Computer-implemented method for reconstructing at least one test bench setup

Abstract

The invention describes a computer-implemented method (100) for reconstructing at least one test bench setup for determining at least one acoustic measurement signal, wherein the method (100) uses at least one machine learning algorithm that determines at least one output signal from at least one input signal using at least one latent variable (18), comprising at least the following steps: receiving (102) at least one input signal; determining (104) at least one latent variable (18) using the machine learning algorithm; and providing (106) at least one output signal of the machine learning algorithm, which the machine learning algorithm determines from the input signal using the latent variable (18);wherein the input signal is designed as at least one acoustic measurement signal of at least one component, wherein the machine learning algorithm reconstructs at least one test setup for determining the acoustic measurement signal (10) from the at least one input signal, and the output signal is designed as at least one image signal (16) that displays the at least one setup. The computer-implemented method (100) provides high accuracy and ease of use.

Description

[0001] The invention relates to a computer-implemented method for reconstructing at least one test rig setup.

[0002] When developing components, especially those that generate noise, test benches equipped with microphones for recording the sounds can be used. This allows the noise generated by components, such as motors or electric motors, to be taken into account during the development process. In other words, the components can be designed in such a way as to avoid unwanted noise or to promote a specific noise level.

[0003] Subsequent inspection of the components may require reconstructing the test rig used for their development. This can, however, necessitate a time-consuming search for the relevant documentation. Alternatively, the required documentation may no longer be available.

[0004] The idea of ​​identifying rooms based on audio data, also using machine learning, is known from "ZHAO, Hong; MALIK, Hafiz: Audio recording location identification using acoustic environment signature. In: IEEE Transactions on Information Forensics and Security, 2013, Vol. 8, No. 11, pp. 1746-1759" and "AU-YEUNG, Justin; BANAVAR, Mahesh K.; VANITHA, M.: Room Classification using Acoustic Signals. In: 2020 International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE). IEEE, 2020. pp. 1-4". "AHMED, Waqas W., et al.: A generative deep learning approach for shape recognition of arbitrary objects from phaseless acoustic scattering data. In: Advanced Intelligent Systems, 2023, Vol. 5, No. 5, pp. 1-11." deals with the determination of object geometries using acoustic measurement signals and AI methods and latent variables. “CASTELLINI, Paolo, et al.“A neural network based microphone array approach to grid-less noise source localization. In: Applied Acoustics, 2021, Vol. 177, pp. 1-10.” describes the localization of sound sources using acoustic measurement signals and AI methods.

[0005] The object of the invention is therefore to provide a computer-implemented method for reconstructing at least one test rig setup that has high accuracy and is easy to use.

[0006] The problem is solved by the features of the independent claims. Advantageous further developments are the subject of the dependent claims and the following description.

[0007] A computer-implemented method for reconstructing at least one test bench setup for determining at least one acoustic measurement signal, wherein the method uses at least one machine learning algorithm that determines at least one output signal from at least one input signal using at least one latent variable, comprising at least the following steps: receiving at least one input signal; determining at least one latent variable using the machine learning algorithm; and providing at least one output signal of the machine learning algorithm, which the machine learning algorithm determines from the input signal using the latent variable;wherein according to the invention it is provided that the input signal is designed as at least one acoustic measurement signal of at least one component, in particular a motor, preferably an electric motor, wherein the learning algorithm reconstructs at least one setup of a test bench for determining the acoustic measurement signal from the at least one input signal and the output signal is designed as at least one image signal that displays the at least one setup.;

[0008] The computer-implemented method described above allows for the reconstruction of the test setup used to acquire the acoustic measurement signal from the corresponding component simply by providing that signal. The reconstruction represents the most probable setup determined by the machine learning algorithm. The image signal can display the positions of the acoustic sensors and the measured component. By considering latent variables, which are typically unknown to a user of the computer-implemented method and do not necessarily become known during the execution of the method, the probability of reconstructing a correct setup can be significantly increased. Therefore, latent variables are only used in a latent manner within the method. At least one latent variable can be determined by the machine learning algorithm.Furthermore, latent variables can include, for example, an assumed number of sensors, particularly microphones, an assumed level of ambient noise, or assumed distances between the sensors and the measured component. Alternatively or additionally, latent variables can be predicted by a predictor algorithm. The computer-implemented method is easy to use, as only the acoustic measurement signal needs to be provided as input. Furthermore, the computer-implemented method exhibits high accuracy because the use of latent variables provides additional information that can be used to determine the output signal.

