DK181393B1 - Establishing health indicator of a rotating component - Google Patents

Abstract

The invention relates to a method and a system for remote monitoring of a wind turbine component of a first wind turbine, the method comprises the steps of: from a vibration sensor establish a vibration signal of the wind turbine component. Provide the vibration signal to a data processing device located remote to the wind turbine component, and by the data processing device: transforming the vibration signal into a frequency domain representation of the vibration signal, perform an autocorrelation of the frequency domain representation, determining, based on the result of the autocorrelation, the power of at least one expected fault fundamental frequency, and provide a health indication of the wind turbine components based on the determined power.

Description

DK 181393 B1 1

ESTABLISHING HEALTH INDICATOR OF A ROTATING COMPONENT

Field of the invention

[0001] The invention is related to a method and a system for establishing a health indicator value of a rotating component preferably a rotating component of a wind turbine.

Background of the invention

[0002] Condition monitoring of components of a wind turbine has been known for several years. An example of a method for performing condition monitoring in a wind farm is found in WO2012000506. WO2012000506 disclose how a vibration signal is obtained from a wind turbine component and based on which a faulty frequency index is generated for the component. Variations in the vibration signal from the rotating shaft are filtered out before comparison of the faulty frequency index of a component of one wind turbine is compared to a faulty frequency index of the same component in another wind turbine. Based on the comparison the condition of the component is evaluated.

[0003] Another example is found in EP3421786, where vibration data relating to a bearing is collected and stored. Harmonics, in the vibration data, indicating a damaged bearing are identified and harmonics of specific frequencies having high energy content are eliminated. Based on increased energy or variance of energy in each remaining harmonic a bearing damage factor is calculated. The damage factor is compared to a threshold value to determine if the bearing is damaged.

Yet another example is found in EP2543977 which disclose to detect rubbing abnormalities of a sliding bearing in a diesel engine. This is done by detecting a waveform which represents an acceleration of a vibration, transforming the acceleration waveform data into an acceleration spectrum of a frequency domain, quantifying a plurality of peak information which occurs at a rotational frequency interval of a shaft to be measured in the acceleration spectrum, obtaining a

DK 181393 B1 2 characteristic value and monitoring if this value has exceeded a predetermined threshold value and thus if an abnormality has occurred in the slide bearing.

Yet another example is the publication made in Mechanical Systems and Signal

Processing, 20 March 2012, VOI- 31, pages 428-446 by LI et al, titled "Multi-scale autocorrelation via morphological wavelet slices for rolling element bearing fault diagnosis". This publication discloses to use multi-scale autocorrelation via morphological wavelet transformation in fault diagnosis for rolling element bearings.

The Fourier spectrum representation is enhanced by using autocorrelation to get rid of background noise.

[0004] A general problem with the prior art monitoring systems is that some types of faults are not detected before they have evolved to become serious faults which then may cause shut down or require immediate servicing of the wind turbine.

DK 181393 B1 3

Summary of the invention

[0005] To solve the above problem, the present invention suggests a method for remote monitoring of a wind turbine component of a first wind turbine, the method comprises the steps of: from a vibration sensor establish a vibration signal of the wind turbine component. Provide the vibration signal to a data processing device located remote to the wind turbine component, and by the data processing device: transforming the vibration signal into a frequency domain representation of the vibration signal, perform an autocorrelation of the frequency domain representation, determining, based on the result of the autocorrelation, the magnitude of at least one expected fault frequency peak, and establish a health indication of the wind turbine component based on the determined magnitude.

[0006] The present invention is advantageous in that it has the effect that, independent of where power, originating from fault monitored by a vibration sensor, is in the harmonic family of the expected fault fundamental frequency, the fault can be identified automatically. More specifically, by the present invention it is possible to extract the power of harmonic families related to component faults, even though the fundamental frequency is only approximately known, and thereby identify faults automatically. Thereby, servicing of the faulty component of the wind turbine can be planned without risking the faulty component causes faults in other parts of the wind — turbine (also referred to as limiting the impact of aggregate damages) or the wind turbine need to be derated or stopped completely as consequence of the fault.

[0007] Determining the magnitude at an expected fault frequency in the autocorrelation or in a narrow range / band of frequencies centered around the expected fault frequency of the autocorrelation is advantageous in that the magnitude of power / energy at such fault fundamental frequency including its harmonics from the frequency spectrum, is contained in that narrow range / band in the autocorrelation.

Hence, the magnitude originates from a family of harmonics which is an integer multiple of the fault fundamental frequency in the frequency spectrum. This is in contrary to known systems, where power is found at each of a plurality of harmonics of one expected fault frequency in the frequency spectrum and where the range around

DK 181393 B1 4 that frequency is multiplied with the order of the harmonic which leads to considerable ranges at high harmonics. Note that typically frequencies which are not caused by the component fault will not influence the calculated power.

[0008] The power / energy / magnitude of the peak at a given frequency in the autocorrelation (in the result of the autocorrelation) originates from a family of harmonics which is an integer multiple of that frequency in the frequency spectrum (the signal before the autocorrelation is made). Hence, this is a result of the analysis of the vibration signal. More generally, the result is any signals in the frequency spectrum which are periodic with a given frequency including those originating at non-integer — location (e.g. a series of peaks at 0.3*expected fault frequency, 1.3*expected fault frequency, 2.3*expected fault frequency, etc. may also give rise to a peak in the autocorrelation at the expected fault frequency). It should be noted, that looking for a fault by analysing the magnitude of a peak at an expected fault frequency in the autocorrelation is preferred, however at least to some degree similar information can be derived in the 2nd or 3rd harmonic peak of the fundamental fault frequency in the autocorrelation. Of cause the information derived from e.g. 2nd or 3rd harmonics is not quite as good as information derived from the fundamental fault frequency as it will only see 2x / 3x periodicities in the signal. Therefore instead of analysing or looking solely for energy in the expected fundamental fault frequency and its individual harmonics the present invention suggest to analyse the autocorrelation and thereby analyse the fundamental fault frequency and its harmonics simultaneously.

[0009] Accordingly, the present invention is advantageous in that, the task of monitoring components especially of wind turbines, but also components of other machines, which in the past mostly has been carried out manually, can be automated to ensure cheaper, more extensive, faster and more reliable monitoring.

[0010] The vibration signal is provided to the data processing device in the time domain i.e. the vibration signal comprises vibrations measured during a time period such as continuous measurements of vibrations or discrete measurements of vibrations during predefined time periods.

DK 181393 B1

[0011] Note that the method steps c)-f) of claim 1 is preferably made on the same data processing device, however these steps, may within the scope of the invention, also be performed on two or more different data processing devices.

[0012] Further note that the expected fault fundamental frequency in this document 5 is also simply referred to as fundamental frequency and that a reference to a fundamental frequency may also include a reference to harmonics of this fundamental frequency. More specifically, when referring to the result of the Fourier transformation i.e. the vibration signal in the frequency domain, a fault is expected at a fundamental fault frequency which repeats “x” times that fundamental frequency and therefore — together referred to as the harmonic family. In the result of the autocorrelation, the energy of the harmonic family is found at the fault frequency. Hence, the magnitude of a peak in the autocorrelation originates from the total energy in the harmonic family.

