EP3857822A1

EP3857822A1 - Method and device for determining a control signal

Info

Publication number: EP3857822A1
Application number: EP19755586.5A
Authority: EP
Inventors: Michael Klar; Zhongyu LOU; Duc Tam Nguyen
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2018-09-27
Filing date: 2019-08-14
Publication date: 2021-08-04
Also published as: DE102018218834A1; US20210294278A1; WO2020064211A1; CN112740625A; CN112740625B

Abstract

A computer-implemented method for classifying an input signal (x), which is determined in particular according to an output signal (S) of a sensor (30), in respect of whether or not the input signal has an anomaly. According to an output signal (y) of an autoencoder (6) to which the input signal (x) is fed, it is decided whether the input signal (x) has or does not have the anomaly, the autoencoder (60) comprising at least one encoder (61) and at least one decoder (631; 632, ...), wherein an intermediate variable (z) is determined by means of the encoder (61) according to the input signal (x), and the output signal (y) is determined on this basis by means of the decoder (631; 632,...). The autoencoder (60) provides a plurality of hypotheses of formula (I) for the input signal (x) fed to it in order to reconstruct the input signal and determines the output signal (y) according to said plurality of hypotheses of formula (I).

Description

description

title

Method and device for determining a control signal

The invention relates to a method and a device for classifying an input signal and determining a control signal, in particular for

Driving an actuator, a method for training the machine

Learning system, a training system, a computer program and a

machine-readable storage medium.

State of the art

A method for determining whether there is an anomaly in a communication network is known from the unpublished DE 10 201 721 0787.1, wherein at least one message transmitted via the communication network is analyzed by depending on the, i.e. especially from the

transmitted message, an input variable of an encoder is determined, an intermediate variable of reduced dimensionality being determined by means of the encoder, and depending on the intermediate variable using a

Decoder an output variable with the dimensionality of the input variable is determined, it being decided depending on the input variable and output variable whether or not the anomaly is present. Such a combination of encoder and decoder such that the input variable has one or more intermediate variables of reduced dimensionality, i.e. with a

Loss of information connected, converted into the output size, can also be referred to with the word "Autoencoder".

A "Variant Autoencoder" is known from "Auto-Encoding Variational Bayes", arXiv preprint arXiv: 1312.6114vlO, Diederik P Kingma, Max Welling, 2014. From “Learning in an Uncertain World: Representing Ambiguity Through Multiple Hypotheses”, arXiv preprint arXiv: 1612.00197v3, Christian Rupprecht et al., International Conference on Computer Vision (2017) and “Uncertainty Estimates and Multi-Hypotheses Networks for Optical Flow”, arXiv preprint

arXiv: 1802.07095v3, Eddy Ilg et al. (2018) are known as Convolutional Neural Networks (CNN), which predict a number of hypotheses.

From “Autoencoding beyond pixels using a learned similarity metric”, arXiv preprint arXiv: 1512.09300v2, Anders Boesen Lindbo Larsen, Soren Kaae Sonderby,

Hugo Larochelle, Oie Winther, 2016 and “Generating Images with Perceptual Similarity Metrics based on Deep Networks”, arXiv preprint arXiv: 1602.02644v2, Alexey Dosovitskiy, Thomas Brox, 2016, methods for training Variational Autoencoders with a 2-player game are known .

Advantages of the invention

The object of the invention has the advantage over that known from the prior art that the anomaly detection improves

Has extrapolation capability, so that the resulting control method is more robust, or alternatively achieves comparable robustness with less training data.

Disclosure of the invention

In a first aspect, the invention therefore relates to the method with the features of claim 1. Further aspects are the subject of

sibling claims. Advantageous further developments are the subject of the dependent claims.

In the first aspect, the invention therefore relates to a method for classifying an input signal, which is determined in particular as a function of a sensor signal, that is to say an output signal from a sensor. In this case, depending on an output signal of an auto encoder to which the input signal is fed, a decision is made as to whether the input signal has an anomaly or not, the auto encoder having at least one encoder and has at least one decoder, an intermediate variable being determined as a function of the input signal by means of the encoder, and the output signal being determined as a function of this using the decoder, the autoencoder providing a plurality of different hypotheses for the input signal supplied to it for the reconstruction of the input signal and as a function of the output signal is determined from this plurality of different hypotheses.

