Jtam-A4.dvi JOURNAL OF THEORETICAL AND APPLIED MECHANICS 53, 2, pp. 395-407, Warsaw 2015 DOI: 10.15632/jtam-pl.53.2.395 INTRODUCTION AND EVALUATION OF A DAMPING DETERMINATION METHOD BASED ON THE SHORT-TERM FOURIER TRANSFORM AND RESAMPLING (STFR) Olaf Peter Hentschel, Lars Panning-von Scheidt, Jörg Wallaschek Leibniz Universität Hannover, Institute of Dynamics and Vibration Research, Hannover, Germany e-mail: hentschel@ids.uni-hannover.de; panning@ids.uni-hannover.de; wallaschek@ids.uni-hannover.de Markus Denk ALSTOM Power, Steam Turbines and Generators, Baden, Switzerland e-mail: markus.denk@power.alstom.com In the present paper, a frequency domain method for damping determination is presented. The described method is especially developed for low damped systems with well separa- ted eigenfrequencies. Using the Short-TermFourier transform andResampling (STFR) of the signal, decay curves of several mode shapes can be identified and amplitude-dependent damping values can be calculated. Additionally, two common methods for damping deter- mination are explained briefly. Finally, the quality of the introduced method is evaluated comparing the variances of the identified damping values bymeans of differentmethods. In this context, the damping for beams clamped in a suspendedway is analyzed. Stainless steel is used as the specimenmaterial. Keywords: damping determination, parameter identification, resampling, nonlinear dynamic analysis 1. Introduction In fields of mechanical engineering, High Cycle Fatigue (HCF) risk assessment is of great im- portance. To reduce this risk, the reduction and prediction of vibration amplitudes is a primary objective. Using nonlinear calculation tools (Petrov and Ewins, 2006; Siewert et al., 2008) to predict vibration amplitudes, the specification of contact properties as well as structural dam- ping (among other mechanisms of damping, Rao and Saldanha, 2003) are of great importance. To specify, for example, structural damping values, a suitable experimental setup and an appro- priate methodology for damping identification are elementary. In this context, Plunkett (1959) gives an overview of possiblemethods (time and frequency domain) for damping determination. Furthermore, Bert (1973) offers an overview of methods for modeling, experimental determi- nation and parameter identification of material damping. An extended Half-Power method is described by Jinting et al. (2013), whereby the parameter identification is performed in the fre- quency domain. In addition, Yang et al. (2003), Meissner (2012) and Feldman (1994) describe a method applying theHilbert transform to calculate decay curves. ByYang et al. (2011), Lardies and Gouttebroze (2002), Lamargue et al. (2000), Hans et al. (2000), Le and Argoul (2004) and Slavi et al. (2003), the wavelet transform is applied for parameter identification. The methodology which is presented in this paper is limited to structures excited by an impact. Additionally, the eigenfrequencies, respectively the resonant frequencies have to bewell separated. The described method includes the Short-Term Fourier transform and Resampling (STFR) of the analyzed signal. This procedure makes it possible to get adequate information concerning the vibration amplitude and the current eigenfrequency which varies as a function 396 O.P. Hentschel et al. of amplitude for nonlinearmechanical systems. On this basis, it is possible to identify the decay curve of one selected mode shape of interest. Using a decay fit (Rice et al., 2007; Siewert et al., 2008), it is possible to identify the damping value of one selectedmode shape. Subsequently, two other commonmethods for parameter identification arepresented.Furthermore, an experimental setup is introduced briefly, which provides the experimental data for damping determination. Finally, the presented methods are used to evaluate the developed STFR method in the last part of this paper. 2. STFR method The presented method is used to identify damping values of mechanical systems at decaying free vibrations. This method is restricted to impulse excited structures. In this context, only the output signal is needed for damping identification. To achieve an adequate identification concerning the eigenfrequency and amplitude, the signal is additionally resampled. From this, tracking of the eigenfrequency depending on the amplitude is also possible. 