Adaptively cleaning different types of interference in measurement data

Abstract

Method for training a trainable module for purifying a recording with an interference of a measurement signal, with the following steps: providing a trainable module; selecting a de-interfering module or a combination of de-interfering modules with the trainable module for at least one learning recording of the measurement signal; applying the selected de-interfering module or the selected combination to the learning recording to obtain a de-interfered result; determining an evaluation of how far the interference is suppressed in the de-interfered result with a predefined cost function; optimizing a parameter set at least characterizing the behavior of the trainable module with the aim of improving the evaluation of the then obtained de-interfered result by the cost function when reprocessing the same learning recording and / or other learning recordings.

Description

TECHNICAL FIELD

[0001] The present invention relates to signal processing for physical measurement data with noise and other types of disturbances. BACKGROUND

[0002] The automation of many physical processes that have been guided manually so far is premised on an automatic evaluation of physical measurement data obtained in the observation of the process. Manipulation signals are generated from the results of the evaluation, with which the process is then intervened.

[0003] An important example of such a process is at least partially automated driving, in which measurement data from the observation of the vehicle's environment is the basis for each action that the vehicle then performs.

[0004] The challenge here is that the physically collected measurement data is mostly with noise and other types of disturbances. In the case of automatic evaluation, for example in the case of classification and semantic segmentation by means of artificial neural networks, this disturbance can be amplified and can lead to false classifications. Therefore, the measurement data is usually pre-processed. However, a mismatched pre-processing can lead to artifacts and further worsen the influence of the disturbances.

[0005] WO 2019 / 007 632 A1 discloses a customization of the pre-processing of images so that a downstream evaluation module configured to recognize specific objects or features can evaluate the pre-processed images in the best possible way with respect to these objects or features. SUMMARY

[0006] In the framework of the present invention, a method for training a trainable module is developed. The module is used to clean records of measurement signals with disturbances.

[0007] Here, in addition to statistical disturbances such as fluctuations or noise, the term "disturbance" also includes systematic errors, for example image soft focus effects (Weichzeichnen) due to defocus blur or motion blur.

[0008] The measurement signals are obtained by a physical measurement process, and / or by a partial or complete simulation of such a measurement process, and / or by a partial or complete simulation of a technical system that can be observed with such a measurement process. This results in measurement signal records with disturbances, unlike purely synthetic data, which are produced by technical manufacturing.

[0009] The trainable module is in particular to be seen as a module which embodies a function parameterized with adaptable parameters with a great deal of generality. In training the trainable module, the parameters can in particular be adapted such that, in the case of inputting learning input variable values into the module, the associated learning output variable values are reproduced as well as possible. The trainable module can in particular comprise an artificial neural network KNN, for example a convolutional neural network (English: convolutional neural network), and / or the trainable module can be a KNN.

[0010] In the method, a trainable module is provided which is configured, in a trained state, to select a deinterfering module or a combination of deinterfering modules from a predefined catalog of deinterfering modules for a recording of the measurement signal, to apply the deinterfering module to the recording, which is expected to improve the suppression of the interference contained in the recording. The trainable module thus obtains a recording of the measurement signal as input and provides the selection as output.

[0011] The recording of the measurement signal can in particular comprise, for example, an image which represents a spatial distribution of at least one physical measurement variable and / or a time series which represents a time course of at least one physical measurement variable.

[0012] The selection can state, for example, for each available deinterfering module whether the deinterfering module should act on the recording of the measurement signal and, if necessary, to what extent, in order to suppress the interference contained in the recording. The selection of each deinterfering module can state, for example, a numerical evaluation of the extent to which the application of the respective deinterfering module is considered meaningful. This numerical evaluation can then be further refined, for example, by selecting only the deinterfering module with the best evaluation without taking into account all other deinterfering modules.