[0013] Further note, that in the art strength of a signal, frequency component or harmonic family is also sometimes referred to as “energy”, “power” or the more generic terms magnitude or amplitude. Hence, in this document the term energy could therefore replace the term power without changing scope of the claimed invention. {03141 According to an exemplary embodiment of the invention, wherein the health indication is furthermore stablished based on the magnitude of a multiple of the fault — frequency in the autocorrelation. This is advantageous in that it has the effect, that faults can be detected earlier and more reliable than if only power of the fundamental frequency is determined.

[315] Preferably, the health indication is established based on the maximum value of the at least one expected fault fundamental frequency. Detenmining the health — iadication based on magnitude Le maximum amplitude of the autocorrelation at a particular frequency is advantageous if the frequency at which a fault is reflected is known leading to a precise and early error detection. It should be noted that the health indication may alternatively or in combination also be determined by average values of an expected fault fundamental frequency established over time.

DK 181393 B1 6

[0016] According to an exemplary embodiment of the invention, the health indication is established based on a sum of energy in a range of expected fault fundamental frequency.

Establish the health indication based on a sum of energy / magnitudes in a number of frequencies of the autocorrelation is advantageous if faults of a component is expected to be reflected at several frequencies leading to a precise and early error detection. It should be noted that the health indication may be determined based on any features derivable from the result of the autocorrelation such as based on a maximum power, mean value of power, etc. Note that when referring to e.g. the sum of energy a reference is made to one of the list comprising RMS (RMS; Root Mean Square) values, maximum values or sum.

[0017] According to an exemplary embodiment, the health indication is established by using a regression model. This is advantageous in that it has the effect, that using a regression model allows prediction of trends in the dataset. Using an automatic method as presented in this document is advantage in that it is cheaper than having manual fault monitoring and the fault monitoring according to the present invention can be continuous in contrary to manual fault monitoring.

[0018] According to an exemplary embodiment, the regression model is selected from the list comprising linear fit, polynomial fits, or more advanced machine learning techniques such as regression trees / forests, neural networks, etc. This is advantageous in that the data processing device performs regression with supervised machine learning which is advantageous in that it has the effect, that once learned by the machine learning algorithm of the data processing device, future monitoring and establishing of health indication of the same wind turbine component preferably in a similar wind turbine configuration can be made fast, with limited or no reference data i.e. with no or limited manual verification of the health indication established by the data processing device.

[0019] According to an exemplary embodiment of the invention, the health indication is established based on a comparison of magnitude in the at least one

DK 181393 B1 7 expected fault frequency and magnitude at the same frequency established from historic recordings of faulty wind turbine components.

[0020] According to an exemplary embodiment of the invention, the health indication is a health indicator value, wherein the health indicator value is established and associated with a vibration signal based on a comparison of magnitude in the at least one expected fault frequency and magnitude at the same frequency established from historic recordings of faulty wind turbine component. 19021] According to an exemplary embodiment of the invention, a health indication is established for each of a plurality of expected fault fundamental frequencies. This is advantageous in that it has the effect that one wind turbine component may be monitored in detail te. more than one fault at one component can be momtored. An example to ilfustrate this is a rolling element bearing such as roller bearing or ball bearing which may be subject to faults at the outer ring, inner ring and at the rolling elements each of which faults provides different fault frequencies.

[0022] According to an exemplary embodiment of the invention, the configuration of the wind turbine of which the health indication is established and the configuration of the wind turbine of which the historic recording was made is the same.

[0023] Preferably the historic recordings are established on similar wind turbines i.e. having one or more of the following in common; same design, same mechanical — construction, same age, exposure to same environmental impact, etc. The more the wind turbines from which historic reference data is recorded are similar to the monitored wind turbine, the higher precision can be expected in the health indication / health indicator. Note that the health indication may also be a visual alarm or audio alarm informing a user of the existence of a fault or an evolving fault.

[0024] According to an exemplary embodiment of the invention, the first wind turbine is part of a first wind park

DK 181393 B1 8

[0025] According to an exemplary embodiment of the invention, the data processing device is providing an alarm to a user or a digital signal to a second data processor if the value of the health indicator reaches a threshold value.

[0026] If the data processing device establishes a value of the health indicator that is higher than or equal to a threshold value it provides an output. Further, an output may be provided based on a comparison of new and historic magnitudes. The output may either be to a monitoring system which may be connected to a service system handling service and maintenance of the wind turbine and / or to a second data processor such as the wind turbine controller. In this way, appropriate action can be taken either in — form of planning service or adjusting control of the wind turbine.

[0027] According to an exemplary embodiment of the invention, the threshold value is specific for a specific fault of a specific wind turbine component.

[0028] Tt is advantageous to establish threshold values for each / a plurality of the expected fault fundamental frequencies and in embodiments also their harmonics in that it then is possible to monitor for a plurality of errors of one or more wind turbine components.

FOG29] According to an exemplary embodiment of the invention, the threshold value is adjustable. Adjustable threshold values are advantageous in that it has the effect, that it is possible to change the level of when to receive an alarm or change control of the wind turbine, In embodiments, the threshold may be overruled so as to continue operation with a components that is likely to completely break within near future such as within a couple of weeks or months.

[0030] According to an exemplary embodiment of the invention, the method further comprises a pre-processing step applied to the frequency domain representation of the vibration signal prior to step d), the pre-processing step comprises filtering the frequency domain representation of the vibration signal so as to dampen undesired frequencies.

DK 181393 B1 9 10031] According to an exemplary embodiment of the invention, the damping is performed by applying a high-pass filter such as a simple cutoff filter or Laplacian filter. This filtering step is advantageous in that peaks of the signal is enhanced by using a Laplacian operation whereas the igh pass filter removes slowly varying trends inthe frequency spectnum filter. Other types of high-pass filters may also be used. In this way peaks of the frequency domain representation of the vibration signal passes through the filtering and thereby the peaks of the frequency domain representation is enhanced. In this way energy from the wide features in the spectrum close to the frequencies of interest is removed.

[0032] According to an exemplary embodiment of the invention, the method comprises a further pre-processing step applied to the frequency domain representation of the vibration signal prior to step d) wherein the further pre-processing step comprises using the frequency domain representation to establish a baseline frequency spectrum for the first wind turbine, wherein establishing the baseline frequency spectrum includes: Establishing a frequency spectrum of vibration signals from vibration sensors of a plurality of wind turbines having a similar configuration as the first wind turbine. Based on a comparison of the frequency spectra from the plurality of wind turbines, determining a plurality of frequencies that is considered normal for the plurality of wind turbines, and establishing the baseline frequency spectrum for the first wind turbine, by damping the normal frequencies.

[0033] This is advantageous in that it has the effect, that a frequency spectrum for the first wind turbine is obtained favouring frequencies that are not normal occurring in the frequency spectra of the plurality of similar wind turbines e.g. wind turbines having the same configuration. Hence, favouring frequencies not normally occurring i.e. possible fault indicating frequencies includes removal of some frequencies and / or damping some frequencies. The frequencies which often is present with high amplitude is preferably damped the most. Hence, most of the power in frequencies of the frequency spectrum of the baseline vibration signal (also referred to simply as baseline frequency spectrum) now originates from vibration signals that are unique to the first wind turbine. Therefore, when the baseline frequency spectrum is

DK 181393 B1 10 autocorrelated, preferably only power from such unique vibration signals of the first wind turbine is present.

[0034] Similar configuration should be understood as wind turbines of the same design with the same type of components, etc. An example could be a typical wind park, which typically comprises wind turbines of the same type i.e. similar wind turbines. The vibration signals from the plurality of wind turbines are then measured by similar sensors at similar wind turbine components such as a bearing.

