Method and apparatus for processing data associated with radar system

Abstract

A method and apparatus for processing data associated with a radar system. A method for processing data associated with a radar system, such as a vehicle or an infrastructure device, to reduce interference, the method having: providing a generative model of the variational auto-encoder type, the generative model having a first artificial neural network and a second artificial neural network; providing training data having a plurality of different training data sets for training the generative model; training a first artificial neural network and a second artificial neural network using the first set of training data; the second artificial neural network is further trained using the second set of training data.

Description

Technical Field

[0001] This disclosure relates to a method for processing data associated with a radar system.

[0002] This disclosure also relates to an apparatus for processing data associated with a radar system. Summary of the Invention

[0003] Some examples relate to a method, such as a computer-implemented method, for processing data associated with radar systems, such as vehicles or infrastructure devices, to reduce interference. This method includes: providing a generative model of the variational autoencoder type, wherein the generative model has a first artificial neural network and a second artificial neural network; providing training data with multiple different training datasets for training the generative model; training the first and second artificial neural networks using the first set of training data; and further training the second artificial neural network using a second set of training data. In some examples, this enables more efficient reduction of radar signal interference with radar systems, for example, compared to certain conventional methods.

[0004] In some examples, it is specified that: a first artificial neural network has a first network architecture 𝜑 and is designed to: receive, for example, a first radar signal, such as a potentially erroneous or interfered first radar signal, as first input data, for example, in the form of a spectrum, such as in the form of a range Doppler representation; and generate first output data based on the first input data, which characterize a feature representation r associated with a probability distribution, for example, according to... Where, represents the trainable parameters of the first network architecture , such as weights, where, Let represent the mean of the probability distribution, and where , The standard deviation represents the probability distribution, where, for example, the method includes: receiving a first radar signal; and generating first output data by means of a first artificial neural network.

[0005] For example, a first artificial neural network has multiple layers, wherein at least some layers have and / or represent at least one of the following elements: a) a linear layer; or b) a convolutional layer; or c) a non-linear activation layer and / or a self-attention layer, such as a self-attention layer. Similar considerations can be applied to a second artificial neural network in other examples.

[0006] In some examples, the second artificial neural network has a second network architecture λ and is designed to be based on data obtained through the mean. and standard deviation Samples x drawn from the represented probability distribution generate second output data, which characterize the second radar signal, for example, in the form of a spectrum, for example, in the form of a range Doppler representation, for example, according to... Where, for example, η represents the trainable parameters of the second network architecture ᝜�, such as weights, and where, for example, the method has: providing samples x, for example, according to A second output data is generated using a second artificial neural network. Thus, in other examples, a synthetic radar signal can be generated that has a lower level of interference than the first radar signal.

[0007] In some examples, the training and / or the further training has the following features: using a fully supervised training method, such as based on at least one interference-free signal and at least one interference-present signal, such as a potentially interference-present signal, such as a first radar signal.

[0008] In some examples, the training and / or the further training have the following characteristics: using a loss function that follows the formula... ,in, Characterizes the loss function, where ,in, The entire uncertainty network associated with the model is represented, where λ represents the hyperparameters, where λ>0, and KL represents the Kullback-Leibler divergence, which is determined according to... ,in, Characterize a specifyable distribution, such as a "target distribution," or the feature space distribution of a radar signal, where, for example, a loss function can be adapted to another specifyable target distribution, for example, by... Replace it with another distribution, such as a uniform distribution, for example, on an n-dimensional hypercube with a side length of 1.

[0009] In some examples, the method has the following features: for example, in the case of using a simulation environment, determining at least one training dataset by means of simulation, for example, determining three different training datasets by means of simulation, wherein, for example, these different training datasets have different complexities, for example, in terms of possible perturbations.