2.1. Short-Term Fourier transform and Resampling The basis for the used method is the measured output signal. In this case, the specimen is excited by an impact, whereby decay curves are recorded. The signal x can be specified by the following data sequence withNs values. x[n] = { x(t=nT0) for 0¬n¬NS −1 0 otherwise (2.1) wherein T0, t and n denote the sample rate, the time and the value number. From Fig. 1, the superimposed character of the signal (response of severalmode shapes due to impact excitation) can be clearly seen. Fig. 1. Measured free vibration In the next step, the signal has to be extracted with respect to one mode shape of interest. Therefore, the signal is analyzed stepwise. The signal is equidistantly divided into U = NS NW −1 (2.2) Introduction and evaluation of a damping determination method... 397 parts. For these parts of thedecaying signal, theDiscreteFourierTransform(DFT) is performed. NW describes the number of measurement points of the analyzed signal part. The following u signal portions with 0¬u¬U are resulting in xu[w] = { x ( t=(uNw+w)T0 ) for 0¬w¬Nw−1 0 otherwise (2.3) whereinw denotes the value number of the signal portion. Next, the data portion xu[w] has to be transformed into frequency domain Xu[k] = Nw−1∑ w=0 x ( (uNw+w)T0 ) e −j2π kw Nw (2.4) wherein k is the spectral line number. Using this formulation, the frequency resolution reads ∆ωf = 2π T0Nw (2.5) This leads to the problem that the amplitude at the eigenfrequency is underestimated, in par- ticular for weakly damped structures. Performing a resampling of the analyzed data portion solves this problem. The goal in this context is the identification of the required sample rate in a way that the frequency resolution ∆ωf,n is equal to an integer numbered multiple g of the eigenfrequency of interest ωd,int ωd,int = g∆ωf,n (2.6) To achieve the required sample rate corresponding to the criteria in equation (2.6), a real num- bered alternation d of the sample rate T0 is necessary ∆ωf,n = 2π T0,nNw = 2π dT0Nw (2.7) From this, an integral numbered decimation using the factor a followed by an integral numbered interpolation using the factor b is performed. Hereby, a resampling using the real factor d= a/b is realized. To avoid aliasing, theNyquist-Shannon sampling theorem has to be satisfied. In this respect, a lowpass filter regarding the varied sample rate is used. This lowpass filter avoids the aliasing effect at decimation and interpolates the signal at interpolation. The resampling process (Oppenheim and Schafer, 2010) is shown in Fig. 2. Fig. 2. Resampling 2.2. Methodology The presented methodology is based on the assumption of well separated eigenfrequencies. Thus, the knowledge concerning the position of one eigenfrequency, respectively the frequency of themode shape of interest, is important. In general, the exact eigenfrequency, which is the crite- rion for the described resamplingmethodology, is not known. For this purpose, an identification of the exact eigenfrequency using an optimization method (such as the Nelder-Mead algorithm, 398 O.P. Hentschel et al. Fig. 3. Methodology Nelder and Mead, 1965) is necessary. The resampling factor d is optimized to detect the maxi- mum amplitude within the analyzed signal portion u. In Fig. 3, the described methodology to calculate the decay curve is shown for one mode shape of interest. After calculating the eigenfrequency using the describedmethodology, themodal amplitude for each part of the decay curve can be isolated. Calculating the average time tm[u] = (u−1)NwT0+ Nw 2 duT0 (2.8) for each part of the decay curve u (Fig. 4), it is possible to reconstruct the decay curve for one mode shape of interest using the identified amplitude xa[u]. Fig. 4. Identified decay curve To enable a quantification of a damping parameter, an analytical formulation of the decay curve is necessary. Using the least squares method, the decay coefficient δ can be determined U∑ u=0 ( x [tm[u]]− x̂ae −δtm[u] )2 →min (2.9) In equation (2.9), x̂a denotes the initial amplitude of the decaying signal. Performing the signal resampling, an additional advantage is present. In this context, the eigenfrequency for each time Introduction and evaluation of a damping determination method... 