The hypotheses are preferably different in pairs, i.e. no two of the hypotheses are the same. This can preferably be ensured in a training phase of the auto encoder in which parameters of the auto encoder are varied in order to minimize a predefinable cost function, the cost function is selected such that it is punished if two hypotheses are too similar, in particular the same, to one another.

In a further development it can be provided that the car encoder has a

Comprises a plurality of decoders, each of the decoders providing output variables, one of the plurality of different hypotheses depending on the output variables each providing one of the decoders.

This is particularly easy to do with e.g. the approach presented above prevent two hypotheses from being the same, for example by structurally choosing the decoders and / or their parameters

initialized differently. Of course, it is possible that the decoders are partially identical. If the decoders are implemented, for example, by artificial neural networks, it can be provided that the first layers are shared by all decoders and that the decoders only differentiate in the last layer or the last layers.

Provision can preferably be made for the output signal to be selected as a function of, in particular the same, that of the hypotheses that the

Input signal is closest.

In a further development, a control signal can be provided

which depends on whether anomaly was detected or not. This control signal can in particular be used to control a Actuator, preferably for controlling an at least partially automated vehicle.

In a further aspect, a system is provided which is set up to carry out one of the aforementioned methods. This system therefore also includes the aforementioned auto encoder.

In a still further aspect, a method for training this car encoder is therefore provided. This is done using a discriminator. The

The discriminator will either be supplied with an input signal, in particular determined from sensor signals, or an output signal generated by the auto-encoder, whereby the discriminator is trained to distinguish whether the signal supplied to it is the input signal or the output signal generated by the auto-encoder, and the auto-encoder is trained that

Generate output signal so that the discriminator cannot distinguish between the input signal and the generated output signal. I.e.

Discriminator and Autoencoder play a two-player game similar to the game that is used in so-called Generative Adversarial Networks (short: GAN).

To this well-known two-player game for the present

To make Autoencoder applicable, it is therefore proposed that the intermediate variable be generated randomly and by means of the decoder

Output signal is determined, which is then fed to the discriminator for differentiation.

In particular, the intermediate size can be predefined from a fixed value

Probability distribution is drawn, only the discriminator being trained depending on the output signal determined as a result.

I.e. this output signal is included in a term of the cost function so that, depending on the gradient of this cost function, only immediately

Parameters that characterize the discriminator are varied. In the

Variation of further parameters is only included in this cost function insofar as the discriminator parameterized with varied parameters changes behavior which, in a later iteration of the training method, can lead to parameters of the auto-encoder being varied in a changed manner.

It is also possible for the encoder to determine, depending on the input signal, parameters which characterize a statistical distribution, in particular a normal distribution, of the intermediate variable for a given input signal and the intermediate variable from this statistical

Distribution is drawn.

To ensure that the anomaly detection is minimized

If an input signal is erroneously identified as an anomaly, it can be provided that from the majority of the hypotheses generated by the auto-encoder, the one that is most similar to the input signal is selected, the

Discriminator only similar to the one selected depending on this one

The hypothesis tries to decide whether it was generated by the auto-encoder or not, and an adjustment of parameters which characterize the discriminators and / or the auto-encoder is dependent on the result of this attempted decision.

However, this can lead to those generated with the rest

Hypotheses associated gradients disappear, leaving the rest

Hypotheses do not improve to the extent desired

Extrapolation ability of the proposed auto encoder contribute. To counteract this effect, it can be provided that the discriminator tries to decide for each hypothesis generated by the auto-encoder depending on this intermediate variable, whether it was generated by the auto-encoder or not, and with a variation of parameters that affect the discriminator and / or characterizing the autoencoder is dependent on the results of all of these attempted decisions.

Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. The drawings show:

Figure 1 shows schematically an auto encoder according to the invention; Figure 2 schematically shows a structure of an embodiment of the invention;

Figure 3 schematically shows an embodiment for controlling a

at least partially autonomous robot;

Figure 4 schematically shows an embodiment for controlling a

Manufacturing system;

Figure 5 schematically shows an embodiment for controlling a

personal assistant;

Figure 6 schematically shows an embodiment for controlling a

Access system;

Figure 7 schematically shows an embodiment for controlling a

Surveillance system;

Figure 8 schematically shows an embodiment for controlling a

medical imaging system;

FIG. 9 schematically shows a training system for training the car encoder;

