Method for training and applying machine learning model for determining rotational speed

Abstract

The invention relates to a method for training and applying a machine learning model for determining a rotational speed. Wherein the component is vibrated during operation of the technical system by means of a rotational movement, and wherein the training data comprises a selection frequency range of at least one order of the rotational speed of the component in a first spectrogram (204), the first spectrogram (204) having been determined on the basis of a first set of sound measurement data (200), and wherein the first set of sound measurement data (200) has been determined on the basis of a second set of sound measurement data (200). The sound measurement data has been detected during operation of the technical system by means of a sound sensor (160) arranged on or in the vicinity of a component of the technical system.

Description

Technical Field

[0001] The present invention relates to a method for training a machine learning model, wherein the machine learning model is used to determine the rotational speed of a component of a technical system based on operating parameters of the technical system, wherein the component is in vibration due to rotational movement during operation of the technical system, and the present invention also relates to an application and a computing unit of such a machine learning model and a computer program for executing the method. Background Art

[0002] A turbocharger, ie an exhaust gas turbocharger (ATL) or a turbine as it is commonly known, is used to increase the power or efficiency of an internal combustion engine. In many cases, it is necessary to determine the rotational speed of the exhaust gas turbocharger or other rotating parts. Summary of the invention

[0003] According to the invention, a method and a computing unit according to the invention for training and applying a machine learning model as well as a computer program for executing the method are proposed. Advantageous embodiments are the subject of the following description.

[0004] The invention relates to components of technical systems which, for example an exhaust gas turbocharger of an internal combustion engine, are subjected to vibrations due to rotational movements during operation of the technical system.

[0005] An exhaust gas turbocharger (ATL) is an auxiliary device in an internal combustion engine and is used there to compress the delivered air. Due to the higher air mass flow and thus the greater amount of oxygen, the internal combustion engine can be operated more efficiently, i.e. the engine power is increased. In its simplest form, a turbocharger consists of a turbine wheel, which is driven by the energy of the exhaust gas, and a compressor wheel, which uses the drive energy to increase the air pressure in the intake system. The turbine wheel and the compressor wheel are fixed to the shaft in a rotationally fixed manner and are hydrodynamically slidably supported in the housing.

[0006] Further components of a turbocharger are, for example, a charge air cooler and a wastegate valve arrangement which serves to control the charge pressure and to prevent overloading of the exhaust gas turbocharger.

[0007] Contactless sensors based on magnetic or electrodynamic principles can be used, for example, to determine the rotational speed of a rotating machine element. In simple terms, a rise or fall in the structure of the machine element influences the magnetic field, so that the rotational speed can be correlated. Conventional sensors usually reach a maximum rotational speed of approximately 30,000 rpm, while special ATL rotational speed sensor systems, for example, reach a maximum rotational speed of 400,000 rpm, but require retrofit measures on the component that carries the sensor (such as the housing of a turbocharger).

[0008] If no sensor is available, then, for example, a solution can be used in which the rotational speed is determined using sound signals, in particular airborne or structure-borne sound signals, and which is based on a spectral analysis. In this case, for example, the rotational speed can be determined from a structure-borne sound spectrogram by marking individual points and by path tracing. However, this is usually only effective if the frequency is dominant and is always visible in the spectrogram, which is usually not the case.

[0009] Within the scope of the invention, a possibility is now proposed for ascertaining or determining a rotational speed in a technical system, the structure of which (or at least one of its components) is excited by the rotational movement of a machine element to form vibrations. For this purpose, sounds, such as structure-borne sound signals or airborne sound signals, which can be detected on the component (e.g. its surface) or in the case of airborne sound in the vicinity of the component, are used as learning variables on the one hand and as specific variables describing the load state and / or operating state on the other hand, which are also referred to below as operating variables of the technical system.

[0010] In this case, a functional mapping of measurable operating variables onto the rotational speed of the exhaust gas turbocharger or other components, which is usually not directly measurable, can be achieved. In this context, a "virtual sensor system" can also be mentioned. The proposed method is primarily suitable for applications in which the structure-borne sound signature of the object under examination (here ATL) is strongly superimposed by other excitations (such as engine noise, fans, etc.) and the relevant orders (fundamental frequencies and harmonics) of the object under examination are not presented as the clearest signature in the spectrum or spectrogram.

