Method and system for monitoring a plain bearing

Abstract

The present invention relates to a method (10) for monitoring a plain bearing (50) of a wind turbine, a steam turbine or a compressor. The method comprises at least the following steps: - detecting (12) a structure-borne-sound signal of the plain bearing (50) by means of a structure-borne-sound sensor (52); - transforming (16) the detected structure-borne-sound signal into a time-frequency domain, in particular by continuous wavelet transform; - identifying (18) anomalies in the transformed structure-borne-sound signal by means of an autoencoder; - generating (20) anomaly data on the basis of the identified anomalies. The invention also relates to a system (40) for monitoring a plain bearing (50).

Description

[0001] Method and system for monitoring a plain bearing

[0002] Technical area

[0003] The present invention relates to a method for monitoring a journal bearing of a wind turbine, a steam turbine, or a compressor. Furthermore, the invention relates to a system for monitoring a journal bearing of a wind turbine, a steam turbine, or a compressor.

[0004] State of the art

[0005] Wind turbines, steam turbines, and compressors typically use plain bearings, for example, to support a rotor and / or as part of a planetary gear system. Due to the frequent and often unpredictable start-up and shutdown processes, particularly in these systems and devices, the plain bearings are often in a critical operating condition in terms of wear and fatigue. As a result, an unexpectedly early failure of the plain bearing can occur. This results in unwanted downtime, which can incur high costs and requires flexible repairs. However, especially with offshore wind turbines, timely repairs are usually not possible due to bad weather and long distances. To prevent unexpected failure, the corresponding plain bearings can be replaced frequently and on schedule.However, this also results in the replacement of plain bearings that are not at all worn and are therefore still suitable for extended use. Furthermore, a planned replacement also prevents use for a certain period of time, resulting in high costs.

[0006] Another option is condition monitoring of the plain bearing to detect wear and fatigue at an early stage. This allows for needs-based maintenance before or immediately after a bearing failure and before further consequential damage occurs. However, condition monitoring can be difficult and unreliable due to dynamically changing operating conditions and vibrations. In addition, the analysis of large amounts of data may be necessary, which can make the monitoring itself very complex. Bearing monitoring can, for example, be based on temperature monitoring. However, this form of monitoring is only suitable for certain bearing types. For example, condition monitoring of plain bearings with a polymer running layer can be very unreliable because the plastics used in these bearings have very poor thermal conductivity.As a result, areas with excessive temperatures that are relevant for failure cannot be detected or can only be detected very poorly, and despite monitoring, the plain bearing may fail without warning.

[0007] Description of the invention

[0008] A first aspect of the invention relates to a method for monitoring a plain bearing of a wind turbine, a steam turbine, or a compressor. A plain bearing can be a bearing type in which parts moving relative to one another can have direct contact depending on the operating point. At high relative speeds, the parts moving relative to one another can be separated by a lubricating film. The plain bearing can, for example, rotatably support a rotor of the wind turbine, the steam turbine, or the compressor. The plain bearing can, for example, be a main bearing, a gearbox bearing, or a peripheral bearing. The plain bearing can, for example, be part of a transmission, in particular a planetary gear. For example, the plain bearing can rotatably support a planetary gear on a planetary pin of a planetary carrier. A wind turbine can be a system that converts kinetic energy from wind into electrical energy.For example, the wind turbine can have a tower with a rotor rotatably mounted thereon. A steam turbine can be a heat engine for converting thermal energy from steam into rotational energy. This can be used to drive a generator to generate electricity. The steam turbine can, for example, be designed as part of a power plant. The steam turbine can also be designed as a gas turbine for other fluids. A rotor of the steam turbine can, for example, have respective turbine blades. A compressor can be a machine that supplies mechanical work to a gas, for example by compression. The compressor can be designed as part of a power plant or an aircraft engine. The journal bearing can, for example, also form part of a turbine of a hydroelectric power plant. The method comprises a step of detecting a structure-borne sound signal of the journal bearing using a structure-borne sound sensor.Structure-borne sound can be sound that propagates through a solid body. For example, the structure-borne sound can be in the form of a longitudinal wave and / or a transverse wave. The structure-borne sound can correspond to a vibration of at least a portion of the plain bearing or of the entire plain bearing. The structure-borne sound signal can contain frequency information and / or amplitude information. The structure-borne sound signal can, for example, be recorded continuously or discontinuously. The structure-borne sound signal can be used to draw conclusions about the operating condition, wear, and even damage to the plain bearing. During analytical evaluation, however, it may be necessary to separate components of the structure-borne sound signal containing information regarding the bearing condition from components of the structure-borne sound signal that are caused by noise or irrelevant environmental influences.