[0013] For at least one learning recording of the measurement signal, a deinterfering module or a combination of deinterfering modules is selected using the trainable module. The selected deinterfering module or the selected combination is applied to the learning recording to obtain a deinterfered result. Using a predefined cost function, an evaluation is determined of the extent to which the interference is suppressed in the deinterfered result.

[0014] A parameter set is optimized in the framework of the training. The parameter set characterizes at least the behavior of the trainable module. The optimization is intended to improve the evaluation of the deinterfered result then obtained by the cost function when the same learning recording and / or other learning recordings are reprocessed.

[0015] It has been recognized that in this way, even if each existing disturbance fluctuates in terms of temporal course, the purification of the measurement signal can be adapted to the disturbances in an adaptive manner. Different disturbances can be based on very different physical mechanisms and thus also exhibit different countermeasures to suppress these disturbances. Thus, for example, the influence of noise can be at least partially eliminated by soft focus effects, while the influence of defocus blur or motion blur is exacerbated by soft focus effects.

[0016] Furthermore, by the adaptive selection it can be possible, for example, to avoid the de- disturbance module acting when it is not actually needed. Thus, for example, filtering an image in the spatial frequency space by means of Fourier analysis or wavelet analysis can largely eliminate noise in a noisy image, while the same filtering in the case of a noise-free image only leads to a loss of detail. If, for example, there is only a temporary malfunction, for example when a vehicle drives through a poorly lit area while recording images, the corresponding de-disturbance module only acts on the images recorded in this period of time.

[0017] Finally, the use of the trainable module for selecting a de-disturbance module to be applied can achieve greater freedom and also greater transparency in terms of the de- disturbance methods that can be applied to the recordings compared to using the trainable module directly as a de-disturbance module.

[0018] If the trainable module, for example, contains a KNN with a pre-given structure, the behavior of the trainable module is characterized by a set of parameters, it is difficult to see which values for the parameters are considered by the KNN for a specific desired de- disturbance method, for example a specific filtering. By decoupling the actual de- disturbance process from the selection, it is possible to specifically pre-give a catalog of de- disturbance methods to be considered.

[0019] Furthermore, this decoupling prevents strong nonlinearities or even non-differentiable transfer functions in the de-disturbance module from directly breaking the final dependence of the cost function on the parameters characterizing the behavior of the trainable module. The cost function, which is then expressed as a function of these parameters, is "well-behaved", so that, for example, the optimization of the parameters converges better using gradient descent methods.

[0020] Thus, each de-disturbance module can convert a measurement signal recording obtained as input into a de-disturbance result using an arbitrary transfer function and provide the de- disturbance result as output.

[0021] Here, for example, there can be transfer functions that depend only on the records obtained as input, for example, filters with fixed, pre-specified filter properties that subtract the mean of all values contained in the records, for example, intensity values of image pixels, or copy the records unchanged into the de-interference result. The last-mentioned "virtual de-interference modules" can advantageously be included in the pre-specified catalog, whereby the trainable modules can signal that the de-interference process has converged to a final result by choosing this de-interference module exactly in the recursively performed de-interference process.

[0022] However, for example, there can also be transfer functions with free parameters. Instead of many filters each with fixed, pre-specified filter properties, for example, only one filter can then be stored in the pre-specified catalog, wherein the filter properties are set by still free parameters. For example, the de-interference modules can also be configured as autoencoder networks that first convert the input records into an intermediate state with low dimensionality and then reconstruct from this intermediate state again. Such autoencoder networks have very many free parameters that are suitable for training.

[0023] Therefore, in a particularly advantageous configuration, the parameter set optimized in terms of improving the evaluation by the cost function additionally characterizes the behavior of at least one de-interference module. The selection of a de-interference module and the de-interference module itself can then be trained together, for example.

