Method for computer-implemented monitoring of components of a wind turbine

Abstract

A method for computer-implemented monitoring of a component of a wind turbine. The invention relates to a method for computer-implemented monitoring of a component of a wind turbine, wherein the wind turbine is a first wind turbine (1) and the component is a first component (6, 7, 8), and wherein the method accesses a trained machine learning model (ML), which has been trained for one or more second components (6', 7', 8') of the same type of one or more second wind turbines (1'). The trained machine learning model (ML) is configured to provide an output (OU) indicating a predetermined fault (FT) occurring at a second component (6', 7', 8') of a second wind turbine (1') by processing a vibration signal (VS') in a predetermined domain measured near the second component (6', 7', 8') during operation of the second wind turbine (1'). In step i), a vibration signal (VS) in the predetermined domain measured near the first component (6, 7, 8) during operation of the first wind turbine (1) is provided. In step ii), based on one or more given kinematic parameters (KP1) of the first component (6, 7, 8) and one or more given kinematic parameters (KP2) of the second component (6', 7', 8'), the vibration signal (VS) is mapped to a corresponding vibration signal (VS') valid for the second component (6', 7', 8'). In step iii), a machine learning model (ML) is applied to the vibration signal (VS') valid for the second component (6', 7', 8'), resulting in an output (OU) indicating said predetermined fault (FT) occurring at the first component (6, 7, 8).

Description

Technical Field

[0001] The invention refers to a method and an apparatus for computer-implemented monitoring of components of a wind turbine. Furthermore, the invention refers to a corresponding computer program product and a corresponding computer program. Background Art

[0002] The present invention relates to the field of fault detection for components within wind turbines based on vibration analysis. It is known from the prior art to use machine learning models, such as neural networks, in order to detect faults in wind turbines. Such machine learning models receive as input a vibration signal of a monitored component in a wind turbine and provide an output indicating the presence or absence of a specific fault occurring in the component.

[0003] Machine learning models for detecting specific faults need to be trained with suitable training data, which includes fault data, i.e., vibration signals present when a specific fault occurs in a wind turbine. However, such fault data is rare or needs to be generated based on expensive test benches where faults (e.g., damage) are introduced to the components to be monitored. Furthermore, a machine learning model trained for one turbine type only performs well when used to monitor this turbine type and not other turbine types. Summary of the invention

[0004] It is an object of the present invention to enable monitoring of components of a wind turbine by using a machine learning model that has been trained for one or more wind turbines of another type than the monitored wind turbine.

[0005] This object is achieved by a method according to the first aspect of the invention and an apparatus according to the second aspect of the invention. The method of the invention enables computer-implemented monitoring of components of a wind turbine. A wind turbine having a monitored component is designated as a first wind turbine, and the monitored component is designated as a first component. The method accesses an already trained machine learning model, which has been trained for one or more second components of the same type of one or more second wind turbines, the one or more second wind turbines being another type of wind turbine other than the first wind turbine. That is, the training data is operational data from the one or more second components. In a preferred embodiment, each of the one or more second components of the one or more second wind turbines has a function equivalent to that of the first component of the first wind turbine. However, the function of the second component and the function of the first component may also be different. This difference is reflected by the mapping of step ii) described below. For example, the first component may refer to a planetary gear, while the type of the second component is a high-speed pinion.

[0006] The trained machine learning model is configured to provide an output referring to a predetermined fault occurring at a second component of a second wind turbine by processing a vibration signal in a predetermined domain measured near the second component during operation of the second wind turbine. This second component can be any second component among the one or more second components. In step ii) described below, reference is made to this second component. The trained machine learning model can be any known data-driven model that has been trained by machine learning using training data including vibration signals. Depending on the circumstances, machine learning can be supervised learning or unsupervised learning. In supervised learning, for the vibration signals of the training data set, it is known whether the predetermined fault exists. For unsupervised learning, this is not the case.

[0007] In step i) of the method according to the invention, vibration signals in said predetermined domain measured (being measured or measured) in the vicinity of the first component during operation of the first wind turbine are provided. The process of measuring those vibration signals may be part of step i). However, it is also possible that the vibration signals have been measured and stored in advance, so that step i) only comprises the step of reading the vibration signals that have been previously measured and stored. The vibration signals provided in step i) are based on the outputs of one or more vibration sensors. However, the vibration signals need not refer directly to the sensor outputs, but rather to signals obtained from these outputs. In particular, the vibration signals may be obtained from the sensor outputs by a transformation, for example a Fourier transform.