[0009] The structure-borne sound sensor can be designed, for example, as a piezoelectric element and / or a contact microphone. To detect the structure-borne sound signal, the structure-borne sound sensor can be attached to a part of the plain bearing or another component in contact with it. Attaching it to the plain bearing can result in particularly good signal quality. For example, the structure-borne sound sensor can be mounted on one of the bearing pins of the wind turbine's planetary gear or on a gear housing of the wind turbine's planetary gear.

[0010] The method can also detect multiple structure-borne sound signals, for example, from different plain bearings. For example, a dedicated structure-borne sound sensor can be provided for each plain bearing. The method can also be configured to monitor multiple plain bearings simultaneously.

[0011] The method comprises a step of transforming the acquired structure-borne sound signal into a time-frequency domain. The transformed structure-borne sound signal can, for example, allow an amplitude to be assigned to a respective time-frequency or a respective frequency range within a specific time period. The signal can be acquired with high locality in the time and frequency domains. The transformed structure-borne sound signal can allow a visual representation of the acquired structure-borne sound signal and / or processing and storage of the acquired signal in an image format. The transformation can significantly compress the respective structure-borne sound information. This can significantly reduce the effort required for data processing and / or data transmission. The transformation can be performed, for example, using continuous wavelet transformation. Other transformations are also possible, for example, generally using wavelet transformation.The wavelet transform can be a linear time-frequency transform. The wavelet transform can be composed of a wavelet analysis, which describes a conversion of a time representation into a wavelet representation, and a wavelet synthesis, which describes an inverse transformation of the wavelet transform into the time representation. The wavelet transform can be used for data compression and signal processing. The time-frequency domain can greatly simplify the automatic evaluation of structure-borne sound signals. Compared to a Fourier transform, a continuous wavelet transform is capable of capturing signals with a high degree of locality in the time and frequency domain, as is typically the case with structure-borne sound signals, with high quality and efficiency.

[0012] The method includes a step of identifying anomalies in the transformed structure-borne sound signal using an autoencoder. Using the autoencoder can eliminate the need for complex analytical filtering of the structure-borne sound signal, for example, to separate portions of the structure-borne sound signal containing information regarding a bearing condition from portions of the structure-borne sound signal that correspond to noise or are caused by irrelevant environmental influences. An anomaly can, for example, be an operating condition that does not correspond to a normal condition and / or an operating condition with increased wear. An anomaly can also be a condition in which the plain bearing has failed and / or is beginning to wear severely. The anomaly can be a wear-critical and / or fatigue-critical condition of the plain bearing.For example, an anomaly can be a condition in which the relatively moving parts of the journal bearing are in a state of solid friction with each other, even though the journal bearing's rotational speed should already generate a stable oil film. An autoencoder can be an artificial neural network. The autoencoder can be designed to reduce the dimension of the transformed structure-borne sound signal. The autoencoder can be trained to ignore insignificant data when identifying anomalies. For example, the autoencoder can only consider representations of respective anomalies in the transformed structure-borne sound signal. The autoencoder can, for example, be designed as a deep convolutional autoencoder, which has a CNN as a neural network. The CNN can, for example, learn a structure in the input data.Convolutional filters can be used to capture local patterns, and the transformed structure-borne sound signals can be treated like images by the CNN. Respective features can be correlated locally, i.e., in terms of time and frequency. The autoencoder can be configured to classify the transformed structure-borne sound signal as an anomalous state or a normal state in order to identify anomalies. The autoencoder can receive the transformed structure-borne sound signal as input data. The autoencoder can reconstruct the input data in its neural network, with significant deviations in the result indicating an anomaly. These deviations can be logged and / or output. A degree of deviation can be output as an anomaly characteristic. Identification can involve detecting anomalies and / or classifying anomalies.The autoencoder can be designed as an unsupervised trained autoencoder, which makes its training and thus its implementation particularly cost-effective.