[0009] According to some embodiments, it is conceivable that the image signal can additionally display text modules describing the structure.

[0010] Using descriptive text modules, a user can receive annotations about the reconstructed structure. These descriptive text modules, in combination with the displayed image data of the structure, can lead to a more comprehensive understanding of the reconstructed structure.

[0011] According to some embodiments, it is conceivable that the learning algorithm can have at least one prediction function for the output signal based on the latent variable, wherein the learning algorithm can preferably have at least one Joint Embedding Predictive Architecture, at least one Image Joint Embedding Predictive Architecture and / or at least one Latent Variable Generative Energy-Based Model.

[0012] The prediction function can, for example, be a predictor that can predict at least one output signal based on at least one latent variable. That is, the predictor can, for example, predict the output signal from a combination of the input signal and the latent variable.

[0013] Furthermore, the predictor can be designed in such a way that it can provide an embedded output signal from an embedded input signal, which, for example, was processed by means of an autoencoder of the learning algorithm, and which can be converted into an image signal, for example, using a decoder of the learning algorithm.

[0014] According to some embodiments, it is conceivable that the learning algorithm can have at least one Latent Variable Energy-Based Model.

[0015] According to some embodiments, it is conceivable that the latent variable may comprise at least a number of sensors, at least a distance between the component and an acoustic sensor and / or information about ambient noise, in particular a level of ambient noise.

[0016] According to some embodiments, it is conceivable that the latent variable can lie within a range of values ​​that is usable for setting up a test bench to determine at least one acoustic measurement signal.

[0017] For example, the number of sensors used as latent variables can be limited to a range that makes engineering sense for a test bench. Similarly, the distances and / or information about ambient noise, such as its level, can be limited to an engineeringally sensible range.

[0018] According to some embodiments, it is conceivable that the learning algorithm, before performing the steps of: receiving at least one input signal; determining at least one latent variable; and providing at least one output signal of the learning algorithm, may have at least the following steps: receiving at least one training signal, which is designed as at least one acoustic measurement signal of at least one component, and at least one target output signal associated with the training signal, which is designed as at least one image signal, preferably displaying at least one text element describing the image content of the image signal, wherein the target output signal displays at least part of a test bench setup with which the training signal was determined; determining at least one latent variable;Provide at least one output signal of the machine learning algorithm, which the machine learning algorithm derives from the training signal using the latent variable; check whether a cost value of a cost function derived from the output signal and the target output signal lies within a target range; and if the cost value lies outside the target range: modify the machine learning algorithm and repeat the two previous steps with the modified machine learning algorithm.

[0019] In this embodiment, the machine learning algorithm can be trained to derive an image signal of the corresponding test bench setup from the acoustic measurement signal. Pairs of acoustic measurement data and image data, showing at least part of the setup, can be used as training data. The goal of the training is to train the machine learning algorithm to generate the image data of a pair of acoustic measurement data as an output signal with a high probability. The latent variables, which were explained above, are also used during training. Their selection by the machine learning algorithm for the operation described above can be influenced by the training process.

[0020] According to some embodiments, it is conceivable that an energy function can be determined from the training signal, the latent variable and the target output signal, such that a lower energy value is assigned to a first combination based on the training signal and the target output signal associated with the training signal, and a higher energy value is assigned to a second combination based on the training signal and a target output signal not associated with the training signal.

[0021] The energy function thus provides a means by which previously unknown acoustic measurement signals can exhibit low energy values ​​with a plurality of possible output signals if these output signals indicate a setup that is highly likely to match the unknown acoustic measurement signal. Output signals indicating setups that are highly unlikely to match the unknown acoustic measurement signal can then exhibit a higher energy value.

[0022] According to some embodiments, it is conceivable that the learning algorithm can provide the at least one output signal taking into account the energy function, wherein the provided output signal in combination with the input signal and the latent variable represents a minimum of the energy function.

[0023] The adaptive algorithm therefore selects the output signal from among several suitable output signals that has a minimum energy value when using the energy function.

[0024] According to some embodiments, it is conceivable that the acoustic measurement signal can be converted into an image signal that displays a spectrogram from the acoustic measurement signal.

[0025] In this way, the machine learning algorithm can only process image data. The text modules described above can also be provided as image data, either alternatively or additionally.

[0026] According to some embodiments, it is conceivable that the learning algorithm can be trained before performing the steps of: receiving at least one input signal; determining at least one latent variable and providing at least one output signal of the learning algorithm, in such a way that the learning algorithm reconstructs complete image data, in particular of the spectrogram or the setup of the test bench, from subsections of image data, in particular of a spectrogram or of the setup of the test bench.