[0008] In step ii) of the method according to the invention, the vibration signal is mapped to a corresponding vibration signal valid for the second component based on one or more given kinematic parameters of the first component and one or more given kinematic parameters of the second component.

[0009] In step iii) of the method according to the invention, the machine learning model is applied to the vibration signal determined in step ii) to be valid for the second component. Due to the mapping in step ii), this will result in an output referring to the predetermined fault occurring at the first component. This output may be stored in a storage device and / or may be provided on a user interface (e.g., a display) in order to inform the user of the occurrence of a fault.

[0010] The present invention is based on the finding that vibration signals of different components can be mapped to each other by taking into account the kinematics of the different components. Therefore, a machine learning model applied to the mapped vibration signals will provide a correct output for the vibration signals of the monitored turbine. The inventors have found that good results can be achieved regardless of the type of mapping provided that a mapping of the vibration signals can be found. However, particularly good results can be achieved in case a linear mapping is used between the vibration signal of the first component and the vibration signal of the second component.

[0011] In a particularly preferred embodiment, the vibration signal provided in step i) is determined in the frequency domain, ie the vibration signal comprises the frequency spectrum of the vibration. However, the vibration signal can also be provided in the known cepstrum-domain or any other domain.

[0012] In another preferred embodiment of the present invention, the one or more given kinematic parameters of the first component are described by the same function type as the one or more given kinematic parameters of the second component but with different function parameters.

[0013] In another variation of the present invention, the one or more given kinematic parameters of the first component are one or more specific values ​​within the predetermined domain contained in the vibration signal measured near the first component when the predetermined fault occurs. Similarly, the one or more given kinematic parameters of the second component are one or more specific values ​​within the predetermined domain contained in the vibration signal measured near the second component when the predetermined fault occurs. For example, the one or more specific values ​​may refer to a specific frequency in the frequency domain or a corresponding value in the cepstrum domain. It is known in advance that a specific fault is associated with such a specific value.

[0014] In another preferred embodiment, the monitored component is part of a drive train of the monitored wind turbine. For example, the component may refer to a bearing supporting the rotor or to a gearbox or to a generator providing the electrical power generated by the wind turbine.

[0015] In another preferred embodiment, the predetermined fault refers to damage of a gear or damage of a bearing race or damage of a ball or roller in a ball or roller bearing.These elements may for example be comprised in one of the above components of a drive train in a wind turbine.

[0016] In another preferred variant of the invention, the machine learning model is based on one or more neural networks or principal component analysis. However, any other machine learning model may also be used.

[0017] In addition to the above method, the present invention also relates to a device for monitoring components of a wind turbine, wherein the device comprises a device for performing the above step i), a device for performing the above step ii) and a device for performing the above step iii). The device for performing step i) may include a measuring device for acquiring a vibration signal. That is, the device may include one or more vibration sensors. However, in the case where the vibration signal has been pre-stored, the device may only be implemented as a device for reading the vibration signal from a storage device. This type of device is used in a computer program product and a computer program as defined below.

[0018] The apparatus of the invention is preferably configured to perform one or more preferred embodiments of the method according to the invention.

[0019] The invention also relates to a computer program product having a program code stored on a non-transitory machine-readable carrier for performing the inventive method or one or more preferred embodiments of such method when the program code is executed on a computer.

[0020] Furthermore, the invention relates to a computer program having a program code for executing the method according to the invention or one or more preferred embodiments of such a method when the program code is executed on a computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In the following, embodiments of the present invention will be described with reference to the accompanying drawings, in which

[0022] Figure 1 shows an example of a wind turbine monitored by an embodiment of the present invention;

[0023] Figure 2 A wind turbine is shown, based on which a wind turbine has been trained to monitor Figure 1 Machine learning models of turbines;

[0024] Figure 3 is a flow chart illustrating the steps of a method according to an embodiment of the present invention; and

[0025] Figure 4 It is used to execute Figure 3 Schematic diagram of components of the method shown in FIG. DETAILED DESCRIPTION

[0026] The method described below is used to monitor components in a wind turbine. An example of such a wind turbine is shown in Figure 1 A wind turbine is designated with reference numeral 1 and comprises a tower 2 on top of which a nacelle 3 is arranged. The nacelle is connected to a rotor having a hub 4 to which three rotating blades 5 are attached. A drive train for the wind turbine is housed in the nacelle 3. Figure 1 In the embodiment of the invention, this transmission system comprises a main bearing 6 for supporting the hub of the rotor, and a generator 8 and a gearbox 7 for converting the mechanical power generated by the rotation of the rotor into electrical power. Inside the nacelle 3, three vibration sensors 9, 10 and 11 are provided. The vibration sensor 9 is positioned adjacent to the bearing 6, the vibration sensor 10 is positioned adjacent to the gearbox 7, and the vibration sensor 11 is positioned adjacent to the generator 8. The present invention is not limited to the use of exactly three vibration sensors. That is, in other embodiments, a higher or lower number of vibration sensors may be used.