[0013] Using the autoencoder to identify anomalies in the transformed structure-borne sound signal can enable efficient automatic monitoring of the sliding bearing. Only small amounts of data and low computing power are required. Complex human evaluation, human monitoring, and analytical data analysis of the structure-borne sound signal are not necessary, for example. The autoencoder can be particularly efficient at evaluating signals in the time-frequency domain. Furthermore, it is not necessary to provide features in the structure-borne sound signals that can be used to identify anomalies. Rather, the autoencoder can detect and / or learn anomalies based on known normal conditions. No extraction of features from the data set itself is necessary to identify anomalies.The autoencoder can automatically extract relevant information from the transformed structure-borne sound signal, despite numerous interfering signals. Early damage detection based on structure-borne sound signals allows for very early detection of wear in the plain bearing. For example, damage detection is possible using the following options in descending order: structure-borne sound-based, vibration-based, bearing oil contamination-based, airborne sound-based, temperature-based, smoke-based, and based on an actual failure. Clearly, structure-borne sound-based monitoring enables particularly early warning of an impending failure, especially earlier than with other methods.

[0014] The method is particularly suitable for wind turbines, steam turbines, and compressors. Due to external and / or random influences, such as changes in the fluid properties in the compressor and steam turbine, as well as turbulence in the wind turbine, analytical evaluation is extremely complex and time-consuming, and in some cases even impossible. Furthermore, structure-borne noise is often very device-specific. For example, two identical wind turbines may exhibit different structure-borne noise signals in normal operation under the same operating conditions. Nevertheless, the autoencoder can easily learn to identify anomalies in each case. Until now, structure-borne noise-based monitoring of plain bearings in such systems and devices was either not possible or would not have been economically viable.

[0015] The method comprises a step of generating anomaly data depending on identified anomalies. The anomaly data can, for example, be the data logged and / or output by the autoencoder. The anomaly data can be or comprise a set of transformed or untransformed structure-borne sound signals in which an anomaly was identified. The anomaly data can comprise a number of identified anomalies, i.e., a type of counter. The number can be cumulative. The anomaly data can comprise the degree of deviation as an anomaly characteristic. The anomaly data can, for example, comprise information regarding the number, time, nature, type, and / or severity of the identified anomaly. The anomaly data can be output by the autoencoder or by an evaluation device controlled depending on the anomalies identified by the autoencoder.The anomaly data can correspond to the output data of the autoencoder. The anomaly data can include warning information regarding a bearing condition, outputs of bearing failure probabilities, and / or other metrics.

[0016] The system can have a control device that generates control data depending on the anomaly data. The control device can be configured to use the control data to control the wind turbine or, in general, the device whose plain bearings are being monitored. Alternatively or additionally, the control device can be configured to control an output device using the control data. For example, a shutdown or a reduction in maximum power can occur in the event of certain anomaly data. The output device can be configured to control a display depending on the control data. For example, the anomaly data can be output visually or acoustically.

[0017] In one embodiment of the method, the autoencoder is designed as a deep convolutional autoencoder. This type of autoencoder can be particularly efficient in evaluating the transformed structure-borne sound signals. Furthermore, this type of autoencoder can learn particularly reliably to ignore interfering signals.

[0018] In one embodiment of the method, the autoencoder is trained to identify deviations from a system-specific normal state as anomalies. The normal state can be a state that is non-critical for wear and / or fatigue of the plain bearing. For simple implementation, it can be assumed, for example, when training the autoencoder that a newly installed or freshly serviced plain bearing is in its normal state. The autoencoder can therefore be trained, for example, directly after commissioning of the wind turbine or other device with the plain bearing to be monitored, with the first minutes, hours, days, or even weeks of operation. By taking a system-specific normal state into account, a respective system-specific structure-borne sound transmission behavior can be taken into account in order to identify anomalies particularly reliably.For example, the autoencoder can be trained only with the normal state. Transformed structure-borne sound signals from a system-specific normal state of the plain bearing can be used to train the autoencoder. Alternatively, the autoencoder can be trained on a test bench, with historical data from other systems and / or devices, and with simulated data. The autoencoder can be trained, or have been trained, using a backpropagation method, for example. The input data for training can also be initially weighted to improve training.

[0019] In one embodiment of the method, the method further comprises a step of digitizing the structure-borne sound signal prior to its transformation. The structure-borne sound signal can, for example, be acquired as an analog signal. Digitization simplifies further data processing. Digitization can, for example, occur prior to transforming the structure-borne sound signal into the time-frequency domain.

