Parameter Identification for Nonlinear Dynamic Systems Via Multilinear Least Square Estimation

Sushil Doranga and Christine Q. Wu1 Department of Mechanical Engineering, University of Manitoba Winnipeg, MB, R3T 5V6, Canada

Abstract
The Nonlinear Resonant Decay Method is an approach for the nonlinear system identification of continuous, multimode and lightly damped systems. In this method the identified nonlinear model is expressed in a linear modal space with additional terms representing the nonlinear behaviour. These additional nonlinear terms are obtained by performing least square regression in a modal space. Implementing this methodology in a structure, a large number of these nonlinear terms may exist but not all of them are significant. The problem of deciding which and how many terms are required for an accurate identification is still challenging. The challenges are (i) a term which is deemed insignificant at an early stage of regression may actually prove to be significant in later steps when the model size is smaller. (ii) A term which has been removed cannot be reinstated and the identified model will be sub-optimal. This paper attempts to address these challenges by using a least square regression algorithm in a statistical sense. A Coefficient of Determination criterion is used to select the essential nonlinear terms. Only those terms which have the largest correlation with the nonlinear restoring force are entered into the model. The proposed methodology is demonstrated through numerical simulations in a five degree-of freedom system and in a cantilever beam. Simulation results shows that this methodology is robust for estimating the local nonlinearity of the system. Keywords: nonlinear parameter identification, multilinear least square estimation, coefficient of determination

Introduction

Nonlinear resonant decay method is one of the methods for representing a nonlinear system in a modal space. It aims to simplify the unknown nonlinear model structure, with too many parameters by using a multi-exciter force pattern to excite one mode at a time. This approach is in essence based upon the idea of force appropriation which permits the determination of a monophase force vectors from the measured frequency response functions (FRFs) matrix. These force vectors will induce single mode behaviour when applied to a linear system at the relevant undamped natural frequency. A single mode may thus be tuned and measured. Scaling these force vectors at a level great enough to excite the nonlinearity present in the structure yield two possible results: (i) the excited mode dominate the response in the steady-state phase. In this case the mode is nonlinearly uncoupled. (ii) Other modes also exhibit a significant response. This is due to the nonlinear cross coupling between the modes. A suitable curve fit may then be used to identify the linear and nonlinear modal parameters based on restoring force-state mapping. For uncoupled nonlinear modes, the curve fitting involves the modal response associated with that mode. For coupled linear or nonlinear modes, the curve fitting includes all modes coupled to the mode of interest. It should be noted that appropriate basis functions need to be choose in curve fittings. The problem
1 For correspondence (email: cwu@cc.umanitoba.ca)

associate with the curve fitting is to determine which of the candidate basis functions are actually required for an accurate representation of the nonlinear system. Some terms may simply not be present in the system, some may not be significant at the level of the measured responses, some may be masked by noise or other measurement errors and some may not have been adequately excited by the forces applied to the structure. It is important that these terms need to be removed to ensure that the final model is parsimonious and truly represent the dynamics of the nonlinearity. To address these issues Platten et.al [3-5] proposed three algorithm which are Forward Selection (FS) algorithm, Backward Elimination (BE) algorithm and Genetic Algorithm optimization. However, the problem associated with implementing these algorithms are that (i) a term which is deemed insignificant at an early stage of regression may actually prove to be significant in later steps when the model size is smaller, (ii) a term which has been removed cannot be reinstated and the identified model will be sub-optimal, (iii) there are no exact guidelines for selecting a suitable term and (iv) there is no basis to deal with the noisy signal. So, there is a necessity of improving the algorithm such that the above mentioned problems can be addressed. In this paper a statistical based multilinear least square algorithm is proposed and is applied to a simulated five degree of freedom system and in a cantilever beam. A statistical based least square curve fitting has been widely used in the field of electrical engineering for time series modeling [4, 6]. Similarly a multilinear regression model along with artificial neural network (ANN) has been used successfully for nonlinear modeling of time series data [7]. It has been shown in [8] that maximum likelihood estimation is efficient in estimating the parameter of a nonlinear system with polynomial nonlinearities. The paper is organised as follows. The NLRDM approach is outlined first. A statistical based least square estimation algorithm is proposed for nonlinear parameter estimation. Following the least square estimation algorithm, two nonlinear parameter selection techniques are presented which are solely based on the correlation coefficient. The proposed parameter selection algorithm is demonstrated through a simulated five degree-of freedom lumped parameter nonlinear system and in a cantilever beam.

