You are on page 1of 8

Applying static and dynamic test responses

for defect prediction in wind turbine blades
using a probabilistic approach.
Marcin Luczak, Maciej Kahsin, Branner, Kim, Wieslaw Ostachowicz, Tomasz
Bart Peeters Martyniuk, Karolina Wandowski, Pawel Malinowski

LMS International Wind Energy Division, Institute of Fluid Flow Machinery
Risø DTU Polish Academy of Sciences

Abstract numerical models based on reliable test data to
limit blade weight without compromising the
This paper presents a novel approach in the field needed dynamic behavior of the blades.
of experimental and numerical investigation of Structural health of the large future blades will be
mechanical properties of composite structures. It monitored for reduction of failure risk. This
takes into account test data variability resulting requires robust identification of the anisotropic
from structural dynamic properties measurement. material properties and structural dynamic’s of
The main goal of the conducted research is to the blades.
investigate the dynamic and static properties of Experimental modal analysis, structural
fiber reinforced composite structure towards dynamic and static testing are core methods
assessment of accuracy of the damage providing characteristics of mechanical system.
detection. Non-destructive experimental and Several limiting factors are limiting their
numerical simulation methods are used hereto. In application for the wind turbine blades. Static and
the experimental part static and dynamic test dynamic tests preformed on the full scale blade
were performed. The dynamic excitation was are expensive and difficult to perform due to size
performed by means of random single point of the investigated object [1, 2, 3, 4]. Therefore
stimuli while the response measurement was subcomponent are more often investigated to
done through contact acceleration sensing. In the develop measurement, modeling and analysis
static test four point bending configuration was methods. The usefulness of structural dynamics
implemented. Applied force and strain response test and analysis results for solving noise and
was measured. The test results are applied in vibration problems or for performing a damage
two ways: for the structural identification of the assessment, depends largely on the confidence
object and for non-deterministic updating of the that one can have in these results. In other
numerical model according to a range of words, the results must be characteristic for the
experimental models obtained from test. The actual problem and the models must be
sources of the test data variabilities were related representative for the actual behavior of the
to the specimen-to-specimen and test-to-test of investigated structure (s). Essentially, two types
the investigated object. Non–deterministic model of problems are distinguished: (1) the test and
updating and validation included uncertainties of modeling data are subject to particular
its parameters by means of probabilistic experiment technique related systematic errors
methods. A number of variable test modal and analysis errors and (2) the tested (or
models were statistically assessed to investigate modeled) structure is not representative for the
impact of variability source onto clarity of damage actual structure.
identification. Then, for each of investigated The first problem is this of experimentation and
specimens an individual damage scenario was analysis uncertainty [5, 6, 7, 8, 9]. The “true” test
introduced. Two different vibration based result can in principle never be achieved. The
methods were applied for the damage detection. level of the uncertainty associated with the test
The results are presented and compared. The result is however not easy to quantify. Also in
research was conducted in the context of the FP7 building numerical simulation models (e.g. based
project PROND. on the Finite Element approach), uncertainty is
KEYWORDS: probabilistic model updating, introduced by discretization effects, through
uncertainty, inverse problems, imperfectly known material, geometry or loading
1. Introduction parameters, or through uncertainty in the
applicable boundary conditions [10, 11].
One of the major trends in wind turbine The second problem is this of product
technology is the increase in size of wind turbines variability [12], introducing changes in the
and thereby also increase of composite blade structural dynamics characteristics because of
sizes, especially for off-shore sites. Design of differences in material, geometric, manufacturing
future generation’s blades requires adequate or even operational use (loading, temperature…)

