Method and apparatus for processing data associated with a radar system

Abstract

Methods for processing data associated with a radar system, for example for a vehicle or infrastructure facility, to reduce interference, comprising: providing a generative model of type variational autoencoder, wherein the generative model comprises a first artificial neural network and a second artificial neural network, providing training data comprising several different training datasets for training the generative model, training the first artificial neural network and the second artificial neural network with a first set of training data, and further training the second artificial neural network with a second set of training data.

Description

State of the art

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

[0002] The disclosure also relates to a device for processing data associated with a radar system. Disclosure of the invention

[0003] Some examples refer to a method, for example a computer-implemented method, for processing data associated with a radar system, for example for a vehicle or infrastructure facility, to reduce interference, comprising: providing a generative model of type variational autoencoder, wherein the generative model comprises a first artificial neural network and a second artificial neural network; providing training data comprising several different training datasets for training the generative model; training the first artificial neural network and the second artificial neural network with a first set of training data; and further training the second artificial neural network with a second set of training data. In some examples, this enables, for example,Compared to some conventional approaches, it offers a more efficient reduction of interference with radar signals from the radar system.

[0004] In some examples, the first artificial neural network is designed to have a first network architecture φ and to receive, for example, a potentially faulty, for example, interference-laden, first radar signal, for example, in the form of a spectrum, for example, in the form of a distance Doppler representation, for example, as first input data, and to generate first output data based on the first input data, which is a feature representation r associated with a probability distribution, for example, according to r = (µ r , σ r ) = φ θ (S) ∈ R n × R n , characterize, where θ represents trainable parameters, for example weights, of the first network architecture φ, where µ rrepresents a mean value of the probability distribution and where σ r represents a standard deviation of the probability distribution, where, for example, the procedure includes: receiving the first radar signal, generating the first output data using the first artificial neural network.

[0005] For example, the first artificial neural network has several layers, where at least some layers contain and / or represent at least one of the following elements: a) a linear layer, or b) a convolutional layer, or c) a nonlinear activation and / or self-attention layer, for example, a self-attention layer. Similar considerations can apply to the second artificial neural network in further examples.

[0006] In some examples, the second artificial neural network has a second network architecture ψ and is designed to operate based on one derived from the mean µ. r and the standard deviation σ r to generate second output data from a sample x characterized by a probability distribution, which characterizes a 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 S̃ = ψ η (x), where, for example, η represents trainable parameters, such as weights, of the second network architecture ψ, where, for example, the procedure includes: providing the sample x, for example according to x~N(μr,diag(σr2)), Generating the second set of output data using the second artificial neural network.

[0007] This allows for the generation of a synthetic radar signal in other examples, which, for instance, exhibits a lower degree of interference than the first radar signal.

[0008] In some examples, the training and / or further training involves the use of a fully supervised training procedure, for example based on at least one non-interference signal and at least one potentially interference signal, such as the first radar signal.

[0009] In some examples, training and / or further training indicates the use of a loss function according to L(S,S0)=‖S˜−S0‖22+λ⋅(‖μr‖2+〈σr,1〉−〈ln(σr2),1〉−n)︸KL on, where L(S, S0) characterizes the loss function, with (µ r , σ r ) = φ θ (S) where S̃ = f θ,η(S) characterizes an entire nondeterministic network associated with the model, where λ, with λ>0, characterizes a hyperparameter, where KL is a Kullback-Leibler divergence according to KL[N(μr,diag(σr2))‖N(0,diag(1))] characterized, where N(0,diag(1)) characterizes a predefined distribution, for example a “target distribution”, e.g. a distribution of the feature space for radar signals, where, for example, the loss function can be adapted to a predefined other target distribution, for example by replacing N(0,diag(1)) with another, for example uniform, distribution, for example over an n-dimensional hypercube with edge length 1.

[0010] In some examples, the procedure involves: determining at least one training data set of the training data by means of simulation, for example using a simulation environment, for example determining three different training data sets by means of simulation, where, for example, the different training data sets each have a different complexity, for example with regard to possible interference.

