Vibration-based Fault Detection of Wind Turbine Gearbox using Empirical Mode Decomposition Method

Transcription

1 International Journal of Science and Advanced Technology (ISSN -8386) Volume 3 No 8 August 3 Vibration-based Fault Detection of Wind Turbine Gearbox using Empirical Mode Decomposition Method E.M. Ashmila Faculty of Science- Zliten Department of Physics Al-Mergab University Alkums-Libya eashmila@yahoo.co.uk Alsdeg A. Abohnik School of Computing, Science & Engineering Acoustics Research Centre Salford University Manchester, United Kigdom alsdegstudent@gmail.com Andy T. Moorhouse School of Computing, Science & Engineering Acoustics Research Centre Salford University Manchester, United Kigdom a.t.moorhouse@salford.ac.uk Abstract Vibration based monitoring techniques are widely adopted for monitoring the condition of rotating machinery. However, in the case of wind turbines the measured vibration is complex due to the high number of vibration sources and modulation phenomenon. Signals are generated by tooth meshing, shaft rotation, gearbox resonance vibration signatures and a substantial amount of noise. Therefore, extracting condition related information of a specific element e.g. gears condition is very difficult. In this paper, a single stage gearbox and generator was manufactured to simulate a small horizontal-axis wind turbine that mounted with three blades. One accelerometer used to extract vibration data contains information about wind turbine gearbox health condition. Vibration signals were collected for healthy gears and gear suffering from a tooth breakage created by removing 3%, 6% 9% of gear tooth to simulate three faults at different rotational speeds;, and 5 rpm. Gear fault detection method based on Empirical Mode Decomposition (EMD) that combined with Total energy calculation (TE) technique is presented. Healthy vibration signal has been used as baseline data and analyzed to be compared with faulty signals that may occur in gear to provide a comparison for assessing gear condition. The results showed that proposed method of vibration based condition monitoring is a promising technique for detecting the presence of the faults in gear. Moreover, it successfully differentiated the signals from healthy system and system containing damaged gear. Keywords- Empirical mode decomposition method (EMD); Total Energy; Gearbox. I. INTRODUCTION Accelerometers are the most widely used for measuring vibrations. An accelerometer is a full-contact transducer mounted directly on a system. They can measure low frequencies from a few Hertz and high frequencies to tens of kilohertz. Their benefits include linearity over a wide frequency range and a large dynamic range. Transducers are sensitive to the direction of the vibration source and need to be attached or mounted on the surface of machines. Xueli An et al.[]are proposed a fault diagnosis model of a direct-drive wind turbine based on back propagation neural network parameters of the horizontal vibration and the vertical vibration of the wind turbine main shaft are comprehensively considered. It has been known for some time through both analytical and experimental investigations that some machine faults can be directly related to acoustic and vibration harmonics []. Because of the drive signals are noise rich and difficult to remove it by using conventional filters with fixed cut-off frequencies, some techniques are commonly used to remove the noise beside a Matlab program, among them the discrete wavelet transforms (DWT), which has been applied out for noise cancellation, whilst Continuous Wavelet Transforms (CWT) is then used for feature extraction. Yang et al. [3] developed a new method to deal with non-stationary and non-linear signals. Empirical mode decomposition (EMD) was used to analysis the feature intensity level power signals which is measured from the terminals of 3-phase wind turbine induction generator. The result showed that the intrinsic mode functions (s) are always able to give an obvious indication of change of machine running condition. Moorhouse et al [] have investigated a methodology for prediction of structure-borne sound and vibration inside attached dwellings. The prediction methodology was verified in a field survey of existing installations. In this study the source strength was established as a function of rotor speed although a general relationship to wind speed could not be established. Moreover, the influence of turbulence has been investigated. Based on obtained results, it has found that there were vital differences in behavior and subjective character between the airborne and structure-borne noise from BMWTs. Baydar and Ball [5] used the instantaneous power spectrum (IPS) to diagnose faults in a helical gear. The study showed that useful information on the progression of damage in the gear can be obtained by changing in the feature distribution (IPS). Moreover, they stated that the

