Methods for detecting the health status of a sound-emitting component of a technical system and methods for multi-stage training of a machine learning model for use therein

Abstract

The invention relates to a method for detecting the health status of at least one sound-emitting component of a technical system. In the method, input data is provided for a multi-stage machine learning model comprising a plurality of sub-machine learning models (200), wherein each sub-machine learning model is configured to evaluate input data in a predetermined operating range, the input data being determined based on sound measurement data and at least one operating parameter of the technical system. Furthermore, output data (210) are determined based on the input data and the machine learning model, the output data comprising information about the number of occurrences of at least one characteristic parameter for damage to the technical system in the input data.Finally, the output data for detecting the health status of at least one component of the technical system are provided (220). The invention further relates to a method for training a machine learning model for use in detecting the health status of at least one sound-emitting component of a technical system.

Description

[0001] The present invention relates to a method for detecting a health condition of at least one sound-emitting component of a technical system and a method for training a machine learning model for use therein, as well as a computing unit and a computer program for carrying it out. Background of the invention

[0002] Exhaust gas turbochargers (EGTs) are auxiliary components in internal combustion engines, used to compress the intake air. The increased airflow and thus greater oxygen content allows the engine to operate more efficiently, resulting in increased engine power. A turbocharger typically consists of a turbine wheel, driven by the energy of the exhaust gases, and a compressor wheel, which uses this energy to increase the air pressure in the intake system. The turbine wheel and compressor wheel are fixed to a shaft and hydrodynamically supported, usually by plain bearings, within a housing.

[0003] Other components of the turbocharger include, for example, an intercooler and a wastegate, which serves to regulate boost pressure and protect the turbocharger from overload. Besides the turbochargers mentioned, other types of turbochargers also exist, such as turbochargers with variable turbine geometry (VTG).

[0004] In noise-emitting technical systems, such as the aforementioned internal combustion engine, knock sensors are conventionally installed. These sensors detect the structure-borne noise of the system or one or more of its components, such as the exhaust gas turbocharger or parts thereof. Their function is to acquire the signal and transmit it to the engine control unit (ECU). If the air-fuel mixture ignites prematurely, this leads to undesirable engine knocking, which can damage engine components. Conversely, if ignition occurs too late, this results in suboptimal energy efficiency. The current goal of existing knock sensor signal acquisition is therefore to prevent engine knocking in order to reduce fuel consumption and increase engine torque.

[0005] Internal combustion engines may still contain airborne sound sensors, such as microphones, which record the sound emitted by sound-emitting components of the technical system.

[0006] Furthermore, airborne sound, liquid sound, structure-borne sound and / or knock sensors can also be installed in other technical systems, such as hydraulic pumps. Disclosure of the invention

[0007] According to the invention, a method for detecting a health condition, a method for training a machine learning model, and a computing unit and computer program for carrying them out, comprising the features of the independent claims, are proposed. Advantageous embodiments are the subject of the dependent claims and the following description.

[0008] The invention relates to a technical system with at least one sound-emitting component, which is operated, in particular, at variable or dynamic operating points. Examples of such technical systems are internal combustion engines, especially those comprising an internal combustion engine and an exhaust gas turbocharger, or hydraulic pumps. These technical systems typically incorporate airborne sound, liquid sound, structure-borne sound, and / or knock sensors, which detect the sound emitted by the sound-emitting components and transmit it, for example, to an engine control unit.

[0009] Since the knock sensor, as described, detects the characteristic structure-borne sound signature of the engine or internal combustion engine, or a part thereof, and transmits it to the engine control unit, the signal can also contain information that, in addition to knocking, allows conclusions to be drawn about other forms of force excitation. These can be, for example, excitation patterns that describe the engine's operating behavior. Furthermore, it is possible to determine the health status of a component, subsystem, or system.

[0010] In the event of damage, an additional force excitation, transmitted through the structure and manifested as a measurable change in surface vibration acceleration, velocity, or movement of the structure, can be measured or determined through time integration, thus enabling the detection of a fault pattern during operation. Using signal analysis and a multi-stage machine learning process, an indicator for the presence of a fault or for the health status of the system under consideration can be determined.

[0011] The presence of characteristic sound signatures can also be observed in the event of damage by airborne sound sensors that measure the airborne sound emitted by the technical system.

[0012] Against this background, it is proposed to use sound measurement data to determine the health status of a technical system or individual components of the technical system.