[0035] The damping can be made by applying a mask weighting predetermined frequencies. A mask may be designed to assign a weight to each individual frequency or to each of a plurality of bins of frequencies in a spectrum or spectra. In an embodiment, a low weight indicates that high energy values for the frequencies / bins are expected i.e. that the frequency / frequency bin is normal across all wind turbines in a site. Accordingly, the baseline vibration signal (also referred to as baseline frequency spectrum) is the frequency spectrum measured from a wind turbine component after the mask have been applied. Preferably, a mask is applied for each incoming frequency spectrum measurement before the autocorrelation is performed and features are extracted for analysis such as regression and analysis relative to a defined threshold.

[0036] According to an exemplary embodiment of the invention, the damping is — established by applying a mask to the frequency spectrum, the mask is established by the following steps: - establish a plurality of frequency spectra from wind turbine components across the first wind park, - divide the individual frequencies of the plurality of frequency spectra in a plurality of frequency bins, - apply log transformation to each of the plurality of frequency bins, - calculate a median and standard deviation for each frequency bin,

DK 181393 B1 11 - calculating the weight assigned to each of the plurality of frequency bins by taking the reciprocal value of exp(median plus standard deviation).

[0037] As mentioned, the generation of the mask is done initially when monitoring is commenced, and can be updated, if needed after months/years of operation.

Preferably, all available spectra across the site such as a wind farm is used in generating the mask. Note that, the application of the log transformation to each frequency bin is made under the assumption that the frequency bins are approximately log-normal distributed.

[0038] According to an exemplary embodiment of the invention, the method comprises yet a further pre-processing step of establishing order-tracking of the received vibration signal, wherein the order-tracking filters out variations of the rotation speed of a shaft.

[0039] Order tracking is advantageous in that it prepares the vibration signal for analysis by filtering out variations in the signal. The order-tracking could be made on — the vibration signal before or after FFT. The shaft could be main shaft or other rotating shafts in the wind turbine.

[0040] According to an exemplary embodiment of the invention, a health indication in a second wind turbine is established based on the baseline signal, wherein the second wind turbine is located in a second wind park.

[0041] This is advantageous in that it has the effect that monitoring of components of the second wind turbine can be established based on the already existing baseline signal i.e. very fast. The first and second wind turbine are of the same type and configuration. The second wind park may be a site having only a few wind turbines such as not enough wind turbines to establish a baseline signal for that second wind park. The second wind park may be remote to the site of the plurality of wind turbines of which vibration signals was used to establish the baseline signal.

DK 181393 B1 12

[0042] According to an exemplary embodiment of the invention, the mask used to calculate the baseline vibration signal is updated at least once every month, once every year, once every second year or every third year.

[0043] Updating the baseline vibration signal is advantages to be able to filter or damp normal frequencies as these may change over time due to wear. However, it may turn out, that update of the baseline signal is not necessary or only necessary once over the lifetime of the component or of the wind turbine. The update may include re- evaluating the normal frequencies used to calculated the baseline vibration signal or establish a new mask i.e. perform the above mentioned steps. With this said, it may sometimes be chosen not to update the mask in that updating the mask if the whole fleet of wind turbines of the wind park gets more worn the ability to determine the faulty outliers is diminished.

[0044] According to an exemplary embodiment of the invention, the method comprises yet a further processing step applied to the frequency domain representation of the vibration signal prior to step d) (of claim 1), this further processing step comprises normalizing the frequency domain representation by normalizing the frequency domain representation according to rotation speed of a component mechanically connected to the monitored component.

[0045] This is advantageous in that in this way it is possible to distinguish between variations in the recorded vibration signal if these originates from failure in the component or from variation in the operation of the machine. This is possible in that frequency components get distinct and are present at the same relative frequency no matter the rotation speed of the machinery mechanically connected to the monitored component. As an example, a frequency domain representation of a vibration signal recorded by a vibration sensor at a main bearing of a wind turbine may be normalised according to the rotation speed of the main shaft.

[0046] According to an exemplary embodiment of the invention, the vibration signal is established by a vibration sensor mechanically connected to the wind turbine component, such as an accelerometer.

[0047] According to an exemplary embodiment of the invention, the vibration signal is established by a non-contact vibration sensor, such as an ultra-sonic based, laser based or microphone based measuring device. Hence, the term vibration sensor may be understood as any measuring device that is able to convert a vibration or displacement of a component into an analogue or digital representation. Such representation may require further processing (such as analog to digital conversion) before the data processing device is able to analyze the frequencies contained in the representation of the measured vibrations. In fact, the vibration sensor may also include a current sensor.

[0048] According to an exemplary embodiment of the invention, the vibration sensor is communicatively connected to the data processing device.

[0049] Data may be sent from the local data processing device to the central server when the local processing is made or it may be sent at predetermined time intervals or when otherwise suitable. Alternatively, the server may read or request data from the local data processor when needed.

[0050] According to an exemplary embodiment of the invention, the data processing device is the wind turbine controller.

[0051] According to an exemplary embodiment of the invention, the data processing device is a local generic computer dedicated to monitor the wind turbine component.

[0052] A generic computer may be implemented as a PLC or a robust industrial embedded computer having relevant data processing means such as a data processing core, RAM, storage, etc. Such generic computer may communicate directly to a central server or communicate via a wind turbine controller to a central server.

[0053] According to an exemplary embodiment of the invention, at least one of the — steps c) —f) is processed at a central server.

[0054] The steps that is most advantageous to execute on the central server is the steps that requires comparison of data from different wind turbines. In addition, it

DK 181393 B1 14 should be mentioned, that information from many sources may be processed distributed by exchanging some or all relevant information.

[0055] According to an exemplary embodiment of the invention, the pre-processing step(s) is processed locally by the data processing device at the site of the wind turbine of which the vibration signal is obtained.

[0056] Such local processing may include edge computing. Local processing is advantages in that load on the central server is reduced.

[0057] According to an exemplary embodiment of the invention, the transformation of the vibration signal to a frequency spectrum is made by using Fourier

Transformation principles.

[0058] According to an exemplary embodiment of the invention, the at least one expected fault frequency is a predefined frequency established based on frequencies caused by expected errors of the wind turbine component. Note that predefined here should be understood as within a margin / range and therefore it is an approximate of a predefined frequency.

[0059] A bearing as example, is expected to suffer from certain errors over time due to wear, construction failure, etc. The frequency response of each individual of such errors can be established based on knowledge of e.g. mechanical design of the bearing (such as size and diameter), expected rotation speed, empiric / historic data, etc.

Accordingly, the frequency at which an error may be detected is approximately predictable which is why the frequency at which the power is determined can be predefined.

[0060] It should be noted that several expected fault frequencies can be monitored simultaneously with the method of the present invention. Hence, the at least one expected fault frequency may be predefined and may include a predefined set of different expected fault frequencies.

DK 181393 B1 15

[0061] According to an exemplary embodiment of the invention, the at least one frequency is a center frequency of a frequency range within which the sum of energy is determined

[0062] To be sure to capture an error from the autocorrelation, a frequency range may be specified around a center frequency (typically the expected fault frequency, which may in this document also be referred to as the expected fault fundamental frequency), where the center frequency is the frequency at which a specific error is expected to be visible in the autocorrelation of frequency spectrum of the vibration signal. Suitable ranges may vary from less than + (plus/minus) 0.1Hz of the center — frequency to + (plus/minus) SHz or even more.

