificial intelligence) software.

In further training, speed is considered an anomaly in a traffic scenario. Excessive speeds are known to be a common cause of accidents.

Furthermore, the object is achieved by a device for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, based on provided measurement data and/or simulation data from traffic scenarios or vehicle events,
wherein the device has a computing unit, wherein the computing unit is designed to determine anomalies by determining outliers in the measurement data and / or simulation data provided as anomalies as part of a statistical evaluation,
and wherein the computing unit is further designed to apply the extreme value theory to the anomalies and to estimate a probability of occurrence of an even more extreme anomaly based on the applied extreme value theory, wherein a more extreme anomaly has a higher extreme value than the previously known anomaly, so that the more extreme anomaly is a previously unknown one corresponds to a critical traffic scenario or critical vehicle event.

In particular, the device is suitable for carrying out the method according to the invention. The advantages of the method can also be transferred to the device.

A computing unit can be a processor, for example.

In particular, the device can be designed to generate simulation data and/or have a receiving unit for receiving the measurement data. Furthermore, the outliers can be stored in a memory for processing, for example can be saved. This means that the device can be operated virtually independently without manual input.

In a further embodiment, the computing unit is designed to estimate the probability of occurrence of an even more extreme anomaly by generating a limit value distribution using the extreme value theory of the detected anomalies, wherein the computing unit is further designed to estimate the probability of occurrence of an even more extreme anomaly based on the generated limit value distribution .

In a further embodiment, the computing unit is designed to generate the limit value distribution using extreme value theory, where the extreme value theory can be a Pareto distribution, Gumbel or Frechet distribution or generalized extreme value theory.

Furthermore, the object is achieved by using the method as described above and/or the device as described above to estimate the probability of occurrence of previously unknown critical speeds in a traffic scenario and/or to estimate the probability of occurrence of a failure and/or a disruption of one or more Sensors in a sensor system and / or a failure and / or a malfunction in a vehicle integrated AI software (artificial intelligence) and / or for estimating a malfunction or a failure of one or more vehicle components of an at least partially autonomously operated vehicle and for using the probability of occurrence Validating an operating mode of at least one partially autonomously operated vehicle and/or its vehicle components.

This makes it possible, for example, to demonstrate that anomalies that violate the maxima/minima upwards or downwards occur so rarely that they can be neglected.

This can, for example, advance the validation of a driver assistance system/driving functions etc. in an autonomously operated vehicle.

Further properties and advantages of the present invention result from the following description with reference to the accompanying figures.

• 1: ein erfindungsgemäßes Verfahren schematisch,
• 2: ein Beispiel mit extrahierten Ausreißern (Maxima),
• 3: ein Beispiel mit extrahierten Ausreißern (Minima),
• 4: eine erste Anwendung des erfindungsgemäßen Verfahrens,
• 5: eine durchschnittliche Verteilungsfunktion der Geschwindigkeiten,
• 6: einen quantile-quantile Plot als Q-Q -Diagramm und als P-P-Diagramm einen Probability-Probability-Plot,
• 7: ein return value Diagramm,
• 8: schematisch die fünf niedrigsten Maxima bzgl. der Geschwindigkeit,
• 9: schematisch die fünf höchsten Maxima bzgl. der Geschwindigkeit,
• 10: ein Diagramm mit minimaler Geschwindigkeit als Anomalien,
• 11: schematisch eine erfindungsgemäße Vorrichtung.

The figures show schematically:

• 1 : a method according to the invention schematically,
• 2 : an example with extracted outliers (maxima),
• 3 : an example with extracted outliers (minima),
• 4 : a first application of the method according to the invention,
• 5 : an average distribution function of velocities,
• 6 : a quantile-quantile plot as a QQ diagram and a probability-probability plot as a PP diagram,
• 7 : a return value diagram,
• 8th : schematically the five lowest maxima in terms of speed,
• 9 : schematically the five highest maxima in terms of speed,
• 10 : a graph with minimum velocity as anomalies,
• 11 : schematically a device according to the invention.

1 shows schematically the method according to the invention for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components.