[0013] The proposed method for detecting the health status of at least one noise-emitting component of a technical system will be explained below, using the exhaust gas turbocharger as a particular example. However, the method can also be applied to other noise-emitting components of technical systems.

[0014] This process utilizes a machine learning model that can or must be trained in multiple stages before deployment. The machine learning model comprises a multitude of sub-models, each configured or trained to evaluate input data within a predetermined operating range. These operating ranges, or rather the criteria for classifying data within them, are learned during the training of the machine learning model, particularly during the clustering of sound measurement data based on operating parameters. The following section provides a comprehensive description of the training and application of the machine learning model.

[0015] In the inventive method for detecting the health status of at least one sound-emitting component of a technical system, input data for the machine learning model are first provided. This data is determined based on sound measurement data and at least one operating parameter of the technical system. The sound measurement data corresponds to a sound signal that has been detected by a sensor from the at least one sound-emitting component of the technical system (e.g., the exhaust gas turbocharger) during operation of the technical system. As mentioned, an airborne sound sensor, a structure-borne sound sensor, and / or a knock sensor can be used as the sensor. In general, however, other types of sound can also be detected, for example, with correspondingly different sound sensors. Advantageously, the sensor is arranged near the sound-emitting component. The sound measurement data can also have already been processed, i.e.,For example, interference may have been isolated from the sound measurement data using a filtering method. The sound measurement data may have been upsampled, downsampled, bandpass filtered, highpass filtered, lowpass filtered, amplified, interpolated, pre-cleaned, or otherwise modified.

[0016] When applying the machine learning model, i.e., to determine the health status of one or more sound-emitting components of the technical system, output data is determined based on the input data and the machine learning model. The output data contains information about whether and / or to what extent a data point of the sound measurement data from the input data deviates from a known pattern. Subsequently, the output data is provided for recognizing the health status of at least one component of the technical system. In particular, the health status of the at least one sound-emitting component can be determined using the output data.

[0017] The information in the output data can, for example, indicate whether a data point of the sound measurement data follows the known pattern (e.g., anomaly value equal to 0) or deviates from the known pattern (e.g., anomaly value equal to 1). Subsequently, the health status can be determined based on the number of sound measurement data points that deviate from the known pattern (e.g., a value of 1), particularly per unit of time. Alternatively, a moving average, for example, of the values ​​for the last 50 data points, can be determined as the anomaly value, and the health status can be determined based on a value of this anomaly value, for example, by comparison with one or more predefined anomaly thresholds.

[0018] It is also conceivable that the values ​​contained in the information can range between 0 and 1, thus indicating how much a data point of the sound measurement data deviates from the known pattern, with an increasing value signifying an increasing deviation from the known pattern. The deviation could, for example, take the form of a distance from the known pattern, an angular spectrum of the data point, or a reconstruction error of an autoencoder. The values ​​determined in this way can also be scaled or normalized to achieve greater comparability. Subsequently, the course of the value can be observed over a predefined period, and if it is found that the value exhibits an elevated value during this period—for example, a value above a predefined threshold—it can be determined that an anomaly exists in at least one sound-emitting component.

[0019] It should be mentioned that the initial provision of the sound measurement data and the preprocessing, until the input data for the machine learning model is obtained, can also be carried out separately, i.e. on a separate computing unit, while the application of the machine learning model itself then takes place on another computing unit.

[0020] Through this method, i.e., the analysis of sound emission (structure-borne sound, airborne sound, liquid-borne sound), the health of at least one component can be assessed without having to carry out a complex disassembly, thereby simplifying damage detection or allowing it to take place earlier, so that the service life of the technical system can be extended, as consequential damage can be avoided.

[0021] By determining the health status of one or more sound-emitting components of the technical system based on sound measurement data, existing sensors can be used for the determination and the model can be applied, for example, in older technical systems, such as older internal combustion engines.

[0022] Because the method uses classical signal analysis techniques as well as unsupervised learning, there is no need for complex labeling (assigning class memberships) of the sound measurement data. Thus, in a field application, for example, training at the beginning of the product lifecycle can be performed exclusively with the data collected in good condition, assuming that the technical system has 100% wear reserve upon delivery, i.e., no damage.

