Toward an online strategy for mechanical failures diagnostics inside the wind turbine generators based on spectral analysis

Abstract

Mechanical faults in wind turbine generators lead to a huge problems of breaking electricity production, increase the maintenance cost and can damage the wind turbine generator, which caught fire in the nacelle. Gearbox installed in the nacelle is the important component in wind turbine, it composed by shafts bearings and gearings. Always bearings and gearings failures are related together. Therefore, to control these rotating components we need to analysis their vibration signature. This work was done with Tunisian Electricity and Gas Company (STEG) it’s about mechanical failures diagnostics which are based on three temporal vibration signals which are made by three sensors installed also in three positions axial, vertical, and horizontal. In this work we present a Fast Fourier Transform which is applied on raw mechanical vibration signals by the vibration department of STEG, which are compared with a new strategy of envelope analysis is commonly used to obtain the mechanical faults harmonics from the envelope signal spectrum analysis and has shown a more suitable results of diagnostics which is applied on the same data sets from Fast Fourier Transform. In this work, we illustrate a squared envelope based on spectral kurtosis approach to determine optimum envelope analysis parameters including the filtering band and center frequency through a short time Fourier transform. This method proves a better diagnosis results using real vibration data set from vibration department of STEG.

Keywords

Wind turbine generator bearing vibration spectral envelope analysis spectral kurtosis kurtogram diagnostics

Introduction

In nature there are several renewable and sustainable energies for life, each country has its own geographic, climatic, and geologic characteristics. The wind energy is one of these natural springs that engineers and scientists seek to make inventions in order to convert it into green energy. Electrical energy is an important solution versus global warming. Electricity production is based on the optimal yield of wind turbines, exceptionally on mechanical rotating components from the gearbox. This multiplier increase the rotation speed of the rotor blades which is about 18-1500 rpm in order to made electricity in the generator which converts mechanical rotation into electrical energy.

This work was done on real wind turbine generator (WTG) Gamesa AE61 1.320 kW it was installed onshore wind farm in the north-east of Tunisia, supervised and managed by the vibration department of the Tunisian electricity and gas company (STEG).

The research framework in this paper is to enhance performance of wind turbine by the controlling the raw mechanical vibration signals collected from sensors installed on the gearbox body in order to make a preventive and corrective maintenance. It should be noted that mechanical components are characterized by a low failure rate compared to electrical components. However, failure in mechanical components in drive train introduces a high repair costs and revenue loss due to long down-times in WTG.

The gearbox also called multiplier speed receives as input the rotation speed of propeller rotor and increases to the desired speed to output (generator shaft). It is thought that the WTG’s useful life is predicted for 20 years.

Rolling element bearings (REBs) are the most existent components used in wind turbine generator, they ensure the rotation guiding in all turning equipment like gearings and shafts. The wind farm park is installed in north-east of Tunisia at Sidi-Dawoud (governorate Nabeul). This wind farm park is localized near the sea and characterized by a strong wind variation and humidity. The north-west dominant wind is 17 m/s which can raise the risk of damage in bearings and gearings. In addition, the radial and axial loads in the rotor produce the periodic load fluctuations. That is why bearings and gearing damage can be noticed in the gearbox (McDade and Munch, 2010; Saidi et al., 2014).

According to the literature, most of bearing fault diagnosis methods which based on vibration analysis signal can be classified into time domain, frequency domain, and time-frequency domain. In the recent research works there are several work based on bearing vibration signal.

Saidi et al. (2014) proposed a method of high order statistics (HOS) which are merging together to introduce a new based approach named bi-spectrum based empirical mode decomposition (BSEMD).

Empirical mode decomposition (EMD) is a self-adaptive signal processing approach that can be used to preprocess the non-stationary and nonlinear signals completely. By analyzing the outer race bearing signals from experiment rig, the experimental results show that this method could identify the fault accurately and highlight the visibility of outer race fault (ORF) signals. The BSEMD method is proved a promising way to fault diagnosis of rotating machinery. In addition, these results also show the advantages of the BSEMD method as applied to: noisy data, data which are unsuitable for frequency-domain analysis.

In addition Saidi et al. (2016) the authors applied squared envelope based spectral kurtosis (SESK) method to raw vibration bearing signal for fault diagnosis, the main step of this methods is to find the filtering frequency band and center frequency that include faulty bearing signal component through a short time Fourier transform.