[0011] By using machine learning methods and machine learning models, the learning process of the virtual sensor system can be automated to a certain extent, so that the user, for example, only marks a few samples from the sound signal with the existing rotation speed at the beginning of the process and presets the value range to be expected. Subsequently, the method itself learns the previously unknown relationships. Thus, the information from the air system is correlated, which is therefore responsible for ensuring the ATL or component rotation.

[0012] In this case, firstly a first machine learning model can be trained. For this purpose, first training data are provided, which include operating variable data of one or more operating variables of the technical system other than the rotational speed of the component. The operating variable data are detected or used during the operation of the technical system. They can be measured data, but also other values, for example, which are predetermined within the scope of the operation of the technical system. In this case, thermodynamic variables are considered in particular as operating variables.

[0013] The training data also include a selected frequency range of at least one order (e.g. fundamental frequency and / or harmonics) of the rotational speed of the component to be expected in the first spectrogram. The first spectrogram has been determined based on a first set of sound measurement data which has been detected during operation of the technical system by means of sound sensors arranged on or near the component of the technical system. In particular, a smaller part can be selected from a larger set of sound measurement data for this purpose.

[0014] To this end, in an embodiment, a first set of sound measurement data may be provided, in particular by selecting from a larger set of sound measurement data. The sound measurement data of the first set is then transformed into a frequency range in order to obtain a spectrogram, which is then provided, for example, to a user, for example by display on a display. A selected frequency range in the spectrogram is then selected, in particular based on an obtained user input.

[0015] The training data also includes one or more time ranges of at least one order of the rotational speed of the identifiable or identified component in the first set of sound measurement data. The time ranges can in particular also be provided based on the obtained user input; in this case, in particular, so-called labeling or marking of these time ranges is involved. In order to increase accuracy, labeling can also be performed in these frequency ranges.

[0016] It is mentioned in this case that the operational data is not necessarily required for training the first machine learning model.

[0017] Subsequently, the first machine learning model is trained based on the first training data, so that based on the sound measurement data and possibly the operating data as input data for the first machine learning model, at least one of the following output variables is determined: a validity variable, an upper frequency limit of a frequency range for at least one order of rotational speed of the component to be expected, and a lower frequency limit of a frequency range for at least one order of rotational speed of the component to be expected, the validity variable indicating whether at least one order of rotational speed of the component can be identified in the input data. The first machine learning model thus trained is then provided for further use.

[0018] In one embodiment, the first machine learning model includes three sub-models, wherein each sub-model determines a different output variable, and wherein the sub-models are in particular each configured as a classifier model or a regression model. For example, a so-called random forest classifier is considered for the validity variable, and a random forest regressor is considered for the limit frequency.

[0019] Thus, a first machine learning model may be used to identify or select relevant data within a large collection of sound measurement data and related operational data.

[0020] In an embodiment, a second machine learning model is trained using the trained first machine learning model, which is used to determine the rotational speed of a component of the technical system. For this purpose, second training data are provided, which include operating variable data and the rotational speed determined by the trained first machine learning model based on the operating variable data. In other words, the operating variable data is directly used at least for the determined or selected operating variable and is based on this (or, if necessary, also the rotational speed determined more or less from the operating data).

[0021] In order to determine the rotational speed, operating parameter data can be provided as input data for a first machine learning model. Subsequently, with the aid of the first machine learning model, a time range of at least one order of the sound measurement data in which the rotational speed of the component can be identified or exists is determined based on a validity parameter as an output parameter. These time ranges of the sound measurement data are provided. Subsequently, based on the sound measurement data of a second set corresponding to these time ranges, a second spectrogram is determined with the aid of the first machine learning model and then provided. Then, based on applying a search function to the spectrogram in a selected frequency range, the rotational speed of the component is determined and also provided. A Python implementation is considered, for example, as a search function.

[0022] Subsequently, a second machine learning model is trained based on the second training data so that the rotational speed of the component is determined based on the operating parameter data as input data of the machine learning model; and then the trained second machine learning model is provided.

[0023] This second machine learning model can be used in particular if the sound sensor is no longer available or at least is not to be used later in order to determine the rotational speed.

[0024] The application of the machine learning model involves determining a rotational speed of a component of the technical system based on operating variables of the technical system using the machine learning model, wherein the component is in vibration due to the rotational movement during operation of the technical system.

