Non-parametric identification of a non-linear buckled beam using discrete-time Volterra Series

Porto, Portugal, 30 June - 2 July 2014 A. Cunha, E. Caetano, P. Ribeiro, G. Müller (eds.) ISSN: 2311-9020; ISBN: 978-972-752-165-4 Non-parametric identification of a non-linear buckled beam using discrete-time Volterra Series Cristian Hansen 1, Sidney Bruce Shiki 1, Vicente Lopes Jr 1, Samuel da Silva 1 1 Departamento de Engenharia Mecânica, Faculdade de Engenharia de Ilha Solteira, UNESP - Univ Estadual Paulista, Av. Brasil 56, 15385-000, Ilha Solteira, SP, Brasil email: engcristianhansen@gmail.com, sbshiki@gmail.com, vicente@dem.feis.unesp.br, samuel@dem.feis.unesp.br ABSTRACT: The consideration of nonlinearities in mechanical structures is a question of high importance because several common features as joints, large displacements and backlash may give rise to these kinds of phenomena. However, nonlinear tools for the area of structural dynamics are still not consolidated and need further research effort. In this sense, the Volterra series is an interesting mathematical framework to deal with nonlinear dynamics since it is a clear generalization of the linear convolution for weakly nonlinear systems. Unfortunately, the main drawback of this non-parametric model is the need of a large number of terms for accurately identifying the system, but it can be overcomed by expanding the Volterra kernels with orthonormal basis functions. In this paper, this technique is used to identify a Volterra model of a nonlinear buckled beam and the kernels are used for the detection of the nonlinear behavior of the structure. The main advantages and drawbacks of the proposed methodology are highlighted in the final remarks of the paper. KEY WORDS: Volterra series, detection of nonlinearities, nonlinear buckled beam, orthonormal basis function 1 INTRODUCTION In real-world engineering structures it can be necessary to take into account the nonlinearities depending on the level of excitation, type of material, joints, level of displacement, and applied load [1,2]. Different experimental effects can be observed analyzing the response of the system operating in nonlinear conditions, as jumps, harmonics, limit cycle, discontinuities, etc. [3]. Unfortunately, even with the recent development, the nonlinear identification approach is not mature as the classical linear techniques. Thus, there is a growing search for general and easy methods able to treat the most common nonlinearities in engineering structures, mainly the polynomial stiffness effect [4]. Several methods have been used for detection, identification and analysis of nonlinear behavior in structural systems, as for example: nonlinear time series [5], nonlinear normal modes [6], harmonic probing method using Volterra series [7], principal component analysis [8], extended constitutive relation error [9], frequency subspace methods [10], harmonic balance method [11], and others. However, all of these methods have drawbacks in some aspects. So, there is no general approach for the analysis nonlinear systems. Among these methods the Volterra series is very powerful because it is a clear generalization of the convolution of the input force signal with the impulse response function and several properties of the linear systems can be extended for nonlinear systems [12,13]. However, problems associated with overparameterization, ill-posed convergence, etc. have motivated criticism to this approach [14]. Partially this is a true statement; although, there exist some techniques to overcome this inconvenience. In order to make easier the practical application of this technique for engineers, the present work shows the use of discrete-time version of the Volterra series expanded with orthonormal functions [15]. The main contributions of this paper is to detect and identify, in the same time and with the same experimental data, the nonlinear behavior in function of the level of input signal applied in a geometrically nonlinear buckled beam. The paper is organized in four sections. First of all, a background review in the Volterra series expanded in orthonormal functions is discussed. Next the description of the test rig and the experimental tests performed are described in full details. Nonlinear features detection using classical metrics and the Volterra kernels contribution in the experimental output are also presented. Finally, the concluding remarks are presented. 2 DISCRETE-TIME VOLTERRA SERIES The output of a system can be described by a combination of linear and nonlinear contributions given by [12]: y( k) y ( k) y ( k) y ( k) (1) 1 2 3 where y ( ) 1 k is the linear term of the output yk ( ) and 2 3 y ( k) y ( k) are the nonlinear terms of the response. Each th order component can be written using a multidimensional convolution: N N 1 1 (2) i n1 0 nm 0 i1 y ( k) ( n,, n ) u( k n ) where ( n,, n ) 1 is the th Volterra kernel considering N,, 1 the memory length of each kernel and N uk ( ) is the input signal. The order of the model can generally be truncated in 3 to represent most part of structural nonlinearities with smooth behavior. 2013