The review of relevant scientific literature in the field points out that in 2. In the destructive E33 13. models from a measurement of G23 4. composite materials. The nominal thickness Figure 1 The wind turbine E-glass composite blade. The overall objective the three individual composite plates was tested is introduction of a novel approach in the field of repeatedly for a number of times to acquire experimental and numerical investigation of variable test data collection. Object of the investigations samples and with both types of excitation signal. by hand lay-up and cured with manufacturers recommended cure cycle.3 [-] and afterwards.42 [-] based methods applied for the damage detection was made. B and C. 16].2. Each of for damage detection [1-16]. Random signal was mechanical properties of composite structures.7 %. Modal test results for the intact samples yielded glass composite material samples.6 [GPa] damaged structure are confronted with the model G31 5.0 [GPa] testing stage the comparison of two vibration- v23 0.1. The reason for this a different level (severe. static four multi-run modal analysis.parameters when compared to the “nominal” of the samples was 20 mm. The objects of the investigation were three E. These two methods are based on the v31 0. Total number of point bending tests and damage detection measurement points was 49 measurement points vibration based measurements. the length and width case. 15. The presence of test data G12 5. The uncertainties in model parameters is inherent of samples were named A. gives the FRF’s of the whole sample. The acceleration was measured in 14 points in one 2. variability is non-fully repeatability of the Problem of test data variability and composite manufacturing process [6]. For each of these specimens dimensions variability (especially local value of identical damage scenario was introduced but to thickness) at level of ± 0. Experimental campaign set.0 [GPa] variability is accounted to prevent “false alarm” Table 1 E-glass composite material due to the non-identical nature of three nominally properties. Different reference based methods of E11 46. Experimental Modal Analysis most practical engineering application there is a Experimental modal analysis was carried out to serious lack of experimental data that would estimate modal model of the structure. Then responses from the four sets were Measurement activity on the plates covered the analyzed individually and later on merged in the modal analysis of the intact plates. . Acquisition of the response signal evaluate the clearly identifiable damage of wind from all measurement points was done in sets to turbine blade component without false alarms. During the enable a correct quantification and measurement campaign the LMS SCADAS characterization of stochastic system properties Mobile and Test Lab software was used. Table 1 the magnitude of these changes and their impact presents the values of the E-glass composite on the final product behavior when assessing material properties [17]. The material first important conclusion that even if the samples is similar to the one typically used in load carrying were cut from the same construction the behavior panel of a wind turbine blade (Figure 1).0 [GPa] Health Monitoring [14. equal composite panel specimens under Important is the presence of plates geometrical investigation. The of each sample under the investigation was individual samples were cut from plates produced different.3 [-] model identification of the undamaged structure v12 0. used to drive the electro-dynamic shaker exciting The discussion is conducted to assess and the plates.0 [GPa] of the intact structure.0 [GPa] damage detection can be proposed for Structural E22 13. It is important to have at least an idea on had the same dimensions of 320 mm. intermediate and low). damage presence based on reference method Property Value [13]. decrease the sensor mass influence effect. The same procedure was repeated for the rest of the 2.

3. Left . This is the result of variability of the composite five first mode pairs was above 65%. onto the surface of the Consistency of modal vectors estimated and specimen. four areas were shape from test (left) and simulation (right) chosen to deploy the strain gauges. It is very important the good accordance in the mode shape there is to notice that if differences in behavior occurs in a difference in frequency value of the the intact sample. Values of MAC for The definition of regions S1 to S4. Static Four point bending test system. denoted as before FE model updating S1. Static four point bending test was performed on but a range of frequencies values (Figure 2). the plates to measure strains resulting from a given load and support configuration and 1st mode frequency test variability furthermore to introduce the damage (Figure 4).Lab 8B software and the strain gauge rosettes locations. calculated was compared by means of the Modal Assurance Criterion. . Despite materials production process. development of a structure health monitoring 2. To observe adequate strain levels it is important to fix the 566 Hz 523 Hz strain gauges in the areas of tension or Figure 3 Comparison of the second mode compress only. and C in first measurement point. experimental and initial numerical mode shape. Figure 5 presents the results of the pre-test made with Virtual. Figure 3 presents the examples of the Also the numerical pre-test was made. S3 and S4. Right .principal strain direction C22 Figure 5 The numerical pre-test result (left) and location of the strain gauges rosettes (right) at the sample A surface. B damage implementation. must take into account not one value of the frequency. based on reference modal model.principal strain direction C11. Differences between basic theory of the four point bending stress and strain distribution of the considered sample and the pre-test results are caused by the orthotropic properties of the multilayer composite material used for making the samples. 388 386 384 Frequency [Hz] 382 380 378 376 374 372 370 test_02 test_03 test_04 test_08 test_06 test_36 test_11 test_40 test_39 test_38 test_23 test_43 test_42 test_29 test_28 test_27 test_12 test_33 test_18 test_15 test_17 test_22 test_20 test_21 test_16 test_19 test_31 test_32 test_13 test_26 test_14 test_30 test_44 test_25 test_24 test_37 test_41 test_09 test_10 test_34 test_35 test_07 st 05 1 _0 test_ te Experiment No Figure 4 Schema (left) and the setup (right) Figure 2 Measured values of the first natural for the four point bending experimental frequency values for the intact samples A. On that base. All specimens were tested up to damage. S2. then further research in and investigated plates.