FIG. 10 schematically shows an auto encoder for use in the training system;

Figure 11 schematically shows a structure of an embodiment of a

Evaluation unit of the training system;

FIG. 12 shows a flowchart of an embodiment of a method for training the auto-encoder;

Figure 13 is a flowchart showing an embodiment of part of the

Method for training the car encoder; FIG. 14 shows in a flowchart a second embodiment of part of the method for training the auto-encoder;

Figure 15 is a flowchart of a third embodiment of part of the

Method for training the car encoder;

Figure 16 is a flowchart of a fourth embodiment of part of the

Method for training the car encoder;

FIG. 17 shows in a flow chart an embodiment of a method for

Operation of the control system.

Description of the embodiments

Figure 1 shows an actuator 10 in its environment 20 in interaction with a control system 40. Actuator 10 and environment 20 are collectively referred to as an actuator system. At preferably regular time intervals, a state of the actuator system is detected with a sensor 30, which can also be given by a plurality of sensors. The sensor signal S - or in the case of several sensors, one sensor signal S - from the sensor 30 is transmitted to the control system 40. The control system 40 thus receives a sequence of sensor signals S. The control system 40 uses this to determine control signals A, which are transmitted to the actuator 10.

The control system 40 receives the sequence of sensor signals S from the sensor 30 in an optional receiving unit 50, which the sequence of

Converts sensor signals S into a sequence of input signals x (alternatively, sensor signal S can also be directly adopted as input signal x). The input signal x can be, for example, a section or further processing of the sensor signal S. The input signal x can comprise, for example, image data or images, or individual frames of a video recording. In other words, input signal x is determined as a function of sensor signal S. The input signal x is an auto encoder 60, such as it is illustrated by way of example in FIG. 2. The input signal x is also fed to an anomaly detector 80.

The autoencoder 60 is preferably parameterized by parameters f, Q, which are stored in a parameter memory P and are made available by the latter.

The auto encoder 60 determines x output signals y from the input signals.

The output signals y are supplied to the anomaly detector 80.

Output signal y is a reconstruction of the input signal x by the auto-encoder 60. Anomaly detector 80 determined from the input signal x and

Output signal y is a measure of similarity between input signal x and

Output signal y. Depending on this determined degree of similarity, the anomaly detector can determine control signals A, which are fed to the actuator 10, in order to control the actuator 10 accordingly. For example, it can be decided whether the actuator 10 is controlled in a normal mode or in a safety mode. An embodiment of this method is illustrated in FIG. 17.

The actuator 10 receives the control signals A, is controlled accordingly and carries out a corresponding action. The actuator 10 can in this case comprise a (not necessarily structurally integrated) control logic, which determines a second control signal from the control signal A, with which the

Implementation of the control signal A is effected.

In further embodiments, the control system 40 includes the sensor 30. In still further embodiments, the control system 40 alternatively or additionally also includes the actuator 10.

In further preferred embodiments, control system 40 includes one or more processors 45 and at least one

machine-readable storage medium 46 on which instructions are stored which, when executed on the processors 45, the

Control system 40 cause the inventive method

to execute. In alternative embodiments, a display unit 10a is provided as an alternative or in addition to the actuator 10.

FIG. 2 shows an exemplary structure of the auto-encoder 60. The auto-encoder 60 comprises an encoder (61), with the aid of which an intermediate variable z is determined from the input signal x. The dimensionality of the intermediate variable z is reduced compared to the input signal x. Of the

Autoencoder 60 also includes decoders 63 _! , 63 ₂ , ..., which are used to determine reconstructed input signals x ₁ , x ₂ , ... from the intermediate variable z. These reconstructed input signals x ₁ , x ₂ , ... are also called "hypotheses" and preferably have the same dimensionality as the input signal x. An evaluator 65 determines from the hypotheses x ₁ , x ₂ ,

... the output signal y. For example, it can be provided that the evaluator 65 selects from the hypotheses x ₁ , x ₂ ,... As the output signal y that most similar to the input signal.