[0010] In some examples, the method is specified as follows: generating multiple, e.g., batches ("batch") pairs of potentially distorted signals and undisturbed signals based on training data; replacing at least some, e.g., all of the distorted signals with corresponding undisturbed signals with a specifyable probability; updating the trainable parameters of at least one of the two artificial neural networks based on the gradient of a loss function (e.g., the loss function), wherein the loss function evaluates the fit between the output signal of the generative model and the undisturbed signal; and optionally, repeating at least one of the following aspects: a) generation, or b) replacement, or c) update, e.g., continuously repeating until, e.g., a convergence criterion is met based on a validation dataset, wherein, e.g., the training method or the training method has at least one of the following aspects: a) generation, or b) replacement, or c) update, or repeating.

[0011] For example, the method has at least one of the following elements: a) initializing the parameters of the generative model, such as weights; or b) training the trainable parameters φ of the first network architecture φ and the trainable parameters φ of the second network architecture φ, using the training method and the first training dataset; or c) further training the trainable parameters φ of the first network architecture φ and the trainable parameters φ of the second network architecture φ, using the training method and the second training dataset; or d) further training the trainable parameters φ of the second network architecture φ, for example, further training only the trainable parameters of the second network architecture.

[0012] Other examples involve an apparatus designed to perform the methods described in accordance with this disclosure.

[0013] Other examples relate to a vehicle having at least one device as described in this disclosure.

[0014] Other examples relate to an apparatus, such as an infrastructure apparatus, such as a road infrastructure apparatus, such as a roadside unit, which has at least one device according to this disclosure.

[0015] Other examples relate to a computer-readable storage medium that includes instructions that, when executed by a computer, cause the computer to perform the methods described in accordance with this disclosure.

[0016] Other examples relate to a computer program that includes instructions that, when executed by a computer, cause the computer to perform the methods described in accordance with this disclosure.

[0017] Other examples involve a data carrier signal that transmits and / or represents a computer program as described in this disclosure.

[0018] Other examples relate to uses of methods and / or devices and / or vehicles and / or apparatuses and / or computer-readable storage media and / or computer programs and / or data carrier signals according to this disclosure for at least one of the following elements: a) operating a radar system, such as a vehicle and / or apparatus, such as a fixed radar system; or b) reducing interference, such as interference healing; or c) improving training efficiency; or d) reducing the amount of information required for training.

[0019] Other features, applications, and advantages will be derived from the subsequent description of examples presented in the accompanying drawings. All features described or shown herein, either alone or in any combination, form the subject matter of this disclosure, regardless of their generalization in the claims or their references thereto, and regardless of their expression or presentation in the specification or drawings. Attached Figure Description

[0020] In the attached diagram: Figure 1 A simplified flowchart is shown schematically; Figure 2 A simplified block diagram is shown schematically; Figure 3 A simplified flowchart is shown schematically; Figure 4 A simplified flowchart is shown schematically; Figure 5 A simplified flowchart is shown schematically; Figure 6 A simplified flowchart is shown schematically; Figure 7 A simplified flowchart is shown schematically; Figure 8 A simplified block diagram is shown schematically; Figure 9 An example of its use is illustrated. Detailed Implementation

[0021] Some examples, see, for instance. Figure 1 , Figure 2 This relates to a radar system 12, 22 for processing (e.g., vehicle 10 or infrastructure device 20). Figure 2 Methods to reduce interference by associating data, such as computer-implemented methods, which have the following characteristics: providing 100 ( Figure 1 A variational autoencoder type generative model MOD-VAE, wherein the generative model MOD-VAE has a first artificial neural network ANN-1 and a second artificial neural network ANN-2; providing 102 with multiple different training datasets. , , The training data TD is used to train the generative model MOD-VAE; using the first set of... , Using the training data TD, train 104 first artificial neural network ANN-1 and second artificial neural network ANN-2; using the second set The training data TD, for example, after training 104, further trains 106 of the second artificial neural network ANN-2, for example, only further trains the second artificial neural network ANN-2. In some examples, such as compared with some conventional methods, this makes it possible to reduce interference with radar signals of radar systems 12, 22 more efficiently.

[0022] In some examples, see Figure 2 The radar system 12 may be, for example, an automotive radar system 12, used in vehicles 10, 10a, for example, to identify objects such as other vehicles 10a and / or obstacles in the environment UM of vehicle 10.