2.0

Nonlinear Resonant Decay Method (NLRDM)

The NLRDM technique is used to identify the nonlinear system with Multi-Degree-Of-Freedom. It has been successfully applied for a number of simulated and experimental structures [3-5]. The NLRDM technique consists of: (i) linear system identification, (ii) Force vector approximation (linear), (iii) excitation with approximated scaled force vectors to excite nonlinearities of the structure (iv) data processing (integration and transforming into modal space) and (v) least square estimation to obtain the nonlinear parameters. During the linear system identification (first stage), the linear parameters of the structure such as mode shapes, modal mass, modal damping ratios and modal stiffness are identified using the conventional experimental modal analysis technique (EMA). The second stage of this methodology consists of approximating the appropriate force vectors required to excite the mode of interest. This can be achieved by using a multivariate mode indicator function (MMIF)[9-12]. The third stage is the excitation with the appropriate force vector obtained from the second stage. In this stage a scaled force vector is used to excite the particular mode of interest. The excitation force vector is scaled such that it is high enough to induce nonlinear behaviours. The fourth stage is the data processing stage in which the measured response data along with the applied force vector are transformed to the modal space. The final stage of the NLRDM is the suitable curve fitting on an excited mode to obtain the nonlinear parameters.

2.1

Least Square Estimation

In this section, the least square regression technique originally proposed in [3-5] is discussed first. The potential issues and drawbacks associated with the technique proposed in [3-5] are summarized. To overcome these drawbacks a statistical least square estimation algorithm is proposed. This estimation technique shows the potential nonlinear variable which has a first priority to enter into the model and the variable which has a least priority to enter into the model. This section also shows the new criteria for the measurement of the Goodness-OfFit of the identified coefficients. The proposed statistical least square estimation methodology is based on the assumption that a stochastic relationship exist between the nonlinear restoring force and the nonlinear variables [7-8]. After the successful implementation of the fourth stage of NLRDM , the final stage consists of the curve fittings of the excited nonlinear mode to the measured modal force and response data in a least squares sense. Previous authors [3-5] have shown the least square curve fitting algorithm in the form of the changes in the mean square error. Here the same concept is addressed first and the new concept of coefficient of determination in terms of statistical least

square estimation is introduced in the later section. The nonlinear modal equation for the excited mode in a modal space can be written as, (1) Where and are the modal mass, damping and stiffness for the mode. is the modal force for the mode. is the nonlinear modal restoring force. where is the value of (modal response) at the time instance denotes the value of the modal displacement and denotes the value of the modal force. A candidate set of basis function is selected such that, (2) Where is the column vector containing the basis functions. is the column vector containing the coefficients of the basis functions for the mode. The superscript T denotes the transpose. Combining the force and the response data from all time points and moving the linear terms to the right hand side, we have,

(3)

Equation (3) can be reformed as, Where is named as the design matrix and is the right hand side vector for the mode. The right hand side of the equation is known from the linear identification stage and nonlinear excitation stage, Equation (3) can be solved in the least squares sense to estimates the coefficients, . There are various algorithms to estimate the coefficients in the least squares sense. The commonly used one in the literature is the mean square error criteria, and the second algorithm introduced over here is the statistical least square algorithm which defines the best model in terms of the coefficient of determination. At first the existing methodology described in [3-5] is used. The mean square error (MSE) between the reconstituted restoring force and the measured modal restoring force is defined as a measure of the Goodness-of-Fit of the identified coefficients [9-12]. The error vector can be defined by the following equation.