In the exemplary embodiment shown, the autoencoder 60 is designed as a so-called “Variational Autoencoder” (short: VAE). The UAE is characterized in that the encoder 61 determines, depending on the input signal x, parameters m _{Z | C} , S _{Z | C} that have a statistical distribution q (z | x) of the intermediate variable z associated with the input signal x for a given input signal x

characterize. For example, the statistical distribution qz | x) can be given by a normal distribution, and the parameters m _z \ _c , S _z \ _c are (possibly vector or matrix values, provided the input signal x is multidimensional) mean values m _{Z | C} and covariance matrix S _{Z | C.}

From the statistical distribution qz | x) characterized by these parameters m _{Z | C} , S _{Z | C} ), a sampler 62 then draws one or more intermediate variables z as statistical samples.

These are fed to the decoders 63 ₁ , 63 ₂ , ..., each of them

Parameters m _c \ _z, · i, S _c \ _{z; 1} , m _c \ _z · ₂ , S _c \ _z, · 2. ··· determined, each statistical

Characterize distributions p ₁ (x | z), p ₂ (x | z), ... of the hypotheses x ₁ x ₂ , ... These statistical distributions p ₁ (x | z), p ₂ (x | z),), ... are

for example, the same function by different Parameters m _c \ _{Z) 1} , ^ _c \ _{Z) 1} , m _c \ _z, · 2 _> ^ _c \ _z , · 2. ··· can be characterized. For example, the statistical distributions p ₁ (x | z), p ₂ (x | z), ... each by

Normal distributions are given, and the parameters m _c \ _{z; 1} , S _c \ _{z; 1} , m _c \ _{z; 2} , S _c \ _{z; 2.} ··· (may be vector or matrix values, provided the intermediate size e.g.

is multidimensional) mean _values m _c \ _{Z] 1} , p _{X | Z.2} , ... and covariance matrices, S _C \ _{Z; 1} , S _c , _z.2 , .... Determine sampler

From the parameters m _c \ _{z; 1} , S _c \ _{z; 1} , m _c \ _{z; 2} , S _c \ _z, · ₂ . ··· characterized statistical distributions p ₁ (x | z), p ₂ (x | z), ... samplers 64-L, 64 _{2 J} ... then draw one or more hypotheses x ₁ x ₂ as statistical samples , ....

Encoders 61 and / or decoders 63 ₁ , 63 ₂ , ... are preferably neural networks parameterized by parameters f, Q.

FIG. 3 shows an exemplary embodiment in which the control system 40 is used to control an at least partially autonomous robot, here an at least partially autonomous motor vehicle 100.

The sensor 30 can be, for example, one or more video sensors and / or one or more radar sensors and / or one or more ultrasound sensors and / or one or more LiDAR sensors and / or one or more position sensors (for example, preferably arranged in the motor vehicle 100) GPS) act. Alternatively or additionally, the sensor 30 can also include an information system that determines information about a state of the actuator system, such as a

Weather information system that determines a current or future state of the weather in the environment 20.

The input signal x can contain, for example, information about detected objects in the vicinity of the at least partially autonomous robot.

The actuator 10, which is preferably arranged in the motor vehicle 100, can be, for example, a brake, a drive or a steering of the

Act motor vehicle 100. The control signal A can then be determined in this way that the actuator or the actuators 10 are controlled in the normal mode in such a way that the motor vehicle 100, for example, prevents a collision with identified objects, in particular if it is

Items of certain classes, e.g. pedestrians. in the

In safety mode, the actuator or actuators 10 can be controlled such that e.g. a speed of motor vehicle 100 is limited to a predeterminable speed range.

Alternatively, the at least partially autonomous robot can also be another mobile robot (not shown), for example one that moves by flying, swimming, diving or walking. The mobile robot can also be, for example, an at least partially autonomous lawn mower or an at least partially autonomous cleaning robot. In these cases too, the control signal A can be determined in the normal mode in such a way that the drive and / or steering of the mobile robot are controlled in such a way that the at least partially autonomous robot prevents, for example, a collision with the identified objects,

In a further alternative, the at least partially autonomous robot can also be a garden robot (not shown), which uses an imaging sensor 30 to determine a type or a state of plants in the environment 20. The actuator 10 can then be an applicator of chemicals, for example. The control signal A can be determined in the normal mode depending on the determined type or the determined state of the plants in such a way that an amount of the chemicals corresponding to the determined type or the determined state is applied. In the safety mode, the control signal A can be selected such that the amount of chemicals applied is limited to a predeterminable range of chemicals.