[0020] In one embodiment of the method, it is provided that only structure-borne sound signals in a frequency range greater than 30 kHz, in particular greater than 50 kHz and / or less than 1.3 MHz, in particular less than 1 MHz, are used to detect anomalies. Below 30 kHz, in particular below 50 kHz, structure-borne sound signals are not very suitable for identifying anomalies. Above 1.3 MHz, in particular above 1 MHz, the structure-borne sound signals contain hardly any or no additional information relating to anomalies. By only taking a corresponding frequency range into account, the reliability of anomaly identification can be particularly high. In addition, the amount of data to be evaluated can be kept small. The frequency range can be selected, for example, during digitization and / or transformation.However, the frequency range can also be specified directly during the acquisition of the structure-borne sound signal, for example, through filtering. Signals outside the frequency range, for example, are no longer fed to the autoencoder as input data. Alternatively, the structure-borne sound sensor can be designed to only capture structure-borne sound signals in the desired frequency range.

[0021] In one embodiment of the method, the structure-borne sound signals are recorded discontinuously at specific time intervals. Discontinuous recording minimizes the data volume and the effort required for analysis. With continuous recording or monitoring, however, anomaly identification by the autoencoder can be difficult.

[0022] For example, the time interval is 0.1 to 120 seconds, especially 1 second to 60 seconds. With such a measurement duration, for example, one anomaly can be easily detected and two anomalies can be easily prevented from being detected simultaneously, especially under typical operating conditions of wind turbines. This facilitates further analysis. Furthermore, it can minimize monitoring effort and power consumption.

[0023] For example, a measurement duration ranges from 0.1 to 64 milliseconds, in particular from 1 to 32 milliseconds. With such a measurement duration, anomalies in the operation of plain bearings, especially in wind turbines, can typically be detected particularly well. For example, such a measurement duration is sufficient to reliably detect wear- and fatigue-critical conditions in the recorded structure-borne sound signal. At the same time, this short measurement duration reduces the risk of changes in the operating condition, for example, due to gusts, falsely indicating an anomaly.

[0024] In one embodiment of the method, it is provided that the method comprises a step of calculating wear and / or fatigue of the plain bearing using a recurrent neural network. The recurrent neural network can receive the generated anomaly data as input data and generate a wear characteristic value and / or a fatigue characteristic value as output data. A wear calculation can, for example, be designed to output a remaining service life of the plain bearing. A fatigue calculation can, for example, be designed to output a probability of failure due to material fatigue of the plain bearing. Using the recurrent neural network, a forecast can be made, for example, as to when the plain bearing will fail, and in particular with what probability.On this basis, maintenance and / or replacement of the plain bearing can be planned to enable cost-effective and reliable operation. In one embodiment of the method, the recurrent neural network receives additional condition data, such as an operating temperature and / or a rotational speed, as input data. The determined wear characteristic and / or the determined fatigue characteristic can thus be particularly meaningful. This allows operating conditions of the plain bearing to be taken into account. The condition data can, for example, contain information about an environmental condition of the plain bearing, an operating condition of the plain bearing, and / or an operating condition of the system or device in which the plain bearing is used.

[0025] In one embodiment of the method, it is provided that the recurrent neural network has been trained using a training method. The training method can comprise a step of providing anomaly data. The anomaly data can be, for example, historical and / or simulated anomaly data, including from other systems. The anomaly data provided for training can relate to a plain bearing of the same type or a different type than the monitored plain bearing. The training method can comprise a step of providing wear and / or fatigue values ​​corresponding to the anomaly data. The wear and / or fatigue values ​​corresponding to the anomaly data can be, for example, corresponding historical and / or simulated wear and / or fatigue values.The provided data and parameters can also have been determined on a test bench. The training method can include a step of generating a training dataset. The training dataset can include the provided anomaly data as input data. The training dataset can include the provided wear parameters and / or the provided fatigue parameters as output data. The training method can include a step of training the recurrent neural network to learn a relationship between the anomaly data and the wear parameters and / or fatigue parameters.