In this work, we propose an effective fault diagnosis approach to detect the bearing failure as fast as possible in gearbox, to increase the reliability of the gearbox, and reduce maintenance cost. This prevents further damage to wind turbine and reduces gearbox system downtime. The first step in reacting to a high-speed shaft bearing in the gearbox is to detect quickly the occurrence of fault event. The proposed fault diagnosis schemes for bearing is discussed in the rest of this article and this paper is the extended version of Kramti et al. (2020).

The remainder of this work is formulated as fellow: section “Mechanical faults detection in wind turbine generator and Description of wind turbine Gamesa AE61-1.320 Kw” reports the mechanical faults detection in wind turbine generator, the Fast Fourier Transform is represented in sections “First strategy: Fast Fourier Transform (FFT) and “FFT experimental results”. In section “Second strategy: Squared envelope based Spectral kurtosis (SESK)” we introduce Bearing Envelope Analysis using Squared envelope based Spectral kurtosis and finally we conclude this paper.

Mechanical faults detection in wind turbine generator and description of wind turbine Gamesa AE61-1.320 kW

This work was done on real wind turbine generator (WTG) made Gamesa AE61-1.320 kW it was installed onshore wind farm in Tunisia. Supervised and managed by the vibration department of the Tunisian Company of Electricity and Gas (STEG). The main WTG as shown in Figure 1 is a Spanish model the nacelle weight 49 t, the rotor and hub weighs 23 t, and the tower weighs 89.5 t.

The height of tower is 58.5 m, the diameter rotor is 61 m, and the rotation speed of rotor is limited between 13 and 18.8 rpm. The minimum operating wind speed is 3.5 m/s, the nominal operating wind speed is 17 m/s and the maximal operating wind speed is 25 m/s. The gearbox have a three speed, which is made by Flender. The gearbox increases the main rotor speed from (18.8/13 rpm) to the generator (1520/1012 rpm). The ratio gearbox is about 81. The lubricant tank contains 190 L of synthetic oil (Castrol tribol 1510/320). The generator type is asynchronous, made by Siemens, produceing 690 V.

Our wind turbine generator was installed onshore wind farm in Sidi-Daoud north east of Tunisia. For more details, please visit this link https://www.thewindpower.net/turbine_fr_49_made_ae-61.php.

Experimental steps

The data acquisition was done on the gearbox PE1080 (see Figure 2) in real operating condition. There are three vibrating sensors installed in the body of the gearbox near the output shaft in three different positions (axial, vertical, and horizontal) under operating condition 1448.195 rpm. The vibrating signal was saved as a time domain series from the supervisory control and data acquisition (SCADA). On the high-speed shaft there are two bearings are shown in Figure 2. The high-speed shaft bearing (rotor side) has a single row cylindrical bearing (type SKF NJ2324-ECMA-C3). The outer race ring diameter is 260 mm.

The inner race diameter is 120 mm, the thickness is 86 mm, the weight is approximately 22.8 kg, and the limiting speed is 5000 rpm. The high speed shaft bearing (generator side) is a single row cylindrical roller bearing, typed SKF NJ328-ECJ, with an outer race ring diameter of 300 mm, an inner race ring of 140 mm. The thickness is 62 mm, the weight is 20 kg, the reference speed is 2400 rpm, and the speed limit is 2800 rpm. A cylindrical bearing is shown in Figure 3.

Bearing failures may be occur on the outer ring, inner ring, cage, and roller. Generally, faults of rolling element bearings operating under normal condition are caused by wear on running surface, material fatigue, or contamination. Early failures can be made by many factors such as such as corrosion, plastic deformation, faulty installation, and no maintenance i.e., old oil lubrication as mentioned by Randall et al. (2011).

In case of a healthy bearing, there is usually no information of faults frequencies in the vibration spectrum. The only information is related to the shaft frequency speed and its harmonics. Any other frequencies could be related to noise or other rotating components at the same operating time with the bearing under test (Barszcz et al., 2011; Randall et al., 2011; Sawalhi et al., 2007; Urbanek et al., 2014).

In case of defective bearing there are failure frequency added to the healthy spectrum. These harmonics can be related to BPFI (Ball Passing Frequency Inner Race), BPFO (Ball Passing Frequency Outer Race), BFF (Ball Fault frequency), and CPF (Cage Pass Frequency). Their mathematical equation defined as fellow

BPFO = \frac{f_{r}}{2} N_{b} (1 - \frac{D_{b} \cos β}{D_{c}})

(1)

BPFI = \frac{f_{r}}{2} N_{b} (1 + \frac{D_{b} \cos β}{D_{c}})

(2)

BFF = \frac{f_{r}}{2} \frac{D_{c}}{D_{b}} (1 - {(\frac{D_{b} \cos β}{D_{c}})}^{2})

(3)

CPF = \frac{f_{r}}{2} (1 - \frac{D_{b}}{D_{c}} \cos β)

(4)

Where: $f_{r}$ is the Shaft frequency speed, $N_{b}$ is the Number of rolling components, $D_{b}$ is the Ball diameter, $D_{c}$ is the Pitch diameter, and $β$ is the Ball contact angle.