46E-04 Figure 7 Plate deformation (z-direction) under enforced displacement of 1mm along Table 2 Comparison of strain values in negative z axis of two middle half-cylinders.09E-05 15mm with local refinement S2 3. In each of these cases mapped mesh.04E-04 linear elements of the isotropic material S4 -2. 15mm (555 nodes. size S1 2.17mm.47E-05 -5. size S1 2. In this type of contact friction (1122 nodes.4 has been applied 1056 elements) (classical Coulomb friction).81E-05 for different types of support obtained under 1mm Triangle.30E-06 -1. In this case structure.37E-04 -3. In the case “d” which represents The numerical modeling study aimed into three linear-gap contact support it could be seen that goals: continuity of the second derivative of deflection is .41E-04 -3. and supporting on Triangular. linear-gap contact.51E-04 . it is investigated how reliable an maximum displacement is 1. and 1.20E-04 -3. This situation orthotropic shell model with averaged values (as is caused by the fact that equidistant node presented in Table. material properties. S2 2.12E-03 S4 4.46E-05 node distance that is defined on basis of the Triangular. S3 5. free mesh.20E-04 to model an interaction between the mesh.81E-04 direction). size 7mm composite plate and steel shafts during four point (2278 nodes.95E-04 1022 elements) S3 7.35E-04 -3.25E-06 -3.91E-04 2mm (1153 nodes.52E-04 pressure equals zero on the contacting interface 10mm if separation occurs. that is.34E-04 -3. S1 2.70E-04 -3. In the linear gap S3 4. S2 1.54E-04 support.20E-04 the model’s mesh density influencing strain result.48E-04 2218 elements) S4 -2.03mm.53E-04 8606 elements) simulations has been modeled with use of 3D S3 4. realized by rigid bar elements.21E-04 -2. free nodes initial position.79E-06 -5.2mm. Quad mapped S1 2. normal mesh. S3 4. and how to average these results regi average average Description which should be furthermore used in the updating on C11 C22 process.54E-04 -3.16E-06 properties (steel).11E-04 -3. 2184 elements) During the study three numerical models of S3 4. The DOF restrain type of mesh.1) is both for the static and assumption introduces additional stiffness dynamic simulations.13E-04 presented in (Figure 7). S4 4.36E-04 supports have been analyzed: shell-to-solid S4 -2.35E-04 -3. .21E-04 -1.48E-04 with static friction coefficient 0. b – support basing on the DOF restrain.3.17E-06 contact. and restrains (1215 nodes. and DOF restrain support modeling for static and dynamic respectively) shows mutually the same results.65E-05 -5. and lastly. S2 1.90E-04 -3. size S1 2.54E-04 bending test simulation. The plate deformations in z-direction S4 3. The shell-to-solid contact is based on the Quad mapped standard unilateral contact model.51E-04 plate has been modeled with the same linear size 5mm shell elements of the same size and orthotropic (4432 nodes.80E-04 -2. b – shell-to-solid contact simulation at load of 3827N type support.51E-04 five movements on two inner plate-shafts contact 2289 elements) lines allowing for lateral displacement (y.91E-04 Triangle. a direction C11 and C22 for numerical – unreformed model.37E-06 -6.29E-04 -3. basis of DOF restrain. Parametric study of FE The deflection in case of “b” and “c” (shell-to- solid contact.75E-04 -3. holds all the six possible movement on 10mm two outer plate-shafts contact lines. free enforced displacement of upper (inner) shafts is mesh.60E-04 -7.97E-04 contact there is assumption of constant the case of an unknown inner composite not held which is unrealistic. analysis The maximum deflection is respectively 1. S2 3. size S1 2. b – linear-gap contact type support. cylinders in contact S2 1. .