[0024] Via the cost function, any type of prior knowledge ("ground truth") available in terms of the learning records can be fed into the training process. Based on a sufficient number of learning records, the trainable modules can then learn to evaluate any record in terms of which types of disturbances are present and which measures can be used to suppress these disturbances. The trainable modules can then fulfill a function comparable to a human expert for image processing. If a photo is presented to such an expert, the expert directly identifies which problems affect the quality of the photo. The expert can then purposefully employ tools in image processing software that eliminate these problems.

[0025] For example, the cost function can depend on a comparison between the deinterference result and a known learning deinterference result. Learning data for the trainable module can be obtained, for example, by subjecting an undisturbed recording of a measurement signal, such as an image, to a disturbance of defined type and intensity. The recording with the disturbance is then the learning recording for which the trainable module is to select one or more matching deinterference modules for the deinterference. The undisturbed original state of the recording is the learning deinterference result which ideally should be reproduced when applying the selected deinterference modules to the learning recording.

[0026] Alternatively or in combination, the cost function can depend, for example, on an automated evaluation of the deinterference result based on at least one pre-given criterion. For example, the noise intensity in an image can be determined based on a histogram of intensity values and / or based on the standard deviation. For this purpose, it is not necessary to have an undisturbed version of the recording available.

[0027] Alternatively or in combination, the cost function can depend, for example, on an evaluation of the deinterference result obtained from at least one external evaluator. In this way, the knowledge of an expert in the quality of the deinterfered recording can be retained and converted into a machine-usable form. Hitherto, this type of knowledge has been difficult to put down in writing.

[0028] In a further advantageous configuration, the deinterference result is processed into a work result by a pre-given processing module. The cost function then depends on a comparison between the work result and a known learning work result. The processing module can be configured, for example, as a classifier and / or regressor, which provides a classification and / or regression of the deinterference result as a work result.

[0029] In the example mentioned - in which an initially undisturbed recording is subjected to a disturbance of defined type and intensity - the undisturbed recording can be fed, for example, to the processing module to obtain the learning work result. If the disturbed recording has been deinterfered by the one or more deinterference modules selected by the trainable module and the deinterference result obtained therefrom is likewise fed to the processing module, the processing module should provide a result as close as possible to the learning work result.

[0030] In a particularly advantageous configuration, the trainable module is selected which comprises an artificial neural network. In such a network, the task of selecting one or more deinterference modules matching a given recording for the measurement signal can be particularly simply decomposed into two subtasks. In a first step, the disturbance contained in the recording can be characterized in more detail, and in a second step, the one or more deinterference modules can be selected on the basis of this knowledge.

[0031] The interference can be characterized, for example, by a first sequence of convolutional layers which successively reduce the dimensionality of the recordings. For example, this sequence can comprise between 2 and 7 convolutional layers, wherein one or more pooling layers can optionally also be arranged between these convolutional layers.

[0032] The selection of the one or more de-interference modules can be carried out, for example, by a second sequence of fully cross-linked layers. For example, this sequence can comprise between 2 and 4 fully cross-linked layers.

[0033] In another particularly advantageous configuration, an intermediate result provided as output by at least one convolutional layer is evaluated using a pre-training cost function. In order to improve the value of this pre-training cost function, parameters characterizing the behavior of the first sequence of convolutional layers are optimized in advance. The intermediate result can be obtained, for example, from the penultimate convolutional layer, i.e. the convolutional layer adjacent to the convolutional layer immediately adjacent to the first fully cross-linked layer. In order to exemplarily apply de-interference of an image, the intermediate result can comprise, for example, approximately 500 numerical values. This type of pre-training essentially focuses on the interference itself and largely neglects the semantic content of the image. Thereby, the following tendency of subsequently training the complete parameter set is suppressed, namely always also to link the recognized interference to the semantic content of the measurement signal recording.

[0034] The pre-training cost function can in particular depend on a comparison between the intermediate result and known additional information about the interference. In the example mentioned, in which a defined interference is applied to an initially undisturbed recording, it is thus possible, for example, to assess whether the interference was recognized correctly.