In still further alternatives, the at least partially autonomous robot can also be a household appliance (not shown), in particular a washing machine, a stove, an oven, a microwave or a dishwasher. With the sensor 30, for example an optical one Sensor, a state of an object treated with the household appliance can be detected, for example in the case of the washing machine, a state of laundry that is in the washing machine. A type or a state of this object can then be determined from this. The control signal A can then be determined in the normal mode in such a way that the household appliance is controlled as a function of the determined type or the determined state of the object. For example, in the case of the washing machine, this can be controlled depending on the material from which the laundry contained therein is made. Control signal A can then be selected depending on which material of the laundry has been determined. Control signal A can in

Security mode can be selected such that the household appliance,

for example, the washing machine is turned off.

FIG. 4 shows an exemplary embodiment in which the control system 40 is used to control a production machine 11 of a production system 200 by actuating an actuator 10 that controls this production machine 11. The production machine 11 can, for example, be a machine for punching, sawing, drilling and / or cutting.

Sensor 30 is then, for example, an optical sensor that e.g. Properties of finished products 12 recorded. It is possible that the actuator 10 controlling the production machine 11 is controlled in the normal mode depending on the determined properties of the production product 12, so that the production machine 11 accordingly carries out a subsequent processing step of this production product 12. It is also possible for the sensor 30 to determine the properties of the manufactured product 12 processed by the manufacturing machine 11 and, depending on this, to control the manufacturing machine 11 for a subsequent one

Adapted manufacturing product. In the safety mode it can be provided that the actuator 10 is controlled in such a way that the production machine is switched off.

FIG. 5 shows an exemplary embodiment in which the control system 40 is used to control a personal assistant 250. The sensor 30 is, for example, an acoustic sensor, the speech signals of a user 249 receives. Alternatively or additionally, the sensor 30 can also be configured to receive optical signals, for example video images of a gesture of the user 249.

Depending on the signals from the sensor 30, the control system 40 determines a control signal A of the personal assistant 250 in the normal mode, for example by the machine learning system performing a gesture recognition. This determined control signal A is then transmitted to the personal assistant 250 and is therefore controlled accordingly. This determined control signal A ist can in particular be selected in such a way that it corresponds to a presumed desired activation by the user 249. This presumed desired activation can be determined depending on the recognized gesture. The control system 40 can then select the control signal A for transmission to the personal assistant 250 depending on the presumed desired control and / or select the control signal A for transmission to the personal assistant in accordance with the presumed desired control 250.

This corresponding control can include, for example, that the personal assistant 250 retrieves information from a database and reproduces it in a way that the user 249 can receive.

In the safety mode, it can be provided that the personal assistant 250 is switched off or outputs an error message.

Instead of the personal assistant 250, a household appliance (not shown), in particular a washing machine, a stove, an oven, a microwave or a dishwasher, can also be provided in order to be controlled accordingly.

FIG. 6 shows an exemplary embodiment in which the control system 40 is used to control an access system 300. Access system 300 may include physical access control, such as door 401. The sensor 30 can be, for example, an optical sensor (for example for capturing image or video data), which is set up to grasp a face. Control system 40 can interpret this captured image. For example, the identity of a person can be determined. The actuator 10 can be a lock which, in normal mode, releases the access control depending on the control signal A or not, for example opens the door 401 or not. For this purpose, the control signal A can be selected depending on the interpretation, for example depending on the determined identity of the person. In the safety mode, it can be provided that the actuator 10 is, for example, permanently activated so that the door is permanently open (or alternatively also permanently closed). Instead of the physical access control, a logical access control can also be provided.

FIG. 7 shows an exemplary embodiment in which the control system 40 is used to control a monitoring system 400. This exemplary embodiment differs from the exemplary embodiment shown in FIG. 6 in that, instead of the actuator 10, the display unit 10a is provided, which is controlled by the control system 40. For example, from

Anomaly detector 80 can be determined whether one from the optical sensor

recorded object is suspicious. If this is the case, the

Security mode can be activated. In the normal mode, the object can then be reproduced normally by the display unit 10a, whereas in the safety mode and the control signal A can then be selected such that this object is highlighted in color by the display unit 10a.

FIG. 8 shows an exemplary embodiment in which the control system 40 is used to control a medical imaging system 500, for example an MRI, X-ray or ultrasound device. The sensor 30 can be provided, for example, by an imaging sensor, and the control system 40 controls the display unit 10a. For example, the anomaly detector 80 can determine whether an area recorded by the imaging sensor is noticeable. If this is the case, the

Security mode can be activated. In normal mode this area can then be reproduced normally, in safety mode the control signal A can be selected such that this area is highlighted by the display unit 10a.