In a first step S1, real measured data and/or simulation data from traffic scenarios or vehicle events are provided. If the validation concerns, for example, a sensor system and/or its evaluation software, these measurement data can be generated by real sensors and/or traffic monitoring systems. For example, if a transmission element is affected, these measurement data can come from the relevant actuators and sensors.

In a second step S2, the anomalies are determined by identifying outliers in the actually recorded measurement data and/or simulation data as anomalies as part of a statistical evaluation. Parameters can be defined in advance, based on which the outliers should be determined. For example, in a traffic scenario, this could be the speed.

The outliers can be extracted depending on the anomalies to be determined using block maxima/block minima in a block length.

With this method, for example, daily measured data, ie the time series, can be divided into time blocks and the maxima/minima can be filtered out. The block duration can be So the time period varies depending on the type of outlier (speed, sensor error). Based on the quality of the result, for example using a Kolmogorov-Smirnov test, the block length can be varied until an optimal block length has been achieved. An optimal block length is a block length in which the block length is not too small, i.e. too much data remains after determining the maximum, and is not too large, i.e. not many large data points are filtered out, which are then ignored for a distribution assumption would.

2 shows an example of extracted outliers based on a block duration of measurements from the simulation and/or from real driving tests/real measurements in which the speed was recorded, for example at an intersection. These are examined for block maxima, ie outliers, here higher speeds than assumed or permitted, are extracted from this block duration. These outliers/anomalies are then examined in more detail.

3 shows an example of extracted outliers based on a block duration of measurements from the simulation and/or from real driving tests/real measurements in which the km (distance) up to a crash were recorded. The outliers here are the block minima. These outliers/anomalies are then examined in more detail.

Alternatively, the anomalies to be determined can be determined using the Peak Over Threshold (POT) method. With Peak Over Threshold, the peak values that exceed or fall below a certain threshold are extracted from a continuous recording.

The use of POT or the block maxima/block minima method can be selected depending on the outliers to be determined.

In a third step S3, the extreme value theory is applied to the anomalies (outliers). Depending on the anomalies, this can be the generalized extreme value theory and/or Pareto distribution and/or Gumbel distribution and/or Frechet distribution.

By applying the extreme value theory, a limit value distribution (probability distribution) of the anomalies, i.e. the maxima/minima of the occurrence of rare events, is generated.

In a fourth step S4, the probability of the occurrence of a more extreme anomaly is estimated, for example by the convergence behavior of the limit value distribution, with a more extreme anomaly having a higher extreme value than the previously known anomaly, so that the more extreme anomaly corresponds to a previously unknown critical traffic scenario or critical vehicle event .

In a fifth step S5, the probability of occurrence can be used to validate a driver assistance system/driving function in an autonomously operated vehicle.

Likewise, an operating mode of a vehicle component such as the sensor system, for example/radar/lidar sensor system, and its evaluation software, for example image processing software, can be validated. Driver assistance systems/driving functions can also be validated with AI (artificial intelligence) software.

The more measurement data/simulation data there are, the more accurately the probability of occurrence can be estimated. From this limit probability distribution, the mean or expected value can also be determined, as well as the variance in order to obtain further information.

The process does not depend on any clustering algorithm. The method according to the invention even allows for scenario-independent validation.

Using the method according to the invention, scenario-based testing can be accomplished in that the method estimates the probability of occurrence of new anomalies, on the basis of which further decisions and analyzes can be made.

4 shows a first application of the method according to the invention.

For this purpose, the measurement data is first provided. The measurement data here are traffic records from traffic detection. In this example, the "speed" parameter is to be examined for anomalies. The recorded measurement data reflects the measured speeds at various measuring stations in the years 2015 - 2020. This results in an average speed at a measuring station in these five years, which is in 5 will be shown.

5 shows the average distribution function of the speeds in the upper diagram as speed plotted against its frequency (probability density function). This shows that the average The speed that is driven at this point is around 70 km/h.

Furthermore, a block length is then determined. This is here 30 days. The block length can be set arbitrarily. The maximum is extracted in these blocks. 4 show the extracted maxima as points on the respective measured peaks.