The Table 1 illustrate a failure frequencies characteristic of two bearings related to high-speed shaft of the gearbox. These frequencies are required for diagnostics steps to analyze the spectrum of the vibration signal (Kramti et al., 2020).

Table 1.

High speed shaft bearings failures frequencies.

	Bearing rotor side	Bearing generator side
Speed (rpm)	1448.195	1448.195
Bearing ref	NJ2324ECMA/C3	NJ328ECJ
F (Hz)	24.137	24.137
BPFI	188.137	201.171
BPFO	125.994	137.162
BFF	117.322	123.166
CPF	8.164	8.195

First strategy: Fast Fourier transform (FFT)

The presented fault diagnosis approach is built on the analysis of the vibrating signal. The analysis of the vibrating signal is achieved through frequency analysis. Typically, frequency analysis transforms the signal from time domain into frequency domain, where; the frequencies are used as the fault characterization principle. In this work we use the Fast Fourier Transform (FFT) as a frequency analysis tool (Kramti et al., 2020).

The main advantage of this technique over other methods explored in the literature is the reduced computational cost and time. Also this approach was used by the maintenance engineers of STEG.

It is hard to think of condition monitoring without FFT, which is a mathematical approach for mapping a function of time into a function of frequency. It is characterized as transforming from time domain to the frequency domain. The FFT is an evolution of the Discrete Fourier Transform (DFT) which eliminates duplicated terms in the mathematical algorithms to decrease the number of mathematical operations executed. In this way, it is probable to use a great number of samples without adjusting the speed of the transformation. While the DFT is an N² operation, the FFT diminishes computation by a factor of N/(log₂ (N)).

The FFT calculate the DFT and makes absolutely the same result as evaluating the DFT, but it is faster this is the most important difference (Jones and Selesnick, 2010).

Evaluating this formula directly involves N² operations, there are N outputs of X_k, and every output necessitates a sum of N terms. The FFT is an approach to compute the same result in N × log (N) operations. All admitted algorithms of FFT involve N × log (N) operations.

Let’s say X_{0 …}X_N₋₁ is a complex number, the DFT is defined as fellow

X_{k} = \sum_{n = 0}^{N - 1} x_{n} e^{- i 2 π k \frac{n}{N}} k = 0, \dots, N - 1

(5)

To clarify the savings of an FFT, consider the count of complex multiplication and addition. Evaluating the Discrete Fourier Transform’s sums exactly requires N² complex multiplications and N(N − 1) complex additions. The FFT algorithms can calculate the same results with just (N/2) × log₂ (N) complex multiplication and N × log₂ (N) complex addition.

In practice, existent efficiency on new modern computers is generally dominated by factors other than the speed of arithmetic process and the analysis is a complicated point, but in general the improvement from N² to N × log₂ (N) persist.

The only disadvantage of FFT that the data samples are length, which for most applications is the power of 2 (like 256, 512, 1024…, although other radix or prime factor can be used). Apart from that, the result looks like the same as for the DFT.

Many amplitude scales of FFT can reveal more about the vibration signal if it is used perfectly. Linear amplitude scale makes the best view of maximum peaks in the signal, logarithmic scale can present more invisible peaks and signal noise but gives a bad comparison of high and low peaks. Scale in dB gives the best estimation of signal noise if 0 dB is maximum determinable value and is also used in noise measurements, where the dB scaling is absolutely the result since the human ear has logarithmic sensitivity to noise.

The abscissa scale can be either logarithmic or linear. Linear scaling is the right illustration of the mathematic transformation and generally gives the best information for analysis. Sometimes it is nice to show the abscissa axis in logarithmic values since most interesting frequencies are in lower region, also it is more a holdover from older sensor system, or application, where lower frequencies were more important. We have to know that just to set the abscissa scale to logarithmic can’t improve the results in the lower region, so the resolution will be better in the upper region, since there are more frequency lines available there.

FFT experimental results

On the plots below there are three different raw vibrating signals (axial, vertical, and horizontal). Each data is original of a vibration signal in time domain (unit in s), and data after applying the FFT in frequency domain (unit in Hz; Kramti et al., 2020).