[0023] In one embodiment, providing the input data for the machine learning model includes acquiring the sound measurement data from at least one sensor during operation of the technical system, and the values ​​of at least one operating parameter of the technical system, wherein the at least one operating parameter and / or its values ​​are acquired time-synchronously with the sound measurement data. Subsequently, the sound measurement data and the values ​​of the at least one operating parameter are divided into time-synchronous data windows. The length of the data windows depends on a variety of factors, such as the required frequency resolution in the spectra of the sound measurement data and the dynamics of the technical system. Time-synchronous in this context means that the values ​​of the at least one operating parameter were acquired within the same time interval as the sound measurement data.The data window length can be, for example, 0.1 seconds, but other durations are also possible. Subsequently, for each data window, an operating point range is determined based on at least one operating parameter. This range represents the operating point range in which the technical system was operating during the acquisition of the sound measurement data. The operating point range is a range of similar operating points. Similar operating points occur, for example, at similar rotational speeds and / or torques. The criteria for the operating point ranges are learned during the training of the machine learning model, particularly during the clustering of the training data. For example, if the operating point range is determined based on the rotational speed of the internal combustion engine, different speed ranges, such as 0-1500 revolutions per minute, 1501-3000 revolutions per minute, 3001-4500 revolutions per minute, etc., can be used as the operating point range.The operating point ranges can depend on a multitude of operating parameters, i.e., they can be multidimensional. In particular, not all recorded operating parameters need to be used to determine the operating point range.

[0024] This allows the sound measurement data of each data window to be easily assigned to the individual sub-machine learning models that are trained to evaluate input data in the specific operating point range.

[0025] Operating parameters have a significant influence on sound behavior. Therefore, by dividing the sound measurement data into different operating point ranges and evaluating it using sub-machine learning models for these ranges, early damage detection is provided for sound-emitting components of a technical system that are exposed to variable and dynamically changing operating points.

[0026] In one embodiment, a spectral power density is determined for each data window based on the sound measurement data to provide the input data. Depending on the type of machine learning model used, the input data can vary, which will be explained in more detail below. In particular, the spectral power density, which is conventionally plotted one-dimensionally against frequency, can also be plotted two-dimensionally against time, thus enabling the use of further analysis methods, especially from image recognition. This allows the temporal evolution to be taken into account, and in particular, unlike a pure FFT (Fast Fourier Transform), which assumes a stationary operating point, dynamic, non-stationary changes can be considered.By plotting the spectral power density against frequency and time, it becomes possible to visually represent the effect over extended periods and to analyze it using a machine learning model, such as a convolutional neural network (CNN). Patterns that are not apparent in a time-based or pure frequency representation can thus be detected, such as sudden frequency changes or periodic noises.

[0027] An inventive method for training a machine learning model for use in detecting the health status of at least one sound-emitting component of a technical system comprises providing training data. The training data includes sound measurement data and at least one operating parameter of the technical system, wherein the sound measurement data and the values ​​of the at least one operating parameter have been recorded synchronously, and information on the health status of the technical system during the recording of the sound measurement data.

[0028] The training data is then preprocessed by dividing it into data windows with a predefined duration, a process known as windowing. For each data window, it is determined, based on at least one operating parameter, whether the technical system was operating in a steady state within that window. The duration depends on a variety of factors, such as the required frequency resolution in the spectra and the system's dynamics. Furthermore, it must be noted that the same duration can result in a different number of data points in each data window at different sampling rates. For example, a data window with a duration of 0.1 s for a sensor with a sampling rate of 100 kHz will contain 10,000 data points, while for the ECU torque parameter with a sampling rate of 100 Hz, it will contain 10 data points, and so on.

[0029] In the next step, the preprocessed sound measurement data from the data windows, for which steady-state operation was determined, are clustered into different operating point ranges depending on at least one preprocessed operating parameter. Subsequently, the machine learning model is trained for each cluster based on the clustered sound measurement data. This training determines output data from the sound measurement data as input for the machine learning model. This output data includes information about whether and / or to what extent a data point in the input sound measurement data deviates from a known pattern. By training on different clusters, the machine learning model exhibits a multitude of sub-machine learning models, each of which is suitable for evaluating input data within the operating point range for which it was trained.For training the sub-machine learning models, only sound measurement data recorded when the technical system is in good working order, i.e., free of errors or damage, can be used. Furthermore, various methods can be used to train the machine learning model or sub-machine learning models, of which only the following are examples: Gaussian Mixture Models, One-Class Support Vector Machine, Isolation Forest, Angle-Based Outlier Detection, and Autoencoder.