[0011] In some examples, the procedure is designed to: generate a number, for example a stack (e.g.(“batch”), of pairs consisting of a potentially perturbed signal and an unperturbed signal based on the training data, exchange, with a predefinable probability, of at least some, for example all, perturbed signals of the number by a respective unperturbed signal, update trainable parameters of at least one of the two artificial neural networks based on a gradient of, for example, the, loss function, wherein the loss function evaluates a match between an output signal of the generative model and an unperturbed signal, and optionally repeat at least one of the aspects a) generating, or b) exchanging, or c) updating, for example repeating until a convergence criterion, for example based on a validation dataset, is satisfied, wherein, for example, a orThe training procedure must include at least one of the following aspects: a) generating, b) exchanging, c) updating, or repeating.

[0012] For example, the procedure includes at least one of the following elements: a) initializing parameters, such as weights, of the generative model; or b) training, using the training procedure and a first training dataset, the trainable parameters θ of the first network architecture φ and the trainable parameters η of the second network architecture ψ; or c) further training, using the training procedure and a second training dataset, of the trainable parameters θ of the first network architecture φ and the trainable parameters η of the second network architecture ψ; or d) further training, using the training procedure and a third training dataset, such as alone, of the trainable parameters η of the second network architecture ψ.

[0013] Further examples relate to a device designed to carry out the method according to the disclosure.

[0014] Further examples relate to a vehicle having at least one device according to the disclosure.

[0015] Further examples relate to a facility, for example infrastructure facility, for example road infrastructure facility, for example roadside unit, comprising at least one device according to the disclosure.

[0016] Other examples relate to a computer-readable storage medium comprising instructions which, when executed by a computer, cause it to perform the procedure according to the disclosure.

[0017] Other examples relate to a computer program, comprising commands which, when the program is executed by a computer, cause it to perform the procedure according to the disclosure.

[0018] Further examples relate to a data carrier signal that transmits and / or characterizes the computer program according to the disclosure.

[0019] Further examples relate to the use of the method according to the disclosure and / or the device according to the disclosure and / or the vehicle according to the disclosure and / or the installation according to the disclosure and / or the computer-readable storage medium according to the disclosure and / or the computer program according to the disclosure and / or the data carrier signal according to the disclosure for at least one of the following elements: a) operating a radar system, for example for a vehicle and / or for an installation, for example stationary, or b) reducing interference, for example interference healing, or c) increasing the efficiency of a training, or d) reducing the amount of information required for a training.

[0020] Further features, applications, and advantages will become apparent from the following description of examples illustrated in the figures of the drawing. All described or illustrated features, individually or in any combination, constitute the subject matter of the disclosure, irrespective of their aggregation in the claims or their cross-reference, and irrespective of their formulation or representation in the description or in the drawing.

[0021] The drawing shows: Fig. 1. A simplified flowchart (schematical). Fig. 2. A simplified block diagram (schematically). Fig. 3. A simplified flowchart (schematically). Fig. 4. A simplified flowchart (schematically). Fig. 5. A simplified flowchart (schematically). Fig. 6. A simplified flowchart (schematically). Fig. 7. A simplified flowchart (schematically). Fig. 8. A simplified block diagram (schematically). Fig. 9 schematic examples of uses.

[0022] Some examples, e.g. Fig. 1, Fig. 2, refer to a method, for example a computer-implemented method, for processing with a radar system 12, 22 ( Fig. 2), for example for a vehicle 10 or an infrastructure facility 20, associated data, to reduce interference, showing: providing 100 ( Fig. 1) a generative model MOD-VAE of type variational autoencoder, wherein the generative model MOD-VAE comprises a first artificial neural network ANN-1 and a second artificial neural network ANN-2, providing 102 several different training datasets D0, D1, D2 for training the generative model MOD-VAE, training 104 the first artificial neural network ANN-1 and the second artificial neural network ANN-2 with a first set D1, D2 of the training data TD, further training 106, for example after training 104, of the second artificial neural network ANN-2, for example only of the second artificial neural network ANN-2, with a second set D0 of the training data TD. In some examples, this allows for a more efficient reduction of interference with radar signals of radar system 12, 22 compared to some conventional approaches.

[0023] In some examples, Fig. 2, the radar system 12 can be, for example, an automotive radar system 12 for use in vehicles 10, 10a, for example for detecting objects such as other vehicles 10a and / or obstacles in an environment UM of the vehicle 10.