Extreme value theory is then applied to the extracted maxima. This can, for example, be designed as a generalized extreme value theory and/or a Pareto distribution and/or a Gumbel distribution and/or a Frechet distribution, or a generalized extreme value distribution can thus be approximated.

6 shows a quantile-quantile diagram as a QQ diagram (left) and as a PP diagram a probability-probability plot (right) between observed and theoretical probability values. So-called “plotting positions” are shown in QQ and PP diagrams. These are sets of empirical coefficients that determine how probabilities are assigned to extreme values, which are then used to represent extreme values in the probability diagrams. In the two diagrams of the QQ and PP plots, observed extreme values (black points) are overlaid with the theoretical estimates (by fitting a distribution).

The Q-Q and P-P plots can be used to determine a return period. The return period indicates the period of time that corresponds to the probability that a certain value, for example here the average speed, will be exceeded at least once within 30 days.

This probability is called the exceedance probability and is related to the return period as 1/p, where p is the return period. In short, this number indicates the period of time, here 30 days, which corresponds to the probability that the associated maximum will be exceeded at least once within the observation period.

7 shows a return value plotted over the return period.

In 8th The five lowest maximums in relation to the speed of this road segment are shown. The probability that the smallest “maximum” of 89.6 km/h will be exceeded is quite high and is around 98.7%. However, the probability of the next maxima occurring decreases.

In 9 The five highest maxima in relation to the speed of this road segment are shown. The highest measured maximum of 139.1 km/h is exceeded with a probability of “only” approx. 1.33%, which is small but not insignificant. One statement that can be derived is that this anomaly cannot be neglected in autonomous driving. This means that when simulating this road segment to validate autonomous driving, it is not enough to just consider the range between 0 km/h and 140 km/h, as these numbers are not “small enough”.

Another statement is that longer observation periods are needed in order to be able to make even more precise statements.

The method according to the invention can be used to support the validation of a driver assistance system/driving function. For this purpose, for example, measured data from real driving tests are used and divided into a block length, and the maxima (minima) within the block length are determined. The probability of an anomaly occurring in relation to the maxima (minima) is then estimated using extreme value theory.

The more data and observations there are, the more precisely the probability of occurrence of anomalies or rare events can be estimated.

The method according to the invention can therefore be used to qualitatively evaluate the validation process for releasing autonomous driver assistance systems/driving functions and release can be achieved or rejected more quickly.

The same can be done with anomalies, which are designed as minimum values.

10 shows a diagram in which the anomalies are a minimum velocity.

Here, for example, the minimal outliers (minima) are important. To support the overall validation, the measured data from real driving tests are used, the minima are determined over the block length and, with the help of extreme value theory, the probability of occurrence of the anomalies or rare events is estimated.

The probability of the occurrence of an unknown, more extreme anomaly is then estimated.

Anomalies can be estimated, for example in relation to the functioning of sensors/radar/lidar sensors, or the functioning of AI solutions, the functioning of image processing, etc., which could and should definitely be taken into account to support validation.

The method according to the invention can now be used to estimate the probability of the next anomaly, i.e. exceeding the highest or lowest maximum (minimum).

The method according to the invention and the device according to the invention do not depend on any clustering algorithm. The method/device according to the invention therefore allows for scenario-independent validation.

11 shows a device 1 for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, based on provided measurement data and / or simulation data from traffic scenarios or vehicle events. The device 1 has a computing unit 2.

This can be designed as a processor.

The computing unit 2 is designed to determine anomalies by identifying outliers in the measurement data and/or simulation data as anomalies as part of a statistical evaluation.

Furthermore, the computing unit 2 is designed to apply the extreme value theory to the anomalies and to estimate a probability of occurrence of an even more extreme anomaly based on the applied extreme value theory.

A more extreme anomaly has a higher extreme value than the previously known anomaly, so that the more extreme anomaly corresponds to a previously unknown critical traffic scenario or critical vehicle event.

Reference symbol list

S1-S5

Procedural steps

1

contraption

2

Computing unit

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 assumes no liability for any errors or omissions.

