DE102020203370A1 - Method and device for detecting sensor blindness in a radar sensor of a vehicle - Google Patents

Abstract

The present invention relates to a method for detecting sensor blindness in a radar sensor (104) of a vehicle (100), observations (122) on the radar sensor (104) and on the vehicle (100) as input variables (124) of a Bayesian network ( 126) are processed, and state probabilities (128) of various states (130) of sensor blindness are read out as output variables (132) of the Bayesian network (126), the states (130) at least one distortion state of the radar sensor (104) and one absorption state of the radar sensor (104), the sensor blindness being detected if at least one of the state probabilities (128) is greater than a threshold value (402).

Description

Field of invention

The invention relates to a method and a device for detecting sensor blindness in a radar sensor of a vehicle.

State of the art

In a vehicle with a radar sensor, a perceptual performance of the radar sensor can decrease if, for example, snow or slush is deposited on a surface of the radar sensor. The decrease in perceptual performance can negatively affect the functions of the vehicle. In order to recognize the decrease, sensor data of the radar sensor can be evaluated and the radar sensor can be classified as blind if at least one parameter of the sensor data exceeds a blindness limit value.

the DE 10 2009 001 265 A1 describes, for example, a radar sensor with a blindness detection device.

Disclosure of the invention

Against this background, the approach presented here presents a method for detecting sensor blindness in a radar sensor of a vehicle, a corresponding device, and finally a corresponding computer program product and a machine-readable storage medium according to the independent claims. Advantageous developments and improvements of the approach presented here emerge from the description and are described in the dependent claims.

Advantages of the invention

Embodiments of the present invention can advantageously make it possible to correctly detect sensor blindness of a radar sensor with a high degree of probability.

A method is proposed for detecting sensor blindness in a radar sensor of a vehicle, observations on the radar sensor and on the vehicle being processed as input variables of a Bayesian network, and probabilities of the state of various states of the sensor blindness being read out as output variables of the Bayesian network, the states comprise at least one distortion state of the radar sensor and an absorption state of the radar sensor, the sensor blindness being recognized if at least one of the state probabilities is greater than an assigned threshold value.

Ideas for embodiments of the present invention can be viewed, inter alia, as being based on the thoughts and findings described below.

Sensor blindness can be understood to mean a state of a sensor in which the sensor does not record any or incorrect data. The sensor blindness can be caused, for example, by contamination of the sensor and / or a covering of the sensor. In the case of a radar sensor, the contamination can relate, for example, to an exit area and / or entry area for radar waves, that is to say to a transmission and / or reception path of the sensor. In the case of the radar sensor, absorption effects and / or distortion effects can occur. Absorption effects can occur if a medium in the transmission and / or reception path absorbs the radar waves, that is to say reduces a transmitted and / or receivable power. The range of the radar sensor can be reduced by absorption effects. Distortion effects can occur if a medium in the transmission and / or reception path bends or breaks at least part of the radar waves and thus changes the transmission and / or reception path. Distortion effects can falsely detect objects and show them in a different location than where they are actually located. In some cases, objects in the unchanged send and / or receive path cannot be detected or can only be detected incorrectly. Overall, the uncertainty with regard to the accuracy of the angle estimation increases due to distortion effects.

In addition, interference effects can occur, the interference effects essentially not being caused by the soiling and / or covering. Interference effects can, for example, be caused by other radar sensors and / or unfavorable reflections of the radar waves on objects be evoked. For example, so-called ghost echoes from non-existent objects can be caused by interference effects.

A Bayesian network can also be referred to as a Bayesian network or Bayesian net. The Bayesian network maps the dependencies of observations and states within a system. The observations and states are linked to one another via coefficients. A state can be linked to several observations. An observation can be linked to several states. The observations are usually the input variables of the Bayesian network. The observations are entered as numerical values. The states are usually the output variables of the Bayesian network. The output variables are output as probability values. The probability values represent a probability in percent with which the condition applies to the observations entered.

Further, at least one state from a group of states including yaw rate (yaw), tunnel (bad weather) and bad environment (bad environment) in the Bayesian network can be determined. The condition can be determined using at least one of the observations. The state can also represent a state probability. The state can also be determined as a discrete or binary value, i.e. yes or no or zero and one. The change between yes and no can take place, for example, at a predetermined threshold value.