[0026] In one embodiment of the method, the transformation of the structure-borne sound signal takes place locally, in particular by a local computing device. The local computing device can, for example, be a computer that is installed on the wind turbine, steam turbine, or compressor and / or is directly connected by cable to sensors and / or control elements of this system or device. By transforming, the structure-borne sound signal can, for example, be compressed, whereby data can be easily transmitted for evaluation in another computing device. For example, further evaluation can be carried out by a central server, in particular for a plurality of plain bearings from different systems or devices. This easily enables central monitoring and / or maintenance planning. Digitization can also take place locally, in particular by the local computing device.

[0027] The method can further comprise a step of sending the transformed structure-borne sound signal to a central server, with the autoencoder for identifying anomalies being implemented on the central server. This allows for a particularly high level of computing power to be made available there cost-effectively. Furthermore, the method can be particularly efficient if multiple plain bearings of different systems or devices are being monitored, since the central server can then be fully utilized even if, for example, only intermittent recording is performed. Alternatively, the autoencoder can also be implemented on the local computing device.

[0028] The method may alternatively or additionally further comprise a step of sending the generated anomaly data to the central server. In particular, the recurrent neural network for calculating wear and / or fatigue can be implemented on the central server. Thus, the anomaly data is directly available centrally for further analysis. Alternatively, the recurrent neural network can also be implemented on the local computing device.

[0029] Transmission can occur, for example, via radio, in particular mobile radio. Transmission can also occur via cable. For example, common transmission protocols and networks can be used, such as mobile radio and / or the Internet. A data transmission device can be provided for transmission. The data transmission device can also be designed for reception. For example, respective training data for the autoencoder and / or the recurrent neural network can also be received via the transmission device. A second aspect relates to a system for monitoring a plain bearing of a wind turbine, a steam turbine, or a compressor. The system can be designed to carry out a method according to the first aspect. Corresponding features and advantages of the first aspect can also form features and advantages of the further aspect, and vice versa.

[0030] The system includes a structure-borne sound sensor. The structure-borne sound sensor is designed to detect a structure-borne sound signal from the plain bearing. Multiple structure-borne sound sensors can also be provided, which together detect the structure-borne sound signal from the plain bearing. For example, one or more structure-borne sound sensors can be attached to a bearing housing and / or a planetary pin of a planetary gear of a wind turbine. Respective structure-borne sound sensors can be designed to be attached to a shaft and / or the plain bearing.

[0031] The system comprises a computing system. The computing system may comprise one or more computing devices, which are communicatively connected to one another, for example. A computing device may be embodied, for example, as a computer or microprocessor. The computing system is configured to transform the detected structure-borne sound signal into a time-frequency domain, identify anomalies in the transformed structure-borne sound signal using an autoencoder, and generate anomaly data depending on the identified anomalies. The computing system may comprise a data transmission device, such as a network card, a data storage device, such as a hard disk, and / or an output device, such as a display.

[0032] In one embodiment of the system, the computing system comprises a local computing device installed on the wind turbine, the steam turbine, or the compressor. The local computing device can be compact and cost-effective. Suitable examples are a Raspberry Pi computer and an Arduino development board. The computing system can comprise a central server and a data transmission device for transmitting data from the local computing device to the central server. The local computing device can be configured to transform the detected structure-borne sound signal into a time-frequency domain. The local computing device can be configured to digitize the detected structure-borne sound signal. The central server or the local computing device can be configured to perform a wear calculation and / or a fatigue calculation of the plain bearing using a recurrent neural network.The data transmission device can, for example, transmit respective outputs of the local computing device to the local server, for example, transformed structure-borne sound signals, anomaly data, and / or wear and / or fatigue characteristics. The respective data can be stored on the data storage device. The data storage device can, for example, comprise a memory module of the central server and / or another memory module of the local computing device. Data can be stored locally or centrally, or redundantly both locally and centrally.

[0033] The local computing device can be configured to identify anomalies in the transformed structure-borne sound signal using an autoencoder and to generate anomaly data depending on the identified anomalies. The autoencoder can be implemented on the local computing device. In this case, for example, the respective anomaly data can be transmitted to the central server. Optionally, the recurrent neural network can also be implemented on the local computing device. In this case, for example, the respective wear and / or fatigue values ​​can be transmitted to the central server.

[0034] Alternatively, the central server can be configured to identify anomalies in the transformed structure-borne sound signal using an autoencoder and generate anomaly data depending on the identified anomalies. The autoencoder can be implemented on the central server. In this case, for example, the respective transformed structure-borne sound signals can be transmitted to the central server. Optionally, the recurrent neural network can be implemented on the central server in this case.