FIG. 9 schematically shows an exemplary embodiment of a training system 140 for training the car encoder 60. A training data unit 150 determines suitable input signals x which are fed to a modified form of the car encoder 60, designated by the reference symbol 60 ′. For example, training data unit 150 accesses a computer-implemented database in which a set of training data is stored and selects e.g. randomly input signals x from the set of training data. This will be one

Evaluation unit 180 supplied, the structure of which is illustrated by way of example in FIG. 11.

The modified auto-encoder 60 'is set up to determine training output signals gt, which each include the output signals y, based on an input training signal associated therewith from the input signals x supplied to it. A mode signal m is fed to the modified auto-encoder 60 ', depending on which the modified auto-encoder 60' determines the training output signals yi. These training output signals yi are fed to the evaluation unit 180.

The evaluation unit 180 can, for example, by means of one of the

Input signals x and the training output signals yi dependent cost function (English: loss function) determine the parameter £, which characterizes the performance of the auto encoder 60. The parameters f, Q can be optimized depending on the parameter £. An exemplary method for this is illustrated in FIGS. 12 to 16.

FIG. 10 illustrates an embodiment of modified auto-encoder 60 ', which is largely identical to the auto-encoder 60 shown in FIG.

In particular, the parameterization is identical to parameter f, Q. Only the differences are shown below. Instead of the evaluator 65, the modified auto-encoder 60 'has a modified evaluator 65' which, like the evaluator 65, determines the output signal y. Depending on the mode signal m, a decision is then made as to which of the available variables as the output signal yi are output, ie the difference to the auto-encoder 60 is which sizes are output.

FIG. 11 shows an exemplary embodiment of the evaluation unit 180. Evaluation unit 180 comprises a discriminator 181, to which (selected for example via a switch 183) either the input signal x or the output signal y extracted from the training output signal gt are supplied. The discriminator comprises a mathematical function, for example implemented by a neural network, with the help of which one

A probability quantity p is determined, which characterizes a probability that the signal supplied to the discriminator 181 is an output signal y generated by the modified auto-encoder 60 'and not one

Input signal x. This probability variable p is fed together with the mode signal m to a cost calculator 182, which uses this to determine the characteristic variable L, for example using the methods illustrated in FIGS. 13-16.

FIG. 12 shows a method for training the auto encoder 60 (or

modified auto encoder 60). The method starts (1000) with a selection dependent on the mode signal m, which of the following sub-methods (1001) - (1004) is carried out. Signals x, y are supplied to the discriminator 181 in steps, and the discriminator 181 determines the parameters L _VAE , L _G , L _D depending on these signals x, y. Embodiments of the partial methods (1001) - (1004) are illustrated in FIGS. 13-16.

Then (1200) e.g. with a gradient descent method parameters f, Q, which parameterize the auto-encoder 60, or parameters, which parameterize the discriminator 181, are adapted. The parameters f, Q of the

Auto encoder 60 is adapted so that the parameter L _tot = L _VAE + L _{G is} minimized, while the parameters of the discriminator 181 are adapted such that the parameter L _D = —L _{G is} minimized. The convergence is checked in step (1300), if there is insufficient convergence, the process branches back to the start (1000).

This adaptation can be cascaded in the usual way for GAN, ie it can be provided that in an inner loop (iS) the Autoencoder 60 is trained, and with sufficient convergence of the

Discriminator 181 in an outer loop (aS). The steps of training the car encoder 60 can be skipped here.

If both the training of the auto-encoder 60 and the discriminator 181 have been completed, the method ends (1400).

FIG. 13 illustrates an embodiment of the first partial method (1001).

In this case, input signals x are supplied to the discriminator 181, that is to say, for example, real images.

Discriminator 181 now ascertains (1002) the associated probability variable p for each of these input signals, from which the associated parameter as log (l - Pi) (1) is determined. The sum is carried out over all input signals x, the associated probability quantity p is designated by a subscript. This ends this sub-procedure. The training of the car encoder 60 is omitted in the subsequent pass of step (1200), ie only the discriminator 181 is trained.