[0030] The trained machine learning model is then provided and can be used to evaluate input data and determine the health status of a technical system with at least one sound-emitting component.

[0031] This allows a machine learning model to be provided with which the health status of at least one sound-emitting component of a technical system can be determined, and the advantages previously mentioned in connection with the procedure for determining the health status can be achieved.

[0032] In particular, clustering—that is, grouping sound measurement data from similar operating points—and training the sub-machine learning models on the respective clusters can prevent the machine learning model from being trained on a data distribution that may exhibit a very high degree of inequality, as this inequality could significantly influence or even impair damage detection. This increases the robustness of the machine learning model.

[0033] Furthermore, clustering allows for the identification of load collectives, i.e., load states or operating point ranges, that occur frequently during operation. These load collectives can be used in the design of the components to better optimize them for operation.

[0034] In one embodiment, preprocessing the training data for each data window includes determining a change measure for each of the at least one operating parameter and determining whether the change measure is less than a predefined threshold. Subsequently, if it is determined that the change measure for each of the at least one operating parameter is less than the predefined threshold, it is determined that steady-state operation is present. The data in a data window for which steady-state operation has been determined are hereinafter also referred to as a "sample".

[0035] The change measure can be determined, in particular, from the maximum value of each operating parameter measured in the data window and the measured maximum value of each operating parameter. Expressed as an equation, the change measure can be determined as follows: LPk=logmax(k(t))min(k(t))

[0036] where t ∈ [t i , t i+1 ], where t i and t i+1 the lower and upper limits of the time interval are LP k the change measure of the operating parameter k and k(t) the values ​​of the operating parameter over time.

[0037] Subsequently, for each of the at least one operating parameter, it is determined whether the rate of change is less than a predefined threshold. Steady-state operation is determined if, for each of the at least one operating parameter, the rate of change is less than the predefined threshold. Expressed as an equation, the steady-state condition can be represented as follows: SCk(ti+1−ti)={0 LPk>jk1 LPk≤jk

[0038] With the stationarity condition SC k for the operating parameter k in the time interval of t i are i+1 and the threshold j k for the change measure LP kof the operating parameter k. The threshold value can be different for each operating parameter. For example, the threshold value for the engine torque can be 2%, and for the pressure ratio on the compressor side of the turbocharger, the air mass flow on the compressor side of the turbocharger, the accelerator pedal position, and the engine speed, it can each be 5%.

[0039] Due to the stationary nature of the operation of the technical system, clear patterns and characteristics of the sound measurement data can be identified during the training of the machine learning model, and deviations from the normal (good) state of health can thus be better detected, since changes in the sound measurement data due to a change in the operating point have been filtered out.

[0040] Alternatively, the stationarity of the operation of the technical system can also be determined using an Augmented-Dickey-Fuller test or a Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test.

[0041] In one embodiment, the preprocessing further includes extracting indicator quantities from the sound measurement data, and the machine learning model is trained based on the indicator quantities of the clustered sound measurement data.

[0042] The indicator quantities in the time domain may include one or more of: an arithmetic mean, a median, a standard deviation, skewness, kurtosis, a minimum value, a maximum value, a root mean squared (RMS), zero crossings, and / or a crest factor. In the frequency domain, the indicator quantities may include one or more of a spectrum of the spectral power density of the sound signal, spectral flatness, spectral bandwidth, a spectral centroid, a spectral contrast, a spectral amplitude, a Mel coefficient, a chromagram, and / or an empirical mode decomposition.

[0043] In one embodiment, clustering involves grouping the sound measurement data into a variable or fixed number of operating point ranges using at least one operating parameter.

[0044] If the noise measurement data are to be clustered into a variable number of operating point ranges, a distance-based clustering method (density-based spatial clustering of applications with noise, DBSCAN) can be used. For example, operating parameters for clustering can include engine speed, engine torque, gear selection, pressure ratio on the compressor side of the turbocharger, air mass flow rate on the compressor side of the turbocharger, and vehicle speed. For better comparability, the operating parameters can be normalized to a range of values ​​between 0 and 1. Furthermore, in clusters with a variable number of operating point ranges, each of the at least one operating parameter can be assigned a weighting factor indicating its relevance within the cluster.For example, engine speed can be assigned a weighting factor of 3, engine torque, pressure ratio, and air mass flow a weighting factor of 2, and gear and vehicle speed a weighting factor of 1. Furthermore, additional boundary conditions can be introduced, such as requiring that each cluster contain at least four samples, and further restricting clusters to contain at least four samples with good health, etc. A maximum distance, i.e., a "radius," between the individual samples within a cluster can also be defined.