[0063] More specifically the range or margin may be calculated as percentage based on the expected fault frequency. Such that an expected fault frequency at 90Hz will have a margin of 0.09Hz at 1%. The percentage margins may vary from 2.5% to 10%.

[0064] Note that the center frequency may be predetermined as the described above.

[0065] According to an exemplary embodiment of the invention, the frequency range is defined as plus / minus 2%, plus / minus 4%, preferably plus / minus 6% and most preferably plus / minus 5% of the expected fault frequency.

[0066] According to an exemplary embodiment of the invention, the frequency range is defined by a center frequency and frequencies from 2Hz below the center frequency to 2Hz above the center frequency, preferably by a center frequency and frequencies from 1Hz below the center frequency to 1Hz above the center frequency, most preferably by a center frequency and frequencies from 0.5Hz below the center frequency to 0.5Hz above the center frequency.

[0067] Moreover, the invention relates to a monitoring system configured to monitor an at least partly rotating component of a wind turbine, the system comprises a vibration sensor associated with the at least partly rotating wind turbine component, the vibration sensor is configured to send vibration signals recorded from the at least partly rotating wind turbine component to a data processing device, wherein the data

DK 181393 B1 16 processing device is configured to establish the energy of at least one harmonic family of an expected fault fundamental frequency in the vibration signal, and provide a health indication of the at least partly rotating component based on the established energy.

[0068] According to an exemplary embodiment of the invention, the health indication is established by using a regression model.

[0069] According to an exemplary embodiment of the invention, the regression model is selected from the list comprising linear fit, polynomial fits, or more advanced machine learning techniques.

[0070] According to an exemplary embodiment of the invention, the monitoring — system implements the method described in any of the claims 1-33.

[0071] According to an exemplary embodiment of the invention, the establishing of power includes, by the data processing device, performing a Fourier transformation of the vibration signal and an autocorrelation of the result of the Fourier transformation.

[0072] According to an exemplary embodiment of the invention, the establishing of — health indicator (also referred to as health indication) includes, by the data processing device, performing a comparison of the established energy in the harmonic family at the expected fault fundamental frequency with a reference energy of the expected fault fundamental frequency wherein the reference energy is associated with a reference health indicator and associating the expected fault fundamental frequency with the reference health indicator of the reference power matching the established energy.

[0073] Hence according to the above, based on a comparison of power in a harmonic family including the fundamental frequency with an energy level associated with a health indicator, a health indicator of the at least partly rotating component can be established.

[0074] The health indicator and energy level is associated based on analysis of historic failures of comparable partly rotating components such as bearings, gears, etc.

DK 181393 B1 17

[0075] The vibration sensor is associated with the component in that it is mechanically connected thereto or is monitoring vibrations of the components by light waves, sound waves, etc.

The drawings

[0076] For a more complete understanding of this disclosure, reference to embodiments of the invention described in the following brief description, taken in connection with the accompanying drawings and detailed description, wherein like reference numerals represent like parts:

Fig. 1 illustrates a wind turbine and a wind park,

Fig. 2 illustrates a recorded frequency spectrum,

Fig. 3 illustrates ranges around harmonics in the frequency spectrum,

Fig. 4 illustrates the result of an autocorrelation of the frequency spectrum,

Fig. 5 illustrates health values at different power levels of a harmonic family, and

Fig. 6 illustrates a flow chart of steps of the method according to an embodiment of the invention.

Detailed description

[0077] The invention will now be described with reference to the above-mentioned figures. Note that, in the results of the FFT (FFT, Fast, Fourier, Transformation) i.e. the vibration signal in the frequency domain, the present invention is looking for — patterns of peaks and the strength of said pattern. A typical fault pattern could be a peak at some (approximately) known (fundamental) fault frequency F followed by its harmonics (e.g. peaks 2*F, 3*F, … n*F). In some cases the fault pattern can be modulated due to various effects. In such a case the result of the FFT will also have sidebands around each of the peaks in the harmonics series. The sidebands are a series of peaks distributed with the modulation frequency M around the harmonic n*F (e.g. a pattern of peaks located at n*F'+-1*M, n*F+-2*M, … n*F+-m*M). In the following

DK 181393 B1 18 a harmonic family will be used as reference to such patterns (with or without sidebands). Note that an FFT result may contain multiple harmonic families, if multiple faults are present.

[0078] The severity of a fault is generally related to the strength of such a harmonic family. Extracting the strength of harmonic families are a core component of many condition monitoring programs. One approach to extracting the strength of the harmonics families are to look in the spectrum at the approximate locations of the peaks and sum them together. However, some faults (e.g. early bearing faults) primarily show peaks at the higher harmonics (e.g. 10*F). In such a case small perturbation/uncertainties in the fault frequency dF may cause the peaks to be shifted a lot compared to the expected location (e.g. /0%F +- 70%dF) . Such shifts are quite easy account for by a human analyst — but it does make the automated extraction of such data more difficult. The present invention aims to solve this problem.

[0079] An alternative approach to solving the problem above is presented by the cepstrum. It takes logarithm and the (inverse) FFT of the frequency spectrum. This approach will analyze the signal in time-space, which may be beneficial in certain cases and cause challenges in others. However, as the features of interest in the spectrum are typically narrow features (peaks, or series of peaks) performing the FFT will spread the energy out across the resulting spectrum (of a spectrum). Hence, preferably the present invention keeps the analysis in frequency-space.

[0080] Note that if the cepstrum is used, then taking LOG to the spectrum and then perform the autocorrelation may also be included as a preprocessing step to improve, at least in some cases, the result of the analysis i.e. the determination of existence of a fault.

[0081] As will be described in more details below, according to the present invention, the FFT result will be transformed to a baseline frequency spectrum. The baseline frequency spectrum is then autocorrelated and in the result of the autocorrelation, a measure of the combined strength of the harmonic family is represented by the peak at the fault frequency /%/. Thus, only one reading in the autocorrelation is needed to

DK 181393 B1 19 extract the total strength of the harmonic family. Any perturbation/uncertainties dF in the fault frequency will only shift the peak to F-+dF.

[0082] Note that the autocorrelation also contains other peaks. E.g. typically there will also be a peak in the autocorrelation at 2%F, 3*F, … n*F. They will to a degree contain similar information as the 1*F peak. In the following, any reference to the 1*F peak in the autocorrelation, may be expanded to include the n*F peak.

[0083] If sidebands are present in the harmonic family the 1*F peak in the autocorrelation will still contain an estimate of the combined strength of the harmonic family (including the sidebands) — but the autocorrelation will contain even more — additional peaks.

[0084] Firstly, the fault peak at 1*F (but also the other peaks at n*F) in the autocorrelation will also have sidebands (e.g. the autocorrelation will have peaks at 1*F+-1*M, ... ,1*F+-m*M). These peaks in the autocorrelation also contain information about the strength of the harmonic family. In the following any reference to the 1*F peak in the autocorrelation, may be expanded to include the n*F +- mM peak.

[0085] Secondly, there will be a peak in the autocorrelation at the modulation frequency 1*M, it will measure the strength of the sidebands. Similar to above there may also be peaks at 2*M, 3*M, ..., m*M that contain similar info. In the following, any reference to the 1%F peak in the autocorrelation, may be expanded to the include n*M peak.