2 International Journal of Science and Advanced Technology (ISSN -8386) Volume 3 No 8 August 3 IPS can be used to determine the location of a defect by observing the location of the energy peaks in the IPS. In additional, IPS can correctly indicate faults under different load conditions. Endo et al [6] studied two types of localized gear tooth defects that were crack and spall. To detect theses faults in the system they used the differential diagnostic (DD) technique as a tool. It has been found that there is a difference between results of the experimental study and results of the simulation study when differential diagnostic (DD) technique was used. Belsak and Flasker [7] investigated the influences of a crack in a single stage spur gear on the vibration generated. In this study wavelet de-noising methods has been applied to detect the faults in a gearbox. Bartelmus and Zimroz [8] investigated condition monitoring of planetary gearboxes under different external load. In their study a new approach to condition monitoring using STFT and Wigner-Ville distribution were presented. Lu and Chu [9] examined a number of algorithms and methods based on vibration, noise and AE signals used to diagnose faults in wind turbines. It was established that morphological undecimated wavelet decomposition is more efficient and appropriate for online diagnostics of bearings in rotating machines. It was also established that the time wavelet energy spectrum is efficient in extracting impulse features created by localised gear damage. For identifying and locating gear faults vibration - AE based methods are capable of recognising the type of fault that has occurred and to implement precise diagnostics. Hu [] proposed a novel non-linear method for feature extraction from the time-domain signal using wavelet packet pre-processing and from the corresponding frequency-domain of the signal using the kernel principal component analysis (KPCA), to characterize the condition of a gearbox. Experimental tests on an automobile gearbox showed that KPCA outperformed PCA in terms of clustering capability, and both the two KPCA-based subspace methods were effectively applied to gearbox CM. Hui Li et al [] conducted their experimental study on a single-stage gearbox. The drive pinion had 8 teeth meshing with a 36- tooth wheel and a.5kw AC governor motor. They used EMD and Teager Kaiser Energy operator (TKEO) techniques and also as the Teager-Huang transform (THT) to investigate a fault in a gear. To implement localized wear fault of a gear, a tooth was chipped from zero thickness to 5% thickness at pitch point. They stated that the gear fault could be effectively diagnosed using the THT which has better resolution than that of Hilbert- Huang transform. The THT transform provides a viable processing tool for gearbox defect detection and diagnosis II. VIBRATION SOURCE The vibrations of a wind turbine originate at various sources. Wind turbines can be affected by the different types of vibration generated by such components as blades; generator, gears, bearings and tower see Figure. Within the wind turbine the drive train will generate significant and sometimes substantial vibration levels. Essentially the drive train consists of the drive, gears, bearings and shafts and the major excitation will take place at the gear mesh frequency, f gm : Where n is the number of teeth on the gear and N gear is its rotational speed (in Hz). The generator may be a source of excitation for the drive train at higher frequencies. The vibration signal at any measurement point on a wind turbine will be a mix of many components at different frequencies, different amplitudes and different phases, thus it needs special techniques to analysis obtained signal. Because the vibration signal will vary with time it should be treated as non-stationary. Blades Hub Shafts Bearing () Gears Nacelle Figure : wind turbine components Generator Tower III. EXPERIMENTAL TEST A single stage gearbox was connected directly to the hub with three blades. It placed about two meters on front of wind tunnel with various speeds. The drive pinion had 5 teeth meshing with a 3-tooth wheel. An accelerometer with sensitivity of mv/g and mounted horizontally on the nacelle of the wind turbine and charge amplifier was used to convert the output of the accelerometer to mv. A National Instruments data acquisition card was connected between a PC and the charge amplifier to collect data, see Figure. Tests were carried out using healthy gears and a one gear suffering from a tooth breakage: 3% tooth removal (fault ), 6% tooth removal (fault ) and 75% tooth removal (fault 3), see figure 3, the experimental was carried out at different rotational speeds;, and 5.