FIG. 14 illustrates an embodiment of the second partial method (1002). Here, the modified auto-encoder 60 'is controlled in such a way that the sampler 62 draws an intermediate variable z from a normal distribution J \ T (0.1) with mean value 0 and variance 1 (1012). Then (1022) the hypotheses x ₁ , x ₂ , ... are determined. These are all transmitted with the training output signal gt. Then (1032) these are fed to discriminator 181, which determines the associated probability quantities p. This becomes the parameter log (Pi) (2)

determined. The sum is based on all of the hypotheses generated and the associated probability quantity p is by a subscript designated. This ends this sub-procedure. The training of the car encoder 60 is omitted in the subsequent run through step (1200), ie only that

Discriminator 181 is trained.

FIG. 15 illustrates an embodiment of the third partial method (1003).

Here, the modified auto-encoder 60 ″ input signals x are supplied (1013), and an intermediate variable z from the corresponding one

Probability distribution drawn (1023). Then (1033) a predefinable number of hypotheses x ₁ , x ₂ , ... are drawn from each of the samplers 64 ₁ 64 ₂ , ... (if the autoencoder 60 is a UAE, the predefinable number is

preferably a plurality, if it is a deterministic auto-encoder, only one hypothesis is generated). Then (1043) for each of the

Input signals selected from the total of the generated hypotheses that hypothesis x _k that is the smallest distance from the

Has input signal x. This selected hypothesis x _k and the associated index k are transmitted with the training output signal yi. From this, the associated probability variable p is then determined (1053). This becomes the parameters

determined. The sum is carried out over all of the input signals x, the associated probability quantity p is designated by a subscript. This ends this sub-procedure.

FIG. 16 illustrates an embodiment of the fourth partial method (1004). It begins with steps (1014) and (1024), which are identical to steps (1013) and (1023) of the third sub-method (1003). Then (1034) a hypothesis x ₁ , x ₂ , ... is drawn from each of the samplers 64 ₁ 64 ₂ , ... From the generated hypotheses, the hypothesis x _{k is} selected that has the smallest distance from the input signal x. This selected hypothesis x _k and the associated index k and all other hypotheses x ₁ , x ₂ , ... are all transmitted with the training output signal yi and the index k. Then (1044) the associated probability values p are determined. For each of the input signals x supplied in step (1014), an average is made determined, where n denotes the number of hypotheses generated for input signal x and j indicates the hypotheses. The parameters are determined from this

Ascertained. The sum is carried out over all input signals x, the associated mean p is designated by a subscript. This ends this sub-procedure.

FIG. 17 illustrates an embodiment of the operation of the control system 40. First (2000), the input signal x is fed to the auto-encoder 60. From this, auto-encoder 60 then determines (2100) hypotheses x ₁ , x ₂ , .... From this (2200) the hypothesis x _{k is} selected whose distance from the

Input signal x is lowest. This selected hypothesis x _k is transmitted to the anomaly detector 80 as an output signal y. Anomaly detector 80 now (2300) provides the distance a = \ x— y \ between output signal y and input signal x. This distance a is now compared (2300) with a predefinable threshold value a. If the distance a is smaller than the predefinable threshold value a, then (2400) it is decided whether actuator 10 is operated in normal mode and control signal A accordingly

selected the actuator 10 controlled accordingly. If the distance a is not less than the predefinable threshold value a, then (2500) it is decided whether the actuator 10 is operated in the safety mode and the control signal A is selected accordingly and the actuator 10 is controlled accordingly.

This ends the process. In further preferred embodiments, training system 140 includes one or more processors 145 and at least one

machine readable storage medium 146 on which instructions are stored which, when executed on processors 145, the

Control system 140 cause to execute the method according to the invention.

It is understood that the procedures are not only fully described as described

Software can be implemented. They can also be implemented in hardware, or in a mixed form of software and hardware.

Claims

Expectations

1. Computer-implemented method for classifying a

Input signal (x), which depends in particular on one

Output signal (S) of a sensor (30) is determined, whether it has an abnormality or not, depending on one

Output signal (y) of an auto-encoder (60), to which the input signal (x) is fed, a decision is made as to whether the input signal (x) has the anomaly or not,

wherein the auto encoder (60) has at least one encoder (61) and at least one decoder (63 ₁ , 63 ₂ , ...), depending on the

Input signal (x) is determined by means of the encoder (61) an intermediate variable (z), and depending on this is determined by means of the decoder (63 ₁ 63 ₂ , ...) the output signal (y) is determined,

characterized in that the auto encoder (60) to it

supplied input signal (x) provides a plurality of hypotheses (x ₁ x ₂ ...) for the reconstruction of the input signal (x) and depending on this plurality of hypotheses (x ₁ x ₂ ...) the output signal (y) is determined.