[0045] If the sound measurement data are to be clustered into a fixed number of operating point ranges, the operating parameters are normalized to be mean-free and have a standard deviation of 1. This ensures that different scalings do not influence the clustering result, as each parameter is initially weighted equally. A k-means algorithm is then used to pre-sort the data into the fixed number of operating point ranges, for example, 25. The choice of the number of operating point ranges represents a compromise between the resolution in the multidimensional operating point space and the number of data points that can be used for the model of each cluster. The number of clusters is determined here using an experimental, iterative approach. This involves solving an optimization problem, which can be represented as follows: J=∑i=1k∑xj∈Si‖xj−μi‖2

[0046] The goal of the k-means algorithm is to assign the data points x j to solve the problem of grouping the sound measurement data into k clusters. The aim is to find the sum of the squared deviations (Euclidean distance) of the centers of gravity µ. i between the clusters S i The number of samples required is minimized. Only sound measurement data recorded when the ATL was in good working order is used for clustering. Furthermore, the minimum number of samples required in a cluster can be defined. For example, the minimum number of samples can be 100, with a model only being trained for clusters containing more than 100 samples. The number 100 should be understood in the context of the dimensions under consideration (characteristics, operating parameters).

[0047] In one embodiment, the at least one operating parameter comprises, if the technical system is, for example, an internal combustion engine with a combustion engine and an exhaust gas turbocharger, one or more of an engine torque, a pressure ratio on a compressor side of the exhaust gas turbocharger, an air mass flow on a compressor side of an exhaust gas turbocharger, an accelerator pedal position, and / or an engine speed. If the technical system is, for example, a hydraulic pump, the at least one operating parameter comprises one or more of a fluid pressure of the hydraulic pump, a drive speed of the hydraulic pump, a volume flow through the hydraulic pump, and / or an oil temperature in the hydraulic pump.

[0048] The operating parameters are typically already widely used for monitoring technical systems and are available without further effort or additional components.

[0049] A computing unit according to the invention, e.g. a control unit of a motor vehicle, is, in particular in terms of programming, equipped to carry out a method according to the invention.

[0050] Implementing a method according to the invention in the form of a computer program or computer program product with program code for carrying out all method steps is also advantageous, as this incurs particularly low costs, especially if an executing control unit is already available for other tasks. Finally, a machine-readable storage medium is provided with a computer program stored on it as described above. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical, and electrical storage media, such as hard drives, flash memory, EEPROMs, DVDs, etc. Downloading a program via computer networks (Internet, intranet, etc.) is also possible. Such a download can be wired or wireless (e.g., via a WLAN network, a 3G, 4G, 5G, or 6G connection, etc.).

[0051] Further advantages and embodiments of the invention will become apparent from the description and the accompanying drawing.

[0052] The invention is schematically illustrated in the drawing using an exemplary embodiment(s) and is described below with reference to the drawing. Brief description of the drawings Fig. Figure 1 schematically shows a technical system in which the invention can be used. Fig. Figure 2 shows a flowchart of an embodiment of a method for detecting a health condition, Fig. Figure 3 shows sound measurement data from various sensors as well as the progression of the values ​​of two operating parameters. Fig. Figure 4 shows a flowchart of an embodiment of a method for training a machine learning model, Fig. Figure 5 schematically shows sound measurement data that were clustered based on two operating parameters, and Fig. Figure 6 shows the evaluation of sound measurement data by a machine learning model that was trained using an embodiment of a method for training a machine learning model. embodiment(s) of the invention

[0053] In Fig. Figure 1 schematically shows a technical system 1 designed as an internal combustion engine in which the invention can be used. A section of the internal combustion engine 1 is schematically depicted, including a cylinder or cylinder block 150 and an exhaust gas turbocharger 100 as a component. The cylinder block 150 has an air inlet 151 for supplying fresh air and an exhaust manifold 152 for removing exhaust gas.