We start the first step of diagnostics. The vibration sensor was fixed on axial position of the output gearbox (high-speed shaft). As shown in Figure 4, the axial vibration signal was analyzed using FFT spectrum. In the logarithmic scale plot lower frequencies (25 Hz) are modulated so much which are related to shaft harmonics. The linear frequency scale plot there are shaft harmonics, BPFO of NJ2324ECMA/C3 which is 125.994 Hz as presented in Table 1 this failure frequency is modulated three times. For the gear-mesh failure frequency (482.732 Hz) is given by STEG vibration department, this fault frequency is modulated twice. Others frequencies are unknown: they may be related to noise in vibration signal or other rotating components inside the gearbox

From the second (vertical) axis is illustrated in Figure 5, where we can explore the FFT spectrum. According the logarithmic scale plot lower frequencies (25 Hz) are modulated such that one can see multiple shaft harmonics. In the linear frequency scale plot there are shaft harmonics, such as the gear-mesh failure frequency (482.732 Hz), which is also modulate. The others frequencies are unknown they can be related to noise in vibration signal or other rotating components inside the gearbox.

Viewing the horizontal vibration of the output gearbox (high-speed shaft) in Figure 6, the horizontal vibration signal was evaluated by FFT spectrum. As believed by the logarithmic scale plot lower frequencies (25 Hz) are modulated such that one can see multiple shaft harmonics. In the linear frequency scale plot the shaft harmonics are not modulated, BPFO of NJ2324ECMA/C3 which is 125.994 Hz is modulated more than the first step of diagnostics, equivalent to the above results the gear-mesh failure frequency (482.732 Hz) as well modulated. The others frequencies are unknown they can be related to noise in vibration signal or other rotating components inside the gearbox.

Second strategy: Squared envelope based spectral kurtosis (SESK)

As summarized in Figure 6, the proposed approach applied in this study for mechanical fault detection based on SESK method of the raw vibration signal. The SESK process steps is defined in Figure 6. When the vibration is achieved. Three principal processing steps are ranged as follows:

(1)SK based approach (fast kurtogram) indicates in a colormap, the kurtosis values for different combinations of bandwidth (B_w), and the center frequency (f_c) in a fixed way (Figure 3). Then, the optimum filter, described by B_w and f_c, (with highest kurtosis value), is selected to be applied in the envelope processing.

(2)SESK can be explored as improvement and development for envelope analysis. Generally, it composed by four steps: (1) description of the analysis frequency band, (2) creation of a band-pass filter, (3) determination of the squared band-passed signal, and (4) derivation of the Fourier spectrum for the envelope signal.

(3)At the end, SESK spectrum is performed, if it does not include the failure frequency characteristics, it indicate that the component (bearing/gear) under test is healthy. Otherwise, the component is faulty when the failure harmonics is modulated and identified by a peak around the predicted value.

SESK experimental results

In order to obtain the failure feature, a SK analysis approach based on fast kurtogram (Antoni, 2006, 2007) is applied to each vibration signal (vertical signal, horizontal signal, and axial signal). This mechanical signature is decomposed into four levels of frequency, with a 1/3 binary tree structure (Arturo et al., 2011; Yan et al., 2014). The corresponding kurtogram of the axial vibration signal is illustrated in Figure 7, from which a kurtosis dominant frequency band. With bandwidth of 80 Hz and center frequency f_c of 120 Hz is clearly recognized. With this information, an optimal band-pass filter is more used to extract the impulses from the raw axial vibration signal (Figure 8) presents the filtered signal. Unfortunately, in the axial signal there is no failure signature (see Table 1).

The corresponding kurtogram of the vertical vibration signal is illustrated in Figure 9, from which a kurtosis dominant frequency band. With bandwidth of 426.66 Hz and center frequency f_c of 1066.66 Hz is clearly recognized. With this information, an optimal band-pass filter is more used to extract the impulses from the raw vertical vibration signal (Figure 10) presents the filtered signal. In this case it is clear to find the shaft harmonic which is equal to 24 Hz (see Table 1), these harmonics are modulated so much, due to mechanical looseness, which support the idea that the bearings are damaged. This can indicate poor maintenance practices, high vibration, or failure of a components due to a crack. Rotating looseness generates harmonics. However, not every spectrum with harmonics indicates looseness because the harmonic can be made by noise from sensors or by other rotating components inside the gearbox. Note that the rolling element is on a film of oil, the contact is non-Hertzian. Typically, there is usually about 1% slip. To find bearing data we look for horizontal vibration signal.

The corresponding kurtogram of the horizontal vibration signal is illustrated in Figure 11, from which a kurtosis dominant frequency band. With bandwidth of 320 Hz and center frequency fc of 800 Hz is clearly recognized. With this information, an optimal band-pass filter is more used to extract the impulses from the raw vertical vibration signal.