[0025] Input data for the machine learning model are provided, wherein the input data at least include operating variable data of one or more operating variables of the technical system other than the rotational speed of the component, wherein the operating variable data have been detected or used during operation of the technical system. Subsequently, the rotational speed of the component of the technical system is determined and provided based on the input data and the machine learning model.

[0026] In one embodiment, the machine learning model comprises a (trained) second machine learning model. As mentioned, no acoustic sensor is required here, but rather the rotational speed can be determined solely based on the operating variable or its value.

[0027] In an embodiment, the machine learning model comprises a (trained) first machine learning model. A sound sensor should be used here. Therefore, the input data also includes sound measurement data, which are detected during operation of the technical system by means of a sound sensor arranged on or near a component of the technical system.

[0028] The rotational speed of a component of the technical system is determined based on the input data and the machine learning model as follows: With the aid of a first machine learning model, time ranges of the sound measurement data are determined based on a validity variable as an output variable, in which at least one step of the rotational speed of the component can be identified or is present; these time ranges of the sound measurement data are provided.

[0029] Then, a spectrogram is determined based on the sound measurement data corresponding to the time range with the help of a first machine learning model; and the spectrogram is then provided.

[0030] The rotational speed of the component is then determined based on applying a search function to the spectrogram within the selected frequency range. The search function may be the same search function as used to train the second machine learning model.

[0031] The proposed method and the above-mentioned machine learning model produce various advantages. For example, complex, extensive measurements of the turbocharger system (e.g., the creation of a compressor characteristic curve) are not required, so that the speed can be estimated even without manufacturer-specific development information and without direct measurement of the ATL speed by a physical sensor system. The speed can be determined even in the case of structure-borne sound data that is strongly induced and strongly superimposed by interference variables (e.g., combustion noise). The method can be used as a retrofit solution because it can be applied without geometric adaptation, deformation or chip removal methods (drilling). In addition to placing the sound sensor on the housing surface or near the noise emission device, no additional installation steps are required. If the sensor mass is also selected to be very small, then there is no intervention in the system, so that the system behavior remains at least sufficiently accurate.

[0032] Since the proposed method does not use a path tracking algorithm (as is the case, for example, in certain software programs), it works even when the identification of the ATL or of the components in general is only visible occasionally or aperiodically, i.e. with interruptions of varying lengths. Since the method is modularly constructed, it can be used in its entirety or in part, for example using only the first machine learning model. If the basic conditions change, the validity classifier of the ATL search function or the learned passband limit can also be used alone, for example.

[0033] The marking of the relevant frequency ranges facilitates manipulation and is less prone to errors than setting frequency-specific markings. Since the recognition method (first machine learning model) is based in particular on a rule-based algorithm, the method can be implemented online as a partial or complete solution, in addition to offline / backend applications, also at the control unit level. This is particularly advantageous because after the training phase, all input variables are available as signals on the control unit.

[0034] The method is also able to depict dynamic speed changes well.By using recognition methods, large data sets can already be labeled using small partial data sets (a first machine learning method is trained using a small data set and then used to generate a large training data set for a second machine learning method).

[0035] Automatic recognition of relevant operating states and thus high automation capabilities and integration possibilities are possible in existing software solutions. For example, if two measurements come from the same test device, a direct transfer of the virtual sensor system to the second measurement is also theoretically conceivable.

[0036] Further advantages associated with the use of speed information are, for example, improved collective detection of the turbocharger load (in addition to the usually recorded thermodynamic variables in the air system, the speed can also be recorded). Depending on the frequency of the available speed information, regulation (e.g. boost pressure control) can thus be optimized and overspeed can be prevented (component protection).

[0037] Methods based on rotational speed information (eg order analysis, bandpass filters adapted to the operating point, etc.) can also be used within the scope of condition monitoring or predictive maintenance applications. This can facilitate the identification of damage that specifically shows changes in the frequency associated with the rotational speed.

[0038] A computing unit according to the invention, for example a control unit or a computer of a motor vehicle, is designed, in particular in terms of programming, to carry out the method according to the invention.

[0039] The method according to the invention is also advantageous in the form of a computer program or computer program product with a program code for performing all method steps, because this results in particularly low costs, especially when the control unit to be performed is still used for other tasks and therefore originally exists. Finally, a machine-readable storage medium and a computer program as described above stored thereon are provided. Suitable storage media or data carriers for providing computer programs are especially magnetic, optical and electrical memories, such as hard disks, flash memories, EEPROMs, DVDs, etc. The program can also be downloaded via a computer network (the Internet, intranets, etc.). This download can be realized here by wire or wireless (e.g., by WLAN networks, 3G, 4G, 5G or 6G connections, etc.).