On Three Level Full Factorial to generate sufficient the contrary. coefficients values of 0. Static stains. computed on the DOE results with the regression As the final topology for the static simulation. Comparing deformation (b) and (c) it could be 4. For this purpose the simulations featuring different mesh densities and analysis flow was defined as presented on the element types have been performed. of the strains in longitudinal (C11). Material properties parameters were definition of regions is given in Figure 5. The dynamic simulation has been obtained from particular measurements performed on the quad mapped mesh of size (Figure 10). Probabilistic model updating stated that the DOF restrained model is slightly stiffer than the model with shell-to-solid contact. 10mm. and almost identical results of simulation with use of Figure 8 TASK Front view of the 50Tflop surface-to-surface and DOF restrained support computing power cluster in TASK the latter one has been selected for furthest Supercomputing Center (fot. Keeping in mind that the FE model should be used for optimization purposes (very large number of simulation runs). As it could be seen meshes were compared to the test results. Based on this model a quad mapped mesh of size 5mm (4225 nodes Self Adaptive Evolution optimization was defined and 4096 elements) without refinement has been with built-in target values of the variables as chosen. Next both static and dynamic finite gauges in the static test. The first step of size 10 and less generates almost the same of the analysis was calculation of the Design of results in the principal strains regardless of the Experiment (DOE). Gburski) numerical analyses. In the case of surface-to-surface support more computational costly nonlinear analysis with potential convergence problems must be performed. M. whereas linear contact gap support produce an over-stiff model with unrealistic deformation. Some selected results element models were incorporated. Fifth order RSM model was results.96. The scheme chosen was elements topology and the local refinement. These defined as the variables for the updating regions correspond to the placement of strain procedure. Computations were performed with Optimus Deformation in case of surface-to-surface software on a 50Tflop cluster in TASK Academic support and DOF restrained support is little Computer Centre (Figure 8). different. additional composite material plate. Static and dynamic measurements results were In order to determine the best mesh topology used to update the numerical model of the for the FE model optimization. with the free-free supporting condition Figure 9 Graph of the model updating process . The Figure 9. over-all rough mesh with fine local results for the precise Response Surface Model refinement does not give good principal strains calculation. and lateral natural frequencies and Modal Assurance (C22) directions for different meshes topology are Criterion were selected as output variables which presented in Table 2.

optimum is reliable.62 959. For sample A with the highest load (most severely damaged. for sample C (medium damaged) up to nine poles and for the lowest load for sample B (lowest damaged) the number of poles was the same as before damage implementation – namely eight poles. Next the reliability of the solution correlation line.00 Mode 2 562.49 Mode 4 945.61 834.29 1 118.71 961. comparison of modal vectors from test and Propagating waves interact with material updated model inhomogeneities and can be used for Structural The FE model of the composite plate has been Health Monitoring purposes.49 1 111. Set_1 MIN A MEAN A MAX A Damage A MIN C MEAN C MAX C DamageC Mode 1 375.18 830.63 663.40 554.72 946.42 567. therefore high frequency (short modal analysis results has been carried out by wavelength) waves were used to ensure means of the Modal Assurance Criterion sensitivity.86 832. Piezoelectric transducers were excited approach in order to associate the closest and then responses from their electrodes were numerical and experimental mode shapes. Main diagonal MAC terms present path it collects the signals. particularly in this presented. however not totally failured) the number of poles increased up to ten.34 855. The static and dynamic analysis are better consistent various numbers of the poles were identified for with measurement results (Figure 11).40 Mode 5 1 099.69 1 118. . Table 3 compares the values of natural frequencies for test made for the intact and damaged samples A and C. For these setup as in the modal test of the intact samples updated model parameter values results of both were used for the damaged plates testing. In registration consistency.76 739.08 946. as a consequence it is possible was assessed by means of the probabilistic to assert that within the investigated frequency Monte Carlo method. 20 000 experiments (sets of bandwidth the numerical model presents a input variable values) were calculated and all of dynamical behavior adequately close to the real them were feasible meaning that the found plate.71 378.78 568. Furthermore a approximate size of introduced damage was correlation between numerical and experimental unknown.19 843.52 565.35 929.92 268.61 1 116. The gathered using integrated generation/acquisition updating procedure accounted for both natural device. These mode the gradient based method was applied to fine pair table can be considered close to the optimal tune the values.26 1 117. the different damaged samples.62 378. In this study even the obtained from the model itself. It conducts generation and amplification frequency value and the mode shape of the signals to drive the piezos. piezoelectric transducers were used to excite and register Figure 11 Modal Assurance Criterion for elastic waves in the considered specimens.31 1 103. In order to perform damage detection using piezo–excited elastic waves.19 568.08 564.55 833. Figure 10 Examples of the evolution of updating parameters with the iteration range indicators Based on the global optimum found by the SEA corresponding mode shape pairs.45 378.84 964.60 378.63 375. as far as the results of modal analysis case for damage detection.31 Table 3 Comparison of the natural frequency for the intact and damaged sample A and C.76 392.52 546.34 Mode 3 837.73 375. refines them and very high values and it was possible to trace the sends to a PC via USB. As the result of presented probabilistic model updating procedure the final 5 Damage detection set of material properties values within 20% The same measurement equipment and test difference from initial ones was found.