[0086] Finally, the resulting autocorrelation may also exhibit features (peaks, noisefloor, etc.) related to non-fault features in the FFT used to calculate the autocorrelation. To get a useable autocorrelation it may be necessary to clean the base

FFT for non-fault related features. We propose a method to solve this issue by establishing a baseline frequency spectrum prior to the autocorrelation.

[0087] Fig. 1 illustrates a wind park 1 comprising a plurality of wind turbines 2. At least a subset of wind turbines 2 of the same wind part 1 is similar wind turbines. In

DK 181393 B1 20 this context, similar wind turbines are wind turbines having the same configuration i.e. the wind turbines may have the same type of bearings, shafts, gearbox, generator, etc.

Knowing this, the wind turbines 2 illustrated on Fig. 1 could be the entire wind park or a subset of wind turbines of a wind park having same configuration.

[0088] Each of the wind turbines 2 have a rotor / hub 3, main bearing 4, main shaft 5, gearbox 6, brake 7, high speed shaft 8 and generator 9 which are all mechanically connected (see the enlarged circle of Fig. 1). Note that in some wind turbines, the connection of some components may be hydraulic or even electric. Electric wires 10 connects the generator 9 to the electric system of the wind turbine and thereby to the — electric grid. The wind turbine further comprises a data processing device 12 such as a controller which may be located in the nacelle or in the tower. The controller 12 may be the wind turbine controller or a stand-alone controller which is part of a monitoring system of the present invention.

[0089] The processing of data received from the vibration sensor 11 may at least partly be processed locally at the data processing device 12, at least partly at the central server 13 and / or at a remote data processing device 15. The remote location of the data processing device 12 should be understand as not part of the component that is monitored. Hence, in an embodiment, the processing of data from the vibration sensor 11 and thereby the establishing of health indication may be done complete or in part — at different data processing devices 12, 13, 15 or at one individual data processing device 12, 13, 15. The remote data processing device 15 may be used instead of the central server 13 so that the functions of the central server is carried out on the remoted data processing device 15. Hence in an embodiment a computer (remoted data processing device 15) located e.g. at a maintenance service provider may establish the baseline signal and the health indication. Alternatively, or in addition, the purpose of such computer may merely be displaying results from processing of data e.g. from the central server 13 / local data processing device 12.

[0090] As mentioned above, the baseline signal is a single spectra where frequencies, considered normal across all wind turbines in a park, are removed.

DK 181393 B1 21

[0091] As mentioned above, the present invention relates to a method of monitoring faults in wind turbine components such as the main bearing 4. The present invention can however, be used on all components in a wind turbine where failure modes can be deduced from vibration signals such as all bearings, gears, etc. The monitoring is based — on vibration signals recorded by a vibration sensor 11. The vibration sensor 11 can be any type of vibrations sensors that record vibrations over time i.e. that is able to establish a frequency spectrum of vibrations of the monitored component. As a non- limiting example, the vibration sensor 11 may be based on an accelerometer, a microphone, laser displacement sensors, etc.

[0092] It should be noted, that in some embodiments, vibration analysis can also be performed on electrical signal from generators and motors etc. A fault such as a bearing fault may spread from bearing to affect the electric output e.g. from a generator. Hence, it is possible to analyse mechanical faults on the electric output e.g. the same known fault signals from vibration analysis in the electrical signal. Electric faults such as in windings of a generator may also be analysed in the same way as mechanical vibrations from e.g. a bearing. The invention would work with this kind of electric sensors, as described with respect to mechanical vibration sensors.

[0093] It should be noted that the present invention may be applied to components or to machines of other types than wind turbines having components that include a rotating element such as bearings, gears, etc. Hence, when a wind turbine is mentioned in this application, it could be replaced with other types of machines or components.

[0094] Fig. 2 illustrates an example of a simple frequency spectrum. The dots at the end of the peaks are the best matched failure frequencies found from the autocorrelation at [1, 2, 3] * 214.42 Hz of the illustrated frequency spectrum.

[0095] The vibrations sensor 11 is connected to a data processing device 12. The data processing device 12 may be located in the wind turbine 2 or external to the wind turbine 2. Having a local data processor inside the wind turbine is advantageous in that data recorded from the vibration sensor may be pre-processed before sent to a central server 13 thereby reducing load on the data communication connection between the

DK 181393 B1 22 wind turbine 2 /wind park 1 and the central server 13. The central server may be implemented as one or more computers of which one or more may be cloud based.

Pre-processing may include all necessary steps to perform monitoring according to the present invention or a subset of these steps. Typically, the establishing of the baseline signal is made centrally at or via the central server 13 i.e. it includes input from each of a plurality of wind turbines. Subsequently, the baseline signal may then be distributed back to the processors of the individual wind turbines, so that here the remaining steps may be done.

[0096] In an embodiment, some processing is done locally and some at the central — server. Typically, local processing relates to the vibration signals recorded from the vibration sensor and the non-local processing relates to e.g. establishing and used of baseline signal.

[0097] The data processing device 12 may in an embodiment be an embedded computer 1.e. a standalone processing device capable of analysing, storing, receiving, sending, etc. data such as data from the vibration sensor 11. The data processing device 12 may also be implemented as part of an existing data processing device of the wind turbine such as the wind turbine controller. The data processing device 12 may thus be implemented as an industrial personal computer, industrial programmable logic controller or similar suitable devices.

[0098] The data processing device 12 is communicating with a central server 13 via the wired or wireless communication network 14 of the wind turbine and / or wind park. Hence, the communication and / or server 13 may at least partly be part of a cloud-based communication and / or processing service. In an embodiment, the server 13 is located at a monitoring service provider where the data initially recorded by the — vibration sensor 11 is analysed.

[0099] Accordingly, remote monitoring may include remote to the component i.e. the data processor is located remote the monitored components. As an example, the component is located in the nacelle whereas the data processing device is located in

DK 181393 B1 23 the tower. As another example the data processing device is located external to the wind turbine and implemented at a central server.

[0100] As will be explained below, the present invention includes at least the following three elements 1) Reliable extraction of harmonics of expected fault frequencies, 2) Establishing of health indicators i.e. a relative measure of a component health based on measured vibrations and historic vibration and failure data, and 3) automatic evaluation of component health based on the health indicator.

[0101] The reliable extraction of harmonics and the subsequent establishing of health indicators is made as often as desired. Hence the recording of vibration signals made — by the vibration sensor 11 can be made continuous or at discrete points in time i.e. one or more times every hour, day, week, month or year. The extraction is also made as often as desired based on expected fundamental fault frequencies and according to a baseline signal established based on input from a plurality of the wind turbines of the same site or similar wind turbines from a different site. As mentioned, the baseline — signal can be updated one or more times every hour, day, week, month or year.

[0102] In an embodiment, the baseline signal is a site-specific baseline signal established in the park comprising the wind turbines of which a component is to be monitored.

[0103] In an embodiment, the a value of the health indicator established or updated based on monitoring of components in one wind park may be used to establish threshold values for health indicator values in a second wind park. With this said, it should be noted that the health indicator values / health indications established according to the present invention, may also be provided “raw” to a service provider or other party to whom it may be relevant. Then this party may on its own motion analyse and predict health of the component based on the received information.

[0104] The health indication and automatic determination of health of a component made according to the present invention, may be evaluated by experts. Such expert evaluation may only need to be made once and may only need to be made base on a few wind turbines subject to component failure (complete or partly). Obviously, the

DK 181393 B1 24 precision and accuracy can be increased the more wind turbines having failing components that is analyzed. Therefore, as wind turbines are subject to failing components, the vibration signals of these components may be analyzed and if necessary, the relationship between power in the fault frequency family and the health indication is adjusted manually.