Furthermore, a state probability of an interference state can be read out. Interference can have characteristics similar to distortion. Multiple observations can be very similar or even identical in terms of interference and distortion. In order to be able to distinguish the distortion from the interference, at least one observation that is only relevant for interference can be used in order to determine the probability of the state of the interference.

At least one of the output variables from previous processing in the Bayesian network can be used as an input variable for subsequent processing in the Bayesian network. The Bayesian network can be implemented using a time series of inputs. The current observations and the output variables determined in the past influence the new output variables proportionally.

At least two observations from a group of observations including steering, wipers, outside temperature, angle quality in the horizontal direction (thetaso quality), angle quality in the vertical direction (phi quality), plausibility check of the angle estimate (aei), variance the plausibility check of the angle estimation (aei variance), number of detections (detected targets), number of bins exceeding a limit value (nbins), estimated range (reach range), interference estimation (interference indicator), tunnel indicator and number of detected ones Overground targets can be processed in the Bayesian network. The observation control can be detected using a steering angle sensor of the vehicle, for example. The observation control can also be recorded by an inertial sensor system of the vehicle. The observation of the windshield wipers can be recorded using a control signal from the windshield wipers of the vehicle. The observation of the outside temperature can be recorded by a temperature sensor in the vehicle. The observations angular quality in the horizontal direction, angular quality in the vertical direction, plausibility check of the angle estimate, variance of the plausibility check of the angle estimate, number of detections, number of bins exceeding a limit value, estimated range, interference estimate, tunnel indicator and number of detected radar targets over ground can be from one Data stream of the radar sensor can be determined.

The observations steering, angular quality in the horizontal direction (thetaso quality), angular quality in the vertical direction (phi quality), plausibility check of the angle estimate (aei), variance of the plausibility check of the angle estimate (aei variance), number of detected targets, number of bins that exceed the limit value (nbins), estimated range (reach range), interference indicator, number of detected radar targets above ground (overground targets), windshield wipers (wiper), Outside temperature (temperature) and tunnel indicator (tunnel indicator) can be discretized to value ranges of at least two discrete values before processing. In particular, the observations steering, angular quality in the horizontal direction (thetaso quality), angular quality in the vertical direction (phi quality), plausibility check of the angle estimate (aei), variance of the plausibility check of the angle estimate (aei variance), number of detections (detected targets) , Number of bins exceeding the limit value (nbins), estimated range (reach range), interference indicator and number of detected radar targets above ground (overground targets) are discretized to a value range of ten possible discrete values each before processing. The observations windshield wiper (wiper), outside temperature (temperature) and tunnel indicator (tunnel indicator) can be discretized to value ranges between two and four possible discrete values before processing. By discretizing the observations, the state probabilities can be determined with a reduced computational effort. During the discretization, a large number of possible original values of the observation are reduced to a predetermined number of individual discrete values. For this purpose, the original values can be subdivided using limit values. The original values can also be rounded.

The Bayesian network can be trained using states and / or observations labeled with at least one variant of sensor blindness each. Data for labeling can be recorded by reference runs. The labels can be assigned manually, partially or automatically. During training, the coefficients of the Bayesian network are set. The process of training can be called machine learning.

The method can be implemented, for example, in software or hardware or in a mixed form of software and hardware, for example in a control device.

The approach presented here also creates a device which is designed to carry out, control or implement the steps of a variant of the method presented here in corresponding devices.

The device can be an electrical device with at least one computing unit for processing signals or data, at least one memory unit for storing signals or data, and at least one interface and / or one communication interface for reading in or outputting data that are embedded in a communication protocol, be. The computing unit can be, for example, a signal processor, a so-called system ASIC or a microcontroller for processing sensor signals and outputting data signals as a function of the sensor signals. The storage unit can be, for example, a flash memory, an EPROM or a magnetic storage unit. The interface can be designed as a sensor interface for reading in the sensor signals from a sensor and / or as an actuator interface for outputting the data signals and / or control signals to an actuator. The communication interface can be designed to read in or output the data wirelessly and / or wired. The interfaces can also be software modules that are present, for example, on a microcontroller alongside other software modules.