[0035] Brief description of the figures Fig. 1 illustrates a method for monitoring a journal bearing of a wind turbine, a steam turbine or a compressor.

[0036] Fig. 2 schematically illustrates a system for monitoring a journal bearing of a wind turbine, a steam turbine or a compressor.

[0037] Detailed description of embodiments

[0038] Fig. 1 illustrates a method 10 for monitoring a plain bearing 50 of a gearbox of a wind turbine, a steam turbine, or a compressor, wherein the plain bearing 50 is schematically illustrated in Fig. 2. In step 12, a structure-borne sound signal of the plain bearing 50 is detected by means of a structure-borne sound sensor 52 attached thereto, wherein the structure-borne sound sensor 52 is also schematically illustrated in Fig. 2.

[0039] The recorded structure-borne sound signal is digitized in a step 14 of method 10. The digitized structure-borne sound signal is transformed into a time-frequency domain using a continuous wavelet transformation in a step 16 of method 10. In a step 18 of method 10, anomalies in the transformed structure-borne sound signal are identified using an autoencoder. Depending on the identified anomalies, anomaly data is generated in a step 20 of method 10. This generated anomaly data is provided to a recurrent neural network as input data. The recurrent neural network is trained to generate a wear characteristic value and / or a fatigue characteristic value as output data in a step 22 of method 10 in response to this input data in order to perform a wear calculation and / or a fatigue calculation of the plain bearing 50 using the recurrent neural network.This makes it possible to predict a failure of the plain bearing 50 at an early stage with little effort and / or to calculate a remaining service life of the plain bearing 50 in order to plan maintenance and / or replacement of the plain bearing 50 accordingly.

[0040] Fig. 2 schematically illustrates a system 40 for monitoring the plain bearing 50. The system 40 is designed to carry out the method 10. The system 40 accordingly has the structure-borne sound sensor 52. The system 40 also has a computing system 42. The computing system 40 has a computing device 44 installed locally on the wind turbine, the steam turbine, or the compressor. The local computing device 44 has a module 54 for receiving the detected structure-borne sound signal, which is communicatively connected to the structure-borne sound sensor 52, for example, by radio or cable. The local computing device 44 has a module 56, by means of which the detected structure-borne sound signal received by the local computing device 44 is digitized.In addition, the local computing device 44 has a module 58, by means of which the recorded and digitized structure-borne sound signal is transformed into a time-frequency domain using continuous wavelet transformation. This transformation significantly reduces the amount of data without, or only slightly, reducing the data quality.

[0041] Furthermore, the computing system 42 has a data transmission device 46 with a transmission module 60 connected to the local computing device 44. Furthermore, the computing system 42 has a central server 48, with a further transmission module 62 of the data transmission device 46 connected to the central server 48. The transformed structure-borne sound signal can be transmitted to the central server 48 by means of the data transmission device 46. The data transmission device 46 transmits the respective data, for example, via mobile radio and / or a landline network.

[0042] The central server 48 has a module 64 on which the autoencoder is implemented and by means of which the central server 48 identifies anomalies in the transmitted and transformed structure-borne sound signal. Furthermore, the central server 48 has a module 66 on which the recurrent neural network is implemented and by means of which the central server 48 performs the wear calculation and / or the fatigue calculation of the plain bearing 50.

[0043] By using the central server 48, complex calculations can be performed centrally and cost-effectively, particularly for monitoring different plain bearings. The data transmission interface can also be provided at a different location in the computing system 42. For example, instead of the central server 48, the local computing device 44 can have the module 64 with the autoencoder implemented therein to further reduce the amount of data to be transmitted.