[0040] Further advantages and embodiments of the invention are apparent from the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The invention is schematically illustrated in the drawings by means of exemplary embodiments and will be described below with reference to the drawings.

[0042] Figure 1 A technical system is schematically shown in which the present invention can be used.

[0043] Figure 2 The process of the method in different embodiments is shown.

[0044] Figure 3 , Figure 4 and Figure 5 A diagram for explaining the present invention is shown. DETAILED DESCRIPTION

[0045] Figure 1 A technical system designed as an internal combustion engine in which the present invention can be used is schematically shown. A section of an internal combustion engine 1 having a cylinder or cylinder block 150 and an exhaust gas turbocharger 100 as components is schematically shown. The cylinder block 150 has an air intake 151 for supplying fresh air and an exhaust gas manifold 152 for discharging exhaust gases.

[0046] The exhaust gas turbocharger 100 has a turbine 120 and a compressor 110 driven by the turbine. Fresh air 10 is supplied to the compressor 110, compressed to a boost pressure by the compressor, and supplied to a cylinder block 150 through an optional air cooler 11. As is known, the fuel / air mixture is burned in the cylinder block 150, and the exhaust gas generated thereby is supplied to the turbine 120 through an exhaust manifold 152.

[0047] The turbine 120 has a turbine housing 123 with a turbine inlet 121 and a turbine outlet 122, and between the turbine inlet and the turbine outlet there is a guide device 124 with adjustable guide blades and a turbine impeller 125. The exhaust gas enters the turbine housing 123 under pressure at the turbine inlet and flows through the turbine housing 123 and the guide device 124, wherein the pressure is reduced at the transition between the guide device and the turbine impeller. Before the exhaust gas escapes into the exhaust system 12 at the turbine outlet 122, the exhaust gas drives the turbine impeller 125.

[0048] In this case, the turbine wheel 125 is coupled to the compressor wheel 115 of the compressor 110 via a shaft 126. During operation, the turbine wheel 125 and the compressor wheel 115 rotate at a rotational speed denoted here by n.

[0049] The exhaust gas turbocharger shown here is merely an example; for example, an exhaust gas turbocharger with a wastegate valve can also be used.

[0050] Furthermore, a computing unit 170 embodied as an engine control unit and an acoustic sensor 160 arranged on the exhaust gas turbocharger, in this case by way of example on the turbine housing 123 , are shown.

[0051] Figure 2 The flow of the method in different embodiments is schematically shown. Here, first a first machine learning model can be trained, which can be used later, for example, during operation, or can also be used to obtain additional training data for training a second machine learning model.

[0052] In a first step, a sound sensor, for example a structure-borne sound sensor or an airborne sound sensor, can first be arranged on or near the exhaust gas turbocharger. The signal or signals of the sound sensor can be present, for example, as position, velocity or surface vibration acceleration. In the example explained here, the differential voltage of the piezoelectric knock sensor is measured. Therefore, sound measurement data 200 are usually present.

[0053] exist Figure 3 In FIG. 1 , such sound measurement data with voltage U are shown by way of example in the upper diagram as a curve over a period of time.

[0054] Furthermore, operating variables of the technical system (here the internal combustion engine) are detected. During the operation of the technical system, operating variable data are detected or used. In particular, relevant thermodynamic variables in the air system, such as the air mass flow, the pressure ratio in the exhaust gas turbocharger and the accelerator pedal position, are considered as operating variables. Therefore, operating variable data 202 are usually present. The operating variable data used here are usually present in the engine control unit, i.e., they are detected, processed or used in another way there, for example.

[0055] exist Figure 3 , for this purpose the accelerator pedal position is shown by way of example as a curve over a period of time in the middle diagram and the mass flow dm / dt (as operating variable data) is shown in the lower diagram.

[0056] In particular, during operation of the internal combustion engine, under load, all variables are measured, for example, by a measuring computer, in particular in a time-synchronous manner, such as Figure 3 shown.

[0057] In a further step, the sound measurement data are transformed into a frequency range. This can be, for example, a short-time Fourier transform, for example for a portion of the sound measurement data used as training data, i.e. a first set. Here, for example, a sampling rate of 100 kHz can be used. The transformation results in a spectrogram 204, i.e. a spectrogram of the sound sensor training data.