A computer program product or computer program with program code, which can be stored on a machine-readable carrier or storage medium such as a semiconductor memory, a hard disk or an optical memory, and for performing, implementing and / or controlling the steps of the method according to one of the embodiments described above is also advantageous is used, especially when the program product or program is executed on a computer or device.

It should be noted that some of the possible features and advantages of the invention are described herein with reference to different embodiments. A person skilled in the art recognizes that the features of the device and the method can be combined, adapted or exchanged in a suitable manner in order to arrive at further embodiments of the invention.

Figure list

• 1 zeigt eine Darstellung eines Fahrzeugs mit einer Vorrichtung gemäß einem Ausführungsbeispiel;
• 2 zeigt eine Darstellung eines Bayes-Netzes zum Erkennen von Sensorblindheit gemäß einem Ausführungsbeispiel;
• 3 zeigt eine Darstellung eines dynamischen Bayes-Netzes zum Erkennen von Sensorblindheit gemäß einem Ausführungsbeispiel; und
• 4 zeigt eine Darstellung von zeitlichen Verläufen von Beobachtungen und einer Zustandswahrscheinlichkeit während einer Fahrsituation.

Embodiments of the invention are described below with reference to the accompanying drawings, wherein neither the drawings nor the description are to be interpreted as restricting the invention.

• 1 shows a representation of a vehicle with a device according to an embodiment;
• 2 shows an illustration of a Bayesian network for detecting sensor blindness according to an embodiment;
• 3 shows an illustration of a dynamic Bayesian network for detecting sensor blindness according to an embodiment; and
• 4th shows a representation of temporal progressions of observations and a state probability during a driving situation.

The figures are only schematic and not true to scale. In the figures, the same reference symbols denote the same or equivalent features.

Embodiments of the invention

1 shows a representation of a vehicle 100 with a device 102 according to an embodiment. The vehicle 100 has a radar sensor 104 on. A perception performance of the radar sensor 104 can be affected by snow, ice and / or mud. Perceptual performance is impaired, for example, if a radoma 106 of the radar sensor 104 for example through a layer of ice 108 is covered or partially covered. The ice sheet 108 can lead to distortion effects and alternatively or additionally to absorption effects by using radar waves 110 flexes, breaks and / or dampens. Absorption effects occur particularly when the radome 106 extensively from the ice sheet 108 is covered. Distortion effects occur in particular when parts of the radome 106 free from the layer of ice 108 are, while other areas of the radome 106 from the ice sheet 108 are covered.

Using radar data 112 of the radar sensor 104 become driver assistance systems 114 of the vehicle 100 operated. When the detection quality of the radar sensor 104 the driver assistance systems can decrease 114 work incorrectly. Hence the device 102 designed for a sensor blindness of the radar sensor 104 to recognize. The device 102 has a processing device 116 , an execution facility 118 and an evaluation device 120 on. The processing facility 116 reads observations 122 on the radar sensor 104 and / or vehicle 100 and prepares them for input variables 124 for a Bayesian network 126 on. The Bayesian network 126 is in the execution facility 118 executed. The evaluation device 120 reads state probabilities 128 of states 130 sensor blindness as output variables 132 of the Bayesian network 126 off and sets a blindness signal 134 ready if at least one of the state probabilities 128 greater than one of those state probability 128 associated threshold is.

In one embodiment, the evaluation device reads 120 State probabilities 128 a state of distortion of the radar sensor 104 and an absorption state of the radar sensor 104 from the Bayesian network 126 the end.

2 shows a representation of a Bayesian network 126 for detecting sensor blindness according to an embodiment. The Bayesian network 126 essentially corresponds to the Bayesian network in 1 . The Bayesian network 126 here shows links between seven states 130 and 13th Observations 122 on. The conditions 130 and observations 122 can act as nodes of the Bayesian network 126 are designated. At least one of the observations 122 a cause for at least one of the conditions 130 be. Likewise, at least one of the observations 122 an effect of at least one of the states 130 depict. A condition 130 can also be a cause of another condition 130 be.

For example, the Bayesian network forms 126 a state 130 "bad weather" 200 as the cause of a condition 130 "Distortion" 202 and a state 130 "Absorption" 204 away. The state 130 "bad weather" 200 but is also the cause of the observations 122 "Windshield wiper" 206 and "Temperature" 208 pictured.