[0024] A VAE (variational autoencoder) model is a generative model that can combine the properties of an autoencoder with probabilistic modeling. In some examples, a VAE consists of two parts: an encoder that transforms input data into a relatively low-dimensional, latent representation described as a probability distribution, and a decoder that, for example, reconstructs the original data from the latent representation. In some examples, a VAE model can be used to generate new data that are, for example, similar to the training data. Here, the generative model MOD-VAE is used to eliminate unwanted interference from, for example, radar system 12, 22 ( Fig. 2) to remove the received radar signal, for example as part of so-called "interference healing", which can improve subsequent evaluation of the interference-reduced radar signal.

[0025] In some examples, Fig. 2, at least some aspects of the method according to the disclosure can be carried out by a device 200, 200' which may, for example, be assigned to a respective radar system 12, 22, for example, be integrated therein (not shown).

[0026] In some examples, Fig. 2, Fig. 3, it is provided that the first artificial neural network ANN-1 has a first network architecture φ and is designed to receive, for example, a potentially erroneous, for example, interference-laden, first radar signal S (e.g. received from the radar system 12), for example in the form of a spectrum, for example in the form of a distance Doppler representation, for example as first input data ED-1, and to generate, based on the first input data ED-1, first output data AD-1, which is a feature representation r associated with a probability distribution, for example according to r = (µ r , σ r ) = φ θ (S) ∈ R n × R n , characterize, where θ represents trainable parameters, for example weights, of the first network architecture φ, where µ r represents a mean value of the probability distribution and where σ rrepresents a standard deviation of the probability distribution. In other words, the procedure can, in some examples, Fig. 3, exhibit: Receiving 110 of the first radar signal S, Generating 112, using the first artificial neural network ANN-1, the first output data AD-1.

[0027] Example journey, Fig. 2. The first artificial neural network, ANN-1, has several layers, where at least some layers contain and / or represent at least one of the following elements: a) a linear layer, or b) a convolutional layer, or c) a nonlinear activation and / or self-attention layer, for example, a self-attention layer. Similar considerations may apply to the second artificial neural network, ANN-2, in other examples. Generally, different combinations of the aforementioned layers are possible in some examples to implement the first and second network architectures.

[0028] In some examples, Fig. 2, Fig. 3, the second artificial neural network ANN-2 has a second network architecture ψ and is designed to operate based on one derived from the mean µ r and the standard deviation σ r to generate a second output data AD-2 from a sample x characterized by a probability distribution, which characterizes a second radar signal S̃, for example in the form of a spectrum, for example in the form of a distance Doppler representation, for example according to S̃ = ψ η (x), where, for example, η represents trainable parameters, such as weights, of the second network architecture ψ. In other words, in some examples, Fig. 3. The procedure includes: providing 114 of the sample x, for example according to x~N(μr,diag(σr2)), Generate 116, using the second artificial neural network ANN-2, the second output data AD-2. This allows a synthetic S̃ radar signal to be generated in further examples, which, for example, has a lower degree of interference than the first radar signal, but which, for example, has a comparable information content with regard to useful radar signal information, e.g. for object detection.

[0029] In some examples, Fig. 4, the training shows 104 ( Fig. 1) and / or further training 106 using 120 ( Fig. 4) a fully supervised training procedure TV, for example based on at least one non-interference signal (or radar signal) S0 and at least one potentially interference-prone signal, for example the first radar signal S.

[0030] In some examples, Fig. 4, the training shows 104 ( Fig. 1) and / or further training 106 using 122 a loss function according to L(S,S0)=‖S˜−S0‖22+λ⋅(‖μr‖2+〈σr,1〉−〈ln(σr2),1〉−n)︸KL on, where L(S, S0) characterizes the loss function LF, with (µ r ,σ r ) = (φ θ (S) where S̃ = f θ,η (S) an entire model MOD-VAE ( Fig. 2) an associated nondeterministic network characterized, where λ, with λ>0, characterizes a hyperparameter, where KL is a Kullback-Leibler divergence according to KL[N(μr,diag(σr2))‖N(0,diag(1))] characterized, where N(0,diag(1)) characterizes a predefinable distribution, for example a “target distribution”, e.g. a distribution of the feature space for radar signals, where, for example, the loss function LF can be adapted to a predefinable other target distribution, see optional block 124 according to Fig. 4, for example by replacing N(0,diag(1)) with another, for example uniform, distribution, for example over an n-dimensional hypercube with edge length 1, which can lead to a fitted loss function LF'.