[0058] The spectrogram 204 may then be visualized or otherwise provided to a user, for example. Figure 4 There, V 2 The spectral power density L in / Hz is plotted against the frequency f in Hz.

[0059] Then, a user input 206 may be made in which a frequency range is specified within which, for example, the first, second, and kth order of turbocharger speeds are expected (e.g., 1300-3000 Hz corresponds to 78000-180000 1 / min). This frequency range is selected as the frequency range 300 in Figure 4 As shown in FIG.

[0060] In a further step, a further user input 208 can be made, in which time ranges in which the turbocharger speed can be identified are marked (labeled). In this case, those time ranges are found in the time period of the sound measurement data, in which the user can see, for example, that the turbocharger speed dropped. Figure 3 In FIG. 3 , such a time range is indicated by way of example at 300. The background to this is that only those time ranges of the entire data are relevant later in which the turbocharger speed can be primarily identified.

[0061] In order to improve the accuracy, it is also possible to additionally carry out marking or labeling in the frequency range, for example by means of a random forest regressor, so that an adaptive bandpass filter is learned depending on the operating point. This may, for example, help to further narrow the previously defined frequency band (selection of the frequency range) and thus improve the estimate. Figure 4 Such a limited frequency range 402 is shown, which may be used as a selected frequency range.

[0062] A first machine learning model 210 can then be trained and in particular validated, which can include a plurality of sub-models, which are shown here using a random forest classifier 212 and two random forest regressors 214 , 216 .

[0063] The labeled label, i.e., the time range, represents the dependent variable here, and the operating parameter represents the independent variable, i.e., the input parameter. The operating parameter usually exists only within this time range or time period. Other classifier and regressor models are also conceivable. In the preprocessing, other steps are also considered, such as the acquisition of other features from the detected signal parameters and the dimensionality reduction method (principal component analysis PCA, kernel principal component analysis, etc.). It should also be pointed out that in this example, the hyperparameter optimization of the execution system is not performed. However, this hyperparameter optimization can be supplemented when necessary and especially leads to higher accuracy and higher robustness of the first machine learning model.

[0064] The recognition method, i.e., the result of the above-mentioned training, is a trained first machine learning model having three sub-models, which can determine different output parameters, i.e., a validity parameter G (validity flag), an upper limit frequency oGF (passband upper limit) of the frequency range of at least one order of the rotational speed of the expected component, and a lower limit frequency uGF (passband lower limit) of the frequency range of at least one order of the rotational speed of the expected component, wherein the validity parameter indicates whether at least one order of the rotational speed of the component can be identified in the input data.

[0065] In particular, these output variables can include a vector for each operating point of the training data set, which has a value for each of the three output variables mentioned. The validity variable or validity flag G can be 0 for invalid, i.e. no flag is seen, or 1 for valid, i.e. a flag is seen. The upper and lower frequency limits can each be specified as a value in Hz.

[0066] Now, in a further step, the first machine learning model, i.e. the classifier and regressor trained with the small training data set, can be used to find or mark those time ranges in the complete data set (or at least one data set larger than that trained for above; the sound measurement data) where a turbocharger speed can be seen or recognized.

[0067] The marking is achieved by predicting a previously defined vector of the first machine learning model, ie, the submodel, based on the unknown data, ie, the sound measurement data 200. Specifically, therefore, a time range of at least one order of the rotational speed of the component can be identified or present in the sound measurement data 200 can be determined first. Figure 3 As shown, this can only be achieved through the first machine learning model.

[0068] In a further step, a spectrogram or spectrum can be determined or extracted from the time range considered to be valid in the sound measurement data, which is here an operating point segment, by means of a time stamp, for example. Figure 4 The spectral power density is shown.

[0069] Therefore, subsequently different methods are possible which are briefly explained below.

[0070] The sound sensor can remain, ie remain mounted, on the exhaust gas turbocharger. Then a search function 218 is used, which searches in the sound spectrum in a selected frequency range 400 (see Figure 4 ) or in the frequency bands defined by the regressors (oGF and uGF, see Figure 4 The turbocharger frequency or turbocharger speed is determined within a selected frequency range 402 in the frequency range 402 and outputted. The output can occur in Hz or 1 / min.