Unfortunately, there are several problems associated with ill-posed problems of convergence due to the high number of terms N,, 1. An effective way to overcome these N problems is to reduce the number of terms in the Volterra kernels by using orthonormal Kautz functions [16]. Thus, the Volterra kernels can be described by: J J 1 1 i1 1 i 1 j1 1 (3) ij j ( n,, n ) i,, i ( n ) where J,, 1 is the number of samples in each J orthonormal projection of the Volterra kernels i,, i 1 and ( n ) are the Kautz functions that are applied in the i representation of underdamped oscillatory systems. Details about the application of Kautz functions can be found in [17] and [18]. In this sense, the Volterra series can be rewritten as: 3.1 Buckled beam description The test structure is composed by an aluminum beam with dimensions of 460 18 2 mm in a clamped configuration with a preload applied in the top. The structure can be axially loaded by a compression mechanism until the buckling of the beam. An electrodynamic shaker is attached 65 mm from the clamped end of the beam. A laser vibrometer is focused in the center of the beam measuring the velocities in this position. Four accelerometers were also used for monitoring the vibration in the length of the beam, but these signals are not used in this work. Figures 1 and 2 shows the experimental setup and a schematic figure of the test rig respectively. J J 1 1 (4) y ( k) i,, i l ( k) i1 1 i 1 j1 which is now a multidimensional convolution between the i,, i 1 and l ( k ) which is the input orthonormal kernel signal uk ( ) filtered by the Kautz function ( n ): V 1 i i ni 0 l ( k) ( n ) u( k n ) (5) i Figure 1. View of the test rig. where V max{ J1,, J }. Since the Volterra series is linear with relationship to the parameters of the model, the terms of the orthonormal kernels can be grouped in a vector and can be estimated by a classical least-squares approximation: T 1 T ( ) y (6) where the matrix contains the filtered input l ( k ) and y y(1),, y( K). The estimation of the orthonormal kernels can also be done in two steps since the linear and nonlinear components of the response are represented in a separated way. It is possible to estimate the linear kernel using a dataset where it is known that the system behaves in a linear way. A second dataset can then be used to estimate the nonlinear kernels, considering that the time series shows some relevant nonlinear behavior. This two step methodology is applied in this work in order to ensure a coherent model. 3 EXPERIMENTAL APPLICATION IN A NONLINEAR BEAM This section shows the main results of the application of Volterra series in a test rig. The results show the applicability of this technique in the identification of nonlinear structures. Figure 2. Schematic figure showing the test rig. This structure was tested in growing levels of input amplitude configured in the signal generator in order to observe the nonlinear behavior of the system. It was used a chirp input sweeping up the frequency range from 20 to 50 Hz and a white noise sequence filtered in the same range of frequency in order to excite essentially the first mode of the system (around 37 Hz). A sampling frequency of 1024 Hz was used and 4096 samples were recorded. Stepped sine and random excitation were also applied in order to observe some interesting nonlinear phenomena in the studied system. The next section shows the effects of the nonlinear behavior. 3.2 Detection of nonlinear behavior The signal data measured in the tests can be analyzed in order to detect some nonlinear behavior through the violation of the superposition principle. The presence of harmonic frequencies 2014

in the response of the system is the simplest way to do it [1]. Figure 3 shows the spectrogram of the response of the structure when a chirp excitation is applied. In the time from 0 to 2 seconds it is possible to observe a stronger diagonal line that follows the input frequency. However, some parallel lines starting around 1 s can be also observed representing the second and third harmonics of the input frequency. This clearly shows the nonlinear behavior of the system. Figure 3. Spectrogram of the output signal to a chirp input (input at 0.1 V). Figure 4 shows the frequency response function (FRF) in three different levels of excitation for the chirp signal. The levels of input correspond to 0.01 V (low input), 0.05 V (medium input) and 0.1 V (high input) applied in the signal generator. The natural frequency of the system presents a stiffening effect since it moves to the right. This is another violation