3 Amplitude International Journal of Science and Advanced Technology (ISSN -8386) Volume 3 No 8 August 3 Figure represents the time waveform of the overall experimental vibration signal for the healthy case. Then EMD method applied on each collected signal for both the healthy and faulty conditions to decompose the signals into a number of. Fast Fourier Transform (FFT) applied on each () to produce its spectra. MATLAB software version used to analyze All Signals. Wind tunnel Accelerometer IV. EMPIRICAL MODE DECOMPOSITION METHOD (EMD) Empirical mode decomposition (EMD) is a powerful technique for improving signal to noise ratio of the measured corrupted vibration. EMD is an adaptive technique for signal decomposition with which complex signals can be decomposed into finite set of signals called intrinsic mode functions (). It is defined by a process called sifting. It decomposes a given signal X(t) into a finite set of signals called s, to give K modes ( ) and a residual term r(t)[]: ( ) ( ) ( ) () Charge amplifier Data acquisition card PC Figure : Experimental test rig The EMD algorithm is summarized as following:. Start with the signal ( ). Sifting process ( ) ( ). Identify all local extrema of ( ). 3. Compute the upper and the lower envelopes ((EnvMax and EnvMin respectively) by cubic spline interpolation of the maxima and the minima.. Calculate the mean of the lower and upper envelopes, ( ) ( ( ) ( )) (3) 5. Extract the detail ( ) ( ) ( ) 6. If ( ) is an, go to step 7, else, iterate steps to 5 upon the signal ( ) 7. Extract the mode ( ) ( ). () 8. Calculate the residual. ( ) ( ) ( ) (5) 9. If ( ) has less than minima or extrema, the extraction is finished ( ) ( ), else iterate the algorithm from step upon the residual ( )..5 Figure 3: 3%, 6% and 75% teeth removal (chipping on one tooth) V. THE PROPOSED METHOD By collecting vibration data from wind turbine, EMD method decomposes the measured signal into its fundamental frequency components. FFT is applied on each to produce spectra that originated from various sources. Total Energy is calculated for spectra related to gear meshing and its sidebands. Faulty signals will be compared with baseline data collected for healthy gear. Figure 5 illustrates the flowchart of the proposed algorithm based on EMD Time (Sec) Figure : Overall healthy vibration signal 3

4 5 Amplitude International Journal of Science and Advanced Technology (ISSN -8386) Volume 3 No 8 August 3 Start Vibration data measurement Decomposition of signal using EMD Calculation of Fast Fourier Transform of frequency Apply curve fitting and total energy determination Total Energy> baseline threshold Yes Confirm fault presence Figure 5: Flowchart of the proposed monitoring technique Calculation of total energy in proposed method uses curve fitting for the FFT spectrum in the region of meshing frequency and its sidebands zones, If the fitted curve for the signal is S( ) the Total energy (TE) of the signal in the frequency band from to, can be expressed as: ( ) (6) In equation (6) and are upper and lower frequencies of the frequency band. The integration may be done for the whole range of frequencies obtained using the FFT. The area contained between adjacent points on the spectral envelope was calculated using the trapezoidal rule via MATLAB. VI. RESULT AND DISCUSSION The time-domain vibration signals collected from wind turbine were analyzed using EMD method. EMD was used to decompose vibration signals into a finite set of signals called intrinsic mode functions () and each represents a different vibration source. Fast Fourier Transform (FFT) was applied to each in turn to produce the spectra produced from wind turbine components and then the Total Energy (TE) contained within certain of interest frequency bands is calculated. No Baseline TE threshold These bands were meshing frequency and its local sideband zones. Analyzing data shows different frequencies; high and low frequencies are produced from different sources such as noise, gears, bearings, fan pass frequency and the shafts. The first mode contains noise-contaminated signals. The second mode is associated with meshing frequency signal and contains a spectral peak in the region to meshing frequency and its local sideband zones. Fast Fourier Transform (FFT) applied on each to produce spectra related to the gear signature. Moreover, the spectra based on FFT method that produced from intrinsic mode functions () for healthy and faulty signals were compared for fault detection. It has been found that although there are changes to the frequency spectrum with the presence of gear faults, the changes did not appear consistent and significant. Therefore, the FFT is not suitable for analysing these signals and it has not provided consistent gearbox s condition related information. Figure 6 show s at rotational speed rpm for healthy vibration signal using EMD method, while figure 7 shows the spectra related to this signal after applying the FFT method Samples Figure 6: Decomposition of experimental vibration signals for healthy signal at speed () rpm Frequency Figure 7: FFT of each at speed () rpm