[0023] A VAE (variational autoencoder) is a type of generative model that combines the characteristics of an autoencoder with probabilistic modeling. In some examples, a VAE consists of two parts: an encoder that transforms the input data into a relatively low-dimensional latent representation, which is described as a probability distribution; and a decoder that, for example, reconstructs the original data from this latent representation. In some examples, VAE models can be used to generate, for example, new data similar to the training data. In the current context, the generative model MOD-VAE is used to: extract data from, for example, radar systems 12, 22 (…). Figure 2 From the received radar signals, interference that is not in line with expectations is removed, for example, within the framework of so-called "interference healing," thereby improving the subsequent evaluation of the radar signals after the interference has been reduced.

[0024] In some examples, see Figure 2 At least some aspects of the method according to this disclosure can be performed by devices 200, 200', which may be assigned to, for example, a corresponding radar system 12, 22, or integrated into such radar system (not shown).

[0025] In some examples, see Figure 2 , Figure 3 The specification states that the first artificial neural network ANN-1 has a first network architecture 𝜑 and is designed to receive 110 first radar signals, for example, in the form of a spectrum, such as in the form of a range Doppler representation. (For example, received from radar system 12), for example, a first radar signal that may be erroneous or subject to interference, for example, as first input data ED-1, and based on the first input data ED-1, first output data AD-1 is generated, these first output data representing a feature representation r associated with a probability distribution, for example, according to Where, represents the trainable parameters of the first network architecture , such as weights, where, Let represent the mean of the probability distribution, and where , This represents the standard deviation of the probability distribution. In other words, in some examples, see [reference needed]. Figure 3 The method can include: receiving the first radar signal from 110. With the help of the first artificial neural network ANN-1, the first output data AD-1 is generated.

[0026] For example, see Figure 2 The first artificial neural network ANN-1 has multiple layers, wherein at least some of the layers have and / or represent at least one of the following elements: a) a linear layer; or b) a convolutional layer; or c) a non-linear activation layer and / or a self-attention layer, such as a self-attention layer. Similar considerations can be applied to the second artificial neural network ANN-2 in other examples. Generally, in some examples, different combinations of the above-described layers can be implemented to achieve the first and second network architectures.

[0027] In some examples, see Figure 2 , Figure 3 The second artificial neural network, ANN-2, has a second network architecture, λ, and is designed to be based on the mean. and standard deviation Samples x drawn from the probability distribution are used to generate second output data AD-2, which characterize the second radar signal. For example, in the form of a spectrum, for example, in the form of a distance-Doppler representation, for example, according to Where, for example, η represents the trainable parameters of the second network architecture ᝜�, such as weights. In other words, in some examples, see Figure 3 This method can have the following features: providing 114 samples x, for example, according to... Using a second artificial neural network (ANN-2), 116 second output data points (AD-2) are generated. Thus, in other examples, synthetic data can be generated. The synthesized radar signal has a lower level of interference than the first radar signal, but it has a comparable amount of information about available radar signals, such as available radar signal information for object detection.

[0028] In some examples, see Figure 4 The training 104 ( Figure 1 ) and / or the further training 106 has: using 120 ( Figure 4 Fully supervised training methods (TV), such as those based on at least one interference-free signal (or radar signal). and at least one interfering signal, such as a potentially interfering signal, such as a first radar signal. .

[0029] In some examples, see Figure 4 The training 104 ( Figure 1 ) and / or the further training 106 has: using a 122 loss function, which according to ,in, The loss function is LF, where ,in, Characterization and Modeling MOD-VAE ( Figure 2 The entire uncertain network associated with ) where λ represents the hyperparameter, where λ>0, and KL represents the Kullback-Leibler divergence, which is determined according to ,in, Characterize a specifyable distribution, such as a "target distribution," or the characteristic spatial distribution of a radar signal, where, for example, the loss function LF can adapt to another specifyable target distribution, see [see according to...]. Figure 4 Optional block 124, for example by... Replacing it with another distribution, such as a uniform distribution, for example on an n-dimensional hypercube with a side length of 1, yields an adjusted loss function LF'.