399 point of the decay curve can be calculated. Hereby, it is possible to examine the dependency of frequency and time. From the frequency versus time history, nonlinearities with respect to variation in frequency can be identified. In Fig. 5, two possible frequency-time curve shapes are presented. Fig. 5. Frequency versus time 2.3. Amplitude-dependent damping coefficient To identify an amplitude-dependent damping, the identified decay curve has to be divided into several parts p. From this, the decay coefficient δm can be calculated as a function of the mean amplitude xp (see Fig. 6) xp[tm,u] = x̂a,pe −δ[xp]tm[u] (2.10) Fig. 6. Amplitude-dependent normalizedmean damping coefficient In addition to the division of the decay curve to calculate amplitude-dependent damping values, also an analytical approach is proposed. On this basis, the decay coefficient described in equation (2.9) is replaced by a time-dependent, respectively amplitude-dependent decay coeffi- cient δt(t) x[tm,u] = x̂ae −δt(tm[u])tm[u] (2.11) 400 O.P. Hentschel et al. In the next step, the time derivative of the presented decay curve is calculated. Following the calculated time derivative of the decay curve, ẋ is divided by the decay curve itself. Hereby, an inhomogeneous first order differential equation results − 1 t ẋ x = δ̇t+ 1 t δt (2.12) After separation of the variables and the solution of the homogeneous differential equation, a variation of the constants is performed to solve the differential equation. The analytical solution to equation (2.12) is the time-dependant decay coefficient δt(t)=− 1 t ∫ ẋ x dt+ C t (2.13) The constantC can be calculated using a signal part ts of the decay curve. For this purpose, it can be assumed that the average decay coefficient through this signal part is equal to the non- amplitude-dependent decay coefficient δm,ts. The latter can be calculated using the formulation in equation (2.9) applying the least squares method C = ( δm,ts ( max(ts)−min(ts) ) + max(ts)∫ min(ts) (1 t ∫ ẋ x dt ) dt )( ln max(ts) min(ts) ) −1 (2.14) Using the prior identified eigenfrequency ωd(t) as a function of time and the time-dependent decay coefficient, damping can be quantified by different physical magnitudes Λ(t)= 2π δ(t) ωd(t) =2π D(t) √ 1−D(t) (2.15) In equation (2.15), Λ(t) and D(t) denote the logarithmic decrement and the damping ratio. Applying the amplitude formulation in equation (2.10), the time-dependent damping values can be specified in terms of amplitude. 2.4. Example of application To clarify the presented methodology, a predefined amplitude-time curve is analyzed. For this reason, a time-dependent, i.e. amplitude-dependent, damping and frequency value is used x(t)= x̂1e −δ1(t)t cos(ω1(t)t)+ x̂2e −δ2(t)t cos(ω2(t)t)+ x̂3e −δ3(t)t cos(ω3(t)t) (2.16) The signal consists of three frequencies, whereby the second one may be of interest. The other signal parts can be treated as additional mode shapes which are not of interest for the analysis. In this context, x̂i, δi(t) and ωi(t) denote the initial amplitude, the decay coefficient and the vibration frequency of each signal part. In Table 1, the magnitudes of the different values are summarized. Hereby, the units millimeters and seconds are used. In the first step, the presented signal is transformed stepwise into frequency domain. The signal is resampled stepwise with respect to the analyzedmode shape.The considered frequency range has to consist only of the mode shape of interest. From this, an optimization of the amplitude with respect to themaximum amplitude in the analyzed signal portion is performed. In Fig. 7, one signal portion before resampling and after resampling and the resulting time- frequency-amplitude diagram are shown. With respect to the time-frequency axis, the eigenfrequency variation can be identified. In doing so, the time dependency of the identified and the default eigenfrequency of the second signal portion are shown in Fig. 8. Introduction and evaluation of a damping determination method... 401 Table 1. Summarized specifications of equation (2.16) i x̂i [mm] δi(t) [1 s ] ωi(t) [rad s ] 1 10mm 1.2 1 s −0.05 1 s2 t 2π ·400 1 s + 1 s2 t 2 8mm 0.6 1 s −0.045 1 s2 t 2π ·500 1 s +2 1 s2 t 3 11mm 3 1 s −0.05 1 s2 t 2π ·650 1 s + 1 s2 t Fig. 7. Time-frequency-amplitude curve and signal resampling Fig. 8. Eigenfrequency versus time In this figure, it is shown that the eigenfrequency time dependency is estimated very well. Regarding the time-amplitude axis, the decay curve of one mode shape of interest can be ana- lyzed. Applying the developed methodology to identify amplitude-dependent damping values (presented in Section 2.3), the damping-time (respectively decaying amplitude) dependency can be analyzed (see Fig. 9). Accordingly to thewell estimated eigenfrequency behavior, the damping values can be evalu- ated in a suitable way. On this basis, it can be shown that the frequency and damping variation are investigable in their magnitudes and courses. 402 O.P. Hentschel et al. Fig. 9. Decay coefficient versus time 3. Common methods for damping determination In this Section, two common methodologies for damping determination used in modal analysis are presented. For this purpose, an estimation of the transfer function is necessary. The esti- mated transfer function is the basis for the presented methods. In this context, the Rational Fractional Polynomials method (Richardson and Formeti, 1982) and the Least-Squares Com- plex Exponential method (He andFu, 2001) are described. In the final step, the damping values determined by the developed methodology are compared to the damping values identified by the RFP and the LSCEmethod respectively. 