[0071] In this way, the speed 220 of the turbocharger can be determined using the current sound measurement data (using the installed sound sensor) while using the first machine learning model 210. Figure 4 In FIG. 4 , the possible turbocharger speeds are marked with a cross, but only one turbocharger speed is located within the selected frequency range 402 , which is the final speed.

[0072] However, the search function can also, for example, either search only for the maximum amplitude in a frequency band or use additional ascertained variables such as the signal-to-noise ratio, the energy content in adjacent frequency bands, sidebands, etc. as robustness criteria. Similarity matching or pattern recognition with segments (time ranges) determined during the training process is also possible. In the case of strong modulation, broadband noise, etc., the turbocharger speed can, for example, subsequently lose its validity.

[0073] However, the acoustic sensor can also be removed from the exhaust gas turbocharger.

[0074] Here, the same search function 210 can determine the turbocharger frequency or turbocharger speed as before and store it in a temporary memory for the training partial data set together with the associated time stamp. This can be done for a plurality of speeds. These data are then linked to the operating data 202 via the time stamp (e.g. via a table merge), see step 222.

[0075] Subsequently, a second machine learning model 224, for example a further regression model, is trained. This further regression model can be, for example, a support vector regression (kernel = 'rbf', C = 2000, Epsilon = 0.1) according to a Python Sklearn implementation, which receives as input the correlation of the operating variables 202, in particular the air mass flow, the pressure ratio and the accelerator pedal position, and learns the turbocharger speed for this purpose. The operating variables can be selected, for example, based on expert knowledge, but data-driven selection is also conceivable.

[0076] In addition, a standardization process step can be performed before the training, for example, that is, for each feature, the Standardization is performed (standardization at the mean degree of freedom and standardization at the standard deviation to avoid different scale effects). It should also be pointed out that in this example, no systematic hyperparameter optimization is performed. However, this hyperparameter optimization can be supplemented if necessary and leads in particular to a higher accuracy and a higher robustness of the rotational speed estimation.

[0077] Trained second machine learning model 224 may then be implemented in the sense of a virtual sensor for predicting turbocharger speed n based solely on operating variables or operating data 202 .

[0078] The training of the virtual sensor system can be carried out on an end device with a user interface, i.e. on a computer, tablet, mobile phone, etc. However, further modifications and improvements are conceivable, for example variations in the signal preprocessing (different windowing, window lengths, overlaps). The method can be used online or offline in a measuring computer, in a backend, in the cloud or on an end device (sensor, control unit, connection unit). The training of the recognition method for determining the validity of the speed ascertainment can be carried out solely on the basis of structure-borne sound data (the load state is implicitly contained therein) and not on control unit signals or operating data.

[0079] Better dynamic modeling can be achieved by using time series as input parameters (such as recurrent neural networks and long short-term memory (LSTM) networks).

[0080] to this end, Figure 5 The diagram over time t shows, by way of example, operating variables which can be used as input data. These are angle measurement window length knock control W, lambda actual value λ and mass flow dm / dt. For this purpose, the validity variable G ascertained with the aid of the first machine learning model is shown. In addition, the speed n of the turbocharger ascertained in this way is also shown. In particular, in the case of a valid signal (G=1), the speed value is well represented.

Claims

1. A method for training a first machine learning model (210) for determining a rotational speed (n) of a component (100) of a technical system (1) based on operating parameters of the technical system, wherein: The component (100) is subjected to vibrations due to a rotational movement during operation of the technical system, and the method comprises: Provide first training data, wherein the first training data includes the following data: a selected frequency range (400, 402) of at least one order of a rotational speed of the component in a first sound spectrogram (204), wherein the first sound spectrogram (204) has been determined on the basis of a first set of sound measurement data (200) which has been detected during operation of the technical system by means of a sound sensor (160) arranged on or in the vicinity of a component of the technical system; - one or more time ranges (300) in said first set of sound measurement data (200) in which at least one order of rotation speed of said component can be or has been identified; The first machine learning model (210) is trained based on the first training data so that at least one of the following output variables is determined based on the sound measurement data as input data of the first machine learning model: a validity variable (G) which indicates whether at least one step of the rotational speed of the component can be identified in the input data, - an upper frequency (oGF) of the frequency range in which at least one order of rotational speed of the component should be expected, - a lower frequency limit (uGF) of the frequency range within which at least one order of rotational speed of the component should be expected; and Providing the first trained machine learning model (210).