The state 130 "Absorption" 204 again is as the cause of the observations 122 "Number of Detections" 210 , "Number of bins that exceed a limit" 212 and "estimated range" 214 pictured.

The state 130 "Distortion" 202 is also as the cause of the observation 122 "Number of Detections" 210 pictured. In addition, the condition is 130 "Distortion" 202 as the cause of the observations 122 "Plausibility check of the angle estimate" 216 , "Variance of the plausibility check of the angle estimate" 218 , "Angular quality in the vertical direction" 220 and angular quality in the horizontal direction ‟ 222 pictured.

The observations 122 "Plausibility check of the angle estimate" 216 , "Variance of the plausibility check of the angle estimate" 218 , "Angular quality in the vertical direction" 220 and angular quality in the horizontal direction ‟ 222 however, are also as an effect of a state 130 "Yaw rate" 224 mapped, which in turn as the effect of an observation 122 "Steering" 226 is shown.

In addition are the observations 122 "Plausibility check of the angle estimate" 216 , "Angular quality in the vertical direction" 220 and angular quality in the horizontal direction ‟ 222 as effects of states 130 "Interference" 228 and "bad environment" 230 pictured.

Here is the state 130 "Interference" 228 as the cause of an observation 122 "Interference Estimation" 232 pictured while the state 130 "Bad environment" 230 as the cause of an observation 122 "Number of detected radar targets above ground" 234 and as an effect of the state 130 "Tunnel" 236 is shown. The state 130 "Tunnel" 236 is also a cause for an observation 122 "Tunnel indicator" 238 pictured.

The observations 122 can be observed on the vehicle and / or on the radar sensor. The links link the nodes using coefficients and result in state probabilities based on the values of the observations 128 of the states 130 . At least the state probabilities 128 of the states 130 "Distortion" 202 and "absorption" 204 are from the Bayesian network 126 read out.

3 shows a representation of a dynamic Bayesian network 126 for detecting sensor blindness according to an embodiment. The Bayesian network 126 essentially corresponds to the Bayesian network in 2 . In addition to this, the observations made at a previous point in time are shown here 300 resulting state probabilities 128 the states "distortion" 202 , "Absorption" 204 and "bad environment" 230 as input variables for a current point in time 302 used. In this way, the calculated state probabilities can be smoothed 128 can be achieved as the state probabilities 128 at the current time 302 taking into account the state probabilities 128 at the previous point in time 300 be calculated.

4th shows a representation of the time courses of observations 122 and a state probability 128 of a state 130 during a driving situation. The observations 122 and the condition 130 represent values of a small section of the Bayesian network, as shown in 3 is shown. Here is a history of the number of bins that exceeded the limit 212 , a history of the estimated range 214 and a history of the number of detections 210 pictured. The observations 122 are on a radar sensor at the same time in an environment with few objects 400 has been recorded. In time correlation with this is a course of the state probability determined in the Bayesian network 128 of absorption 204 shown.

The course of the number of bins that exceed the limit value 212 drops rapidly here several times. When the number of bins exceeds the limit 212 drops sharply, a fault in the radar sensor is detected in a conventional view. The reason for the rapid decline, however, is the lack of objects in the environment 400 . The approach presented here compensates for this observation 122 especially through observation 122 the estimated range 214 . This range 214 would also drop and / or become very small in the event of an error. In fact, however, is the estimated range 214 very high throughout. This prevents the state probability in the Bayesian network 128 of the state 130 of absorption 204 increases sharply. The state probability 128 always remains smaller than a limit value 402 of 0.8 and the state 130 absorption blindness is not recognized. If the state probability 128 greater than threshold 402 is, becomes the state 130 recognized by absorption blindness.

In one embodiment, the state probability is 128 of absorption 204 only after a waiting period has elapsed 404 determined. While waiting 404 become the values of the observations 122 recorded in order to achieve a smoothing of the strongly fluctuating values. Here is the waiting time 30th Seconds. When determining a single value of the probability of state 128 is always a time series of the recorded values of the observations 122 viewed from the past 30 seconds.

In other words, radar blindness detection via Bayesian networks is presented.