[0035] The application also relates to a further method for purifying recordings of measurement signals with interference.

[0036] Under the framework of this method, at least one trainable module trained in accordance with the aforementioned method is first provided. This provision can comprise, for example, training the trainable module as described above. However, the trainable module can also be obtained from any source in a state in which the training has already been completed.

[0037] At least one recording of the measurement signal is fed to the trainable module, so that the trainable module selects a de-interference module or a combination of de-interference modules from a predefined catalog of de-interference modules. The selected de-interference module or the selected combination is applied to the recording, thereby obtaining a de-interference result. This method thus relates to the effective operation (Wirkbetrieb) of the trainable module in a trained state for the purpose of practical application (inference).

[0038] As explained previously, the training described above has the following effect in this connection: exactly the following deinterfering module is applied to each record, namely, the deinterfering module that is indeed required to suppress the specific interference present in the record.

[0039] Advantageously, the deinterfering result is further processed into a work result by a predefined processing module. As described above, improved deinterfering of the records in this connection has the following effect, namely, that the correct work result is obtained in each application context with greater probability. This applies in particular in the case where the processing module is configured as a classifier and / or regressor, which provides a classification and / or regression of the deinterfering result as a work result.

[0040] In a further particularly advantageous configuration, a control signal is generated from the deinterfering result and / or from the work result. The technical system is controlled with the control signal. In this context, the improved quality of the deinterfering result or the work result leads to the respective technical system performing an action that is optimal from the perspective of the respective application with greater probability. For example, if traffic signs or other road users are semantically classified with higher accuracy in the framework of an object recognition for at least partially automated driving, a vehicle controlled on the basis of the object recognition drives according to the expectations of the traffic situation with greater probability.

[0041] During training and in active operation, the trainable module can provide, for example, weights as a selection of a combination of deinterfering modules, with which the deinterfering results provided by different deinterfering modules are combined. This reflects the insight that many interferences are complex and cannot be eliminated by simple single measures. This applies in particular in the case where the predefined catalog of deinterfering modules is very finely granular, i.e., each individual deinterfering module is configured comparatively simply. On the one hand, this granular structure of the catalog provides greater flexibility for processing larger categories of possible interferences. On the other hand, implementation, for example in the form of software, also becomes easier. Testing a plurality of small modules independently of one another can be significantly simpler than testing a monolithic module that combines the functionality of these small modules.

[0042] However, even if the trainable module always only selects one deinterfering module to be applied, a plurality of deinterfering modules can be applied to the records of the measurement signal. To this end, the deinterfering result can be fed back as a record of the measurement signal with interference to the trainable module. That is, deinterfering can be performed recursively, in which case one aspect is always completed. For example, this recursion can be repeated until the deinterfering result converges according to a predefined criterion.

[0043] As explained above, the trainable module can exhibit the convergence, for example, by proposing a "virtual interference cancellation module" which copies its input identically to its output. The reason can be, for example, that the interference cancellation result has reached a Pareto-Optimum. It is then no longer possible to improve one aspect of the interference cancellation result without deteriorating another aspect.

[0044] In another particularly advantageous configuration both in training and in effective operation, at least one interference cancellation module comprises a plurality of sub-modules and further trainable modules. The further trainable modules control to what extent the action of this interference cancellation module consists of the action of one or more sub-modules. The composition can be done, for example, by the weights explained above or by the recursion explained above.

[0045] It is thus possible to nest arrangements consisting of trainable modules and a plurality of interference cancellation modules which the trainable module has at its disposal. One possible practical application of this nesting is to extend further interference cancellation modules to an already existing arrangement in which the trainable module has already completed training. It is then not necessary to discard the successfully concluded training, but rather it can be fully continued to use the training. In order to bind the further interference cancellation modules to the further trainable modules, only a significantly less complex training of the further trainable modules is necessary.

