A Robust Algorithm for Detecting Wind Turbine Blade Health Using Vibro Acoustic Modulation and Sideband Spectral Analysis

Transcription

1 A Robust Algorithm for Detecting Wind Turbine Blade Health Using Vibro Acoustic Modulation and Sideband Spectral Analysis 1 2 Noah J. Myrent and Douglas E. Adams Vanderbilt Laboratory for Systems Integrity and Reliability, Nashville, TN, Gustavo Rodriguez Rivera 3, Denis A. Ulybyshev 4, and Tomas Kalibera 5 Purdue University, West Lafayette, IN, Jan Vitek 6 Northeastern University, Boston, MA, Ethan Blanton 7 Fiji Systems Inc., South Bend, IN This paper presents a robust crack detection algorithm for wind turbine blades using vibro acoustic modulation and the area under the curve of the frequency spectrum of the sidebands. In a previous paper it was shown how the operational vibration of the turbine could be used as a pumping signal in vibro acoustic modulation. However, the sources of the operational vibration frequencies are many and may change over time due to the change of the wind speed, direction and many other factors. Instead of trying to identify peaks in the spectrum of the sideband (a procedure that is error prone), the area under the curve of the sidebands around the probing signal is obtained. Utilizing the area under the curve allows the consideration of multiple operational vibration frequencies simultaneously in the analysis, making the algorithm more robust. By obtaining data simultaneously in all the turbine blades, the same operational vibration frequencies are applied to all blades. The area under the curve of the spectrum of the sidebands is then obtained and compared. If a large discrepancy exists in the area under the curve of the sidebands among the blades, then a potential crack is detected and signaled. In this paper the algorithm and its implementation are described in detail. FFT = fast Fourier transform F prob, f pr = probing frequency F pump, f pu = pumping frequency MFC = macro fiber composite ε = crack threshold factor Nomenclature 1 Staff Engineer, Wind Energy Research Lead, Vanderbilt Laboratory for Systems Integrity & Reliability (LaSIR), noah.myrent@vanderbilt.edu, AIAA Member 2 Distinguished Professor and Chair, Civil & Environmental Engineering, douglas.adams@vanderbilt.edu. 3 Professor, Department of Computer Science, grr@cs.purduee.edu. 4 Graduate Research Assistant, Department of Computer Science, dulybysh@purdue.edu. 5 Researcher, Department of Computer Science, kalibera@cs.purdue.edu. 6 Professor, Department of Computer Science, j.vitek@neu.edu. 7 Computer Scientist, elb@fiji systems.com. 1

2 I. Introduction W ind power is considered by many to be the most competitive source of renewable power. However, due to the high investment cost it is important to continuously monitor the health of the wind turbine and reduce the cost of unscheduled maintenance 1 2. Additionally, the failure of a blade can cause damage in other subsystems and neighboring facilities 3. Detecting early failure in wind turbines is very important to protect the investment cost of the wind farm. In this paper, a robust algorithm for crack detection is demonstrated using Vibro Acoustic Modulation. In previous papers 4 5 it was shown how to use Vibro Acoustic Modulation to detect cracks in turbine blades using a probing signal (F prob ) and a pumping signal (F pump ). A probing frequency such as a sine signal is introduced in the blade using a Micro Fiber Composite (MFC). In operational conditions, the wind turbine will produce a lower natural frequency called the pumping signal due to the rotation or the natural frequency response of the rotor and blades. Due to the nonlinearity introduced by a crack, modulation will occur at the probing frequency producing sidebands at F prob + F pump and F prob F pump. These sidebands will be accentuated in the cracked blade due to the nonlinearity introduced by the cracked blade. II. Methodology Vibro Acoustic Modulation is one type of crack detection technique. It has been demonstrated in the literature that certain kinds of defects in materials such as cracks increase the nonlinear behavior of the material 6. When the blade material is damaged, the integrity and stiffness of that material changes thereby causing a nonlinear vibrational response. Therefore, by measuring the increase in nonlinearity, damage in the specimen can be ascertained. The Vibro Acoustic Modulation test measures the vibrational response of the specimen when it is excited by two sinusoidal signals at different frequencies. These signals are called the pumping signal and the probing signal. The pumping signal is produced at a low frequency called the pumping frequency, f pu, and the probing signal is produced at a high frequency defined as the probing frequency, f pr. Figure 1. Response to two excitation signals. Figure 1 shows the response when the two excitation signals are theoretically applied to a system. If the system is linear and damped, the response in the steady state is the linear superposition of the responses of each signal and only the linear components of Figure 1 will appear in the frequency spectrum of the response. When cracks cause nonlinear behavior within the system, the response contains both the probing frequency and the pumping frequency in addition to 2