The perception performance of a radar sensor can be disrupted by weather influences such as snow and rain as well as by unfavorable environmental conditions such as interference and multipath propagation. The detection of these disturbances is essential for a degradation of functions. The perception performance can be predicted for driver assistance systems and autonomous driving in order to be able to degrade functions of the driver assistance systems. The driver can also be warned immediately, depending on the degradation and SAE level, or an escalation strategy can be selected.

In the approach presented here, the perception performance is predicted via Bayesian networks or machine learning.

For the validation of radar blindness, measurements with weather effects such as snow, rain and ice are run in and evaluated. This means that sequences or time series are known in which degradation should or should not occur. Time sequences with and without disruptive influences are labeled for this.

With the labeled data, a Bayesian network specially defined for radar blindness is taught-in automatically. Since labeled data is available, machine learning methods for time series can also be used to teach in blindness detection. The approach proposed here is based on a Bayesian network. This can detect the different types of blindness via different states and output their probabilities. New indicators can easily be added. The network also benefits directly from indicators that can recognize a scene in which conventional blindness detection has failed to detect blindness. The indicator is not in itself an indicator of blindness, but prevents incorrect identification via the network structure.

The Bayesian network presented here is directly influenced by indicators that can observe a blindness situation or incorrect identification (“explaining away”). It is easy to expand and improve with new indicators.

In the approach presented here, the generalization error is kept small by means of cross-validation. The Bayesian network enables an “optimal” or best possible classification of labeled blindness data. Automatic learning on the cluster or data center is possible using the labeled blindness data. The learned network can also be evaluated automatically on the cluster. (Compliance with KPIs)

All in all, the approach presented here enables a cost reduction through a reduced expenditure of time and fewer post-simulations on the cluster. Exact labeling of the data is necessary, but the additional effort is negligible compared to the time saved.

The Bayesian network depicts various causes or disturbances as states. In this way, probabilities with regard to the current state can be calculated. These can in turn be taken into account in the sensor data fusion and degradation. For example, the probability density functions (e.g. normal distribution) of estimated object positions can be expanded depending on the state probabilities of the Bayesian network (higher variance). This means that the radar sensor can be weighted differently within a merger.

An automated cross-validation is possible. This reduces the generalization error and improves the quality of the blindness detection.

In principle, the perception performance of a radar can be limited due to weather conditions through distortion and / or absorption. Distortion blindness is often recognized first while a layer of snow is building up. Above a certain thickness, absorption blindness is also recognized. In reality, both effects are blurred and often occur at the same time. In addition, regardless of the weather, interference can occur that increase the noise level and thus have an influence on detection (objects disappear in the noise).

In the case of absorption blindness, a layer of water or snow on the radome or behind a cover / emblem leads to a dampening and thus a restricted perception.

In the case of distortion blindness, partial masking leads to combined attenuation / refraction / diffraction effects and virtually distorts the antenna pattern. Angular errors arise.

Various indicators are defined as observations.
thetaSO quality denotes averaged angular quality in the horizontal direction. The angular quality is a calculated measure of the accuracy of the angle estimate.
phi quality denotes averaged angular quality in the vertical direction.
aei denotes a plausibility check of the angle estimate from the antenna diagram with the relative speed of a stationary target measured via the Doppler effect.
aei variance describes the variance of the aei.
detected targets denotes the number of detections.
nbins denotes the number of DFT bins that exceed the CFAR limit value ("number of detections")
reach range describes the estimated range.
steering refers to the steering angle. A direct signal from the yaw rate sensor can also be used.
Interference indicator refers to the interference estimate from the signal processing component.
tunnel indicator denotes an indicator for the detection of a tunnel and is based on detected radar targets of the "tunnel ceiling".
overground targets denotes the number of detected radar targets over the ground.
wiper describes the activation signal of the windshield wiper.
temperature refers to the outside temperature.

The Bayesian network is modeled in such a way that it calculates a probability with which the sensor is in the state of distortion blindness, absorption blindness or interference. The states depend on observations (indicators), themselves, other states and the past (time series).

All states and observations can be discretized. States can be mapped to the value range [1,2]. With the exception of Temperature [1,2,3], Tunnel indicator [1,2] and Wiper [1,2,3,4], observations can be mapped to the value range [1 to 10].