[0054] The exhaust gas turbocharger 100 comprises a turbine 120 and a compressor 110 driven by it. Fresh air 10 is supplied to the compressor 110, compressed by it to a boost pressure, and supplied to the cylinder block 150 via an optional air cooler 11. A fuel / air mixture is combusted in the cylinder block 150, and the resulting exhaust gas is supplied to the turbine 120 via the exhaust manifold 152.

[0055] The turbine 120 comprises a turbine housing 123 with a turbine inlet 121 and a turbine outlet 122, and between them a guide vane assembly 124 with adjustable guide vanes and a turbine wheel 125. The exhaust gas enters the turbine housing 123 at the turbine inlet under pressure and flows through the turbine housing 123 and the guide vane assembly 124, with the pressure being reduced at the transition between the guide vane assembly and the turbine wheel. The exhaust gas drives the turbine wheel 125 before exiting into an exhaust system 12 at the turbine outlet 122.

[0056] The turbine wheel 125 is coupled to a 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 speed denoted here by n. The shaft 126 is mounted on a hydrodynamic sliding bearing.

[0057] The exhaust gas turbocharger shown here is only an example; for instance, an exhaust gas turbocharger with a wastegate could also be used.

[0058] Furthermore, a computing unit 170 designed as an engine control unit is shown, as well as a structure-borne sound or knock sensor 160, which is arranged on a housing of the exhaust gas turbocharger 100.

[0059] Fig. Figure 2 shows a flowchart of an embodiment of a method for detecting a health condition of at least one sound-emitting component of a technical system. Fig. Figure 3 shows sound measurement data from various sensors as well as the progression of the values ​​of two operating parameters.

[0060] In the procedure for detecting the health status of at least one sound-emitting component of a technical system, input data for a machine learning model is provided in step 200. The machine learning model comprises a multitude of sub-machine learning models, each configured to evaluate input data within a predetermined operating range. The input data is determined based on sound measurement data and at least one operating parameter of the technical system.

[0061] An example set of input data is shown in Fig. Figure 3 shows the horizontal axis (300) representing time, and the vertical axis representing the value of the recorded data. The vertical lines within the graphs illustrate the data windowing, i.e., the division of the measurement data into data windows of a predefined time period. In the example shown, the Fig. For example, in graph 3, the time duration is 0.1 seconds, meaning that 0.1 seconds elapse from time point 301 to time point 302. All data are recorded synchronously, meaning they refer to the same time period. Graphs (a) and (b) show sound measurement data from two knock sensors located on the internal combustion engine 1. Graphs (c) and (d) show sound measurement data from two knock sensors located on the exhaust gas turbocharger 100. Graphs (e) and (f) show data from two microphones located near the internal combustion engine 1. Graph (g) shows the torque curve of the internal combustion engine 1, and graph (h) shows the rotational speed of the internal combustion engine. Graphs (a) to (f) thus represent sound measurement data, while graphs (g) and (h) represent values ​​of two operating parameters.

[0062] In the procedure, in step 201, the sound measurement data of at least the sensors (knock sensors and microphones) are recorded during the operation of the technical system, i.e. the exhaust gas turbocharger 100.

[0063] Simultaneously, i.e., synchronously with time, in step 202 the values ​​of the two operating parameters, i.e., the torque and the speed, of the technical system are recorded.

[0064] In step 203, the sound measurement data and the values ​​of the operating parameters are divided into time-synchronized data windows. The data windows are divided into Fig. 3. This is represented by vertical lines. The data windows in each of the sound measurement data and operating parameters cover the same time period, i.e., they are time-synchronized.

[0065] Subsequently, in step 204, an operating point range is determined for each data window, depending on the operating parameters of the data window. In this case, the operating point range in which the internal combustion engine 1 was located during data acquisition is determined.

[0066] In step 205, a spectral power density is determined for each of the data windows based on the sound measurement data, and in step 206, one or more spectrograms are determined as input data.

[0067] Subsequently, in step 210, output data is determined based on the input data and the machine learning model. This output data includes information about the number of occurrences of at least one characteristic parameter for damage to the technical system in the input data. For example, the output data may contain information indicating whether and / or to what extent a data point of the sound measurement data deviates from the known pattern.

[0068] In step 220, the output data for detecting the health status of at least one component of the technical system is provided and can be retrieved, for example, by an anomaly detection unit, which may be part of the same or another computing unit.

[0069] Then, in step 230, based on the output data, the health status of at least one component of the technical system is determined. If the anomaly value remains above a predetermined threshold for a specified period, it can be detected that a part of the exhaust gas turbocharger is damaged. Depending on the deviation from the threshold, it can be determined, for example, whether the damage is minor, severe, or a total component failure.