2. The method according to claim 1, wherein: The first training data also includes the following data: operating parameter data (202) of one or more operating parameters of the technical system other than the rotational speed of the component, wherein the operating parameter data has been detected or used during the operation of the technical system; wherein the first machine learning model (210) is trained based on the first training data so that at least one of the output parameters is also determined based on the operating parameter data as input data of the first machine learning model.

3. The method according to claim 1 or 2, further comprising: In particular, providing said first set of sound measurement data (200) by selecting from a larger set of sound measurement data; transforming the first set of sound measurement data into a frequency range to obtain the spectrogram (204); Providing the spectrogram (204), in particular providing the spectrogram to a user; providing, in particular based on the obtained user input, a selected frequency range in the spectrogram; Therein, the selected frequency range and the one or more time ranges are provided as part of the first training data.

4. A method according to any one of the preceding claims, wherein: The first machine learning model (210) includes three sub-models (212, 214, 216), wherein each sub-model determines a different output variable, and wherein the sub-models are in particular each constructed as a classifier model or a regression model.

5. A method for training a second machine learning model (224) using a first machine learning model trained according to any one of claims 1 to 3, wherein the second machine learning model is used to determine the rotational speed of a component (100) of the technical system (1) based on operating parameters of the technical system, wherein: The component is subjected to vibrations due to a rotational movement during operation of the technical system, and the method comprises: Providing second training data, wherein the second training data includes the following data: operating parameter data (202) and a rotation speed (210) determined based on the operating parameter data by means of the trained first machine learning model (210); training the second machine learning model based on the second training data so that the rotational speed of the component is determined based on the operating parameter data as input data of the machine learning model; and Provide the second machine learning model.

6. The method according to claim 4, further comprising: determining the rotational speed by means of the trained first machine learning model: Providing the operating parameter data (202) as input data for the first machine learning model; Determining, by means of the first machine learning model (210) and based on the validity variable (G) as an output variable, a time range of the sound measurement data in which at least one step of the rotational speed of the component can be identified or is present; providing said time range of said sound measurement data; determining a second spectrogram with the first machine learning model based on a second set of sound measurement data corresponding to the time range; providing the second spectrogram; determining a rotational speed of the component based on applying a search function to the spectrogram within the selected frequency range; The determined rotational speed is provided.

7. A method for determining a rotational speed of a component of a technical system based on operating variables of the technical system using a machine learning model, wherein: The component is subjected to vibrations due to a rotational movement during operation of the technical system, and the method comprises: providing input data for the machine learning model, wherein the input data include operating parameter data of one or more operating parameters of the technical system other than the rotational speed of the component and / or sound measurement data, wherein the operating parameter data have been detected or used during operation of the technical system, wherein the sound measurement data have been detected during operation of the technical system by means of a sound sensor arranged on or near the component of the technical system; determining a rotational speed (n) of a component of the technical system based on the input data and the machine learning model; and The rotational speed of the component is provided.

8. The method according to claim 7, wherein: The machine learning model comprises a first machine learning model (210) trained according to any one of claims 1 to 4, wherein the input data includes the sound measurement data (200), Wherein, determining the rotational speed of a component of the technical system based on the input data and the machine learning model comprises: determining, by means of the first machine learning model, based on a validity variable as an output variable, a time range of the sound measurement data in which at least one step of a rotational speed of the component can be identified or is present; providing said time range of said sound measurement data; determining a spectrogram based on the sound measurement data corresponding to the time range with the aid of the first machine learning model; providing the spectrogram; Based on applying a search function to the spectrogram within the selected frequency range, a rotational speed of the component is determined.

9. The method according to claim 7, wherein: The machine learning model comprises a second machine learning model trained according to any one of claims 5 to 6, and wherein the input data comprises the operating parameter data (202).

10. A method according to any one of the preceding claims, wherein: An internal combustion engine (1) serves as the technical system, and an exhaust gas turbocharger (100) of the internal combustion engine serves as a component of the technical system, wherein one or more operating variables of the technical system include one or more thermodynamic operating variables. 11 . A computing unit designed to carry out all method steps of the method according to claim 1 . 12 . A computer program which, when executed on a computing unit, causes the computing unit to execute all method steps of the method according to claim 1 .

13. A machine-readable storage medium having stored thereon the computer program according to claim 12.

Images