The common cause of distortion and absorption is Bad Weather. The control of the wiper and the temperature are available as direct observation for bad weather.

The indicators thetaso quality, phi quality, aei, aei variance and detected targets are defined as observations for distortion.

The indicators detected targets, nbins and reach range are defined as observations for absorption.

Since the observations thetaso quality, phi quality, aei and aei variance are also influenced by a high yaw rate (yaw), the yaw rate is also linked as a condition (cause) in order to avoid incorrect detection.

The observations thetaso quality, phi quality and aei also depend on the states of interference and bad environments.

Bad Environments, in turn, depends on the tunnel status (with observation tunnel indicator) and also has the observation of overground targets.

The time dependency arises from the linking of states with themselves at previous points in time. This results in a dynamic Bayesian network.

In words, the link means that it is likely that there is, for example, distortion blindness if it was recognized at the previous point in time.

The dynamic Bayesian network with a set of random variables Z = (Z ₁ , Z ₂ , ..., Z _N ) can be defined as a pair of Bayesian networks (BNs) (B ₁ , B →), where B _{1 is} the The initial distribution is P (Z ₁ ) at time 1, and B → is a 2TBN with the distribution P (Z _t | Z _t-1 ). A 2TBN is a Bayesian network that contains variables from Xt whose parents are variables from Xt-1 and / or Xt, and variables from Xt-1 without their parents. $$\mathrm{P.}\left(Z_t|Z_{t-1}\right)={\displaystyle \sum _{i=1}^N\mathrm{P.}\left(Z_t^i|\mathrm{P.}a\left(Z_t^i\right)\right)}$$

Let X _{t be} the unobservable state at time t (including distortion at time t D _t , absorption at time t A _t , bad weather at time t BW _t , bad environment at time t BE _t , tunnel state at time t TU _t , Interference at time t I _t and yaw angle at time t Y _t ) and E _t the observable evidence or indicators at time t (including steering indicator at time t steer _t , Thetaso quality at time t θ _t , Phi quality at time t φ _t , Angle Error at time t aei _t , Angle error variance at time t aevi _t , overground indicator at time t og _t , Tunnel indicator at time t tu _t , Interference indicator at time ti _t , Detected Targets indicator at time t dt _t , Nbins indicator at time t nbins _t , Wiper indicator at time tw _t , temperature indicator at time t temp _t , reach range indicator at time t rri _t ), the initial distribution of the Bayesian network for blindness detection is given by $$\begin{array}{l}\text{P.}\left(X_1|{\mathrm{E.}}_1\right)=\mathrm{P.}\left(\mathrm{D.}{1}_1|\mathrm{B.}{\mathrm{W.}}_1\right)\ast \mathrm{P.}\left({\mathrm{A.}}_1|\mathrm{B.}{\mathrm{W.}}_1\right)\ast \mathrm{P.}\left(\mathrm{B.}{\mathrm{E.}}_1|T{U}_1\right)\ast \mathrm{P.}\left(Y_1|stee{r}_1\right)\ast \mathrm{P.}\left(T{U}_1\right)\ast \\ \mathrm{P.}\left(\mathrm{B.}{\mathrm{W.}}_1\right)\ast \mathrm{P.}\left(w_1|\mathrm{B.}{\mathrm{W.}}_1\right)\ast \mathrm{P.}\left(tem{p}_1|\mathrm{B.}{\mathrm{W.}}_1\right)\ast \mathrm{P.}\left(rr{i}_1|{\mathrm{A.}}_1\right)\ast \mathrm{P.}\left(stee{r}_1\right)\ast \mathrm{P.}\left({\mathrm{\theta }}_1|{\mathrm{D.}}_1,\mathrm{B.}{\mathrm{E.}}_1,{\mathrm{I.}}_1,Y_1\right)\\ \ast \mathrm{P.}\left({\mathrm{\phi }}_1|{\mathrm{D.}}_1,\mathrm{B.}{\mathrm{E.}}_1,{\mathrm{I.}}_1,Y_1\right)\ast \mathrm{P.}\left({\text{aei}}_1|{\mathrm{D.}}_1,\mathrm{B.}{\mathrm{E.}}_1,{\mathrm{I.}}_1,Y_1\right)\ast \mathrm{P.}\left(aev{i}_1|{\mathrm{D.}}_1,Y_1\right)\ast \mathrm{P.}\left(d{t}_1|{\mathrm{D.}}_1,{\mathrm{A.}}_1\right)\ast \\ \mathrm{P.}\left(nbin{s}_1|{\mathrm{D.}}_1\right)\ast \mathrm{P.}\left(i_1|{\mathrm{I.}}_1\right)\ast \mathrm{P.}\left(t{u}_1|T{U}_1\right)\ast \mathrm{P.}\left(O{G}_1|\mathrm{B.}{\mathrm{E.}}_1\right)\end{array}$$