[0030] according to Figure 4 Optional block 126 indicates that, when using loss function LF or adjusted loss function LF', training method TV may be optionally performed for model MOD-VAE, such as network ANN-1 or ANN-2 of the model.

[0031] In some examples, see Figure 5 This method has the following characteristics: for example, in the case of using the 130a simulation environment SU, the 130 training data TD is determined by means of simulation. Figure 2 At least one training dataset , For example, simulation can be used to determine three different training datasets for 130b. , , Among them, for example, these different training datasets , , They possess distinct complexities, for example, in terms of potential interference. According to Figure 5 Optional block 132 indicates that, based on the corresponding training dataset, the training method TV may be performed optionally.

[0032] Subsequently, other examples and aspects related to the determination of training data TD are described. For example, using the simulation environment SU ( Figure 5 ), can be used in complex environments (UM) Figure 2 Various devices, such as vehicles 10 and 10a, are placed in the radar, and it is possible to determine, for example, record clean (i.e. interference-free) signals and interfered (i.e., non-zero interference) signals, i.e., radar signals.

[0033] In some examples, clean (interference-free) signals can be identified, for example, recorded in two different ways: 1. disable all radar equipment or radar systems of interfering vehicles 10a; or 2. the same as point 1 above, but including ignoring the effects of clutter and multipath effects (e.g., directly in the simulation).

[0034] In some examples, the above method can be used to create at least two different datasets, for example... D 1 and D 2 .

[0035] In some examples, datasets are created additionally. D 0 This dataset may use or simulate real data. In some examples, two types of signals may be considered: 1. Each vehicle 10, 10a uses the same center frequency, such as a mid-frequency, for example, in odd-numbered periods; 2. Each vehicle 10, 10a uses a different center frequency or mid-frequency (for example, in even-numbered periods). Since in some examples the road scene is pre-planned to determine the dataset, the frequency planning can be performed, for example, with precisely synchronized timing, thereby obtaining data that can be used for efficient training of 10⁴, 10⁶ ( Figure 1 The dataset. For example, in the method described above, a clean (interference-free) signal can be obtained in even-numbered periods, and an interfering (interference-existing) signal can be obtained in odd-numbered periods.

[0036] In some examples, the training data TD can be organized or provided, for example, in the form of three datasets, or in terms of increasing complexity: • D 2 It is the simplest dataset; D 1 For example, datasets that are more complex due to their own interference; D 0 It is an even more complex dataset, because, for example, clean and disturbed signals may have tiny time delays.

[0037] In some examples, see Figure 6 The method specifies that it comprises: generating more than 140 batches of training data TD, for example, a batch consisting of potentially interfered signals (e.g., radar signals) and undisturbed signals (e.g., radar signals); replacing at least some, for example, all of the interfered signals in the batches with corresponding undisturbed signals with a specified probability p; and based on a loss function LF (e.g., the loss function itself)... Figure 4 The gradient of ) is used to update the trainable parameters η and η of at least one of the two artificial neural networks ANN-1 and ANN-2, where the loss function LF evaluates the degree of agreement between the output signal of the generative model MOD-VAE and the undisturbed signal; and optionally, 146 is repeated. Figure 6 The method may have at least one of the following aspects: a) generating 140, or b) replacing 142, or c) updating 144, or repeating 146a until, for example, based on the validation dataset, the convergence criterion is met, wherein, for example, the training method or the training method TV has at least one of the following aspects: a) generating 140, or b) replacing 142, or c) updating 144, or repeating 146, 146a.

[0038] In some examples, the interfered signal is replaced with 142 (depending on probability). Figure 6 For example, replacing it with a clean signal allows the MOD-VAE model to learn features that focus not on the disturbances but, for example, on the underlying clean signal.

[0039] For example, see Figure 7 The method has at least one of the following elements: a) initializing the parameters η, e.g., weights, of the generative model MOD-VAE to 150; or b) using the training method TV (see, for example, see...). Figure 6 ) and the first training dataset D 2In the case of training 152 first network architecture 𝜑 trainable parameters 𝜃 and second network architecture 𝜂; or c) using the training method TV and second training dataset. D 1 In the case of further training 154 trainable parameters φ of the first network architecture φ and trainable parameters φ of the second network architecture φ; or d) using the training method TV and the third training dataset. D 0 In this case, further trainable parameters φ of the second network architecture φ156 can be trained, for example, only further trainable parameters of the second network architecture.