[0027] In the embodiment described herein, one of the components 6 to 8 of the transmission system is monitored by processing a vibration signal detected by a sensor adjacent to the component. However, the method described below can be used for each of the components 6 to 8, so that all components in the nacelle are monitored. In another embodiment, the signals of several vibration sensors that detect the vibrations of the respective components can be used to monitor the respective components. Without loss of generality, the invention is described below for monitoring the gearbox 7 based on the vibration signal of the sensor 10.

[0028] As a prerequisite for the method described in this article, a pre-trained machine learning model already exists. This machine learning model has been trained by training data from another turbine than turbine 1. This turbine is shown in Figure 2 and is designated by reference numeral 1'. Figure 1 The turbines in FIG. 1 have the same components. However, turbine 1 ' is another type of turbine than turbine 1, which is Figure 2 Elements of the turbines in are indicated using apostrophes. Turbine 1 ' may be a turbine of the same construction as turbine 1 but from another manufacturer. In a modified embodiment, the machine learning model has been trained based on training data from several turbines 1 ' of the same type.

[0029] For the turbine 1 ', the above-mentioned machine learning model has been trained based on the vibration signals of the sensor 10', where it is known whether the vibration signal refers to the operation of the wind turbine 1 ' with the presence of a specific fault (and in particular, a specific damage) in the gearbox 7'. The training is based on a large number of training data sets, which include a considerable number of vibration signals referring to the operation accompanied by a specific fault in the gearbox 7'. Depending on the situation, the machine learning model can be trained for different fault types. For example, the fault type may refer to damage to the gear or damage to the inner race or outer race of the bearing or damage to the balls or rollers of the bearing. The machine learning model is trained for one of such specific fault types.

[0030] In a preferred embodiment, the machine learning model refers to a neural network. However, other machine learning models may be used, such as principal component analysis. The machine learning model may be trained by any suitable learning method. In the embodiments described herein, the learning method is based on supervised learning, wherein the training data comprises information whether a fault is present in the gearbox for a respective training data set. However, the machine learning model may also be used in conjunction with training based on unsupervised learning.

[0031] In order to use the trained machine learning method for the turbine 1 ′, the vibration signals of the vibration sensor 10 ′ are input into the model during operation of the wind turbine 1 ′, resulting in an output indicative of a specific fault. Depending on the situation, the output may be such that it indicates whether a fault is present. However, the output may also be a probability for the presence of a specific fault.

[0032] The idea of ​​the present invention is to enable monitoring of the turbine 1 based on a trained machine learning model, while the training of the model is performed for another wind turbine 1 ′. This is achieved by taking into account a suitable mapping of the kinematic parameters for both turbines 1 and 1 ′. Figure 3 and Figure 4 Describe this.

[0033] Figure 3 A flow chart illustrating an embodiment of the invention for monitoring a gearbox 7 of a turbomachine 1 is shown. In step S1, a vibration signal VS originating from a sensor 10 is recorded. The vibration signal need not refer to a direct sensor output, but may be a pre-processed signal. In the embodiments described herein, the vibration signal is the sensor data after a Fourier transform has been applied, i.e. the vibration signal is a signal in the frequency domain.

[0034] In step S2, the vibration signal VS valid for the sensor 10 of the turbine 1 is converted by mapping to a vibration signal VS', which is a vibration signal that would otherwise occur in the turbine 1' if the operating state of both the gearboxes 7 and 7' were the same. That is, in the case where the vibration signal VS refers to a fault in the gearbox 7, if the vibration signal VS' is detected, this fault will also be present in the gearbox 7'. In order to perform the mapping, the knowledge of the kinematic parameters KP1 for the wind turbine 1 and the kinematic parameters KP2 for the turbine 1' is used. The kinematic parameters KP1 refer to the component 7 being monitored and the specific fault. Similarly, the kinematic parameters KP2 refer to the component 7' being monitored and the specific fault.