3.1. Estimation of the transfer function Due tonoisewhich occurs in themeasured input y(t) and outputx(t) signal, an estimation of theFrequencyResponseFunction (FRF) is necessary.Hereby, several FRFestimators areusable. To determine the estimators, the cross-spectrum (SXY and SYX) and the auto-spectrum (SYY and SXX) of the input and output signals are used. In Table 2, the common FRF estimation functions are presented. Table 2. FRF estimators Estimator Equation Properties H1(ω) SXY (ω) SYY (ω) underestimation of amplitude, good estimation in anti-resonance H2(ω) SXX(ω) SYX(ω) overestimation of amplitude, good estimation in resonance Hv(ω) √ H1(ω)H2(ω) mix of H1(ω) andH2(ω), minimize input and output noise Resulting from the applicability of the RFP and LSCEmethod to linear systems, the cohe- rence γ of the signals has to be analyzed, too. If the signals are free from noise and error, and a linear system behavior being additionally existent; theH1(ω) estimator is equal to theH2(ω) estimator. In this case, the system causality has a maximum, and the coherence is equal to 1. 3.2. Rational fractional polynomials method The Rational Fractional Polynomials (RFP)method is a frequency domain method. In this case, themodel fit is performed in frequency domain. The basis for thismethod is the estimated FRFwith the proceduredescribed in Section 3.1.At the beginning, the analytical formulation of Introduction and evaluation of a damping determination method... 403 a FRF of amulti degree of freedom system is necessary. This transfer function can be expressed in Laplace domain using the complex argument s= jωi Ha(s)= n∑ k=1 ( rk s−pk + r∗k s−p∗ k ) = m∑ k=0 aks k n∑ k=0 bks k (3.1) wherein, the number of pole pairs pk, i.e. the number of degrees of freedom is described byn. In this context, rk, ai and bi denote complex residues and coefficients of the analytical formulation of the FRF. In the next step, the error between the analytical formulation of the FRF and the measured data has to be calculated ei = m∑ k=0 ak(jωi) k −hi [ n−1∑ k=0 bk(jωi) k+(jωi) n ] (3.2) Using hi, the measured data of the FRF at the frequencies ωi are defined. The square error criterion is specified as follows L∑ i=1 e2i →min (3.3) This errorhas tobeminimizedusingthevariablesai and bi.Basedonthecalculated variables, the constant eigenfrequencies and the damping values can be determined. 3.3. Least-squares complex exponential method The Least-Squares Complex Exponential (LSCE) method is a time domain method. Here, the fit is performed in the time domain. The basis within this method is the Inverse Fourier Transform (IRF) of the FRF formulated in frequency domain Ha(s)= 2n∑ k=1 rk s−pk (3.4) In the next step, the IRF is calculated ha(t)= ifft ( Ha(s) ) = 2n∑ k=1 rke pkt (3.5) This equationdescribes the structural responseafter impulse excitation. If the response ispresent at discrete time points, the exponential expression can be replaced by zmk = e pkm∆t using the z-transform. Based on this, it is possible to rewrite equation (3.5) as follows ha(m∆t)= 2n∑ k=1 rkz m k (3.6) Applying theProny equation (HeandFu, 2001), thepolespk i.e. z m k are estimated.This equation represents zmk as roots of a polynomial with real coefficients c0+ c1z+ c2z 2+ . . .+ c2nz 2n =0 (3.7) This equation is for each root equal to 0. Using the several samples m of the IRF data, it is possible to merge equation (3.6) with equation (3.7) 2n∑ m=0 cmha(m∆t)= 0 (3.8) 404 O.P. Hentschel et al. From this relationship, the coefficients cm can be determined directly. Using the calculated coefficients, the roots zm in equation (3.7) can be calculated. In the final step, the constant eigenfrequencies and the damping values can be determined by the poles of the system zm. It should be noted that in order to apply the RFP and the LSCEmethod, the input signal (excitation) and the output signal (measured decay) are necessary. In a next step, the FRF has to be estimated. From this, two noisy signals are used for parameter identification. 