[0040] In other words, using these three training datasets , , The training can be quite similar, for example, using different methods according to... Figure 6 The process, including, for example, the corresponding training dataset , , They are different from each other, as mentioned above.

[0041] In some examples, see Figure 7 Importantly, in block 156, only the second network, ANN-2, is trained, while the first network, ANN-1, is not trained. In some examples, this helps to use the same features as in the simulation, since these features are, for example, ideal features. For instance, the interpretation of these features will only improve based on new real-world data.

[0042] For other examples, see Figure 8 This relates to an apparatus 200, 200' designed to perform the methods described according to this disclosure. For example, the apparatus 200, 200' can be used in at least one vehicle 10, 10a or its radar system 12, and / or can be used in at least one device 20, such as a fixed device, like a roadside unit, or the radar system 22 of that device.

[0043] In some examples, see Figure 8 The device 200, 200' comprises: a computing device (“Computer”) 202 having at least one computing core 202a; and a storage device 204 allocated to the computing device 202 for at least temporarily storing at least one of the following elements: a) data DAT (e.g., information associated with the generative model MOD-VAE and / or the training method TV thereto, such as data TD); b) a computer program PRG, for example for performing the method described in accordance with this disclosure.

[0044] See other examples. Figure 8 The storage device 204 has: volatile memory (e.g., working memory (RAM)) 204a; and / or non-volatile (NVM) memory (e.g., flash EEPROM) 204b; or a combination thereof or a combination with other memory types not explicitly mentioned.

[0045] For other examples, see Figure 8 The present invention relates to a computer-readable storage medium SM comprising instructions PRG that, when executed by a computer 202, cause the computer to perform the method described herein.

[0046] For other examples, see Figure 8 The present invention relates to a computer program PRG comprising instructions which, when executed by a computer 202, cause the computer to perform the method described herein.

[0047] For other examples, see Figure 8 This relates to a data carrier signal (DCS) that represents and / or transmits a computer program (PRG) according to the present disclosure. The DCS can be transmitted (e.g., transmitted and / or received) via, for example, an optional data interface 206 of the devices 200, 200'.

[0048] For other examples, see Figure 2 The present disclosure relates to a vehicle 10 having at least one device 200 as described herein.

[0049] For other examples, see Figure 2 The invention relates to an apparatus 20, such as an infrastructure apparatus, such as a road infrastructure apparatus, such as a roadside unit, which has at least one device 200' as described in this disclosure.

[0050] Subsequently, other aspects and examples are described, which—in other examples—can be combined with at least one of the above aspects and / or examples individually or in any combination of each other.

[0051] The principles described in this disclosure can be advantageously used to reduce or "repair" interference, for example, from received radar signals (e.g., "signal repair"). Radar is, for example, an important sensor for solving perception tasks, such as in vehicles 10, 10a (…). Figure 2 For example, it can be used for autonomous driving. Furthermore, radar can be used, for example, to control beamforming of a communication base station (not shown), for example by incorporating beamforming data from the environment (UM). Figure 2 The knowledge contained therein can be determined, for example, by means of at least one radar system 12, 22.

[0052] In some examples, signal restoration may include estimating a clean signal by removing aspects from the signal to be restored, such as those originating from other transmitters or other sources of interference. For this purpose, in some examples, a generative model MOD-VAE (Modular Model-VAE) may be used. Figure 1 , Figure 2 The networks ANN-1 and ANN-2 of this generative model can be trained as described above in other examples. Signal repair according to this disclosure is useful because, in real-world scenarios, interference can never be completely avoided.

[0053] In some examples, using a generative model of the VAE type, MOD-VAE, enables the creation of relatively effective features, such as for signal restoration. Furthermore, the principles described in this disclosure make it possible to combine multiple training phases, such as during pre-training (see, for example, according to...). Figure 7 Blocks 152 and 154) and subsequent fine-tuning (see, for example, according to...) Figure 7 In the sense of block 156.