(4)

and the MSE error is defined as (5) The optimum combination of basis function is sought by using three techniques, which are Backward elimination, Forward selection and Genetic algorithm optimization. The ultimate goal of those techniques is to find out which of the candidate basis functions should be used in the construction of design matrix and to eliminate those terms which are irrelevant or insignificant. The major drawback of the three techniques described in [3-5] is that there are no any specific rules for selecting the basis function. One can come up with an infinite number of irrelevant nonlinear terms in the model which satisfy the minimum mean square criteria. The major unanswered question to the methodology proposed in [3-5] is which candidate function is the best and get the first priority to enter into the model. To address this problem, there need to be definite criteria in which the nonlinear variables can entered into the model. This criteria is addressed in Section 2.2.

2.2

Coefficient of Determination Criteria

This section describes the criteria in which the nonlinear variables are entered into the model. Three model selection criteria are also mentioned in this section. The coefficient of determination measures the percentage of variation in the nonlinear restoring force that is explained by the input variables. The input variables are the product of basis functions and its coefficients. In other words the coefficient of determination provides a measure of how well future outcomes are likely to be predicted by the model. The value of the coefficient of determination lies between 0 to 1. A value of 1 indicates that a regression line fits the data well. The coefficient of determination is found through the multiple regression criteria. The mathematical concept behind the coefficient of determination criteria is discussed below. Consider the following stochastic relationship between the output variables (nonlinear restoring force) individual mode and the individuals input variables . In the present study participating modes to form the nonlinear modal restoring force of for an individual mode. , for an are the (6)

Where are the unknown regression coefficients (the coefficients of the nonlinear function in the present study) of the model and it is assumed that the stochastic errors are independent and identically distributed with zero mean and constant variance. In this setting it is optimal to estimate these parameters by the method of least squares which minimizes the following sum of the square error function. (7) It is undoubtly true that one can use the constant coefficient in the Equation (6). In this case the regression line makes an intercept of instead of passing through the origin. Differentiating Equation (7) with respect to , the solution of the Equation (7) is in the form of the following normal equations. (8) (9)

(10) The normal equations represent a system of non-homogeneous, linear equations with unknowns , . These equations can easily be solved by matrix algebra assuming that none of the input variables can be written as a linear combination of the other input variables. Let us denotes these solutions, called the ordinary .... . The fit of Equation (6) is represented by the sum of squared errors (SSE) least square estimates (11) Where denotes the residual of the regression fit of the observation on the output variable for the i-th individual and the fitted value of the observation of the output of the individual is represented by, (12) In this setting, the ordinary least squares estimates are optimal with respect to being the best of all unbiased, linear estimators of the parameters . A measure of the Goodness-Of-Fit of Equation (6) is represented by the so called R-square ( determination: or coefficient of (13)

Where is known as the total sum of square errors. In other words it is the total sum of squares of the observations from their mean and is represented by, (14) is defined in such a way as to satisfy the inequality . In other words variation in (dependent variable) that is explained by the input variables is the percentage of the .

As a measure of goodness-of-fit, some authors [5-6] mentioned that is inappropriate for making a choice between competing regression models with different numbers of input variables because can be arbitrarily grown to be as close to 1(a perfect fit) as one would like by simply adding more and more input variables to the regression Equation (6). An alternative measure of Goodness-of-fit that contains a penalty for growing regression models unnecessarily large is introduced which is called as adjusted criterion by statisticians [7]: (15) The optimum combination of input variables is sought mainly by two techniques. The forward selection technique and the backward elimination technique. In a forward selection technique variable are sequentially entered into the model. The first variable considered for entry into the model is the one with the largest correlation with the dependent variables Equation (16). In the second case, the independent variable which is not in the model and has the largest correlation is entered. The procedure stops when there are no variables that meet the entry criterion. Backward elimination is just opposite of forward selection. All variables are entered into the model and then sequentially removed. The variable with the smallest partial correlation with the dependent variable is considered first for removal. After the first variable is removed, the variable remained in the model with the smallest partial correlation is considered next. The procedure stops when there are no variables in the model that satisfy the removal criteria. The measure of correlation coefficients between the nonlinear restoring force and potential nonlinear variables is given by, (16) Where is the correlation coefficient and is the potential nonlinear terms which is in the form of ....... . The exact terms and their structure are initially unknown. , ,