on the superposition principle since for a linear system the ratio between the response and the excitation is kept constant for different levels of the input signal. One can also observe differences in the peak amplitude of the FRF. Figure 5. FRF of the system in different levels with a random input. Another test performed was done exciting the structure with stepped up sinusoidal excitation. The big advantage of this kind of input signal is that it concentrates the energy in a single frequency and in this case it is possible to clearly observe some nonlinear phenomena like the jump [18]. Figure 6 shows the frequency response using a frequency resolution of 0.1 Hz in three diferent levels of stepped sines with comparison to the FRF based on the random excitation (linear). The frequency response of the stepped sine with a low input stepped sine is very close to the one calculated with random excitation. For higher amplitude levels, it is possible to observe the change in the natural frequency of the system and a sudden drop in the amplitude of the FRF characterizing the jump phenomenon. Figure 6. Frequency response of the stepped sine testing in different amplitude levels. Figure 4. FRF of the system in different levels with a chirp input. The random excitation can also be used to estimate the FRF. This is a common test done in the industry. However, as showed in Fig. 5 there is no clear effect of nonlinear behavior caused by changes in the amplitude of excitation. The random excitation is usually not very appropriate to excite nonlinearities of the system since the input energy is spread in a broadband of frequency [1]. All this results demonstrated the nonlinear effects of the test rig in different conditions of the excitation signal. The presented behavior show some features available in systems with the restoring force of the stiffness with nonlinear polynomial shape. 3.3 Volterra kernel identification In this section the Volterra kernels are computed in order to extract a model describing the behavior of the first mode of 2015

Proceedings of the 9th International Conference on Structural Dynamics, EURODYN 2014 the structure when the nonlinearity is activated. The procedure can also be used to detect the level of nonlinearity presented. A discrete-time Volterra model is truncated until the third kernel to consider even and odd harmonics that are presented in the response of the system as shown in Fig. 3. All datasets used were based on a chirp input and the corresponding response measured. The random input signal was not employed for this task since this kind of signal is not appropriate to excite the nonlinear behavior of the structure as shown in Fig. 5. The low chirp input dataset (0.01 V) was used to calculate the first Volterra kernel and the high input dataset (0.1 V) was used to estimate the second and third kernels using the two step methodology previously described. One pair of Kautz functions was used in order to represent each Volterra kernel. These functions have complex conjugate parameters representing the dynamics of the kernel. These parameters can be represented in terms of the frequency ( ) of the system while the second kernel 2 (n1, n2 ) is a matrix and the third kernel 3 (n1, n2, n3 ) is a tridimensional matrix. This last kernel is illustrated through the main diagonal of the matrix since the full representation is not possible in a conventional way. and the damping ratio ( ). These values are generally close to the values describing the linear dynamics of the system. Methodologies to find these parameters are available [18] but they are generally very limited since previous knowledge of the shape of the Volterra kernel is needed. In this paper the frequencies and damping ratios of the first three kernels were defined using an optimization algorithm with the objective to minimize the prediction error of the model. The algorithm was based on a first global search by a genetic algorithm [20] followed by a local search using sequential quadratic programming [21]. Table 1 shows the results of the optimization algorithm. As stated before, the frequency of the third pole is close to the frequency of the first pole and the second frequency has a value around two times the first one. Figure 8. 2nd Volterra kernel 2 (n1, n2 ). Table 1. Kautz parameters of the orthonormal filters. 1 [Hz] 1 [%] 2 [Hz] 2 [%] 3 [Hz] 3 [%] 35.74 1.04 78.33 1.76 35.90 0.14 Figure 9. Main diagonal of the 3rd kernel Figure 7. 1st Volterra kernel 1 (n1, n2, n3 ). (n1 ). Now it is possible to identify the orthonormal kernels through eq. (6). Figures 7, 8 and 9 show the Volterra kernels 1 (n1 ), 3 (n1, n2, n3 ). The first kernel 1 (n1 ) 2 (n1, n2 ) and is equal to a vector representing the impulse response function 2016 3 Figure 10. Response to chirp excitation of the Volterrra model (input at 0.1 V).