[0031] The optional block 126 according to Fig. 4 symbolizes an optional execution of the training procedure TV for the model MOD-VAE, for example for at least one of its networks ANN-1, ANN-2, using the loss function LF or the adapted loss function LF'.

[0032] In some examples, Fig. 5, the procedure is as follows: Determine at least 130 of a training data set D1, D2 of the training data TD ( Fig. 2. By means of simulation, for example using 130a a simulation environment SU, for example determining 130b three different training datasets D0, D1, D2 by means of simulation, wherein, for example, the different training datasets D0, D1, D2 each have a different complexity, for example with regard to possible interference. The optional block 132 according to Fig. 5 symbolizes an optional execution of the training procedure TV based on the respective training data sets.

[0033] Further examples and aspects relating to the determination of the training data TD are described below. For example, the simulation environment SU ( Fig. 5) various devices, e.g. vehicles 10, 10a, in a complex environment UM ( Fig. 2) can be placed, and both clean (i.e., without interference) and disturbed (i.e., with non-zero interference) signals, i.e., radar signals, can be detected, for example recorded.

[0034] In some examples, a clean (interference-free) signal can be determined, for example recorded, in two different ways: 1. Deactivating the radar devices or radar systems of all interfering vehicles 10a, or 2. as above aspect 1, but additionally ignoring the effects of clutter and multipath effects (e.g. directly in a simulation).

[0035] In some examples, at least two different data sets, e.g. D1 and D2, can be created using the approaches described above.

[0036] In some examples, an additional dataset D0 is created, which uses or replicates real-world data. In some examples, two types of signals can be considered: 1. Each vehicle 10, 10a uses the same central frequency, for example, the center frequency, e.g., in odd-numbered cycles; 2. Each vehicle 10, 10a uses a different central frequency (e.g., in even-numbered cycles). Since a road scenario is pre-planned for determining the datasets in some examples, a frequency plan with precisely synchronized timing can be executed, thus enabling efficient training. Fig. 1) usable data sets can be obtained. For example, with the approach described above, clean (interference-free) signals can be obtained for even cycles and noisy (interference-laden) signals for odd cycles.

[0037] In some examples, the training data TD can be organized or provided in the form of three datasets, e.g., with increasing complexity: · D2 is the simplest dataset, · D1 is more complex, e.g., due to self-interference, · D0 is even more complex, as, for example, the clean and noisy signals may have a slight time delay.

[0038] In some examples, Fig. 6, the procedure is provided to: Generate 140 a number, for example a batch, BACTH of pairs consisting of a potentially perturbed signal, for example radar signal, and an unperturbed signal, for example radar signal, based on the training data TD; replace 142, with a predefinable probability p, at least some, for example all, perturbed signals of the number BATCH with a respective unperturbed signal; update 144 of trainable parameters θ, η at least one of the two artificial neural networks ANN-1, ANN-2 based on a gradient of, for example the, loss function LF ( Fig. 4), where the loss function LF evaluates the agreement between an output signal of the generative model MOD-VAE and an unperturbed signal, and optionally repeat 146 ( Fig. 6) at least one of the aspects a) Generate 140, or b) Replace 142, or c) Update 144, for example Repeat 146a, until a convergence criterion, for example based on a validation data set, is satisfied, wherein, for example, a training procedure TV has at least one of the aspects Generate 140, or b) Replace 142, or c) Update 144, or Repeat 146, 146a.

[0039] In some examples, probability-dependent exchange, for example substitution, 142 ( Fig. 6) of the disturbed signal by the clean signal to enable the MOD-VAE model to learn features that are not focused on the disturbances, but e.g. on the underlying clean signal.

[0040] For example, Fig. 7, the procedure includes at least one of the following elements: a) Initializing 150 parameters, for example weights, θ, η of the generative model MOD-VAE, or b) Training 152, using the training procedure TV (szB Fig. 6) and a first training dataset D2, the trainable parameters θ of the first network architecture φ and the trainable parameters η of the second network architecture ψ, or c) further training 154, using the training procedure TV and a second training dataset D1, the trainable parameters θ of the first network architecture φ and the trainable parameters η of the second network architecture ψ, or d) further training 156, using the training procedure TV and a third training dataset D0, for example alone, the trainable parameters η of the second network architecture ψ.