The horizontal vibration signal as shown in Figure 12, in spectrum there is series of impulse of shaft frequencies, a harmonics of (125.994 Hz see Table 1) corresponding to BPFO of the bearing reference NJ2324ECMA/C3 and the failure frequencies harmonics of BFF of the bearing reference NJ3226ECJ (123.166 Hz see Table 1). By the use the SESK spectrum the diagnosis process result prove that mechanical faults were found in outer race bearing NJ2324ECMA/C3 (rotor side) and ball failure frequency of NJ3226ECJ bearing (generator side). The FFT spectrum of horizontal vibration signal the failure frequency of NJ3226ECJ bearing.

Conclusion

In this work, we present two strategies based on mechanical failure diagnostics of wind turbine generator Gamesa AE61-1.32 kW. The diagnostics is based on raw vibration data from a real degradation of the gearbox. The vibration department of STEG supervise the health state of the wind turbine gearbox by three vibration sensors, which are installed in three different positions (axial, vertical, and horizontal). The three vibration signals were processed using FFT analysis to seek the failure frequencies and this approach was proposed by STEG engineers, as shown the experimental results of FFT spectrum are not accurate and not clear to find directly the failure frequencies. An automatically frequency band selection ability of spectral kurtosis, when applied to three vibration signals, has been tested in this paper on real damaged bearing from real wind turbine generator. The capacity of SESK to enhance the signature of mechanical failures is well known when applied on three vibration signals, From the obtained results in this work, it is easy to conclude that SESK can be very useful tool to reduce the noise and improve the vibration signature obtained from defective rotating elements. From the experimental results, it is possible to observe evidence that the SESK make the vibration spectrum clearer and more efficient to do the diagnostics task.

STEG engineers never use the Prognostics and Health Management (PHM) approaches which is required today for remaining useful life prognostics of mechanical parts of wind turbine generators.

Footnotes

Acknowledgements

The authors of this paper wound like to thank all the agents and engineers from the vibration department of STEG for kindly providing the vibration data analyzed in this work.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

References

Antoni

(2006) The spectral kurtosis: A useful tool for characterizing non-stationary signals. Mechanical Systems and Signal Processing 20: 282–307.

Antoni

(2007) Fast computation of the Kurtogram for the detection of transient faults. Mechanical Systems and Signal Processing 21: 108–124.

Arturo

G-P

Troncoso

RJR

Yepez

, et al. (2011) The application of high-resolution spectral analysis for identifying multiple combined faults in induction motors. IEEE Transactions on Industrial Electronics 58: 2002–2010.

Barszcz

JabŁoński

(2011) A novel method for the optimal band selection for vibration signal demodulation and comparison with the Kurtogram. Mechanical Systems and Signal Processing 25: 431–451.

Jones

Selesnick

(2010) The DFT, FFT, and practical spectral analysis.

Kramti

Ali

Bechhofer

, et al. (2020) Diagnosis of mechanical failures in wind turbine generator using spectral analysis. In: Proceeding of the 11th international renewable energy congress (IREC).

McDade

Munch

(2010) Gearbox reliability collaborative: Gearbox inspection metadata. NREL/TP-500-49133.

Randall

(2011) Vibration-Based Condition Monitoring: Industrial Aerospace and Auto-Motive Applications. Chichester: John Wiley & Sons Ltd, pp.167–227.

Randall

Antoni

(2011) Rolling element bearing diagnostics—A tutorial. Mechanical Systems and Signal Processing 25: 485–520.

10.

Saidi

Ali

Benbouszid

, et al. (2016) The use of SESK as a trend parameter for localized bearing fault diagnosis in induction machines. ISA Transactions 63: 436–447.

11.

Saidi

Ali

Fnaiech

(2014) Bi-spectrum based-EMD applied to the non-stationary vibration signals for bearing faults diagnosis. ISA Transactions 53(15): 1650–1660.

12.

Sawalhi

Randall

Endo

(2007) The enhancement of fault detection and diagnosis in rolling element bearings using minimum entropy deconvolution combined with spectral kurtosis. Mechanical Systems and Signal Processing 21: 2616–2633.

13.

Urbanek

Barszcz

Antoni

(2014) Integrated modulation intensity distribution as a practical tool for condition monitoring. Applied Acoustics 77: 184–194.

14.

Yan

Gao

Chen

(2014) Wavelets for fault diagnosis of rotary machines: A review with applications. Signal Processing 96: 1–15.