According to Markov's assumption, the conditional probability distribution of nodes at any point in time t depends only on the current nodes (point in time t) and previous nodes at point in time t-1. The 2TBN probability distribution results from: $$\begin{array}{l}\text{P.}\left(X_t,{\mathrm{E.}}_t|X_{t-1},{\mathrm{E.}}_{t-1}\right)=\mathrm{P.}\left(\mathrm{D.}{1}_t|\mathrm{D.}{1}_{t-1},\mathrm{B.}{\mathrm{W.}}_t\right)\ast \mathrm{P.}\left({\mathrm{A.}}_t|{\mathrm{A.}}_{t-1},\mathrm{B.}{\mathrm{W.}}_t\right)\ast \mathrm{P.}\left(\mathrm{B.}{\mathrm{E.}}_t|\mathrm{B.}{\mathrm{E.}}_{t-1},T{U}_1\right)\ast \\ \mathrm{P.}\left(Y_t|stee{r}_t\right)\ast \mathrm{P.}\left({\mathrm{I.}}_t\right)\ast \mathrm{P.}\left(T{U}_t\right)\ast \mathrm{P.}\left(\mathrm{B.}{\mathrm{W.}}_t\right)\ast \mathrm{P.}\left(w_t|\mathrm{B.}{\mathrm{W.}}_t\right)\ast \mathrm{P.}\left(tem{p}_t|\mathrm{B.}{\mathrm{W.}}_t\right)\ast \mathrm{P.}\left(rr{i}_t|{\mathrm{A.}}_t\right)\ast \\ \mathrm{P.}\left(stee{r}_t\right)\ast \mathrm{P.}\left({\mathrm{\theta }}_t|{\mathrm{D.}}_t,\mathrm{B.}{\mathrm{E.}}_t,{\mathrm{I.}}_t,Y_t\right)\ast \mathrm{P.}\left({\mathrm{\phi }}_t|{\mathrm{D.}}_t,\mathrm{B.}{\mathrm{E.}}_t,{\mathrm{I.}}_t,Y_t\right)\ast \mathrm{P.}\left({\text{aei}}_t|{\mathrm{D.}}_t,\mathrm{B.}{\mathrm{E.}}_t,{\mathrm{I.}}_t,Y_t\right)\ast \\ \mathrm{P.}\left(aev{i}_t|{\mathrm{D.}}_t,Y_t\right)\ast \mathrm{P.}\left(d{t}_t|{\mathrm{D.}}_t,{\mathrm{A.}}_t\right)\ast \mathrm{P.}\left(nbin{s}_t|{\mathrm{D.}}_t\right)\ast \mathrm{P.}\left(i_t|{\mathrm{I.}}_t\right)\ast \mathrm{P.}\left(t{u}_t|T{U}_t\right)\ast \mathrm{P.}\left(O{G}_t|\mathrm{B.}{\mathrm{E.}}_t\right)\end{array}$$

The inference (inferring from non-observable quantities) takes place with the help of the junction tree and forward / backward pass algorithms.

To teach-in the network, all states in the time series are first labeled, i.e. distortion, absorption, yaw, interference, bad environment, tunnel, bad weather are labeled.

After learning the network on the basis of a 5-fold cross-validation, conditional probability distributions can be displayed for aei. The example of the observation aei shows a connection between distortion and high conditional probabilities with large indicator values. At the same time, however, bad env (unfavorable environment) and interference also shift the distribution. After all, yaw (high yaw rate) has almost the same impact on conditional probability as distortion. This means that this one indicator alone would not be sufficient to differentiate between distortion blindness and high yaw rates. However, since high yaw rates are taken into account as observations in the defined Bayesian network, incorrect identifications through “explaining away” are avoided.