The identified Volterra kernels can be used to predict the output obtained by multiple convolutions using eq. (4). The predicted response to a chirp excitation is shown in Fig. 10. A powerful feature of the Volterra series is that it represents the response of the system separating the linear contribution and many different levels of nonlinear components as showed previously in eq. (1). This can be an interesting tool for the analysis of nonlinear systems. Figures 11 and 12 shows the separated responses of the system to a chirp input in a low (0.01 V) and high level (0.1 V) respectively. Clearly the nonlinear components of the response increase with the input amplitude. While y 2 and y3 are negligible compared to y 1 at a low level of input amplitude as seen in Fig. 11, these components have a relevant effect in the response with a higher level (Fig. 12). This effect is a consequence of the polynomial nature of the stiffness in this kind of structure that was also seen analyzing Fig. 3. It can also be observed that the third kernel has more influence than the second kernel since the contribution of y 2 is lower compared to y 3. (a) Linear component of the response. (b) Second order component of the response. Figure 11. Components of the response of the Volterra model to a low level chirp input (0.01 V). The system was also excited in the natural frequency of the system (around 37 Hz). Figure 13 shows the comparison between the power spectral density (PSD) of the responses of the model and from the experimental data. Some even and odd harmonics appear in the measured response and it is possible to observe that the second and third harmonics are well represented by the Volterra model until the third kernel. Furthermore, the second kernel represents the second harmonics of the response and the third kernel has the third harmonics. This fact shows that the representation of the harmonic content of the response of nonlinear systems can only be achieved using an appropriate nonlinear representation. The fact that the th Volterra kernel represents the th harmonic frequency of the response is directly related to the functional form of the model since the the th component of the response has a power of order in the input. This feature was also observed in [17] in a numerical Duffing oscillator. (c) Third order component of the response. Figure 12. Components of the response of the Volterra model for a high level chirp input (0.1 V). Figure 13. PSD of the measured response in comparison to the one estimated by the Volterra model. 2017

The results showed the capabilities of the Volterra series to model the nonlinear dynamics of a buckled beam considering different levels of nonlinear behavior. The separation of the response of the system in linear and nonlinear components was exploited in this paper and it was shown that it has some physical sense with the level of excitation. This means that with a good model it is possible to observe the degree of nonlinearity in the response of the system by evaluating the response of the model. However, one of the main drawbacks of this technique is the limitation in the representation of effects like jumps and bifurcations that are common in this kind of nonlinear system [22]. Other possible limitation of the method is in the representation of discontinuous nonlinearities (e.g. Coulomb friction, clearance, gaps and others). However, some authors claim that abrupt nonlinearities can be easily approximated using polynomials enabling the Volterra representation with a few terms [7]. Parametric identification (model updating) can be also performed based on the objective function using the identified Volterra kernels [14]. 4 FINAL REMARKS This work showed that discrete-time Volterra series expanded in orthonormal Kautz functions can be used effectively to detect and identify nonlinear features. The approach was validated using experimental the dataset of a buckled beam which gave similar conclusions comparing with classical tools to detect the nonlinear behavior. Additionally, the feature of the separation of the response of the system in linear and nonlinear components is a powerful aspect found in the results. With this property, one can identify and analyze separately the part of the model that corresponds to linear and nonlinear response in a further parametric identification. ACKNOWLEDGMENTS