2. The method of claim 1, wherein the auto-encoder (60) comprises a plurality of decoders (63 _1; 63 ₂ , ...), each of the decoders

(63 _1; 63 ₂ , ...) output variables , ···), whereby one of the plurality of different hypotheses (x ₁ , x ₂ ,) depends on the output variables ) one of the decoders (63 _1; 63 ₂ , ...) is provided.

3. The method according to claim 2, wherein the output variables ,

m _c \ _z; 2, _c \ _z ; 2. ···) each of the decoders (63 ^ 63 ₂ , ...)

Characterize probability distribution (p ₁ (x | z), p ₂ (x | z), and each of the hypotheses (c _{1 c2} , ...) from one of these Probability distributions (p ₁ (x | z), p ₂ (x | z), ...) is drawn.

4. The method according to any one of claims 1 to 3, wherein the output signal (y) is selected depending on that of the hypotheses (c _{1 c2} , ...) that is closest to the input signal (x).

5. Computer-implemented method for determining a control signal (A), in particular for controlling an actuator (10), preferably for controlling an at least partially automated vehicle (100), with the method according to one of claims 1 to 4

Input signal (x) is classified as to whether there is an anomaly or not, and the control signal (A) is selected depending on whether an anomaly was detected or not.

6. The method according to claim 5, wherein depending on the determined

Control signal (A) an at least partially autonomous robot (100) and / or a production system (200) and / or a personal assistant (250) and / or an access system (300) and / or a monitoring system (400) or a medical imaging system ( 500) is controlled.

7. System (40) comprising one or more processors (45) and at least one machine-readable storage medium (46) on which

Instructions are stored which, when executed on the processors (45), cause the control system (40) to carry out the method according to one of claims 1 to 6.

8. Computer-implemented method for training the auto-encoder (60) of the system according to claim 7, by means of a discriminator (181) wherein the discriminator (181) an input signal (x) or an output signal (y) generated by the auto-encoder (60) are supplied , the discriminator (181) being trained to distinguish whether the signal supplied to it is the input signal (x) or the output signal (y) generated by the auto-encoder (60), and the auto-encoder (60) being trained to distinguish the output signal ( y) to be generated in such a way that the discriminator cannot distinguish the input signal (x) and the generated output signal (y) in this way.

9. The method according to claim 8, wherein the intermediate variable (z) is generated randomly and the output signal (y) is determined by means of the decoder (63 ₁ , 63 ₂ , ...).

10. The method according to claim 8, wherein the intermediate variable (z) is drawn from a predeterminable probability distribution, wherein depending on the output signal (y) determined as a function thereof, only the

Discriminator (181) is trained.

11. The method according to claim 9, wherein the encoder (61) depending on the input signal (x) parameters determined the one statistical

Characterize distribution (q (z | x)) of the intermediate variable (z) for a given input signal (x) and the intermediate variable (z) from this

statistical distribution (q (z | x)) is drawn.

The method of claim 11, wherein the discriminator (181) to each

depending on this intermediate variable (z) hypothesis (x ₁ , x ₂ , ···) generated by the auto-encoder (60) tries to decide whether it was generated by the auto-encoder (60) or not, and a variation of parameters, the discriminators (181) and / or auto encoders (60)

characterize is dependent on the results of all of these attempted decisions.

13. The method of claim 12, wherein from the plurality of hypotheses (x ₁ , x ₂ , ···) generated by the auto-encoder (60) that (x _k ) is selected which is most similar to the input signal (x), the Discriminator (181) only tries to decide on the basis of this selected most similar hypothesis (x _k ) to decide whether it was generated by the autoencoder (60) or not, and by adjusting parameters that match the

Characterize discriminators (181) and / or the autoencoders (60) depending on the result of this attempted decision.

14. Training system (140) which is set up to carry out the method according to one of claims 8 to 13.

15. Computer program which is set up to carry out the method according to one of claims 1 to 6 or 8 to 13.

16. Machine-readable storage medium (46, 146) on which the

Computer program according to claim 15 is stored.