[0035] In the following, an example of a mapping from VS to VS' is described. As kinematic parameters for both turbines 1 and 1', so-called damage frequencies occurring at specific faults are used. Those damage frequencies are known in advance and can be represented by the following formula:

[0036] f=n·f c +z·f m (1)

[0037] Frequency f c and f m is completely described by the kinematics of the corresponding components 7 and 7' and the rotational speed of the shaft. In the above formula, f c represents the gear mesh frequency when the gear damage is a specific fault. Similarly, f c The frequency f may refer to the frequency of the inner or outer races in the case of a specific fault indicating a crack in these races, or it may refer to the frequency of ball or roller rotation in the case of a specific fault indicating a ball or roller in a ball / roller bearing. c is known, but is different for components 7 and 7'. In addition, f in the above formula m Represents the known sideband frequency present when a particular fault occurs. Similar to the frequency f c , frequency f m is known, but is different for components 7 and 7'. Furthermore, n, z are integers.

[0038] In the following, frequencies occurring in the gearbox 7 of the wind turbine 1 are designated by index A, whereas frequencies occurring in the gearbox 7 ′ of the wind turbine 1 ′ are designated B. In other words, index A refers to component 7 and index B refers to component 7 ′.

[0039] In the embodiments described herein, a linear mapping T(·) is used to convert the vibration signal VS into the vibration signal VS'. That is, the mapping is described by the function T(x)=ax+b. However, other mappings besides the linear mapping also provide good results if a corresponding mapping can be found.

[0040] Using the above formula (1), the kinematic parameters for the gearboxes 7 and 7' are as follows:

[0041] f A =n·f A c +z·fA m

[0042] f B =n·f B c +z·f B m

[0043] Based on the above equation, the parameters a and b should be determined so that f B =T(f A ). Therefore, the following expressions apply:

[0044]

[0045] To determine a and b, the following two equations (i) and (ii) are solved:

[0046] (i)0=n·(af A c -f B c )+b

[0047] (ii) 0 = z·(af A m -f B m )

[0048] Solving equation (ii) will provide the parameter a as follows:

[0049] (ii) a=f B m / f A m For z≠0

[0050] Solving equation (i) will provide the parameter b as follows:

[0051] (i)b=n·(f B c -a·f A c )=n·(f Bc -f A c ·f B m / f A m )=n·(f B c ·f A m -f A c ·f B m / f A m

[0052] The above mapping depends on n. The mapping is obtained in the frequency domain. However, the mapping can also be obtained in the same way for the well-known cepstrum domain.

[0053] Based on the mapping T, the vibration signal VS is converted into a vibration signal VS' in step S2, such as Figure 3 In the next step S3, the vibration signal VS' is provided as input data to the trained machine learning model, and the trained machine learning model is Figure 3 As mentioned above, the machine learning model provides an output indicating a specific fault, where this fault is Figure 3 FT in the diagram. Therefore, when the machine learning model ML is applied to the vibration signal VS', an output OU is generated which refers to the fault FT. Although the machine learning model has been trained for a different gearbox 7' in another wind turbine 1', due to the mapping, this output will indicate whether there is a fault in the gearbox 7 of the wind turbine 1.

[0054] The invention as described above is based on the knowledge that the vibration signal of a rotating machine is characterized by variations in amplitude and frequency that can be explained by the kinematics of the machine components. When an anomaly occurs, these characteristics change. Therefore, a fault or damage type has a vibration pattern that can be described by the component kinematics. Therefore, for each fault type, the vibration signal from one component can be mapped to the other component using the kinematics of both components.

[0055] Figure 4 Shown is a method for executing Figure 3The device of the method described herein comprises means M1 for providing a vibration signal VS. In the embodiment described herein, the means M1 comprises a vibration sensor 10 for detecting vibrations. However, in a modified embodiment, the vibration sensor is not part of the device. In this case, the vibration signal has been previously acquired and stored, and the means M1 only accesses the storage device to obtain the vibration signal. To achieve this, the means M1 is part of a computer program implemented on a computer. Such a computer program also comprises means M2 and M3 described below.

[0056] Device M2 executes as follows Figure 3 The vibration signal VS' generated by this mapping is processed by the device M3, which performs Figure 3 Step S3. That is, the machine learning model ML is implemented in the device M3, and by applying this model to the vibration signal VS', an output OU indicating a specific fault is generated. This output can be stored in a storage device and / or can be provided on a user interface (e.g., a display) to notify the user of the occurrence of a fault.

[0057] The above described embodiments have several advantages. In particular, a pre-trained machine learning model can be used to monitor wind turbines other than the wind turbine for which the machine learning model has been trained. This can be achieved by a suitable mapping based on known kinematic parameters for the monitored wind turbine and the wind turbine used to train the machine learning model. Due to the use of a pre-trained machine learning method, there is no need to develop fault detection algorithms for new wind turbines. Furthermore, the method of the present invention artificially generates new training data and allows for the further development and design of new machine learning algorithms.