5 Normalized Total Energy Normalized Total Energy Normalized Total Energy International Journal of Science and Advanced Technology (ISSN -8386) Volume 3 No 8 August 3 Meshing frequencies of gear calculated by equation () and were 5, 6 and 75 Hz at rotational speeds, nd 5 rpm respectively. The spectra corresponding to Mode contains a spectral peak in the region of the meshing frequency and its sidebands. Thus for healthy and faulty cases, mode was adopted to calculate total energy at rotational speeds;, and 5 rpm. Figures 8, 9 and represent the total energy calculation for healthy gear (h condition) and damaged gear suffering from different faults f, f and f3 at three rotation speeds RPM.75 h f f f3 normalized TE calculation will be significant for practical applications. VII. CONCLUSION In this paper, vibration signals generated from a wind turbine used to extract useful information related to wind turbine gearbox health condition. In this work, only one accelerometer was used to reduce condition monitoring costs. Presented method used to detect the presence and severity of faults in a gearbox and so help to avoid gear damage. It was effective at revealing faults and can give an indication of change in the wind turbine running condition more than other methods. ACKNOWLEDGMENT The work presented in the paper was funded and supported by Al-Mergab University, Libya and Acoustic Research Centre, Salford University, United Kingdom.. Xueli An, Jiang, D., and Li, S. Application of back propagation neural network to fault diagnosis of direct-drive wind turbine. In: World Non-Grid-Connected Wind Power and Energy Conference (WNWEC),, pp Singh G. K, and Sa'ad, A.S.A.K. (3) Induction machine drive condition monitoring and diagnostic researchâ a survey. Electric Power Systems Research 6, Wenxian, Y., Jiesheng, J., Tavner, P.J., and Crabtree, C.J. (8) Monitoring wind turbine condition by the approach of Empirical Mode Decomposition. In: Electrical Machines and Systems, 8. ICEMS 8. International Conference on, pp Andy Moorhouse, A.E., Graham Eastwick, Tomos Evans, Andy Ryan, Sabine von Hunerbein, Valentin le Bescond, David Waddington. () Structure-borne sound and vibration from building-mounted wind turbines.. Environ. Res. Lett doi:.88/78-936/6/3/35..7 RPM.65 h f f f RPM.7 h f f f3 Figures 8, 9 and, illustrate that the curve of normalized total energy is increased with increasing the fault size. This is a promising method for gear fault detection because it suggests that the rate of increase of 5 5. Baydar, N., and Ball, A. () Detection of gear deterioration under varying load conditions by using the instantaneous power spectrum. Mechanical Systems and Signal Processing, Endo, H., Randall, R.B., and Gosselin, C. (9) Differential diagnosis of spall vs. cracks in the gear tooth fillet region: Experimental validation. Mechanical Systems and Signal Processing 3, Belsak, A., and Flasker, J. (9) Wavelet analysis for gear crack identification. Engineering Failure Analysis 6, R.Zimroz., W.B.a. (9) Vibration condition monitoring of planetary gearbox under varying external load.. Mechanical Systems and Signal Processing, Vol. 3(), pp Wenxiu, L., and Fulei, C. Condition monitoring and fault diagnostics of wind turbines. In: Prognostics and Health Management Conference,. PHM '., pp. -.. Hu, Q., He, Z., Zhang, Z., and Zi, Y. (7) Fault diagnosis of rotating machinery based on improved wavelet package transform and SVMs ensemble. Mechanical Systems and Signal Processing, Li, H., Zheng, H., and Tang, L. Gear Fault Detection Based on Teager- Huang Transform. International Journal of Rotating Machinery.. Huang, N., Shen, Z., Long, S.,Wu, M., Shih, H., Zhen, Q., Yen, N., Tung, C., and Liu, H. (998) The empirical mode descomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc. R. Soc. London A 5,