3 other frequency components such as harmonics of each signal and sidebands around the probing signal. When the system is nonlinear, the stress strain relationship can be expressed as follows: where σ is the stress, ε is the strain, E is Young s modulus, and β 1, β 2 are nonlinear coefficients. When the strain contains two signals at two different frequencies, the strain can be expressed as follows: (1) (2) where and The stress is then expressed as (3) The first term contains the linear response components and these are the only components that linear systems exhibit. The components in the second term contain frequencies of 2ω pu and 2ω pr which are forced harmonics. The third term contains frequencies of ω pu ± ω pr which are called modulation sidebands. The ω pu ± ω pr components can be further expanded into ω pu ± ω pr (where =1, 2, 3, ) components when higher order nonlinear terms are considered. Vibro Acoustic Modulation elicits these modulation sidebands components at f pr ± nf pu. Since nonlinearities due to defects in the system are not limited to quadratic types of nonlinearities, 2 nd and 3 rd order sidebands can also be observed. (a) (b) Figure 2. (a) Change of effective contact area during crack opening and closing. (b) Nonlinear stress strain function. The mechanism of how a crack increases the nonlinear behavior is not clearly understood. However, there are several explanations which suggest why a crack introduces nonlinear behavior. An intuitive explanation for 3

4 understanding the relationship between the crack and nonlinearity in the system is the opening and closing action of the crack, or equivalently the change in contact stress along the interface of the crack as a function of load. When the specimen is excited by the pumping signal the crack in the specimen opens and closes according to the period of excitation. The stiffness changes due to the change in contact stress when the crack opens, which introduces nonlinearity into the specimen response because the effective contact area changes (Figure 2(a)). Figure 2(b) shows a nonlinear function in the stress strain relationship based on the opening and closing effect of the crack. This function is linear when the strain is either positive or negative. However, this function cannot be expressed as a linear function because its slope is discontinuous at the zero strain point. Additionally, there are many sources of pumping frequencies that are not fixed and may change over time due to different factors such as: The different modes of the turbine blades. The rotational speed of the rotor. The aerodynamic loading and how it causes deformations in the blades. These aspects make it difficult to identify a single peak for F pump that can be analyzed. Instead of looking for a single F pump peak, all F pump sources are considered in the analysis by integrating the area in the spectrum over the range where these sidebands are likely to appear (F pump_min, F pump_max ). (4) In Figure 3, two pumping frequencies are shown as sidebands of the probing frequency. Instead of trying to identify a single peak, the area under the curve of all the side band frequencies is computed. The area under the curve of the sidebands for all the blades is then calculated. Since the blades have the same structural properties and they are subject to the same operational frequencies, the area under the curve for each blade will be very similar. If they are different by some predetermined factor ε, then a crack is detected in one of the blades. Figure 3. Based on Vibro Acoustic Modulation, the entire area under the curve is calculated at the locations of the sidebands of the probing frequency in order to determine whether or not a blade is cracked. 4

5 III. Experimental Setup The wind turbine testbed at Vanderbilt s Laboratory for Systems Integrity and Reliability (LASIR) includes two Whisper 100 wind turbines manufactured by Southwest Wind Power. To produce operating conditions of the wind turbine, a wind tunnel was built as shown in Figure 4(a). The wind tunnel consists of six industrial fans at the outlet, a polycarbonate enclosure, and polycarbonate honeycomb at the inlet. The six industrial fans are controlled by an industrial motor controller and the air flow created by the six fans produces wind in the wind tunnel. The dimensions of the wind tunnel are 4.089(w) 3.175(h) (l) m. (a) (b) (c) Figure 4. (a) Experimental setup for Vibro Acoustic Modulation tests; (b) Cup exciter used to provide probing excitation; (c) MFC sensors installed on the blade. One damaged blade and two healthy blades were installed. A Tectonic Elements metal cup exciter and two Micro Fiber Composite (MFC) transducers were installed on each blade the exciter was used for the probing excitation (shown in Figure 4(b)) and the MFCs were used for sensing (shown in Figure 4(c)). An MFC is a low profile transducer that can actuate as well as sense vibration as the transducer expands, bends, or twists. By using these low profile and lightweight MFCs, the effect of the experimental setup on the aerodynamic performance of the blade was minimized. As shown in Figure 5, a wireless data acquisition system was installed on the rotor to eliminate the need to use a slip ring to transmit the MFC signals because slip rings can be susceptible to electrical noise. The data acquisition system includes a Raspberry Pi B ARM board running Raspbian Linux. A Wolfson Rapberri Pi Sound Card is used to produce both the probing frequency, as well as capturing the input signal from the MFC, and this system provided signals with good signal to noise ratios. Figure 6 shows a screenshot of the data acquisition code being run on the Raspberry Pi and called remotely via PC wireless connection. 5