The range indicator rri shows a strong correlation with the labeled absorption states. The conditional probabilities for a long range (high indicator value) are high in the case of absorption blindness. Conversely, the conditional probabilities of a short range in the case of absorption blindness are just as great. The indicator is therefore in principle suitable for predicting absorption blindness.

The DBN uses additional indicators, whereby the use of rri (reach range indicator) in particular can avoid incorrect identification.

In 4th an environment with few objects is shown with the time course of the probability for the state of absorption.

Three indicators (observations) with values between 0 and 1 are shown on the left. The probability of absorption blindness is shown at the top right. In the depicted scene (desert), the two indicators “Nbins” and “Detected Targets” drop sharply, which leads to an increased calculated probability of absorption blindness. The high indicator values of rri reduce the probabilities of absorption blindness again and prevent incorrect identification.

The calculated probability P (absorption | evidence) with which the radar sensor is in the absorption blindness state is advantageous. These probabilities can be taken into account within a degradation system - especially for fusion systems. A binary shutdown of a function can also take place via a parameterizable shutdown threshold of 0.8, for example, e.g. ACC is deactivated above 0.8.

Finally, it should be pointed out that terms such as “having”, “comprising”, etc. do not exclude any other elements or steps and that terms such as “a” or “an” do not exclude a plurality. Reference signs in the claims are not to be regarded as a restriction.

QUOTES INCLUDED IN THE DESCRIPTION

This list of the 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.

Patent literature cited

• DE 102009001265 A1 [0003]

Claims (10)

Method for detecting sensor blindness in a radar sensor (104) of a vehicle (100), observations (122) on the radar sensor (104) and on the vehicle (100) being processed as input variables (124) of a Bayesian network (126), and as Output variables (132) of the Bayesian network (126) state probabilities (128) of different states (130) of the sensor blindness are read out, the states (130) including at least one distortion state of the radar sensor (104) and an absorption state of the radar sensor (104), wherein the sensor blindness is recognized if at least one of the state probabilities (128) is greater than a threshold value (402). Procedure according to Claim 1 further comprising determining at least one state (130) from a group of states (130) comprising yaw rate (224), tunnel (236), bad weather (200) and bad environment (230) in the Bayesian network (126) , the state (130) being used to secure the state probabilities (128). Method according to one of the preceding claims, in which a state probability (128) of an interference state is also read out. Method according to one of the preceding claims, in which at least one of the output variables (132) from previous processing in the Bayesian network (126) is used as the input variable (124) for subsequent processing in the Bayesian network (126). Method according to one of the preceding claims, in which at least two observations (122) from a group of observations (122) comprising control (226), windshield wipers (206), outside temperature (208), angular quality in the horizontal direction (222), angular quality in the vertical Direction (220), plausibility check of the angle estimate (216), variance of the plausibility check of the angle estimate (218), number of detections (210), number of bins that exceed a limit value (212), estimated range (214), interference estimate (232), tunnel Indicator (238) and number of detected radar targets above ground (234) are processed in the Bayesian network (126). Procedure according to Claim 5 , in which the observations (122) control (226), angular quality in the horizontal direction (222), angular quality in the vertical direction (220), plausibility check of the angle estimate (216), variance of the plausibility check of the angle estimate (218), number of detections (210) , Number of bins exceeding the limit value (212), estimated range (214), interference estimate (232), number of detected radar targets above ground (234), windshield wipers (206), outside temperature (208) and tunnel indicator (238) before the Processing to value ranges of at least two discrete values are discretized. Method according to one of the preceding claims, in which the Bayesian network (126) is trained using states (130) and / or observations (122) labeled with at least one variant of the sensor blindness each. Device (102), the device being designed to carry out, implement and / or control the method according to one of the preceding claims in corresponding devices. Computer program product which is set up to instruct a processor when executing the computer program product to implement the method according to one of the Claims 1 until 7th execute, implement and / or control. Machine-readable storage medium on which the computer program product according to Claim 9 is stored.

Images