[0070] Fig. Figure 4 shows a flowchart of an embodiment of a method for multi-stage training of a machine learning model for use in detecting a health condition of at least one sound-emitting component of a technical system.

[0071] In step 400 of the procedure, training data is provided, consisting of sound measurement data, at least one operating parameter of the technical system, and information 200 that includes at least the state of health of the technical system during the recording of the sound measurement data. The sound measurement data and the values ​​of the at least one operating parameter were recorded synchronously; see, for example, also Fig. 3.

[0072] The information that can be included in a training dataset includes, for example, details about the measurement location (on the road or on a roller), the measurement program used, and the condition of the noise-emitting component. For instance, the condition can be categorized as good (i.e., no damage) or light, moderate, and severe damage. Furthermore, the type of damage can also be included in the information. In particular, the noise measurement data can also contain noise measurements recorded under different types of damage, allowing the machine learning model to distinguish between these different types of damage.

[0073] The training data is then preprocessed in step 410. During preprocessing in step 410, the training data is divided into data windows with a predefined time duration, and for each data window, it is determined, depending on at least one operating parameter, whether the technical system was in a steady state.

[0074] In step 411, a change measure is first determined for each of the at least one operating parameter. Equation (1) can be used, for example, to determine the change measure.

[0075] Subsequently, in step 412, it is determined for each of the at least one operating parameter whether the rate of change is less than a predefined threshold, and in step 413, it is determined that steady-state operation exists if, for each of the at least one operating parameter, it is determined that the rate of change is less than the predefined threshold. This condition is, for example, represented by equation (2).

[0076] In step 414, indicator parameters are determined from the sound measurement data. These indicator parameters can include, for example, the flatness of the spectral power density or other indicator parameters.

[0077] In step 420, the preprocessed sound measurement data of the data windows for which it was determined that the operation was steady are clustered into different operating point ranges depending on at least one operating parameter.

[0078] Fig. Figure 5 schematically shows sound measurement data clustered based on two operating parameters: the motor speed (horizontal axis 501) and the torque (vertical axis 502). In the example shown, three different operating point ranges 510, 520, and 530 are marked as examples.

[0079] Subsequently, in step 430, the machine learning model for each cluster is trained based on the indicator sizes of the clustered sound measurement data such that, based on the sound measurement data as input data for the machine learning model, output data is determined that includes information about how far a data point of the sound measurement data of the input data deviates from a known pattern, and in step 440 the trained machine model is output.

[0080] Fig.Figure 6 shows the evaluation of sound measurement data by a machine learning model that was trained using an embodiment of a method for training a machine learning model. Time is plotted on the horizontal axis 601, and the information on whether a data point of the sound measurement data deviates from a known pattern is plotted on the vertical axis 602. This information 603 is represented as a numerical value that takes on values ​​of 0 or 1, where 0 indicates no deviation from the known pattern and 1 indicates a deviation from the known pattern. The anomaly value 604 is the moving average of the values ​​of this information.

[0081] Graph (a) shows information 603 and anomaly value 604 for data used to train the machine learning model and which were labeled "GOOD". It can be seen that anomaly value 604 consistently has a low value (<0.3) and therefore indicates that the health of the sound-emitting component is good.

[0082] Graph (b) shows information 603 and anomaly value 604 for data that was not used to train the machine learning model but was recorded from a sound-emitting component that shows no damage. It can be seen that anomaly value 604 is consistently low (<0.4) and, except for a few peaks, remains below 0.5. Therefore, based on information 603 output by the machine model and the resulting anomaly value 604, it can be concluded that the sound-emitting component is in good condition.

[0083] Graph (c) shows information 603 and anomaly value 604 for data not used to train the machine learning model. This data was recorded from a sound-emitting component that sustained damage between a first time point 605 and a second time point 606. It can be seen that before the first time point 605 and time point 605, anomaly value 604 is mostly below 0.5. After the first time point 605 and time point 605, anomaly value 604 increases sharply, reaching a value of 1 at times. Therefore, information 603, output by the machine learning model, indicates that the sound-emitting component has a serious defect.

[0084] The component for which the sound measurement data were recorded is an exhaust gas turbocharger, and from the first time point 605 in graph (c) there is damage to a bearing of the exhaust gas turbocharger and the shaft has suffered a break.