3.0

Problem Formulation

In order to demonstrate the proposed methodology of statistical least square regression, two case studies are chosen for demonstration. (i) A five-degree-of freedom lumped parameter system with cubic stiffness nonlinearity and (ii) a cantilever beam with a grounded cubic stiffness nonlinearity. For the case study of (i) the system is designed to be symmetric in its linear components so as to yield very simple mode shapes (Figure 1). Hardening cubic stiffness nonlinearity is inserted between masses 2 and 4. If the displacements of masses are denoted by , , , and , the free vibration equations of motion for the system shown in Figure (1) can be written as,

C
1

C 2m 3m K

C 2m K

m K K

m K K

Figure 1. Five degree-of freedom system

. , , and are the damping, linear stiffness and nonlinear stiffness Where respectively. The parameter used for the simulation are =1kg, = 4.8Ns/m, N/m and = N . The resulting natural frequencies and mode shapes of the undamped system are shown in table 1.
Table 1. Natural Frequencies and Mode Shapes of Five degree-of Freedom System Mode Shapes Modal Frequencies (Hz) 1 2 3 4 3.51 1 -1 1 1 8.01 1.88 -1.37 0.87 -0.37 10.72 2.3 0 -1.23 0 15.48 1.88 1.37 0.87 0.37 15.63 1 1 1 -1 Let denotes the mode shape matrix, and then the mode shapes matrix can be written as,

5 1 -0.41 0.16 -0.41 1

Equation (17) can be represented in the form, (18) Introducing the term and multiplying both sides by , Equation (18) becomes, (19) Equation (19) represents the linear equation of system in the modal space with the nonlinearities in the physical coordinates. Utilizing Equation (19) one can write the system of equations as, (20) (21) (22) (23)

Equations (20) to (24) show that some of the modes are completely uncoupled. Only the second and fourth modes are coupled to each other. If the nonlinear terms in Equations (21) and (23) are transformed to the modal coordinate system, the equations of motion for Mode 2 and 4 become,

An optimum force pattern for each mode is obtained through the normal mode tuning procedure [9-12]. To excite the nonlinearity in an equal amount, a scaling factor is applied to the obtained force vector. The representation of the linear dynamics of a system in a modal space is shown in Figure 2. In the physical coordinate system the force is applied to all the mass (Figure 1), however due to the optimum force vector, Mode 1 is only excited at the modal space. (Figure 2). The nonlinear response of Mode 4 is shown in Figure 3. In Figure 3, it is seen that, there is some response in Mode 2 without any excitation force in Mode 2. The response in Mode 2 is due to the cross-coupling nonlinearity between Mode 2 and Mode 4. The cross-coupling nonlinearity is shown in Equations 25 and 26 respectively.

0.6 0.4

Excitation at Mode 1
1.5 1 x 10
-3

Response at Mode 1

Modal Force (N)

0.2

Displacement (m)

0.5 0 -0.5 -1

0 -0.2 -0.4

10

15

20

25 30 Time (Sec)

35

40

45

50

-1.5 0

10

15

20

25 Time (Sec)

30

35

40

45

50

1.5
0.6 0.4 0.2 0 -0.2 -0.4 Excitation at Mode 2

x 10

-3

Response at Mode 2

1 0.5 0 -0.5 -1 -1.5 0

Modal Force (N)

Displacement (m)

10

15

20

25 30 Time (Sec)

35

40

45

50

10

15

20

25 30 Time (Sec)