[0046] For example, if the already existing arrangement comprises 50 interference cancellation modules and three further interference cancellation modules are to be extended, it is not necessary to train completely from scratch using now a total of 53 interference cancellation modules. Now, the newly added trainable module has the option of selecting the previous arrangement of the 50 interference cancellation modules as a whole or of selecting one of the three newly added interference cancellation modules. The selection between only four modules is much easier to train than the selection between 53 modules. If the already existing arrangement is selected, it is decided within the arrangement by the original, already trained trainable module to what extent which of the original 50 interference cancellation modules are used.

[0047] The invention also relates to a parameter set with parameters which characterize the behavior of a trainable module and which are obtained with the aforementioned training method. These parameters can be, for example, weights with which the inputs of neurons or other computing units in a KNN are calculated for the activation of these neurons or computing units. The parameter set embodies the costs which were invested in the training and is thus an independent product. That is, a person who provides this parameter set in similar circumstances does not need to invest the costs for training anew.

[0048] According to the above, the present application also relates to a device for performing one of the above methods. The device comprises a plurality of interference cancellation modules, each interference cancellation module obtaining a record of a measurement signal as input and being configured to output a version of said record in which at least part of the influence of a particular interference is removed. The device further comprises at least one trainable module, which also obtains said record of said measurement signal as input and controls to what extent the interference cancellation result provided by the device as a whole consists of versions provided by the individual interference cancellation modules.

[0049] In particular, the device can further comprise a processing module configured to further process the interference cancellation result into a working result. The processing module can in particular be configured as a classifier and / or a regressor.

[0050] The methods can be implemented in whole or in part in a computer, e.g. in the form of software. Thus, the present application also relates to a computer program having machine-readable instructions which, when executed on one or more computers, cause the one or more computers to perform one of the described methods. In this sense, also a control device for a vehicle and an embedded system for a technical device which are capable of executing machine-readable instructions are considered to be computers.

[0051] The present application also relates to a machine-readable data carrier and / or a download product having said computer program. A download product is a digital product which can be transmitted via a data network, i.e. which can be downloaded by a user of said data network, which can for example be sold in an online shop for immediate download.

[0052] Furthermore, a computer can be equipped with said computer program, said machine-readable data carrier or said download product. BRIEF DESCRIPTION OF DRAWINGS

[0053] Further measures to improve the present application are shown in detail below together with a description of preferred embodiments of the present application based on the drawings.

[0054] Figure 1 An embodiment of a method 100 for training a trainable module 3 is shown;

[0055] Figure 2 An embodiment of a method 200 for purifying a record 11 of a measurement signal 10 is shown;

[0056] Figure 3 An embodiment of a device 1 is shown;

[0057] Figure 4 A detailed view of a trainable module 3 having a neural network 30 is shown;

[0058] Fig. 5 shows an exemplary elimination of Gaussian noise Figure 5a ) and shot noise Figure 5b ). DETAILED DESCRIPTION

[0059] Figure 1 is a flow chart of an embodiment of the method 100. In step 110, the trainable module 3 is provided. In step 120, learning records 11a of the measurement signal 10 are fed to the trainable module 3. These learning records 11a are with disturbances 11b. For each learning record 11a, the trainable module 3 selects from a pre-given catalog 2 of de- disturbance modules 2a-2d those de-disturbance modules 2a-2d that are best suited to remove the disturbances 11b from the learning record 11a from the perspective of the trainable module 3. If necessary, weights 5a-5d are assigned to the selected de-disturbance modules 2a-2d that should be used by said de-disturbance modules.

[0060] In step 130, the selected de-disturbance module 2a-2d or each de-disturbance module 2a-2d is applied to the learning record 11a. A de-disturbed result 12 is produced.

[0061] Optionally, it is checked in step 180 whether the de-disturbed result 12 has converged according to a pre-given criterion. If this is not the case (logical value 0), the de-disturbed result 12 is re-fed in step 190 to the trainable module 3 as a record 11a of the measurement signal 10 with disturbances 11b. If, in contrast, the de-disturbed result 12 has converged (logical value 1), it is final and is further evaluated.