6 Figure 5. Wireless data acquisition system mounted to the rotor of a Whisper 100 HAWT. Figure 6. A PC is used to remotely send commands to the Raspberry Pi to provide the probing frequency actuation and MFC data acquisition. The screenshot above is for 30 seconds of data acquisition while the Raspberry Pi outputs a 5 khz probing signal to the cup exciter. To maximize the nonlinear behavior of the blade, the location of the crack was carefully chosen such that the dynamic stress on the blade in operation is the maximum at the crack. In previous studies7 10, it was shown that alternating bending stress due to gravity is the major source of vibration of wind turbine blades in operation. Therefore, only the vibration due to the gravitational force of the blades was considered as the source of dynamic stress in operation. Figure 7(a) shows the dynamic stress distribution along the span direction of the blade while in operation. A theoretical centrifugal force acting at the frequency of 3 Hz was plotted for comparison. The stress caused by bending is much larger than tensile or shear stress as commonly shown in other beam structures. The bending stress is even larger than the centrifugal force over most of the range. Figure 7(b) shows the stress distribution in the cross section of the blade in which the highest stress is caused by bending. Due to the twisting angle, the local maximum points may not exactly be the leading and trailing edges. However, because the twisting angle is small enough in this cross section, the leading and trailing edges are local maxima in the cross section in this case. This result also indicates that the Vibro Acoustic Modulation method should be sensitive to cracks along the trailing edge of the airfoil near the mid span area of the blade. This characteristic sensitivity corresponds well with reports that blades often fail in the mid span area of the blade2. Another noticeable feature in this result is that a crack located at points where the dynamic strain is large will grow more rapidly under fatigue loading. In other words, cracks that are not as easily detected using Vibro Acoustic Modulation should be less likely to grow under fatigue (periodic) loading than the cracks that can be detected. This stress distribution can be useful for pinpointing the location of the crack with modal analysis after the existence of damage is verified. 6

7 Based on this result, the trailing edge of the blade at meters from the tip of the blade was chosen for the crack location where the maximum stress occurs as shown in Figure 7(a). The crack was induced by applying a sudden load to the tip while the blade was clamped at the intended crack location. (a) (b) Figure 7. Stress distribution on the blade. (a) Stress distribution along span direction. (b) Stress distribution on a cross section airfoil surface. IV. Experimental Results Figures 8 and 9 show the frequency spectrum of the healthy and cracked blades, respectively, in the range of the probing frequency at 7 KHz. The y axis in both figures is the amplitude of signal normalized with the amplitude of the probing signal computed as Amplitude(F)/Amplitude(F probing ). It can be observed that the sidebands of the cracked blade are larger and more abundant than the ones of the healthy blade due to the nonlinear combination of the probing frequency at 7 khz and the basebands due to the rotational and natural frequencies of the blade. Figure 8. Frequency spectrum of healthy blade at the probing frequency and sidebands. 7