[0041] In other words, training with the three training datasets D0, D1, D2 can be relatively similar, for example, each following the procedure according to Fig. 6, where, for example, the respective training datasets D0, D1, D2 are each different from each other, as described above.

[0042] In some examples, Fig. 7. It is important that in block 156 only the second network, ANN-2, is trained, and not, for example, the first network, ANN-1. This helps in some examples to use the same features as in the simulation, since these are, for example, ideal features. For instance, only the interpretation of these features will improve based on the new real-world data.

[0043] Further examples, Fig. References 8 to a device 200, 200' configured for carrying out the method according to the disclosure. For example, the device 200, 200' is usable for at least one vehicle 10, 10a or its radar system 12 and / or for at least one, for example stationary, installation 20, such as a roadside unit, or its radar system 22.

[0044] In some examples, Fig. 8, it is provided that the device 200, 200' comprises: a computing device (“computer”) 202 having at least one computing core 202a, a storage device 204 associated with the computing device 202 for at least temporary storage of 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, for example, data TD), b) computer program PRG, for example, for executing the method according to the disclosure.

[0045] For further examples, Fig. 8, the memory device 204 includes volatile memory (e.g., RAM) 204a, and / or non-volatile (NVM) memory (e.g., Flash EEPROM) 204b, or a combination thereof or with other memory types not explicitly mentioned.

[0046] Further examples, Fig. 8, refer to a computer-readable storage medium SM, comprising instructions PRG which, when executed by a computer 202, cause it to execute the procedure according to the disclosure.

[0047] Further examples, Fig. 8, refer to a computer program PRG, comprising commands which, when the program PRG is executed by a computer 202, cause it to execute the procedure according to the disclosure.

[0048] Further examples, Fig. References to reference 8 refer to a data carrier signal DCS, which characterizes and / or transmits the computer program PRG according to the disclosure. The data carrier signal DCS can be transmitted (e.g., sent and / or received) via an optional data interface 206 of the device 200, 200'.

[0049] Further examples, Fig. 2, refer to a vehicle 10 having at least one device 200 according to the disclosure.

[0050] Further examples, Fig. 2, refer to a facility 20, for example infrastructure facility, for example roadside facility, for example roadside unit, comprising at least one device 200' according to the disclosure.

[0051] Further aspects and examples are described below, which – in the case of further examples – can each be combined individually or in any combination with at least one of the aspects and / or examples described above.

[0052] The principle according to the disclosure can be advantageously used to reduce or “cure” interference, e.g., of received radar signals (e.g., “signal healing”). Radar is, for example, an important sensor for solving perception tasks, such as for vehicles 10, 10a ( Fig. 2), for example, for autonomous driving. Furthermore, radar can be used, for example, to control the beam shaping of a communication base station (not shown), e.g., by acquiring knowledge from the surrounding environment ( ). Fig. 2), e.g., detectable by means of at least one radar system 12, 22, is included.

[0053] In some examples, signal healing might involve estimating a clean signal by removing aspects from the signal to be healed that originate, for example, from other transmitters or other sources of interference. For this, the generative model MOD-VAE ( Fig. 1, Fig. 2) can be used, whose networks ANN-1 and ANN-2 can be trained in further examples as described above. Signal healing according to the disclosure is useful because, in real-world scenarios, interference, for example, can never be completely avoided.

[0054] The use of the generative model MOD-VAE of the VAE type makes it possible in some examples to create comparatively meaningful features, for example, for signal healing. Furthermore, according to the disclosure, the principle allows for a combination of several training stages, e.g., in the sense of pre-training (see blocks 152, 154 according to...). Fig. 7) and subsequent fine-tuning (e.g., block 156 according to Fig. 7).