[0027] This allows for pre-training, in which, for example, a module or part of the machine learning algorithm is trained to reconstruct abstracted data. This means that the machine learning algorithm learns, for instance, to infer the whole image from parts of an image. Pre-training increases the probability that the machine learning algorithm will provide a correct reconstruction.

[0028] The invention further relates to a computer program product comprising instructions which, when the program is executed by a computer, cause it to perform the steps of the method according to the preceding description.

[0029] The advantages, effects, and further developments of the computer program product result from the advantages, effects, and further developments of the method described above. Therefore, reference is made to the preceding description in this regard. A computer program product can be understood, for example, as a data carrier on which a computer program element is stored, containing instructions executable by a computer. Alternatively or additionally, a computer program product can also be understood, for example, as a permanent or volatile data storage medium, such as flash memory or main memory, that contains the computer program element. However, this does not exclude other types of data storage media that contain the computer program element.

[0030] The invention is described below with reference to an exemplary embodiment and the accompanying drawing. The drawing shows: Fig. 1. A flowchart of the process; Fig. 2 a schematic representation of an exemplary implementation of the learning algorithm; Fig. 3 a schematic representation of the conversion of acoustic measurement data into image data; Fig. 4 a schematic representation of the pre-training of the learning algorithm; Fig. 5 a schematic representation for creating the energy function; and Fig. 6 A schematic representation of the training of the learning algorithm.

[0031] The computer-implemented procedure will be referred to in the following as a whole by the reference symbol 100, as shown in the flowchart according to Fig. 1 is shown.

[0032] Method 100 is designed to reconstruct at least one test bench setup for determining at least one acoustic measurement signal. Method 100 uses at least one machine learning algorithm. The algorithm determines at least one output signal from an input signal. In determining the output signal, the algorithm uses at least one latent variable 18, as described in Fig. 2 shown.

[0033] A latent variable 18, for example, is a variable that is only used within the algorithm and does not necessarily have to be transmitted to a user or provided by them.

[0034] The latent variable can, for example, represent a probable number of sensors, a probable distance between the component whose acoustic measurement signals are used as the input signal and the sensors that recorded the measurement signals, and / or a probable level of ambient noise. The latent variable can be estimated using a method that ensures a high probability of the latent variable corresponding to the measurement signal. Alternatively or additionally, the latent variable can be randomly determined from an engineering-sensible range of values. This means that the range of values ​​for the setup of a test bench can be used to determine at least one acoustic measurement signal.

[0035] The latent variable can thus, for example, parameterize a set of plausible predictions.

[0036] The machine learning algorithm can, for example, include a Latent Variable Energy-Based Model, a Joint Embedding Predictive Architecture, an Image Joint Embedding Predictive Architecture, and / or a Latent Variable Generative Energy-Based Model. In some of these embodiments, the machine learning algorithm can include a predictor 14 or a prediction function for the output signal, as exemplified in Fig. 2 is shown.

[0037] Method 100 can include a step 102 in which at least one input signal can be received. The input signal is designed as an acoustic measurement signal 10. For example, the acoustic measurement signal 10 can indicate noise from a running engine or an electric motor.

[0038] In an optional step 120, the acoustic measurement signal 10 can be converted into an image signal. For example, a spectrogram can be generated from the acoustic measurement signal 10. This is exemplified in Fig. 3 shown.

[0039] The acoustic measurement signal 10 or the image signal which displays the spectrogram can be converted into an embedded audio signal by means of an autoencoder 12 of the learning algorithm.

[0040] In a further step 104, at least one latent variable 18 can be determined. The latent variable 18 can be determined by the machine learning algorithm.

[0041] Using the latent variable 18 and the received input signal, which may be an embedded audio signal, the machine learning algorithm can determine an output signal. This output signal can be determined, for example, using the predictor 14 explained above, or the prediction function based on the at least one audio signal (which may be embedded) and the at least one latent variable.

[0042] The output signal is designed as at least one image signal 16, which displays the setup that was most likely used to determine the acoustic measurement signal 10 of the input signal. This allows the machine learning algorithm to reconstruct, from the input signal and the latent variable 18, the setup with which the acoustic measurement signal 10 can most likely be determined when the corresponding test bench is implemented.

[0043] In a further step, at least one output signal of the machine learning algorithm can be provided. This can be done, for example, via an output interface. The output interface can, for example, be connected to a screen on which a user can view the image data.