The authors acknowledge the financial support provided by the following funding agencies: São Paulo Research Foundation (FAPESP) by grant number 12/09135-3 and the National Council for Scientific and Technological Development (CNPq) grant number 47058/2012-0. The first and second authors acknowledge FAPESP for their scholarships by grant number 13/09008-4 and 12/04757-6 respectively. The authors also thank the CNPq and FAPEMIG for partially funding the present research work through the National Institute of Science and Technology in Smart Structures in Engineering (INCT-EIE). demonstration, Mechanical Systems and Signal Processing, 25, n. 4, pp. 1227-1247, 2011. [7] A. Chaterjee, Identification and parameter estimation of a bilinear oscillator using Volterra series with harmonic probing, International Journal of Nonlinear Mechanics, 45, pp. 12-20, 2010. [8] A. Hot, G. Kerschen, E. Foltête and S. Cogan, Detection and quantification of non-linear structural behavior using principal component analysis, Mechanical Systems and Signal Processing, 26, pp. 104-116, 2012. [9] I. Isasa, A. Hot, S. Cogan and E. Sadoulet-Reboul, Model updating of locally non-linear systems based on multi-harmonic extended constitutive relation error, Mechanical Systems and Signal Processing, 25, n. 7, pp. 2413-2425, 2011. [10] J. P. Nöel and G. Kerschen, Frequency-domain subspace identification for nonlinear mechanical systems, Mechanical Systems and Signal Processing, 40, n. 2, pp. 701-717, 2013. [11] S. Meyer and M. Link, Modelling and updating of local non-linearities using frequency response residuals, Mechanical Systems and Signal Processing, 17, n. 1, pp. 219-226, 2003. [12] M. Schetzen, The Volterra and Wiener theories of nonlinear systems, Wiley, New York, US, first edition, 1980. [13] W. J. Rugh, Nonlinear System Theory - The Volterra/Wiener Approach, The John Hopkins University Press, 1991. [14] S. da Silva, Non-linear model updating of a three-dimensional portal frame based on Wiener series, International Journal of Non-Linear Mechanics, 46, n. 1, pp. 312-320, 2011. [15] S. da Silva, S. Cogan and E. Foltête, Nonlinear identification in structural dynamics based on Wiener series and Kautz filters, Mechanical Systems and Signal Processing, 24, pp. 52-58, 2010. [16] P. S. C. Heuberger, P. M. J. Van den Hof and B. Wahlberg, Modelling and identification with rational orthogonal basis functions, Springer, 2005. [17] S. B. Shiki, V. Lopes Jr and S. da Silva, Identification of nonlinear structures using discrete-time Votlerra series, Journal of the Brazilian Society of Mechanical Sciences and Engineering, 2013. [18] M.J. Brennan, I. Kovacic, A. Carrella and T.P. Waters, On the jump-up and jump-down frequencies of the Duffing oscillator, Journal of Sound and Vibration, 318, pp. 1250-1261, 2008. [19] A. da Rosa, R. J. G. B. Campello and W. C. Amaral, Choice of free parameters in expansions of discrete-time Volterra models using Kautz functions, Automatica, 43, pp. 1084-1091, 2007. [20] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, Massachussets, US, 1989. [21] D. G. Luenberger and Y. Ye, Linear and Nonlinear Programming, Springer, 2008. [22] Z.K. Peng, Z.Q. Lang, S.A. Billings and G.R. Tomlinson, Comparisons between harmonic balance and nonlinear output frequency response function in nonlinear system analysis, Journal of Sound and Vibration, 311, pp. 56-73, 2008. REFERENCES [1] K. Worden and G. R. Tomlinson, Nonlinearity in Structural Dynamics, Institute of Physics Publishing, 2001. [2] G. Kerschen, K. Worden, A. Vakakis and J. C. Golinval, Past, present and future of nonlinear system identification in structural dynamics, Mechanical Systems and Signal Processing, 20, pp. 502-592, 2006. [3] L. N. Virgin, Introduction to Experimental Nonlinear Dynamics, Cambridge University Press, 2000. [4] S. da Silva, S. Cogan, E. Foltête and F. Buffe, Metrics for nonlinear model updating in structural dynamics, Journal of Brazilian Society of Mechanical Sciences and Engineering, 31, pp. 27-34. [5] S. D. Fassois and J. S. Sakellariou, Time-series methods for fault detection and identification in vibrating structures, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 365, n. 1851, pp. 411-448, 2007. [6] M. Peeters, G. Kerschen and J. Golinval, Modal testing of nonlinear vibrating structures based on nonlinear normal modes: Experimental 2018