Cited patent literature

• DE 102020207909 A1 [0009]

Claims (14)

Computer-implemented method for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, characterized by the steps: - Providing recorded measurement data and / or simulation data from traffic scenarios or vehicle events, - Determination of anomalies by identifying outliers in the recorded measurement data and / or simulation data as anomalies be determined as part of a statistical evaluation, - applying the extreme value theory to the anomalies and estimating a probability of occurrence of an even more extreme anomaly based on the applied extreme value theory, whereby a more extreme anomaly has a higher extreme value than the previously known anomaly, so that the more extreme anomaly is a previously unknown one corresponds to a critical traffic scenario or critical vehicle event. Computer-implemented method Claim 1 , characterized in that the maxima and/or minima in the recorded measurement data and/or simulation data are determined as anomalies. Computer-implemented method Claim 2 , characterized in that parameters to be examined are defined in advance and the maxima and/or minima in the recorded measurement data and/or simulation data in the parameter to be examined are determined as anomalies. Computer-implemented method according to one of the preceding claims, characterized in that to estimate the probability of occurrence of an even more extreme anomaly, a limit value distribution is generated using the extreme value theory based on the detected anomalies, and the probability of occurrence of an even more extreme anomaly is estimated based on the limit value distribution. Computer-implemented method according to one of the preceding claims, characterized in that the extreme value theory is designed as a generalized extreme value theory and/or Pareto distribution and/or Gumbel distribution and/or Frechet distribution. Computer-implemented method according to one of the preceding claims, characterized in that the outliers are extracted depending on the anomalies to be determined using the Peak Over Threshold method (POT) or using block maxima/block minima in a block length. Computer-implemented method according to one of the preceding claims, characterized in that the malfunction and/or failure of an internal vehicle component and/or vehicle software are considered as anomalies in relation to a vehicle event and/or traffic scenario. Computer-implemented method Claim 7 , characterized in that the vehicle component is a sensor or sensor system and/or its evaluation system and/or the vehicle software is AI (artificial intelligence) software. Computer-implemented method according to one of the preceding claims, characterized in that the speed is considered as an anomaly in a traffic scenario. Device (1) for validating an operating mode of at least partially autonomously operated vehicles or their vehicle components, based on provided measurement data and/or simulation data of traffic scenarios or vehicle events, characterized in that the device (1) has a computing unit (2), wherein the computing unit ( 2) is designed to determine anomalies by determining outliers in the measurement data and/or simulation data provided as anomalies in the context of a statistical evaluation, and wherein the computing unit (2) is also designed to apply the extreme value theory to the anomalies and to estimate a probability of occurrence of one more more extreme anomaly is formed based on the applied extreme value theory, wherein a more extreme anomaly has a higher extreme value than the previously known anomaly, so that the more extreme anomaly corresponds to a previously unknown critical traffic scenario or critical vehicle event. Device (1) according to Claim 10 , characterized in that the computing unit (2) is designed to estimate the probability of occurrence of an even more extreme anomaly by generating a limit value distribution using extreme value theory based on the detected anomalies, wherein the computing unit (2) is further designed to estimate the probability of occurrence of an even more extreme one to estimate more extreme anomalies based on the generated limit value distribution. Device (1) according to Claim 10 or 11 , characterized in that the extreme value Theory is designed as a generalized extreme value theory and/or Pareto distribution and/or Gumbel distribution and/or Frechet distribution. Device (1) according to one of the preceding Claims 10 until 12 , characterized in that the computing unit (2) is designed to extract the outliers depending on the anomalies to be determined using the Peak Over Threshold method (POT) or using block maxima/block minima in a block length. Use of the method according to one of the preceding Claims 1 until 9 and/or the device (1) according to one of the preceding Claims 10 until 13 to estimate the probability of occurrence of previously unknown critical speeds in a traffic scenario and/or to estimate the probability of occurrence of a failure of a fault in one or more sensors in a sensor system and/or a failure and/or a fault in a vehicle integrated AI software (artificial intelligence ) and/or to estimate a malfunction or failure of one or more vehicle components of an at least partially autonomously operated vehicle and to use the probability of occurrence to validate an operating mode of at least one partially autonomously operated vehicle and/or its vehicle components.

Images