33033-0267. “Torsional performance of wind dynamic behavior of the structure in the turbine blades . specimen surface. Stang. Polish drawn that even if the samples were cut from the Academy of Sciences). Conducted mapping development of the structure health monitoring procedure indicated that the greatest differences system. The authors of this work gratefully acknowledge Figure 12 Results of damage detection using support for this research under the project No. It is important to Obtained results were normalised to the notice that if the differences in behavior occur in maximum value. based on reference modal model. Branner. presented and used for modal models estimation. Acknowledgements This research was supported by Marie Curie European Reintegration Grants within the 7th European Community Framework Programme. The while registration was realised in the rest of main conclusion is that the implemented damage transducers. simulation of a 34m. should be underlined that the difference could be Both presented methods successfully identified also a result of transducer debonding caused by the presence of damage based on the reference four–point bending test. 7. but a range of frequencies damage could occur in this area. frequency. Transducer network was same construction the behavior of samples under designed to obtain as much information about the the investigation was different. Composite wind turbine Optimal locations of the strain gauges rosettes blade” Composite Structures 76 (2006) 52–61 were computed in preliminary FE model [2] Berring. Lamb wave propagation for 100 kHz 239191 “PROND” provided by the EU. Also the numerical prediction of the Knudsen. is Research Programme (EFP2007). M. Calculations were performed in the Academic Supercomputing Center TASK. presented partly in this paper. This suggests that “accurate” frequency. It was was obtained from measurement for damaged therefore possible to detect damage in the specimen. must between damaged an intact sample are in its take into account not only one value of the lower half (see Figure 12). References Experimental test data examples were [1] F. The procedure ought to be incorporated in the drawback of this method is more complex test detection procedure. Kyoto (JP). experimental modal analysis. Colour scale is from blue – the intact sample then further research in and minimum to red – maximum.. In result 132 signals were obtained caused a significant change in the sample FRF’s for intact plate and the same number of signals and in the number of poles of FRF’s. 2007. calculation. 8-13 Jul . This approach provided full specimen to introduce damage.W. Poland. Berggreen. 6. Lamb wave based approach not the reason a transducer self–testing provides also the localization of the damage. Gdansk.. Twelve the result of variability of the composite materials transducers were distributed on the whole production process. Further research activity should account for the numerical modeling of the fracture mechanism which was not included in the present investigation. Jensen. K. In order to ensure this is model comparison.Part 1: Experimental frequency range corresponding to global modes investigation” International conference on was carried out.G. J. The supported oriented towards test data variability and EFP-project is titled “Improved design of large estimation of numerical model parameters wind turbine blades of fibre composites – phase uncertainties obtained within the E-glass wind 4” and has journal no. Conclusions The work is also partly supported by the Danish The multidisciplinary and interdisciplinary Energy Authority through the 2007 Energy research. Measured signals were processed investigated composite plates by means of with special signal processing algorithm. turbine reinforcement composite material structure. The concept of this device was born at the Due to the differences of three series of tests Department of Mechanics of Intelligent Structures for each intact sample a first conclusion was (Institute of Fluid–Flow Machinery. “Structural testing and numerical Some general remarks have been formulated. Ankersen. composite materials (ICCM-16). Ch. Falzon. Excitation was control over the load value and led to different applied to each transducer from configuration level of damage in each sample structure. This is probably specimen condition as possible. However it values. H. procedure than in case of experiemtnal modal analysis. This will allow the prediction of the defect before it occurs. However they could not be The damage implementation process was placed uniformly due to the fact that four–point carried out by means of the very precise bending quasi–static test was performed on this hydraulic press. P.. H. B.