Claims

1. A method for computer-implemented monitoring of a component of a wind turbine, wherein the wind turbine is a first wind turbine (1) and the component is a first component (6, 7, 8), and wherein the method accesses a trained machine learning model (ML), the trained machine learning model (ML) having been trained for one or more second components (6', 7', 8') of the same type of one or more second wind turbines (1'), the one or more second wind turbines (1') being wind turbines of another type than the first wind turbine (1), wherein the trained machine learning model (ML) is configured to provide an output (OU) indicating a predetermined fault (FT) occurring at a second component (6', 7', 8') of the second wind turbine (1') by processing a vibration signal (VS') in a predetermined domain measured in the vicinity of the second component (6', 7', 8') during operation of the second wind turbine (1'), the method comprising the steps of: i) providing a vibration signal (VS) in said predetermined domain measured in the vicinity of the first component (6, 7, 8) during operation of the first wind turbine (1); ii) mapping the vibration signal (VS) to a corresponding vibration signal (VS') valid for the second component (6', 7', 8') based on one or more given kinematic parameters (KP1) of the first component (6, 7, 8) and one or more given kinematic parameters (KP2) of the second component (6', 7', 8'); iii) applying a machine learning model (ML) to the vibration signal (VS') valid for the second component (6', 7', 8'), resulting in an output (OU) indicative of said predetermined fault (FT) occurring at the first component (6, 7, 8). The method according to claim 1 , wherein the predetermined domain is a frequency domain or a cepstrum domain.

3. The method according to claim 1, wherein the one or more given kinematic parameters (KP1) of the first component (6, 7, 8) are described by the same function type as the one or more given kinematic parameters (KP2) of the second component (6', 7', 8') but with different function parameters.

4. The method according to any one of the preceding claims 1 to 3, wherein the one or more given kinematic parameters (KP1) of the first component (6, 7, 8) are one or more specific values ​​within the predetermined domain contained in the vibration signal (VS) measured near the first component (6, 7, 8) when the predetermined fault (FT) occurs, and wherein the one or more given kinematic parameters (KP2) of the second component (6', 7', 8') are one or more specific values ​​within the predetermined domain contained in the vibration signal (VS') measured near the second component (6', 7', 8') when the predetermined fault (FT) occurs.

5. Method according to one of the preceding claims 1 to 3, wherein the monitored component is a part of a wind turbine (1) or a drive train.

6. Method according to one of the preceding claims 1 to 3, wherein the predetermined fault (FT) refers to damage to a gear wheel or damage to a bearing race or damage to a ball or roller in a ball or roller bearing.

7. A method according to any of the preceding claims 1-3, wherein the machine learning model (ML) is based on one or more neural networks or principal component analysis.

8. A device for monitoring a component of a wind turbine, wherein the wind turbine is a first wind turbine (1) and the component is a first component (6, 7, 8), and wherein the device has access to a trained machine learning model (ML), the trained machine learning model (ML) having been trained for one or more second components (6', 7', 8') of the same type of one or more second wind turbines (1'), the one or more second wind turbines (1') being wind turbines of another type than the first wind turbine (1), wherein the trained machine learning model (ML) is configured to provide an output (OU) indicating a predetermined fault (FT) occurring at a second component (6', 7', 8') of the second wind turbine (1') by processing a vibration signal (VS') in a predetermined domain measured in the vicinity of the second component (6', 7', 8') during operation of the second wind turbine (1'), wherein the device comprises: - means (M1) for providing a vibration signal (VS) in said predetermined domain measured in the vicinity of the first component (6, 7, 8) during operation of the first wind turbine (1); - means (M2) for mapping the vibration signal (VS) to a corresponding vibration signal (VS') valid for the second component (6', 7', 8') based on one or more given kinematic parameters (KP1) of the first component (6, 7, 8) and one or more given kinematic parameters (KP2) of the second component (6', 7', 8'); - means (M3) for applying a machine learning model (ML) to the vibration signal (VS') valid for the second component (6', 7', 8'), resulting in an output (OU) indicative of said predetermined fault (FT) occurring at the first component (6, 7, 8).

9. The device according to claim 8, wherein the device is configured to perform the method according to one of claims 2 to 7. 10 . A computer program product having a program code for executing the method according to claim 1 , when the program code is executed on a computer.

11. A non-transitory machine-readable carrier storing a computer program having a program code, wherein when the program code is executed on a computer, the program code is used to perform the method according to any one of claims 1 to 7.

Images