[0044] The image signal may also contain text modules that describe the structure in text form.

[0045] The following describes optional steps of procedure 100.

[0046] Before the steps described above, training and, if necessary, pre-training of the learning algorithm can take place.

[0047] In step 122, the machine learning algorithm can receive optional pre-training. In this pre-training, the machine learning algorithm can be trained such that input signals representing subsections 32 and 34 of an image 30 are converted by the machine learning algorithm into an output signal indicating a reconstruction 56 of the image 30 based on subsections 32 and 42. It should be noted that the reconstruction 56 is only highly probable, not absolutely certain, to be the reconstruction of the image. This definition of reconstruction applies analogously to the use of the term "reconstruction" in relation to other contexts in this disclosure.

[0048] An example is the pre-training in Fig. Figure 3 illustrates this process. Subsections 32 and 34 are converted into embedded image signals 38 and 40 using a third encoder 36. These embedded image signals 38 and 40 can be determined by a neural network, which identifies embedded reconstructed image sections 42-44 and transmits them, along with the embedded image signals 38 and 40, to a decoder 48. The decoder 48 can convert the embedded signals 38-44 into image signals displaying subsections 32 and 34, as well as the reconstructed subsections 50-54. A function that calculates the mean square error between the individual reconstructed subsections provides the reconstruction 56 from subsections 32 and 34 and the reconstructed subsections 50-54.

[0049] After completion of step 122, the actual training of the learning algorithm can take place following the pre-training.

[0050] In step 108, at least one training signal 20 and at least one target output signal associated with the training signal 20 can be received. The training signal 20 and the target output signal 22 can form a related pair. The target output signal 22 can contain at least one image signal that shows at least a section of the test bench setup used to record the training signal 20. The goal of the training is for the machine learning algorithm to use the training signal 20 to provide the test bench setup provided by the target output signal 22 as an output signal, or at least to approximate this setup using the target output signal 22.

[0051] The image signals can also contain text modules that describe the image content of the image signal in text form.

[0052] In a further step 110, at least one latent variable 18 can be determined. The latent variable 18 used in step 110 can be referred to as the training variable.

[0053] Using the training signal 20 and the target output signal 22, an energy function 24 can be determined that assigns a low energy value to the pair of training signal 20 and target output signal 22. For this purpose, the training signal 20 can, for example, be converted into an embedded signal using a first encoder 12. Similarly, the target output signal 22 can also be converted into an embedded signal using a second encoder 28.

[0054] The two embedded signals can be used to determine the energy function 24, as shown in Fig. Figure 5 is shown. Additionally, a latent variable 18 can be used to determine the energy function 24.

[0055] By training with a large number of pairs of training signal 20 and target output signal 22, a wide range of values ​​for the energy function 24 can be covered, enabling the energy function 24 to provide low energy values ​​for a large number of training signal 20 / target output signal 22 pairs. The energy function 24 can also be defined between these energy values. In the vicinity of the low energy values, the energy function 24 can then determine output signals with similarly low energy values ​​for a previously unknown acoustic measurement signal. Furthermore, the energy function 24 can be configured to assign higher energy values ​​if an acoustic measurement signal and a intended output signal or latent variable do not match.

[0056] In a further step 112, at least one output signal of the learning algorithm can be provided. The learning algorithm determines this output signal from the at least one training signal 20 and using the at least latent variable 18.

[0057] In Fig. In section 6, the machine learning algorithm features a Joint Embedding Predictive Architecture. In this embodiment, the at least one embedded signal of the training signal 20 and the at least one latent variable 18 can be fed to the predictor 14. The predictor 14 then determines a probable value for an embedded output signal. The predictor 14 can determine the embedded output signal using the energy function 24 described above.

[0058] The embedded output signal can then be fed to a cost function 26.

[0059] Furthermore, the target output signal 22 can be provided to the cost function 26 as an embedded signal. The cost function 26 can then determine a cost value. The target range can, for example, be chosen such that the difference between the embedded target output signal 22 and the embedded output signal is below a predefined threshold. The predefined threshold can be set by a user.

[0060] If the cost value is within the target range, the training can be terminated. If the cost value is outside the target range, step 116 can be performed, which modifies the learning algorithm. This can involve, for example, adjusting predictor 14. Alternatively or additionally, adjustments can be made to one or both encoders 12 and 28. Furthermore, the choice of the latent variable can be adjusted.