[0055] In some examples, aspects of the generative model MOD-VAE can be pre-conceived using simulation data, e.g., obtained using the simulation environment SU ( Fig. 5) are trained, which can be used, for example, to generate comparatively complex urban scenarios with various multiple disturbances. In some examples, for example, after preliminary training (152, 154), only the second part, ANN-2, of the MOD-VAU model or network can be used, for example, for fine-tuning with real data. This reduces the data requirement in some examples, for example, the need for real data, without requiring the MOD-VAE model to be adapted to a potentially smaller dataset of real data.

[0056] The principle according to the disclosure produces comparatively meaningful representations and reduces, for example, overfitting. As a result, the approach according to the disclosure is comparatively well adapted to the real problem of radar interference and can, for example, generate a purer signal than some conventional approaches. This helps, for example, with perception tasks (e.g., object or road user detection), which can increase road safety.

[0057] In general, the principle according to the disclosure can also be applied to image analysis (e.g., digital images, video, radar, lidar, ultrasound, motion, thermal imaging) or video analysis or audio analysis and is not limited to radar signals.

[0058] Further examples, Fig.9, refer to a use 300 of the method according to the disclosure and / or the device 200, 200' according to the disclosure and / or the vehicle 10 according to the disclosure and / or the installation 20 according to the disclosure and / or the computer-readable storage medium SM according to the disclosure and / or the computer program PRG according to the disclosure and / or the data carrier signal DCS according to the disclosure for at least one of the following elements: a) operating 301 a radar system 12, 22, for example for a vehicle 10 and / or for an installation 20, for example stationary, or b) reducing 302 an interference, for example interference healing, or c) increasing 303 an efficiency of a training 104, 106, or d) reducing 304 an amount of information required for a training, for example real data.