[0054] In some examples, aspects of the generative model MOD-VAE can be trained in advance using simulation data, for example, using the simulation environment SU ( Figure 5 In situations where this is achieved, these aspects can be utilized to generate, for example, high-requirement urban scenes with various sources of interference. In some examples, such as after pre-training 152 or 154, only the second part of the MOD-VAU network, ANN-2, can be used, for example, to further train the model or network, for example, to fine-tune it using real data. Thus, in some examples, the data requirements are reduced, for example, the data requirement for real data is reduced, for example, there is no need to adapt the MOD-VAE model to a potentially small dataset of real data.

[0055] The principles described in this disclosure generate a relatively effective representation and, for example, reduce overfitting. Thus, the method according to this disclosure adapts better to real-world radar interference problems and, for example, can generate a cleaner signal than some conventional methods. This, for example, aids perception tasks (such as identifying objects or road users), which can improve traffic safety.

[0056] Generally speaking, the principles described in this disclosure can also be applied to image analysis (e.g., digital images, video, radar, lidar, ultrasound, motion, thermal imaging) or video or audio analysis, and are not limited to radar signals.

[0057] For other examples, see Figure 9This relates to uses 300 of at least one of the following elements: a) operating 301 radar systems 12, 22, such as vehicle 10 and / or device 20, such as fixed radar systems; b) reducing 302 interference, such as interference healing; c) improving 303 the efficiency of 303 training 104, 106; or d) reducing 304 the amount of information required for 304 training, such as the amount of information in real data.

Claims

1. A method, for example a computer-implemented method, for processing data associated with a radar system (12, 22), such as a vehicle (10) or infrastructure device (20) to reduce interference, the method comprising: providing (100) a generative model of the variational autoencoder type (MOD-VAE), wherein, The generative model (MOD-VAE) has a first artificial neural network (ANN-1) and a second artificial neural network (ANN-2); and provides (102) multiple different training datasets ( , , Training data (TD) is used to train the generative model (MOD-VAE); using the first set ( , Using training data (TD), train (104) the first artificial neural network (ANN-1) and the second artificial neural network (ANN-2); using the second set of ( The training data (TD) of the second artificial neural network (ANN-2) are further trained (106).

2. The method according to claim 1, wherein, The first artificial neural network (ANN-1) has a first network architecture φ and is designed to receive a first radar signal, for example, in the form of a spectrum, such as in the form of a range Doppler representation. For example, a first radar signal that may be erroneous or subject to interference is used as first input data (ED-1), and based on the first input data (ED-1), a first output data (AD-1) is generated, wherein the first output data represents a feature representation associated with a probability distribution. For example, according to Where, represents the trainable parameters of the first network architecture , such as weights, where, Let represent the mean of the probability distribution, and wherein, The standard deviation of the probability distribution is represented, wherein, for example, the method includes: receiving (110) the first radar signal ( ); using the first artificial neural network (ANN-1), generate (112) the first output data (AD-1).

3. The method according to claim 2, wherein, The first artificial neural network (ANN-1) has multiple layers, wherein at least some of the layers have and / or represent at least one of the following elements: a) a linear layer; or b) a convolutional layer; or c) a nonlinear activation layer and / or a self-attention layer, such as a self-attention layer.

4. The method according to at least one of claims 2 to 3, wherein, The second artificial neural network (ANN-2) has a second network architecture and a training dataset. , , And it is designed to be based on the mean. and standard deviation Sample x is drawn from the probability distribution to generate second output data (AD-2), which represents the second radar signal. ), for example in the form of a spectrum, for example in the form of a distance-Doppler representation, for example according to Where, for example, η represents the trainable parameters of the second network architecture 𝜓, such as weights, where, for example, the method has the following features: providing (114) the sample x, for example, according to ; Using the second artificial neural network (ANN-2), the second output data (AD-2) is generated (116).

5. The method according to at least one of the preceding claims, wherein, The training (104) and / or the further training (106) have the following characteristics: using a fully supervised (TV) training method (120), for example based on at least one interference-free signal ( ) and at least one interfering signal, such as a potentially interfering signal, such as the first radar signal ( ).