[0105] As already known by the skilled person, a faulty bearing in a wind turbine will exhibit failure frequencies which are approximately — known given the bearing dimensions / kinematic data. The more relevant data available of the component such as a bearing the more accurate the prediction of expected fault frequencies can be made. Note that typically, the health indication does not provide information of type of component failure, only of component health.

[0106] A particular fault in a bearing will exhibit harmonics of the fault frequency (f fai ) associated with that fault i.e. 2% fai, 3% frau, etc. where the earliest detection of the failure is typically found in the higher ranges of harmonics such as e.g. in the range of 20% f fail to 40% f fail. Exactly at which of the harmonics the fault appear / first appear is typically not deterministic. In an aspect of the invention, an expected fault frequency and its harmonics are monitored to observe developments of power in these frequencies.

[0107] It should be noted, that one fault may be detectable at e.g. 2 or 3 different frequencies. Hence, a fault in this document, may be associated to a failure mode of the component and a component may have a plurality of failure modes.

[0108] Following the above, if a fault in e.g. a bearing is expected (calculated) at 75Hz, but in reality is measured to be at 75,085Hz, then at the 35" harmonic, then the fault was expected to be found at 2625Hz (75%35), but is in reality found at 2628Hz —(75,085%35). Therefore, due to the uncertainty of where a fault actually is found in the frequency spectrum, a margin around the expected fault frequency harmonics is needed. Note that the uncertainty may arise form transfer function, modulation and / or interference often found around resonance frequencies. If such margin is e.g. 0.25Hz (0.25Hz above and below) the expected fault fundamental frequency, at the 35%

DK 181393 B1 25 harmonic the margin or range of frequencies monitored around the 35% harmonic is 8.25Hz. Hence, as the harmonics increase, the margin also increase and thereby uncertainty of where power found within the margin originates. This is illustrated in

Fig. 3, where the increasing margin is illustrated as the boxes around harmonics of an — expected fault frequency of approximately 90Hz.

[0109] This problem with the prior art, is solved by the present invention, by the autocorrelation of the frequency spectrum so that the power of each of the scattered harmonics in the frequency spectrum then is found at their shared fundamental in the autocorrelation. In this way, the problem with the increasing margins and following uncertainties are solved

[0110] Another problem with the prior art solved by the present invention is that in prior art monitoring systems a certain amount of energy in the harmonic signal is needed to be able to determine that a fault exists. If the fault only produces vibrations generating only a small amount of energy, then it may take months to identify such — fault with prior art systems. This problem is solved by the present invention by looking at the power level in the entire harmonic family.

[0111] Fig. 4 illustrates the result of an autocorrelation of the frequency spectrum illustrated in Fig. 2. Here the peak marked by the single dot is a measure the energy of the entire harmonic family [1, 2, 3] * 214.42 Hz illustrated on Fig. 2. Compared to the — frequency spectrum illustrated in Fig. 2, from Fig. 4 it is much easier to identify an expected fault fundamental frequency (@214.42Hz) due to the combined power of the harmonic series’ present in the original frequency spectrum illustrated in Fig. 2.

Accordingly, the resulting autocorrelation is easier to interpret than the frequency spectrum.

[0112] As mentioned, e.g. the frequency at which a bearing fault can be found can be calculated or it can be known from historic data of faulty wind turbines. However, since the vibration sensor 11 record vibrations from other components i.e. the frequency spectrum include other vibrations than vibrations from a fault of the wind

DK 181393 B1 26 turbine component, the power from these other components is also present in the frequency spectrum.

[0113] In embodiments, such power may disturb the analysis of the power at the expected fault frequencies and therefore the present invention provides a pre- processing step to dampen such non-fault related power / frequencies. It should be noted that monitoring of a wind turbine component with a known fault frequency, the pre-processing step might not be necessary.

[0114] One pre-processing step is to “clean” the input from the vibration sensor 11 for “normal” frequencies, which in an embodiment is done before the autocorrelation.

With this said, it can be done both before the autocorrelation at the frequency spectrum or after with the result of the autocorrelation. In any case, this cleaning is based on a comparison of input from vibration sensors 11 monitoring the same wind turbine components of wind turbines across the wind park. In this way, what can be considered as normal frequencies / power can be identified and removed to establish a so-called baseline signal for the wind turbine having the particular configuration. Hence, if the recorded frequency spectra from e.g. 10 / a majority different wind turbines all shows high power at 100Hz, then the origin of the 100Hz signal is considered normal. This is because it is unlikely that all 10 wind turbines are subject to the same fault visible at 100Hz. One way of defining normal frequencies may be frequencies that are found in the majority of the plurality of wind turbines. These normal frequencies may be dampened by use of the mask as explained above.

[0115] This cleaning step further has the advantage that now the mask can be used for similar wind turbines located at other sites having the same configuration. This is even true without having historic data from that other site.

[0116] As mentioned above, the baseline signal may be updated (e.g. as consequence of updating the above-mentioned mask) from time to time over the lifetime of the wind turbine component to be able to keep filtering out “normal vibration signals”. Such normal vibration signals may change over time due to wear of the components such as rotor/hub, shafts, gear box, etc. which is why update of the baseline signal may be

DK 181393 B1 27 relevant with a frequency of 6, 12, 18, 24, 30 or even more months. In an embodiment, such update may only be performed every 3th, 4th, 5th, 6th, 7th, 8th, 9th, 10th year or only one or two time of the lifetime of the wind turbine.

[0117] Another pre-processing step is to assign a health indication to a certain energy level of the harmonic family. As an example a health indicator e.g. between 0 (new component) and 1 (broken component) could be determined for energy levels in the harmonic family. However, the range is of course not important, alternatively it could go from 0 — 100 or vice versa. Hence, as an example the energy in the harmonic family at the point in time, where the wind turbine component brake due to a fault may be assigned a health indicator value of 1. At a certain time prior to the complete failure, the power level could be assigned a health indicator value of 0.3 or 0.6 as indicated as the highest health indicator value on Fig. 5.

[0118] Fig. 5 illustrates an example of failure in a specific harmonic family originating from a specific component failure with health indicator values established — by expert on the right hand y-axis, the energy/magnitude on the left hand y-axis, and the time along the x-axis. The curve illustrates the energy in the harmonic family and the horizontal lines / dots indicate health indicator values of this component estimated by the present invention at different times.

[0119] Accordingly, based on experience and analysis of past faults of monitored — wind turbine components a threshold health indicator can be determined. In the example of Fig. 5, if a health indicator threshold for service or replacement of the component was established at a health indicator value of 0.3, a data processing device would alert a service responsible person or system e.g. be sending a message around 2020-05. If the threshold on the other hand was set to 0.5, no reaction is needed and no alert was needed. This threshold value is determined by considering a trade-off between not servicing / replacing a component that can still be used and the risk of the component causing a standstill of the wind turbine.

[0120] Note that in an embodiment, a numeric health indicator value and comparison to numeric threshold value is not established. Instead, a level of the autocorrelation

DK 181393 B1 28 feature is associated with an alarm, notification or the like. When this feature is found in the analysis, an alarm notification or the like is produces and preferably presented to a user.

[0121] With the establishing of a health indication (feature or value) (which is dynamical in that it changes over time, typically monotonously growing until the component is repaired) and threshold thereof, then an alert (as explained above) or a recommendation on e.g. service or replacement can be provided to a site manager. The site manager then uses the information as desired e.g. as input to service planning or site risk mitigation.