[0062] In step 140, an evaluation 14 of the de-disturbed result 12 is determined by means of a cost function 13. In step 150, the de-disturbed result 12 is processed into a work result 15 by a processing module 4.

[0063] Exemplary components of the cost function 13 are broken down in detail within block 13. According to block 13a, the de-disturbed result 12 can be compared to a known learning de-disturbed result 12a. According to block 13b, the de-disturbed result 12 can be automatically evaluated based on a pre-given criterion 18. According to block 13c, an evaluation of the de-disturbed result 12 can be used that is obtained from at least one external evaluator 19, for example from an expert or by Crowdsourcing. According to block 13d, the work result 15 formed in step 150, for example a classification, can be compared to a known learning work result 15a.

[0064] In step 160, the parameter set 33 is optimized using the evaluation 14 determined with the cost function 13, which at least characterizes the behavior of the trainable module 3. Here, according to block 161, the intermediate result 16 representing the disturbance 1 1 b obtained from the trainable module 3 can be evaluated, for example, with a pre-training cost function 17, and according to block 162, a part of the parameters 33 can be pre-optimized with the goal of improving the value of the pre-training cost function 17. The pre-training cost function 17 can contain, for example, according to block 17a, a comparison between the intermediate result 16 and additional known information 16a about the disturbance 1 1 b.

[0065] Figure 2 is a flowchart of an embodiment of the method 200 for purifying records 1 1. In step 210, a trainable module 3 is provided which has been trained with the previously described method 100, wherein the trainable module 3 can be trained up to this point, but can also be obtained from any source in a state in which the training has already been completed.

[0066] In step 220, at least one record 1 1 of the measurement signal 10 is fed to the trainable module 3, so that the trainable module selects one or more de-disturbing modules 2a-2d to be applied to the record 1 1, optionally with weights 5a-5d. In step 230, the de-disturbing modules 12a-12d selected in this way are applied to the record 1 1.

[0067] Optionally, in step 280, it is checked whether the de-disturbed result 12 has converged according to pre-given criteria. If this is not the case (logical value 0), the de-disturbed result 12 is fed to the trainable module 3 again in step 290 as a record 1 1 of the measurement signal 10 with a disturbance 1 1 b. If, in contrast, the de-disturbed result 12 has converged (logical value 1 ), this de-disturbed result is final and is further evaluated.

[0068] In step 240, the de-disturbed result 12 is further processed into a work result 15 by a pre-given processing module 4, for example a classifier.

[0069] In step 250, a control signal 20 is formed from the de-disturbed result 12 and / or the work result 15. In step 260, the vehicle 50 and / or the classification system 60 and / or the system 70 for quality control of mass-produced products and / or the system 80 for medical imaging are controlled with the control signal.

[0070] Figure 3An exemplary embodiment of the device 1 for performing the above described methods 100, 200 is shown. Here, the inventory 2 of deinterfering modules comprises four deinterfering modules 2a-2d, each of which produces a version 2a*-2d* from the recording 11b, 11a with the interference 11b. The trainable module 3 likewise obtains the recording 11, 11a as input and determines the weights 5a-5d with which the versions 2a*-2d* are calculated by the weighting unit 5 to the deinterfered result 12. Furthermore, the processing module 4 is set up in order to further process the deinterfered result 12 to the work result 15.

[0071] The first deinterfering module 2a comprises submodules 2a1-2a3 and further trainable modules 3 which determine how the submodules 2a1-2a3 each act on the version provided by the deinterfering module 2a as a whole. This can take place, for example, with further weighting units 5, analogously to the interaction of the deinterfering modules 2a-2d.