35

40

45

50

0.6 0.4 0.2 0 -0.2 -0.4

Excitation at Mode 3

1.5 1 0.5 0 -0.5 -1

x 10

-3

Response at Mode 3

Modal Force (N)

Displacement (m)

10

15

20

25 30 Time (Sec)

35

40

45

50

-1.5 0

10

15

20

25 30 Time (Sec)

35

40

45

50

0.6 0.4 0.2 0 -0.2 -0.4

Excitation at Mode 4

1.5 1 0.5 0 -0.5 -1

x 10

-3

Response at Mode 4

Displacement (m)

Modal Force (N)

10

15

20

25 30 Time (Sec)

35

40

45

50

-1.5 0

10

15

20

25 30 Time (Sec)

35

40

45

50

0.6 0.4 0.2 0 -0.2 -0.4

Excitation at Mode 5

1.5 1 0.5 0 -0.5 -1

x 10

-3

Response at Mode 5

Displacement (m)

Modal Force (N)

10

15

20

25 30 Time (Sec)

35

40

45

50

-1.5 0

10

15

20

25 30 Time (Sec)

35

40

45

50

Figure 2. Modal Forces and Responses due to Perfect Excitation at Mode 1(a: Mode1, b: Mode2, c: Mode3, d: Mode4, e: Mode5)

Modal Force at Mode 1 20 15 10 5 0 -5 -10 -15 -20 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50 4 3 2

x 10

-3

Modal Response at Mode 1

Displacement (m)

Modal Force (N)

1 0 -1 -2 -3 -4 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50

Modal Force at Mode 2 20 15 10 5 0 -5 -10 -15 -20 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50 4 3 2

x 10

-3

Modal Response at Mode 2

Displacement (m)

Modal Force (N)

1 0 -1 -2 -3 -4 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50

Modal Force at Mode 3 20 15 10 5 0 -5 -10 -15 -20 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50 4 3 2

x 10

-3

Modal Response at Mode 3

Displacement (m)

Modal Force (N)

1 0 -1 -2 -3 -4 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50

20 15 10 5 0 -5 -10 -15 -20 0 5 10 15

Modal Force at Mode 4 4 3 2

x 10

-3

Response at Mode 4

Displacement (m)

Modal Force (N)

1 0 -1 -2 -3

20

25 30 Time (Sec)

35

40

45

50

-4 0

10

15

20

25 30 Time (Sec)

35

40

45

50

Modal Force at Mode 5 20 15 10 5 0 -5 -10 -15 -20 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50 4 3 2

x 10

-3

Modal Response at Mode 5

Displacement (m)

Modal Force (N)

1 0 -1 -2 -3 -4 0 5 10 15 20 25 30 Time (Sec) 35 40 45 50

Figure 3. Nonlinear Responses due to High Level Force Excitation at Mode 4(a: Mode1, b: Mode2, c: Mode3, d: Mode4, e: Mode5)

3.1

Curve Fitting of Numerical problem

A numerical problem of lumped mass system mentioned in Figure 1 is demonstrated through the curve fitting procedure. The response of the system is simulated in MATLAB/SIMULINK using adaptive 4th order Runge Kutta algorithm. For the brevity, in non-linear system analysis only Mode 4 is taken. The force required to excite only Mode 4 for a linear system is obtained through force appropriation technique as described earlier. This force is scaled such that, when it is applied to a structure will induce nonlinearity, if present in the structure. As shown in Figure 3. The response is due to the excitation of the fourth mode. As Mode 4 is connected nonlinearly with Mode 2, there is some response in Mode 2 also. As the input to the system and the linear parameters are known, the nonlinear parameters are extracted in this problem through the curve criteria is used to fitting techniques. To estimate the parameter forward selection technique is applied. An adjusted measure the Goodness-Of-Fit between the identified and true nonlinear coefficients. if and are the nonlinear coefficients corresponding to Mode 4 , the true coefficients corresponding to Mode 4 can be obtained from Equation 26. The identified coefficients are obtained from forward selection technique. The result is presented in Table 2.