[0044] Reference symbol

[0045] 10 procedures

[0046] 12 Step of Capturing

[0047] 14 steps of digitization

[0048] 16 Steps of Transformation

[0049] 18 Step of Anomaly Identification

[0050] 20 Step of generating anomaly data

[0051] 22 Step of determining a wear and / or fatigue characteristic

[0052] 40 System for monitoring a plain bearing

[0053] 42 Calculation system

[0054] 44 local computing devices

[0055] 46 Data transmission device

[0056] 48 central servers

[0057] 50 plain bearings

[0058] 52 Structure-borne sound sensor

[0059] 54 Structure-borne sound signal receiving module

[0060] 56 Digitization module

[0061] 58 Transformation module

[0062] 60 local transmission module

[0063] 62 central transmission module

[0064] 64 Autoencoder module

[0065] 66 Wear and / or fatigue calculation module

Claims

Claims Method (10) for monitoring a sliding bearing (50) of a wind turbine, a steam turbine or a compressor, wherein the method comprises at least the following steps: - Detection (12) of a structure-borne sound signal from the sliding bearing (50) by means of a structure-borne sound sensor (52); - Transforming (16) the detected structure-borne sound signal into a time-frequency domain, in particular by means of continuous wavelet transformation; - Identifying (18) anomalies in the transformed structure-borne sound signal using an autoencoder; - Generating (20) anomaly data depending on identified anomalies. Method (10) according to claim 1, wherein the autoencoder is configured as a deep convolutional autoencoder. Method (10) according to claim 1 or 2, wherein the autoencoder is trained to identify deviations from a plant-specific normal state as anomalies, wherein transformed structure-borne sound signals from a plant-specific normal state of the sliding bearing (50) are used to train the autoencoder. Method (10) according to any one of the preceding claims, wherein the method (10) further comprises a step of digitizing (14) the structure-borne sound signal prior to its transformation (16). Method (10) according to any one of the preceding claims, wherein only structure-borne sound signals in a frequency range greater than 30 kHz, in particular greater than 50 kHz, and / or less than 1.3 MHz, in particular less than 1 MHz, are used for the detection of anomalies.Method (10) according to any one of the preceding claims. wherein the structure-borne sound signals are recorded discontinuously at specific time intervals, in particular wherein the time interval is 0.1 to 120 seconds, for example 1 second to 60 seconds, and / or in particular wherein a measurement duration is 0.1 to 64 milliseconds, for example 1 to 32 milliseconds. Method (10) according to any of the preceding claims, wherein the method (10) comprises a step of a wear calculation and / or a fatigue calculation (22) of the sliding bearing (50) using a recurrent neural network, wherein the recurrent neural network receives the generated anomaly data as input data and generates a wear characteristic and / or a fatigue characteristic as output data. Method (10) according to claim 7, wherein the recurrent neural network receives additional state data, such as an operating temperature and / or a rotational speed, as input data. Method (10) according to claim 7 or 8, wherein the recurrent neural network was trained using a training method comprising the following steps: - Providing anomaly data, in particular historical and / or simulated anomaly data; - Providing wear and / or fatigue parameters corresponding to the anomaly data, in particular corresponding historical and / or simulated wear and / or fatigue parameters; - Generating a training dataset from the provided anomaly data as input data and the provided wear and / or fatigue parameters as output data; and Training the recurrent neural network to learn a correlation between the anomaly data and the wear and / or fatigue parameters. Method (10) according to one of the preceding claims, wherein the transformation (16) of the structure-borne sound signal is performed locally, in particular by a local computing device (44), and the method comprises at least one of the following steps: - Sending the transformed structure-borne sound signal to a central server (48), wherein the autoencoder for identifying anomalies is implemented on the central server (48); and - Sending the generated anomaly data to the central server (48), in particular wherein the recurrent neural network for wear calculation and / or fatigue calculation is implemented on the central server (48). System (40) for monitoring a sliding bearing (50) of a wind turbine, a steam turbine or a compressor, in particular wherein the system (40) is configured to perform a method (10) according to one of the preceding claims 1 to 10, wherein the system (40) comprises a structure-borne sound sensor (52) configured to detect a structure-borne sound signal of the sliding bearing (50), and a computing system (42), wherein the computing system (42) is configured to transform the detected structure-borne sound signal into a time-frequency domain, to identify anomalies in the transformed structure-borne sound signal by means of an autoencoder and to generate anomaly data depending on identified anomalies.System (40) according to claim 10 or 11, wherein the computing system (42) comprises a local computing device (44) installed on the wind turbine, steam turbine or compressor, a central server (48) and a data transmission device (46) for transferring data from the local computing device (44) to the central server (48). wherein the local computing device (44) is configured to transform the detected structure-borne sound signal into a time-frequency domain and wherein the central server (48) or the local computing device (44) is configured to identify anomalies in the transformed structure-borne sound signal by means of an autoencoder and to generate anomaly data depending on identified anomalies.

Images