4. Experimental tests In this Section, an experimental setup for damping evaluation is presented briefly, see Hentschel et al. (2014) for details. This setup delivers the necessary data to compare the methods for parameter estimation. It consists of a specimen with defined eigenfrequencies which is clamped in its nodes of vibration (regarding the analyzed mode shape). The structure is excited by a force impact using a voice coil actuator. Fig. 10. Experimental setup 4.1. Excitation and measurement Theanalyzed specimen is excitedbya force impactgeneratedbyavoice-coil actuator.Hereby, an automated test sequence with a high reproducibility of the excitation force is realizable. In addition, no mass loading between excitation and structure is necessary. From this, a potential source for data distortion is eliminated. The voice-coil actuator is operating in current mode to control the acceleration of the actuator. Thedecaying velocity ismeasuredbya laser vibrometer. The optical measurement technique is used to avoid themechanical coupling between the sensor and the structure. In Fig. 11, the schematic experimental setup is presented. The excitation and themeasurement are carried out at one point of the structure. In doing so, a system reduced to a single input and output is analyzed. 4.2. Specimen clamping The investigated specimens are beams with defined eigenfrequencies. The sample geometry (width, length and thickness) is selected in away so that the eigenfrequencies arewell seperated. To reduce the influence of frictional damping due to clamping and to ensure a linear system behavior, the specimens are clamped in a suspendedway, see Hentschel et al. (2014) for details. The beams are clamped in their nodes of vibration. Bolts with an apex are used to fix the specimens depending on the analyzedmode shape. This allows an evaluation of “free”modes of vibration and prevents rigid bodymode shapes. Introduction and evaluation of a damping determination method... 405 Fig. 11. Experimental setup (schematic) 5. Evaluation To quantify the quality of the presented STFR method, the damping values identified using the different methods are compared. To apply the LSCE and the RFP method respectively, an estimation of the transfer function is necessary. As it is usual for modal analysis, the average of several transfer functions is performed, until a sufficient value of the identified coherence is present. For this purpose, an averaging of five measurements is performed. On this basis, the FRF is estimated using the H2(ω) FRF estimator. Applying the LSCE and the RFP method, the damping values are calculated for the first two mode shapes of the specimen. To calculate a standard deviation of the identified damping values using the two common methods, the described procedure is repeated six times after reassembly of the experimental setup. Using the performed thirty measurements to identify the damping by making use the RFP and theLSCEmethod respectively, the STFRmethod is applied to identify the damping values. Hereby, the decay curve is identified by means of the procedure described in Section 2.2. On the basis of the analytical formulation in equation (2.9), the least square method is applied to identify the decay coefficient. Themean damping values and the standard deviations calculated with the several methods are shown in Fig. 12. Fig. 12. Methods for damping determination It can be seen that nearly the same mean damping values are calculated using the several methodologies. The standard deviations calculated on the basis of the RFP and the LSCE methods show almost equal magnitudes. However, the standard deviation found by using the STFR method is lower in comparison to the other methods. This difference arises from the calculation of the FRFwhich is necessary to apply the LSCEand theRFPmethod respectively. Besides themeasured output signal, themeasured input signal includingnoise andmeasurement errors are necessary, too. This is not the case applying the STFRmethod. 406 O.P. Hentschel et al. 6. Conclusions In this paper, amethod for the evaluation of damping values from experimental tests is presen- ted. The presented method is limited to excited structures by an impulse with well separated eigenfrequencies. On this basis, resampling of the original data is introduced to achieve a suf- ficient quality concerning the calculated amplitude and frequency. An additional benefit of the applied resampling procedure is the identification of the present eigenfrequency depending of the vibration amplitude. To perform the damping identification, only the output signal is ne- cessary in contrast to other common methods. For this purpose, determination of the transfer function is not necessary. In addition, the analysis of a given output signal is performed to clarify the approach for parameter identification. In this context, an amplitude-dependant dam- ping and eigenfrequency is predefined. Hereby, the possibility of non-linear