Claims

[1] Method for detecting the health status of at least one sound-emitting component (100) of a technical system (1), comprising: - Providing (200) input data for a machine learning model, wherein the machine learning model comprises a plurality of sub-machine learning models, each sub-machine learning model being set up to evaluate input data in a predetermined operating range, wherein the input data have been determined based on sound measurement data and at least one operating parameter of the technical system, - Determine (210), based on the input data and the machine learning model, output data, wherein the output data includes information about whether and / or how far a data point of the sound measurement data of the input data deviates from a known pattern, and - Providing (220) the output data for detecting the health status of at least one component of the technical system. [2] The method of claim 1, wherein the provision (200) of input data for the machine learning model comprises: - Recording (201) the sound measurement data of at least one sensor during operation of the technical system, - Recording (202) values ​​of at least one operating parameter of the technical system, wherein the at least one operating parameter is recorded synchronously with the sound measurement data, - Splitting (203) the sound measurement data and the values ​​of at least one operating parameter into time-synchronous data windows, - Determine (204), for each data window, an operating point range depending on at least one operating parameter of the data window. [3] The method of claim 2, wherein the provision (200) of input data for the machine learning model further comprises: - Determine (205), for each of the data windows, a spectral power density, based on the sound measurement data, and - Determine (206), as input data, one or more spectrograms as a function of the spectral power density. [4] Method according to any one of claims 1 to 3, further comprising: Determine (230), based on the output data, the health status of at least one component of the technical system. [5] Method for multi-stage training of a machine learning model for use in detecting the health status of at least one sound-emitting component of a technical system, wherein the component performs a periodic movement during the operation of the technical system, comprising: - Providing (400) training data, wherein the training data includes at least sound measurement data and at least one operating parameter of the technical system, wherein the sound measurement data and values ​​of the at least one operating parameter have been recorded in a time-synchronous manner, - Preprocessing (410) of the training data, wherein during preprocessing the training data are divided into data windows with a predetermined time duration and for each data window it is determined, depending on at least one operating parameter, whether the technical system was in steady-state operation when the training data were acquired, - Clustering (420) the preprocessed sound measurement data of the data windows for which the operation was found to be steady-state, depending on at least one operating parameter into different operating point ranges, - Training (430) the machine learning model for each cluster based on the clustered sound measurement data such that, based on the sound measurement data as input data for the machine learning model, output data are determined which include information about how far a data point of the sound measurement data of the input data deviates from a known pattern, and - Deploying (440) the trained machine model. [6] Method according to claim 5, wherein the preprocessing (410) comprises for each data window: Determine (411) a measure of change for each of the at least one operating parameter, Determine (412) for each of the at least one operating parameter whether the rate of change is less than a specified threshold, and Determine (413) that steady state exists if for each of the at least one operating parameter it is determined that the rate of change is less than the specified threshold. [7] Method according to claim 6, wherein the preprocessing (410) further comprises: Extracting (414) indicator quantities from the sound measurement data, where the training (430) of the machine learning model is performed for each cluster based on the indicator sizes of the clustered sound measurement data. [8] Method according to any one of claims 5 to 7, wherein in clustering (420) the sound measurement data are grouped into a variable or fixed number of operating point ranges using the at least one operating parameter. [9] Method according to claim 8, wherein in clustering (420) the sound measurement data are grouped into a variable number of operating point ranges using the at least one operating parameter and a distance-based clustering method is used for clustering (420). [10] Method according to claim 9, wherein each of the at least one operating parameter for clustering (420) is assigned a weighting factor indicating the relevance of the operating parameter in clustering. [11] Method according to one of the preceding claims, wherein the at least one operating parameter comprises one or more of the following operating parameters: an engine torque, a pressure ratio on a compressor side of an exhaust gas turbocharger, an air mass flow on a compressor side of an exhaust gas turbocharger, an accelerator pedal position, an engine speed, a fluid pressure of a hydraulic pump, a drive speed of the hydraulic pump, a volume flow through the hydraulic pump, an oil temperature in the hydraulic pump. [12] Computing unit configured to perform all the process steps of a process according to any of the preceding claims. [13] Computer program that causes a computing unit to perform all the process steps of a method according to any one of claims 1 to 11 when executed on the computing unit. [14] Machine-readable storage medium with a computer program stored thereon according to claim 13.

Images