Table 2. Identified and True Nonlinear Coefficients of Mode 4 of Five Degree-of Freedoms Simulated System Parameter True Identified Non linear cubic stiffness coefficients

3.2

Cantilever Beam

A cantilever beam made of Aluminium T6061 (Figure 4) is chosen for the second case study. The beam used for the case study has the following dimensions: length: 1 m; width: 0.0254m; and, thickness: 0.00125m. A spring of cubic stiffness nonlinearity is inserted at the node 2 of the beam. The modal damping ratio for each mode is assumed to be 0.001. The detail of the beam is shown in Figure 4, where three degrees of freedom are taken for measurement as indicated by red ovals. Three modes are taken for the analysis, the mode shapes and natural frequencies of the beam are obtained by solving the governing partial differential equation of the beam. The natural frequencies of the beam for first three modes are obtained as, 2.55Hz, 17.1Hz and 47.3Hz respectively. Optimum combinations of three force vector are used to excite the particular mode of interest. The Runge- Kutta integration method is used to find out the nonlinear response data of the beam. The detail implementation of Runge Kutta method is described below.

Excitation Force Excitation Force

Excitation Force

Cubic Stiffness nonlinearity

Excitation Force

Figure 4. Cantilever Beam with a Nonlinear Element at node 2 The basic idea of the Runge-Kutta integration method is to find a solution for the equilibrium equation at a discrete time point. In the case of a dynamic engineering structure, the equilibrium equation can be described by a set of second-order differential equations given by, (27) Where , and is the mass, stiffness and damping matrix of the structure. nonlinear restoring force vector. Multiplying both sides of equation (27) by mode shapes matrix becomes, is the input force vector and is the

and introducing the term

, Equation (27) (28)

For mass normalized mode shapes and proportional damping assumptions, the orthogonal properties of the linear mode shape matrix , can be written as, (29)

(30) (31) Using Equation (29), (30) and (31) in Equation (28) we get, + (32)

Integration of Equation (32) by the Runge- Kutta method involves first reducing the second order differential equation into a first order differential equation by rewriting as two first order equations. (33) + Equation (33) and (34) can be written in matrix form as, (35) The ordinary differential Equation (27) is reduced to a coupled first order differential Equation (35) which has the explicit form shown by Equation (36), (36) Where is known and given by Equation (37), (37) The displacement is given by Equation (38), (38) Implementation of Equation (37) required the mode shapes matrix. The mode shapes matrix is obtained by solving the partial differential equation of motion. Once the mode shapes matrix is obtained, the solution for the Equation (37) is obtained by implementing the MATLAB solver ode45. Figure 5 Shows the mode shapes of the beam, where the first three modes are shown. (34)

3.3

Modal Space Equation

The modal space of the beam is approximated by using the standard Equation (28). Using the response vectors in the physical coordinates are transformed to the modal space. Where Using Equation (28), we can write,

(39)
In Equation (39) is the modal force and is the nonlinear restoring force. The mode shapes matrix is obtained from Figure 5, is the input excitation vector which is obtained from MMIF. is the nonlinear restoring force vector with cubic stiffness nonlinearity at Node 2. Using the numerical values in Equation (39), we can write the modal equation for Mode 1 as, (40)

3.4

Implementation of Curve Fitting Algorithm

The optimum excitation force vector required to excite the first mode of the beam is obtained by using MMIF. These optimum excitation force vector are found to be 3.7090N, 3.6223N and 0.5780 N at the Node 1, 2 and 3 respectively. The

response data of the beam is obtained by solving Equation (32) in MATLAB using ode45. Figure (6-8) shows the response data of the beam by using the above mentioned excitation vector. The response data of the beam are than transformed to modal space by using mode shapes matrix. A least square regression in a modal space for Mode 1 is carried out using the forward selection technique. The results of the estimated and the exact nonlinear parameters are shown in Table 3. The exact nonlinear parameters are given by Equation (40) whereas the true parameters are estimated through regression. As shown in Table 3, the deviation between the estimated and the true coefficient of and are quite high. There are several reasons for this deviation. The main reasons are: (i) only three nodes are taken during simulation and (ii) only three modes are taken for analysis.
0.5 0.4 0.3 0.2 First Mode Second Mode Third Mode