6. The method according to at least one of the preceding claims, wherein, The training (104) and / or the further training (106) uses a loss function (LF) (122), which is based on ,in, Characterizes the loss function (LF), where ,in, The entire uncertainty network associated with the model (MOD-VAE) is characterized, where λ represents the hyperparameters, where λ > 0, and KL represents the Kullback-Leibler divergence, which is determined according to... ,in, Characterize a specifyable distribution, such as a "target distribution," for example, the characteristic spatial distribution of a radar signal, wherein, for example, the loss function (LF) can be adapted to another specifyable target distribution (124), for example, by... Replace it with another distribution, such as a uniform distribution, for example, on an n-dimensional hypercube with a side length of 1.

7. The method according to at least one of the preceding claims, the method comprising: for example, in the case of using a simulation environment (SU) (130a), determining by means of simulation at least one training dataset (TD) of the training data (130) (130) , For example, the simulation was used to determine (130b) three different training datasets. , , ),in, For example, the different training datasets mentioned ( , , They have different complexities, for example, in terms of possible interference.

8. The method according to at least one of the preceding claims, the method comprising: generating (140) a plurality of (BATCH) pairs, for example, a batch of potentially harassed signals and undisturbed signals, based on the training data (TD); replacing (142) at least some, for example all, of the plurality of (BATCH) pairs with corresponding undisturbed signals with a specifyable probability; updating (144) trainable parameters of at least one of the two artificial neural networks (ANN-1, ANN-2) based on a loss function (LF), for example, the gradient of the loss function, wherein, The loss function (LF) evaluates the degree of agreement between the output signal of the generative model (MOD-VAE) and the undisturbed signal; and optionally, repeating (146) at least one of the following aspects: a) generating (140), or b) replacing (142), or c) updating (144), for example, repeating (146a) until, for example, a convergence criterion is met based on a validation dataset, wherein, for example, the training method or the training method (TV) has at least one of the following aspects: a) generating (140), or b) replacing (142), or c) updating (144), or repeating (146a).

9. The method according to claim 8, wherein the method comprises at least one of the following elements: a) initializing (150) the parameters of the generative model (MOD-VAE), such as weights; or b) using the training method (TV) and the first training dataset ( D 2 In the case of training (152) the trainable parameters φ of the first network architecture φ and the trainable parameters φ of the second network architecture φ; or c) using the training method (TV) and the second training dataset ( D 1 In the case of ), further train (154) the trainable parameters φ of the first network architecture φ and the trainable parameters φ of the second network architecture φ; or d) using the training method (TV) and the third training dataset ( D 0 In the case of ), further trainable parameters Ẃ of the second network architecture ẓ are trained (156), for example, only the trainable parameters of the second network architecture are further trained.

10. An apparatus (200; 200') designed to perform the method according to at least one of the preceding claims.

11. A vehicle (10) having at least one device (200) according to claim 10.

12. An apparatus (20), such as an infrastructure apparatus, such as a road infrastructure apparatus, such as a roadside unit, said apparatus having at least one device (200') according to claim 10.

13. A computer-readable storage medium (SM) comprising instructions (PRG) that, when executed by a computer (202), cause the computer to perform the method according to at least one of claims 1 to 9.

14. A computer program (PRG) comprising instructions that, when executed by a computer (202), cause the computer to perform the method according to at least one of claims 1 to 9.

15. A data carrier signal (DCS) that transmits and / or characterizes the computer program (PRG) according to claim 14.

16. The method (300) of at least one of claims 1 to 9 and / or the device (200; 200') of claim 10 and / or the vehicle (10) of claim 11 and / or the apparatus (20) of claim 12 and / or the computer-readable storage medium (SM) of claim 13 and / or the computer program (PRG) of claim 14 and / or the data carrier signal (DCS) of claim 15 for use (300) of at least one of the following elements: a) operating (301) a radar system, such as the radar system of a vehicle (10); or b) reducing (302) interference, such as interference repair; or c) improving (303) the efficiency of training; or d) reducing (304) the amount of information required for training.

Images