8 Figure 9. Frequency spectrum of cracked blade at the probing frequency and sidebands. Figures 8 and 9 also show the need for using the area under the curve for the crack detection comparison instead of using individual peaks as described before. Due to the many peaks in the sidebands, it is difficult to know which of the peaks to use for comparison with the peaks of the healthy blade. Instead, we use the area under the curve or energy of the sidebands. For both blades we have made sure that the amplification of the probing signal and the input signal has been the same in both healthy and cracked blades. The peak at the probing signal of 7 KHz is the same in both blades. To be able to compare both signals we have also normalized the signals by plotting the Frequency vs. Amplitude(F)/Amplitude(F probing ). V. Conclusions Wind turbines require a robust structural health monitoring strategy for wind turbine blades, and this research developed a crack detection technique for wind turbine blades in operation using Vibro Acoustic Modulation. The technique utilizes the structural vibration of the wind turbine blades as a pumping signal when the rotor rotates, and measures the sideband levels which are the result of the modulation between the probing signal generated from the piezoelectric actuator on the blade and the pumping signal. In the case of applying the developed approach for a utility scale horizontal axis wind turbine, modifications may be necessary and the algorithm would certainly need to be adapted for the increased size of the structure and the varying rotational speed of the turbine due to the stochastic input provided by the wind resource. For example, the rotational speed of a utility scale wind turbine will be much lower for a utility scale wind turbine than that of the Whisper 100 turbine. Therefore, increased frequency resolution may be required in order to detect modulation sidebands of the probing frequency, and this feature can be detected as long as the data acquisition system can provide adequate frequency resolution. For example, if a utility scale wind turbine has a rotational speed of 0.5 Hz, then there must be enough frequency resolution to observe peaks at f pr and f pr ± 0.5 Hz. Modern data acquisition hardware is capable of achieving such a frequency resolution. Another area of concern rises from the variable nature of the wind speed and environment outside of the laboratory. The blade rotational speed of the turbine will vary as the wind changes speed and direction. However, the modulation sidebands of the probing signal can be located as long as the location of the pumping frequency is known. Therefore, a measurement of the low frequency vibrational response of the blade would identify the location of the pumping frequency and resultant locations of the modulation sidebands. If such a measurement is not possible, then the rotor encoder signal could be used to determine the rotational speed of the turbine. Another common wind turbine environmental factor is lightning. The data acquisition system used to acquire the vibration data will be susceptible to lightning strikes like the other electronics and controls on the turbine, but this danger can be mitigated through the use of surge suppression and grounding with a lightning protection system. In addition, the MFC transducer is coated in a polyimide film and these materials are known to be good insulators. This technique has many benefits for structural health monitoring of wind turbine blades. First, the technique does not require the wind turbines to be stopped in order to inspect the blades. Therefore, not only is the maintenance cost reduced, but there is also no loss of generated power from the turbines when they are taken offline for inspection. The technique is also beneficial because it is based on nonlinear property changes which are sensitive to small defects. Therefore, the technique can detect small cracks in blades earlier than other techniques normally would. Furthermore, 8

9 the technique is less affected by various environmental and loading conditions such as temperatures, humidity, wind profile, etc. than the presence of the cracks in the blades. Finally, the technique can be performed with cost effective excitation system because the nonlinear characteristics of the excitation system do not affect sidebands levels which present nonlinear characteristics of the material. In this paper, the Vibro Acoustic Modulation technique on an operating wind turbine blade was demonstrated with a Whisper 100, 900 Watt small scale wind turbine. The test results showed that the proposed technique can measure the nonlinearity in the response of the blades using the structural vibration as the pumping signal in operation as effectively as using other conventional Vibro Acoustic methods that are implemented using different types of pumping signals. Acknowledgments Funding for this research was provided by the Division of Computer and Network Systems (CNS) for the National Science Foundation (NSF) (Award Number ). References 1 M. A. Rumsey and A. Paquette, Structural health monitoring of wind turbine blades, Smart Sensor Phenomena, Technology, Networks, and Systems (2008). 2 C. C. Ciang, J. Lee, and H. Bang, Structural health monitoring for a wind turbine system: a review of damage detection methods, Measurement Science and Technology, Vol. 19 (2008). 3 Caithness Windfarm Information Forum, Summary of Wind Turbine Accident data to 31 December 2012, 4 S. Kim, D. E. Adams, and H. Sohn, Crack detection on wind turbine blades in an operating environment using VibroAcoustic Modulation technique, AIP Conf. Proc. 1511, pp (2012). 5 S. Kim, D. E. Adams, H. Sohn, G. Rodriguez Rivera, J. Vitek, S. Carr, A. Grama (2013). Validation of Vibro Acoustic Modulation of Wind Turbine Blades for Structural Health Monitoring Using Operational Vibration as a Pumping Signal. International Workshop on Structural Health Monitoring (2013), Stanford, CA. 6 A. M. Sutin and V. E. Nazarav, Nonlinear acoustic methods of crack diagnostics, Radiophysics and Quantum Electonics, Vol. 38, Nos. 3 4 (1995). 7 S. R. Dana, Detection of blade damage and ice accretion for health monitoring of wind turbines using integrated blade sensors, Thesis, Purdue University, (2011). 8 D. E. Adams, J. White, M. Rumsey, and C. Farrar, Structural health monitoring of wind turbines: method and application to a HAWT, Wind Energy 14.4 (2011): S. Kim, D. E. Adams, and H. Sohn, Crack detection on wind turbine blades in an operating environment using Vibro Acoustic Modulation technique, AIP Conf. Proc. 1511, (2012). 10 S. Kim et al., Crack detection technique for operating wind turbine blades using Vibro Acoustic Modulation, Structural Health Monitoring (2014), accepted. 9