signal evaluation is demonstrated. Finally, the presented methodology is compared to common methods used in linear modal analysis. Hereby, an analysis of experimentally identified damping values is performed. It can be shown that the intendedmethod allows adequate parameter identification, in which only the output signal is used. Additionally, the identification of the amplitude dependent frequency and damping value is possible. References 1. Bert C.W., 1973, Material damping: an introductory review of mathematical models measures and experimental techniques, Journal of Sound and Vibration, 29, 129-153 2. FeldmanM., 1994,Non-linear system vibration analysis usingHilbert transform-II, forced vibra- tion analysis method ‘Forcevib’, Journal of Mechanical Systems and Signal Processing, 8, 309-318 3. Hans S., Irbaim S., Pernot S., Boutin C., Lamarque C.H., 2000, Damping identification in multi-degree-of-freedom systems via a wavelet-logarithmic decrement – part 2: study of a civil engineering building, Journal of Sound and Vibration, 253, 375-403 4. He J., Fu Z.F., 2001, Modal Analysis, Butterworth Heinemann, Oxford, Auckland, Boston, Jo- hannesburg, Melbourne, NewDelhi 5. Hentschel O.P., Panning-von Scheidt L., Wallaschek J., Denk M., Masserey P.A., 2014, Influential parameters on structural damping values of turbine blades,Proceedings of ASME Turbo Expo 2014, Power for Land, Sea and Air, June 16-20,Düsseldorf, Germany, PaperGT2014- 25656 6. Jinting W., Dandan L., Feng J., Chuhan Z., 2013, Accuracy of the half-power bandwidth method with a third-order correction for estimating damping in multi-DOF systems, Earthquake Engineering and Engineering Vibration, 12, 33-38 7. LamarqueC.H., Pernot S., CuerA., 2000,Damping identification inmulti-degree-of-freedom systems via a wavelet-logarithmic decrement – Part 1: Theory, Journal of Sound and Vibration, 253, 361-374 8. Lardies J., Gouttebroze S., 2002, Identification of modal parameters using the wavelet trans- form, International Journal of Mechanical Sciences, 44, 2263-2283 9. LeT.P.,ArgoulP., 2004,Continuouswavelet transform formodal identificationusing freedecay response, Journal of Sound and Vibration, 277, 73-100 10. MeisnerM., 2012,Accuracy issues of discreteHilbert transform in identification of instantaneous parameters of vibration signals,Acoustic and Biomedical Engineering, 121, 164-167 11. Nelder J.A., Mead R., 1965, A simplex method for function minimization, The Computer Jo- urnal, 7, 308-313 Introduction and evaluation of a damping determination method... 407 12. OppenheimA.V., SchaferR.W., 2010,Discrete-TimeSignal Processing,Third edition,Prentice Hall 13. PetrovE.P., EwinsD.J., 2006,Effects of damping and varying contact area at blade-disk joints in forced response analysis of bladed disk assemblies, Journal of Turbomachinery, 128, 403-410 14. PlunkettR., 1959,Measurement ofDamping, StructuralDamping,RuzickaJ., ed.,ASME,Atlan- tic City, NJ USA, 117-131 15. Rao J.S., Saldanha A., 2003, Turbomachine blade damping, Journal of Sound and Vibration, 262, 731-738 16. Rice T., Bell D., Singh G., 2007, Identification of the stability margin between safe operation and the onset of blade flutter, Proceedings of ASME Turbo Expo 2007, Power for Land, Sea and Air, May 14-17,Montreal, Canada, Paper GT2007-27462 17. Richardson M.H., Formeti D.L., 1982, Parameter estimation from frequency responsemeasu- rementsusing rational fractionalpolynomials,Proceedings of theFirst InternationalModal Analysis Conference, 167-180,Orlando, USA 18. Siewert C., Panning L., Gerber C., Masserey P. A., 2008, Numerical and experimen- tal damping prediction of a nonlinearly coupled low pressure steam turbine blading, Proceedings of ASME Turbo Expo 2008, Power for Land, Sea and Air, June 09-13, Berlin, Germany, Paper GT2010-51073 19. Slavič J., Simonovski I.,BoltežarM., 2003,Damping identificationusingacontinuouswavelet transform: application to real data, Journal of Sound and Vibration, 262, 291-307 20. YangY.,CascanteG.,PolakM.A., 2011,Newmethod for the evaluationofmaterial damping using the wavelet transform, Journal of Geotechnical and Geoenvironmental Engineering, 137, 798-808 21. Yang J.N., Lei Y., Pan S., Huang N., 2003a, System identification of linear structures based onHilbert-Huang spectral analysis. Part 1: Normalmodes,Earthquake Engineering and Structural Dynamics, 32, 1443-1467 22. YangJ.N., Lei Y., PanS.,HuangN., 2003b, System identification of linear structures based on Hilbert-Huang spectral analysis. Part 2: Complex modes, Earthquake Engineering and Structural Dynamics, 32, 1533-1554 Manuscript received July 26, 2014; accepted for print November 13, 2014