Worden. “Distribution of Defects in Wind Turbine Blades” Proc. J. Vandepitte. Oxford University Press. E. Lestari “A Combined Static/Dynamic Technique for Damage . S. P. and Elishakoff. C. Part A: Static test and equivalent single layered models” Composites Part A.. I. Vecchio A. and Yamazaki. T.. J. F. M. T. T. [3] D. February 2010. Special Issue dedicated to Professor Bruno Piombo. Casias. (2002)... “The Polymax Frequency-Domain Method: a New Standard for Modal Parameter Estimation?” Shock And Vibration. H. S. Griffith. and Ren. 11..Culshaw. G. Wandowski T. Jan 26-29.G. Joost. Mucchi E.. W.J.J. 2003. EWEC 2009 [10] Shinozuka. 2004. Sørensen J.J..T. van de peer. Oxford. Issue 2. Denmark.Z. of the IMAC XXII Dearborn.. Baumgart. Smart Structures and Materials—Proceedings of SPIE3986(2000)93–103..C. “Estimation of the uncertainty in measurement of composite material mechanical properties during static testing” Proc. Larsen. G. S. “Parameter Uncertainty and Variability in the Structural Dynamics Modeling Process” Proc... EWEC 2009 [14] Ostachowicz W. U. Qiao & K. Pages 257-270 [5] G.T.Staszewski. Florida. Pierro E.K. Orlando. USA. Guild. Philippidis. Short. M. S. The application of ulstrasonic Lamb-wave techniques to the evaluation of advanced composite structures. [6] MI Friswell.Manson. Risø-R-1181. 2004 [13] Stensgaard H.Elsevier Applied Science. Y. [12] Donders. MJ Terrell. T. [7] B. Mechanical Systems and Signal Processing 23 (2009)1805–1829. 2007. Lekou. Pavier 6454 March 2006 SANDIA REPORT “Delaminations in flat and curved composite [4] L. H. Stochastic structural dynamics: Progress in theory and applications”. M. Van Der Auweraer.. 395-409. Leuridan. “Stochastic finite element analysis: an introduction. Overgaard. Peeters.. 23rd International Modal Analysis Conference. JE Coote. Smith. Branner K. Volume 41. 1998 [11] Elishakoff. Carlén. edited by Ariaratnam. [17] G. Malinowski P.J.. Guillaume. USA. “Modal Analysis of wind turbine blades”. School of Modal Analysis XII Conference... “Experimental Modal Analysis of an Aircraft Fuselage Panel Incorporating Test Data Variabilities”. and T. Damage localization in plate- like structures based on PZT sensors. T. Krakow. Lu &W. H. Shigeoka T.. Large Variation “Finite Element Method for Stochastic Problems”. [16] P. G. paper 201. JR Fonseca & NAJ Lieven. Assimakopoulou. M. London. Schueller.. A. [15] S. [9] D. Simmermacher “Modal Testing Experimental Mechanics (2008) 48:17–35 of the TX-100 Wind Turbine Blade” SAND2005. I. [8] Luczak M. Hansen. J. D. Experimental Data for Uncertainty Quantification. S Adhikari. I. P. Kudela P.. Dom. I. laminates subjected to compressive load” Thomsena “Structural collapse of a wind turbine Composite Structures 58 (2002) 249–258 blade. K. Van der Auweraer..Pierce. W. Lunda and O. Detection of Laminated Composite Plates” Reese. February 2005.. B. F.