[0122] In an embodiment, the threshold values defining when a component is faulty is established or adjusted by experts analysing vibration signals of monitoring of components that have failed in the past. By providing the relationship between health indication and energy at the fault frequency in the autocorrelation to the data processing device 12, 13, an algorithm hereof is trained using the health indications — provided by these experts having analysed past faults and therefore is able to spot faults / emerging faults occurring in monitored wind turbine components.. This is provided to the (machine learning) algorithm which based thereon automatically in the future can use the derived relationship to establish a health indicator based on the recorded and processed vibration signal as described below.

[0123] Preferably, the health indicator threshold value is the same so that a recommendation for servicing is e.g. 0.4. However it may be up to the user / owner of the system to override a system warning at 0.4 and apply his own threshold value at e.g. 0.5. The health indicator threshold value may thus be different from wind turbine to wind turbine and from one wind turbine owner to another.. If planned service is imminent, continuing operation with a high health indication (e.g. value or feature) / high risk of catastrophic component failure until the scheduled service event may be preferable to immediate servicing. Accordingly, the adjustable threshold value for component failure alarm is an advantageous feature of the present invention.

DK 181393 B1 29

[0124] If the health indicator threshold value is reached, the wind turbine controller may be informed, and appropriate action may be initiated. Such appropriate action may e.g. be shutting down or derating the wind turbine. Several threshold values may be determined where one activates an alarm to a service person / system, a second activated derating and a third shuts down the wind turbine.

[0125] The invention will now be described with reference to the above described example of similar configured wind turbines 2 of a wind park 1 and the flow chart of

Fig. 6. Below the main bearing 4 is used as a non-limiting example of a wind turbine component. The illustrated flow chart includes the relevant and most common steps of — the extracting of the expected fault frequencies and applying of a health indication.

Note that a health indication may at least be a value or a feature of the autocorrelation.

With this said, it should be noted that some of the steps of the illustrated flow chart may not always be needed and not illustrated steps may be added. An example of such step could be normalizing the vibration signals according to rotation speed of the main shaft.

[0126] As an example, if the component failure is advanced and thus produces a vibration signal with a high energy content, neither the high pass filtering nor establishing of a baseline signal may be needed. A less power containing vibration signal may only need to go through a high pass filtering. It should however be noted — that such advanced failures would have been detected by the method and system of the present invention. However, as mentioned, the present invention may be retrofitted and applied into older wind turbines or at least not have been implemented from commissioning and therefore in such situation a failure may not have been detected before it reaches such advance stage. Steps 1 to 4 relates to establishing the baseline signal. Hence, , in an embodiment, each of the wind turbines 2 comprise a vibration sensor 11 monitoring vibrations at the main bearing 4. Together, these individual vibration signals could be referred to as a frequency spectrum for the whole wind park 1 (sometimes referred to as a site). This park / site frequency spectrum of vibration signals recorded by a plurality of main bearing vibration sensors is established in step 1.

DK 181393 B1 30

[0127] Further, in an embodiment this site frequency spectrum is pre-processed or more precisely, the frequency spectra recorded by the individual vibration sensors 11 are pre-processed. The pre-processing step applied to the individual frequency spectra is in an embodiment a differentiation step such as high-pass filtering. The effect of this pre-processing step is to emphasize peaks and dampen broad features in the frequency spectrum i.e. only peaks remain. Hence, in the frequency spectrum power at frequencies close to frequency having power peaks is damped. This damping or differentiating of the frequency domain representation of the vibration signals from the individual wind turbines is established in step 2.

[0128] In this embodiment yet a further pre-processing step is performed in step 3.

This further pre-processing step is either performed to the signal established in step 1 or in step 2 and includes a step of comparing the individual frequency spectra. As a result of this comparison normal frequencies can be identified and subsequently damped (in step 4). A normal frequency is a frequency that is recorded by the vibration sensor 11 and originates from e.g. the rotation of toothed wheels in the gear box..

Hence, if a frequency peak is found in the frequency spectra of all wind turbines 2 it is considered a normal frequency in that it is not likely that all wind turbines is subject to a fault that generates an identical frequency peak.

[0129] As indicated, in step 4 the frequencies considered normal at the site 1 are identified. When identified, a mask which may be a site specific mask can be established as described above. The effect of the mask when applied to a vibration signal from a wind turbine is a baseline signal as clean as possible (at least as clean as possible close to the expected fundamental fault frequencies) for power at frequencies which are not considered relevant for expected faults that may occur at the monitored component. Hence, the output signal of step 4 is a mask or filter established based on input from a plurality of wind turbines that can be applied to the frequency spectra recorded at the individual wind turbines. When this mask is applied to a frequency spectrum of a wind turbine, the remaining content of the frequency spectrum is or may be relevant for the monitoring of expected faults at the monitored component of the

DK 181393 B1 31 wind turbine. In an embodiment, the established mask is stored to allow future use thereof.

[0130] The establishing of the mask is in an embodiment made centrally such as by a cloud based service or by a central data processing device such as the central server 13. Alternatively, or in addition, the mask is established at a service provider having educated personnel with expertise in monitoring wind turbine components and thereby the expert knowledge needed to analyse all the frequency spectra received from the individual wind turbines. Expert knowledge in this sense includes knowledge of frequencies originating from vibrations considered normal and frequencies from vibrations that originates from defects of a wind turbine component.

[0131] It should be mentioned that the baseline signal may also be established from an RMS approach where average values of signals from a plurality of vibration sensors are analysed. Further, an approach of establishing median value(s) of signals from a plurality of vibration sensors may also be used.

[0132] When the mask has been established, as mentioned, it can be used to establish a baseline signal for the monitoring of the individual wind turbine components. The steps 5-9 describes how a vibration signal from a monitored component is processed.

Hence in step 5, a vibration signal from a vibration sensor 1la i.e. a frequency spectrum of vibrations recorded by the vibration sensor 11a at the component (e.g. main bearing 4a) of the wind turbine 2a, is established by / provided to the data processing device 12.

[0133] In an embodiment, a high-pass filtering as described above under step 2 may also be applied to the frequency spectrum received from the vibration sensor 11a in step 6.

[0134] In step 7, a baseline signal is established for the wind turbine component 4a by applying the mask established in steps 1-4 to the frequency spectrum established in step 5 or 6. The baseline signal established in step 7 is unique for the main bearing 4a of wind turbine 2a with all frequencies considered normal in wind turbines 2 of the wind park 1 damped.

DK 181393 B1 32

[0135] Hence, now the frequency spectrum of main bearing 4a established in step 7 comprise power from potential faults of the main bearing 4a. However, as described above, early detection of a fault may be identified from harmonics of the expected fault fundamental frequency i.e. the frequency at which it is expected that a given fault can be found. Therefore, the present invention suggests analysing the power of a fundamental frequency and its associated family of harmonics. Where the fundamental frequency is an expected fault (fundamental) frequency.

[0136] Therefore, in step 8 an autocorrelation of the baseline signal originating from main bearing 4a is performed. By doing so, energy in the fault frequency and its — associated family of harmonics is found. From the autocorrelated signal it is not possible to determine at which harmonic the power of the family is, just that it is present.

[0137] In this embodiment, in step 9, a value of the health indicator is associated with the energy established at the fault frequency. That health indicator value may then, as — described above, be compared to one or more a health indicator threshold values.