Claims

[1] Method, for example a computer-implemented method, for processing data associated with a radar system (12, 22), for example for a vehicle (10) or an infrastructure facility (20), for reducing interference, comprising: providing (100) a generative model (MOD-VAE) of the variational autoencoder type, wherein the generative model (MOD-VAE) comprises a first artificial neural network (ANN-1) and a second artificial neural network (ANN-2), providing (102) training data (TD) comprising several different training data sets (D0, D1, D2) for training the generative model (MOD-VAE), training (104) the first artificial neural network (ANN-1) and the second artificial neural network (ANN-2) with a first set (D1, D2) of the training data (TD), further training (106) the second artificial neural network (ANN-2) with a second set (D0) of training data (TD). [2] Method according to claim 1, wherein the first artificial neural network (ANN-1) has a first network architecture φ and is configured to receive, for example, a potentially erroneous, for example interference-laden, first radar signal (S), for example in the form of a spectrum, for example in the form of a distance Doppler representation, for example as first input data (ED-1), and to generate, based on the first input data (ED-1), first output data (AD-1) which form a feature representation r associated with a probability distribution, for example according to r = (µ r , σ r ) = φ θ (S) ∈ R n × R n , characterize, where θ represents trainable parameters, for example weights, of the first network architecture φ, where µ r represents a mean value of the probability distribution and where σ rrepresents a standard deviation of the probability distribution, where, for example, the procedure includes: receiving (110) the first radar signal (S), generating (112) the first output data (AD-1) using the first artificial neural network (ANN-1). [3] Method according to claim 2, wherein the first artificial neural network (ANN-1) has multiple layers, wherein at least some layers have and / or represent at least one of the following elements: a) linear layer, or b) convolutional layer, or c) nonlinear activation and / or self-attention layer, for example self-attention layer. [4] Method according to at least one of claims 2 to 3, wherein the second artificial neural network (ANN-2) has a second network architecture ψ training datasets D0, D1, D2 and is configured to perform a training dataset based on a training dataset derived from the mean value µ r and the standard deviation σ rto generate a second set of output data (AD-2) from a sample drawn with a characterized probability distribution, which characterizes a second radar signal (S̃, for example in the form of a spectrum, for example in the form of a distance Doppler representation, for example according to S̃ = ψ η (x), where, for example, η represents trainable parameters, such as weights, of the second network architecture ψ, where, for example, the procedure includes: providing (114) the sample x, for example according to x~N(μr,diag(σr2)), Generating (116) the second output data (AD-2) using the second artificial neural network (ANN-2). [5] Method according to at least one of the preceding claims, wherein the training (104) and / or further training (106) comprises the use (120) of a fully supervised training method (TV), for example based on at least one non-interference signal (S0) and at least one potentially interference signal, for example the first radar signal (S). [6] Method according to at least one of the preceding claims, wherein the training (104) and / or further training (106) involves the use (122) of a loss function (LF) according to L(S,S0)=‖S˜−S0‖22+λ⋅(‖μr‖2+〈σr,1〉−〈ln(σr2),1〉−n)︸KL exhibits, where L(S,S0) characterizes the loss function (LF), with (µ r , σ r ) = φ θ (S) where S̃ = f θ,η(S) characterizes an entire nondeterministic network associated with the model (MOD-VAE), where λ, with λ>0, characterizes a hyperparameter, where KL is a Kullback-Leibler divergence according to KL[N(μr,diag(σr2))‖N(0,diag(1))] characterized, where N(0,diag(1)) characterizes a predefinable distribution, for example “target distribution”, e.g. a distribution of the feature space for radar signals, where, for example, the loss function (LF) is adaptable to a predefinable other target distribution (124), for example by replacing N(0,diag(1)) with another, for example uniform, distribution, for example over an n-dimensional hypercube with edge length 1. [7] Method according to at least one of the preceding claims, comprising: determining (130) at least one training data set (D1, D2) of the training data (TD) by means of simulation, for example using (130a) a simulation environment (SU), for example determining (130b) three different training data sets (D0, D1, D2) by means of simulation, wherein, for example, the different training data sets (D0, D1, D2) each have a different complexity, for example with respect to a possible interference. [8] A method according to at least one of the preceding claims, comprising: generating (140) a number, for example a batch, (BATCH) of pairs consisting of a potentially perturbed signal and an unperturbed signal based on the training data (TD), replacing (142), with a predefinable probability, at least some, for example all, perturbed signals of the number (BATCH) with a respective unperturbed signal, updating (144) trainable parameters of at least one of the two artificial neural networks (ANN-1, ANN-2) based on a gradient of, for example, a loss function (LF), wherein the loss function (LF) evaluates a match between an output signal of the generative model (MOD-VAE) and an unperturbed signal, and optionally repeating (146) at least one of the aspects a) generating (140), or b) replacing (142), or c) updating (144), for example repeating (146a) as long as,until a convergence criterion, for example based on a validation dataset, is met, wherein, for example, a training procedure (TV) has at least one of the aspects of generating (140), or b) replacing (142), or c) updating (144), or repeating (146a). [9] Method according to claim 8, comprising at least one of the following elements: a) initializing (150) parameters, for example weights, of the generative model (MOD-VAE), or b) training (152), using the training method (TV) and a first training data set (D2), the trainable parameters θ of the first network architecture φ and the trainable parameters η of the second network architecture ψ, or c) further training (154), using the training method (TV) and a second training data set (D1), of the trainable parameters θ of the first network architecture φ and the trainable parameters η of the second network architecture ψ, or d) further training (156), using the training method (TV) and a third training data set (D0), for example alone, of the trainable parameters η of the second network architecture ip. [10] Device (200; 200') designed to carry out the method according to at least one of the preceding claims. [11] Vehicle (10) comprising at least one device (200) according to claim 10. [12] Equipment (20), for example infrastructure equipment, for example roadside infrastructure equipment, for example roadside unit, comprising at least one device (200') according to claim 10. [13] Computer-readable storage medium (SM) comprising instructions (PRG) which, when executed by a computer (202), cause it to execute the method according to at least one of claims 1 to 9. [14] Computer program (PRG) comprising instructions which, when the program (PRG) is executed by a computer (202), cause it to execute the method according to at least one of claims 1 to 9. [15] Data carrier signal (DCS) that transmits and / or characterizes the computer program (PRG) according to claim 14. [16] Use (300) of the method according to at least one of claims 1 to 9 and / or the device (200; 200') according to claim 10 and / or the vehicle (10) according to claim 11 and / or the device (20) according to claim 12 and / or the computer-readable storage medium (SM) according to claim 13 and / or the computer program (PRG) according to claim 14 and / or the data carrier signal (DCS) according to claim 15 for at least one of the following elements: a) operating (301) a radar system, for example for a vehicle (10), or b) reducing (302) interference, for example interference healing, or c) increasing (303) the efficiency of a training, or d) reducing (304) the amount of information required for a training.

Images