[0061] With the modified, self-learning algorithm, steps 108, 110, and 112 can then be repeated to determine a cost value again in step 114. In this way, iterative self-supervised training can be carried out.

[0062] The example described above does not in any way limit the invention. Rather, the invention can be modified in numerous ways. All features of the invention described above can be essential to the invention, either alone or in combination. Reference symbol list 10 acoustic measurement signal 12 encoders 14 Predictor 16 Image signal 18 latent variables 20 Training signal 22 Target output signal 24 Energy function 26 Cost function 28 encoders 30 images 32 Subsection 34 Subsection 36 encoders 38 embedded image signal 40 embedded image signal 42 embedded reconstructed image section 44 embedded reconstructed image section 46 embedded reconstructed image section 48 decoders 50 reconstructed subsection 52 reconstructed subsection 54 reconstructed subsection 56 reconstructed subsection

Claims

[1] Computer-implemented method (100) for reconstructing at least one structure of a test bench for determining at least one acoustic measurement signal, wherein the method (100) uses at least one learning algorithm which determines at least one output signal from at least one input signal using at least one latent variable (18), comprising at least the following steps: a. receiving (102) at least one input signal; b. Determining (104) at least one latent variable (18) by means of the learning algorithm; and c. providing (106) at least one output signal of the learning algorithm, which the learning algorithm determines from the input signal using the latent variable (18); characterized bythat the input signal is designed as at least one acoustic measurement signal of at least one component, wherein the learning algorithm reconstructs at least one structure of a test bench for determining the acoustic measurement signal (10) from the at least one input signal and the output signal is designed as at least one image signal (16) which indicates the at least one structure. [2] Computer-implemented method (100) according to claim 1, characterized by that the image signal (16) additionally displays text modules describing the structure. [3] Computer-implemented method (100) according to claim 1 or 2, characterized byin that the learning algorithm has at least one prediction function (12) for the output signal based on the latent variable, wherein the learning algorithm preferably has at least one joint embedding predictive architecture, at least one image joint embedding predictive architecture and / or at least one latent variable generative energy-based model. [4] Computer-implemented method (100) according to one of the preceding claims, characterized by that the learning algorithm has at least one Latent Variable Energy-Based Model. [5] Computer-implemented method (100) according to one of the preceding claims, characterized by that the latent variable (18) comprises at least a number of sensors, at least a distance between the component and an acoustic sensor and / or information about ambient noise. [6] Computer-implemented method (100) according to one of the preceding claims, characterized by that the latent variable (18) lies within a range of values ​​that can be used for test bench setups for determining at least one acoustic measurement signal. [7] Computer-implemented method (100) according to any one of the preceding steps, characterized by that the learning algorithm, before carrying out the steps of claim 1, has at least the following steps: a. Receiving (108) at least one training signal, which is designed as at least one acoustic measurement signal of at least one component, and at least one target output signal associated with the training signal, which is designed as at least one image signal, which preferably displays at least one text block describing the image content of the image signal, wherein the target output signal displays at least part of a structure of a test bench with which the training signal was determined; b. determining (110) at least one latent variable (18); c. providing (112) at least one output signal of the learning algorithm, which the learning algorithm determines from the training signal using the latent variable (18); d. Checking (114) whether a cost value of a cost function from the output signal and the target output signal lies within a target range; and e. If the cost value is outside the target range: Change (116) the learning algorithm and repeat the two previous steps with the changed learning algorithm. [8] Computer-implemented method (100) according to claim 6, characterized by that an energy function is determined (118) from the training signal, the latent variable (18) and the target output signal, such that a first combination based on the training signal and the target output signal associated with the training signal is assigned a lower energy value, and a second combination based on the training signal and a target output signal not associated with the training signal is assigned a higher energy value. [9] Computer-implemented method (100) according to claim 7, characterized bythat the learning algorithm provides the at least one output signal taking into account the energy function, wherein the provided output signal in combination with the input signal and the latent variable (18) represents a minimum of the energy function. [10] Computer-implemented method (100) according to one of the preceding claims, characterized by that the acoustic measurement signal is converted into an image signal (120) which displays a spectrogram from the acoustic measurement signal. [11] Computer-implemented method (100) according to one of the preceding claims, characterized by that the learning algorithm is trained (122) before carrying out the steps of claim 1 in such a way that the learning algorithm reconstructs complete image data, in particular of the spectrogram or of the structure of the test bench, from partial sections of image data, in particular of a spectrogram or of a structure of a test bench. [12] A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method (100) according to any one of the preceding claims.

Images