[0072] Figure 4 A detailed view of the trainable module 3 is shown, which here comprises a neural network 30. The neural network 30 comprises a first sequence 31 of three convolutional layers 31a-31c which obtain the recording 11, 11a as input and which characterise the interference 11b contained in the recording 11, 11a primarily. A second sequence 32 of two fully cross-linked layers 32a, 32b adjoins this first sequence, which determines the selection of the deinterfering modules 2a-2d, if appropriate using the weights 5a-5d. From the penultimate convolutional layer 31b, the intermediate result 16 is taken which is used in the pre-training cost function 17.

[0073] Fig. 5 shows schematically how different interferences 11b in the recording 11 can be eliminated by applying different deinterfering modules 2a-2d.

[0074] Figure 5a A first example is shown in which the recording 11 contains Gaussian noise as interference 11b. Here, the deinterfering module 2a is activated which softens the recording.

[0075] Figure 5b A second example is shown in which the recording 11 contains shot noise as interference 11b. Here, first the first deinterfering module 2c is activated which "opens up morphologically" the recording 11. The result 12 is subsequently introduced into the second deinterfering module 2d in a recursive manner and is "closed up morphologically" here. The two operations together provide the final result 12' which is free of shot noise and which still substantially obtains all the details.

Claims

1. A method (100) for training a trainable module (3) to clean a recording (11) of a measurement signal (10) with an interference (lib), having the following steps: providing (110) a trainable module (3) which is configured to select, in a trained state, a deinterfering module (2a-2d) or a combination of deinterfering modules (2a-2d) from a predefined catalog (2) of deinterfering modules (2a-2d) for a recording (11) of the measurement signal (10), to apply the deinterfering module (2a-2d) to the recording (11) is expected to improve suppression of the interference (lib) contained in the recording (11); selecting (120), with the trainable module (3), a deinterfering module (2a-2d) or a combination of deinterfering modules (2a-2d) for at least one learning recording (11a) of the measurement signal (10); applying (130) the selected deinterfering module (2a-2d) or the selected combination to the learning recording (11a) to obtain a deinterfered result (12); determining (140), with a predefined cost function (13), an assessment (14) of how far the interference (lib) is suppressed in the deinterfered result (12); optimizing (160) at least a parameter set (33) characterizing a behavior of the trainable module (3) with the aim of improving the assessment (14) of a then obtained deinterfered result (12) by the cost function (13) when reprocessing the same learning recording (11a) and / or other learning recordings (11a), wherein the trainable module (3) provides weights (5a-5d) as a selection of a combination of deinterfering modules (2a-2d) with which deinterfered results (12) provided by different deinterfering modules (2a-2d) are combined. The parameter set (33) additionally characterizes a behavior of at least one deinterfering module (2a-2d). The cost function (13) depends on a result of a comparison (13a) between the deinterfered result (12) and a known learning deinterfered result (12a) and / or on an automated assessment (13b) of the deinterfered result (12) based on at least one predefined criterion (18) and / or on an assessment (13c) of the deinterfered result (12) obtained from at least one external assessor (19).

4. The method (100) according to any one of claims 1 to 2, wherein the deinterfered result (12) is processed (150) by a predefined processing module (4) into a work result (15), and the cost function (13) depends on a result of a comparison (13d) between the work result (15) and a known learning work result (15a). The processing module (4) is configured as a classifier and / or regressor which provides a classification and / or regression of the deinterfered result (12) as a work result (15). ​ ​ 2. The method (100) of claim 1, wherein ​ 3. The method (100) according to any one of claims 1 to 2, wherein ​ ​ ​ ​ 5. The method (100) of claim 4, wherein, ​ 6. The method (100) according to any one of claims 1 to 2, wherein The trainable module (3) comprises an artificial neural network (30), wherein the artificial neural network (30) comprises a first sequence (31) of between 2 and 7 convolutional layers (31a-31c) and a second sequence (32) of between 2 and 4 fully connected layers (32a-32b).