Based on the comparison, appropriate actions can be initiated automatically or manually. These actions may include alarming, change control of wind turbine, plan service, etc.

[0138] Alternatively, the health indicator value may simply be provided to a service — person or the like and it is up to the receiver to determine how to react on the health indicator value.

[0139] As mentioned above, a fault in a main bearing may either occur in the outer ring, inner ring, at one of the rolling elements or the train. Accordingly, expected fault frequencies can be calculated based on knowledge of dimensions of the bearing and of rotation speed of the bearing. Hence, at these expected fault frequencies, the power of the autocorrelated signal is analysed.

[0140] From the above it is now clear, that the present invention relates to a method and system for monitoring an at least partly rotating component (i.e. at least part of the component is rotating such as part of a main bearing) e.g. of a wind turbine. A health

DK 181393 B1 33 indication for the component can be establishing by establishing a mask based on vibration signals from a plurality of wind turbines of a particular configuration, such as 2-10 or more wind turbines. Then, establish measurements (vibration signal) from a component of a wind turbine having the same configuration as the wind turbines based on which the mask is made. The mask is applied to the vibration signal from the component that is monitored and a baseline signal / frequency spectrum, where normal frequencies are dampened, is established. Then, perform an autocorrelation of the result of the baseline signal and analyse the result of the autocorrelation to establish energy at an expected fault frequency. This energy includes power in the expected — fault frequency and its harmonic family. Then, based on the established energy and information of power in past vibration signals from a faulty component (similar to the monitored component) of wind turbines having a similar configuration, determine a value of a health indicator.

[0141] The health indicator value of the monitored component can then be compared to a health indicator threshold value and thereby inform or alert systems or persons to change the control strategy for the wind turbine, that maintenance or replacement of the component is required.

[0142] Hence, using component health indicators, made manually by condition monitoring experts, and the corresponding vibration spectra, a feature is computed — from the autocorrelation (i.e. energy / power sum, maximum value, average value etc.) in the ranges of interest as previously specified. Then health indicators are paired with its corresponding feature derived from the autocorrelation. To predict new health indicator values, regression is performed with supervised machine learning. Using the health indicator and feature pairs, the machine learning is trained to find a relation > between the two. The relation can be any type of relationship such as but not limited to linear, logarithmic, quadratic, etc. This relation is used to estimate the health indicator for any unseen spectra originating from a wind turbine and component of similar type.

[0143] Accordingly, in an embodiment the present invention includes pre-processing of each vibration signal received by the data processing device making the following

DK 181393 B1 34 steps relevant for a preferred embodiment of the invention filter spectra along time axis, establish mask, apply mask, apply high pass filter, autocorrelation and establish health indicator.

List 1. Wind park 2. Wind turbine 3. Rotor / hub 4. Wind turbine component such as a main bearing or other partly rotating components 5. Main shaft 6. Gearbox 7. Brake 8. High speed shaft 9. Generator 10. Electric wires 11. Vibration sensor 12. Data processing device 13. Central server 14. Data communication path between wind turbine and central server 15. Remote data processing device

Claims (14)

DK 181393 B1 35 Patent claim 1. Method for remote monitoring of a wind turbine roller bearing element (4) of a first wind turbine, which method comprises the following steps: a) determining, by a vibration sensor (11), a vibration signal for the wind turbine roller bearing element (4), b) providing the vibration signal to a data processing device (12) located remote from the wind turbine roller bearing element (4) and using the data processing device (12): c) transforming the vibration signal into a frequency domain representation of the vibration signal, d) performing an autocorrelation of a baseline frequency spectrum , e) determining, based on the result of the autocorrelation, the order of magnitude of at least one expected failure frequency peak, and f) determining a health indication for the wind turbine roller bearing element (4) based on the determined order of magnitude, which method comprises a further preprocessing step applied to the frequency domain representation of the vibration signal prior to step d), which additional preprocessing step comprises using the frequency domain representation to determine the baseline frequency spectrum of the first wind turbine, — which determination of the baseline frequency spectrum includes: - determining a frequency spectrum of vibration signals from vibration sensors (11) of a plurality of wind turbines having a configuration corresponding to the first wind turbine, - determination, based on a comparison of the frequency spectra of the plurality of wind turbines, of a plurality of frequencies considered normal for the plurality of wind turbines, and DK 181393 B1 36 - determination of the baseline frequency spectrum for the first wind turbine by damping the normal frequencies. 2. Method according to claim 1, where the health indication is further determined based on the order of magnitude of a multiple of the error frequency in the autocorrelation. 3. Method according to any one of the preceding claims, where the health indication is determined based on a sum of energy in a range of expected error fundamental frequency. 4. Method according to any one of the preceding claims, wherein the health indication is determined using a regression model. 5. Method according to any one of the preceding claims, where the health indication is determined based on a comparison of the order of magnitude for the at least one expected failure frequency and the order of magnitude at the same frequency determined from historical records for defective wind turbine roller bearing elements (4). 6. Method according to any one of the preceding claims, where the health indication is a health indicator value, which health indicator value is determined and associated with a vibration signal based on a comparison of the order of magnitude of the at least one expected error frequency and the order of magnitude at the same frequency determined from historical records for defective wind turbine roller bearing elements (4). 7. Method according to any one of the preceding claims, wherein the configuration of the wind turbine (2) for which the health indication is determined and the configuration of the wind turbine for which historical registration of defective wind turbine component (4) was made is the same. 8. A method according to any one of the preceding claims, wherein the data processing device (12) provides an alarm to a user or a digital signal to another data processor if the value of the health indicator reaches a threshold value. DK 181393 B1 37 9. A method according to any one of the preceding claims, which method further comprises a pre-processing step applied to the frequency domain representation of the vibration signal prior to step), which pre-processing step comprises filtering the frequency domain representation of the vibration signal so as to attenuate unwanted frequencies. 10. Method according to claim 9, where the attenuation is established by using a mask on the frequency spectrum, which mask is determined using the following steps: - establishment of a plurality of frequency spectra from wind turbine roller bearing elements across a first wind farm, - division of the individual frequencies of the plurality of frequency spectra in a plurality of frequency bins, - applying log transformation to each of the plurality of frequency bins, - calculating a median and standard deviation for each frequency bin, - calculating the weighting assigned to each of the plurality of frequency bins, by taking the reciprocal of exp(median plus standard deviation). A method according to any one of the preceding claims, which method comprises a further pre-processing step of determining order tracking for the received vibration signal, which order tracking filters out variations in the rotational speed of a shaft. Method according to any one of the preceding claims, which method comprises yet another processing step applied to the frequency domain representation of the vibration signal prior to step d), which further processing step comprises normalizing the — frequency domain representation by normalizing the frequency domain representation according to —the rotational speed of a component mechanically connected to the monitored component. DK 181393 B1 38 13. Monitoring system configured to monitor a roller bearing element (4) of a wind turbine (2), which system comprises a vibration sensor (11) associated with the roller bearing element (4), which vibration sensor (11) is configured to send — vibration signals detected from the roller bearing element (4) to a data processing device (12), which data processing device (12) is configured to determine the energy of at least one harmonic family for an expected error fundamental frequency in the vibration signal and provide a health indication of the roller bearing element based on the determined — energy, which health indication is determined according to the procedure specified in claim 1. 14. Monitoring system according to claim 13, which monitoring system implements the method described in any one of claims 1-12.