7. The method (100) of claim 6, wherein, An intermediate result (16) provided as output by at least one convolutional layer (31a-31c) is evaluated (161) using a pre-training cost function (17), and wherein parameters (33) characterizing the behavior of the first sequence (31) of convolutional layers (31a-31c) are pre-optimized (162) to improve the value of the pre-training cost function (17).

8. The method (100) of claim 7, wherein The pre-training cost function (17) depends on a result of a comparison (17a) between the intermediate result (16) and known additional information (16a) about the disturbance (11b).

9. A method (200) for purifying recordings (11) of a measurement signal (10) with disturbances (11b), having the following steps: providing (210) at least one trainable module (3) trained with the method (100) of any one of claims 1 to 8; feeding (220) at least one recording (11) of the measurement signal (10) to the trainable module (3) such that the trainable module (3) selects a de-disturbing module (2a-2d) or a combination of de-disturbing modules (2a-2d) from a pre-given catalog (2) of de-disturbing modules (2a-2d); applying (230) the selected de-disturbing module (2a-2d) or the selected combination to the recording (11), thereby obtaining a de-disturbed result (12).

10. The method (200) of claim 9, wherein, The de-disturbed result (12) is further processed (240) by a pre-given processing module (4) to a work result (15).

11. The method (200) of claim 10, wherein, The processing module (4) is configured as a classifier and / or regressor providing a classification and / or regression of the de-disturbed result (12) as a work result (15).

12. The method (200) according to any one of claims 10 to 11, wherein, From the de-disturbed result (12) and / or from the work result (15), a maneuvering signal (20) is generated (250), and wherein a vehicle (50), and / or a classification system (60), and / or a system (70) for quality control of mass-produced products, and / or a system (80) for medical imaging is maneuvered (260) with the maneuvering signal (20).

13. The method (100, 200) of claim 1 or 9, wherein, The de-disturbed result (12) is fed (190, 290) back to the trainable module (3) as a recording (11, 11a) of the measurement signal (10) with the disturbance (11b).

14. The method (100, 200) of claim 13, wherein The feeding (190, 290) is repeated until the de-disturbed result (12) converges (180, 280) according to a pre-given criterion.

15. The method (100, 200) of claim 1 or 9, wherein, The at least one interference-removing module (2a-2d) comprises a plurality of sub-modules (2a1-2a3) and further trainable modules (3a) which control to what extent the effect of the interference-removing module (2a-2d) consists of the effect of one or more sub-modules (2a1-2a3).

16. The method (100, 200) of claim 1 or 9, wherein, The recording (11) of the measurement signal (10) comprises an image representing a spatial distribution of at least one physical measurement variable and / or a time series representing a time course of at least one physical measurement variable.

17. A device (1) for performing a method (100, 200) according to any one of claims 1 to 16, the device comprising: a plurality of interference-removing modules (2a-2d), each interference- removing module obtaining a recording (11) of a measurement signal (10) as input and being configured to output a version (2a*-2d*) of said recording (11) at least partially freed from the influence of a specific interference (11a); at least one trainable module (3) which likewise obtains said recording (11) of the measurement signal (10) as input and controls to what extent an interference-removing result (12) provided overall by the device (1) consists of the versions (2a*-2d*) provided by the individual interference-removing modules (2a-2d).

18. The device (1) according to claim 17, further comprising a processing module (4) which is configured to further process the interference-removing result (12) into a work result (15).

19. The apparatus (1) according to claim 18, wherein The processing module (4) is configured as a classifier and / or regressor which provides a classification and / or regression of the interference-removing result (12) as a work result (15).

20. A computer program product containing machine-readable instructions which, when executed on one or more computers, cause the one or more computers to perform a method (100, 200) according to any one of claims 1 to 16.

21. A machine-readable data carrier and / or a download product with a computer program product according to claim 20.

22. A computer which is equipped with a computer program product according to claim 20 and / or with a machine-readable data carrier and / or a download product according to claim 21.

Images