Mode Shapes (phi)

0.1 0 -0.1 -0.2 -0.3 -0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 Length of the Beam (m) 0.7 0.8 0.9 1

Figure 5. Mode Shapes of the Cantilever Beam

Displacement (m)

2 0 -2 -4 -6 -8 0 10 20 30 40 50 Time (Sec) 60 70 80 90 100

Figure 6: Response of the beam at the First Node

4 3 2

x 10

-3

Displacement (m)

1 0 -1 -2 -3 -4 0 10 20 30 40 50 Time (Sec) 60 70 80 90 100

x 10

-3

Figure 7: Response of the beam at the Second Node

0.5

Displacement (m)

-0.5

-1 0

10

20

30

40

50 Time (Sec)

60

70

80

90

100

Figure 8: Response of the Beam at the Third Node Table 3: Estimated and True Nonlinear Coefficients Modal Couplings True Coefficient ( ) Estimated Coefficient ( )

0.91

4.

Concluding Remarks

This paper presents the improved algorithm for nonlinear parameter estimation of nonlinear dynamic system. A statistical based multilinear least square estimation technique has been proposed to estimate the nonlinear parameters in a modal space. The nonlinear resonant decay method is applied to quantify the modes which are behaving linearly and nonlinearly. For

modes which are behaving nonlinearly, the accurate estimation of the nonlinear function is essential. A large number of nonlinear variables may exist but all of them may not represent the dynamics of the system. A term which is deemed to be insignificant may actually proven to be significant while performing multilinear regression. To overcome this scenario a correlation criteria is set to enter the essential nonlinear terms into the model. A coefficient of determination criteria is used to measure the Goodness-Of-Fit of the identified model. The proposed methodology is demonstrated through simulation in a five degree-of freedom lumped parameter system and a cantilever beam with cubic stiffness nonlinearity. The nonlinear parameter estimation is found to be sufficiently accurate with error less than 0.1% for the lumped parameter system, whereas for a continuous system there are some discrepancies. These discrepancies are due to (i) fewer number of modes and (ii) only three nodes are taken during simulations.

References
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] Amauri H., de Souza J, Francesco C, and Guilherme A. B., Robust Regional Modeling for Nonlinear System Identification Using Self-Organizing Maps.Advances in self orgainizing maps.AISC 198, PP 215-224. Platten, M.F., Wright, J.R., Dimitriadis G., and Cooper J.E., 2009. Identification of Multi-Degree of Freedom Nonlinear Systems Using an extended Modal Space Model, Mechanical Systems and Signal Processing.,23, pp. 8-29. Naylor, S., 1998. Identification of Non-proportionally Damped Systems using a Force Appropriation Technique, PhD Thesis, University of Manchester, UK. Sarmast, M.., 2005. Identification of Nonlinear Dynamic Systems Using the Nonlinear Resonant Decay Method. Ph.D. thesis, University of Manchester, UK. Platten, F.M., Wright, J. R., Worden K., Dimitrladis, G., and Cooper, J. E., 2007. Nonlinear Identification in Modal Space Using a Genetic Algorithm Approach for Model Selection, Int. J. Of Appl. Math and Mech., 3(1), pp. 72-89. M, Ezekiel and K.A., Fox. Methods of correlation and regression analysis: Linear and curvilinear. Oxford, England: John Wiley. (1959). xv 548 pp. Suresh S.,, I, Sabahattin., Srivastava P., and Kalin, L, Deriving Spatially Distributed Precipitation Data Using the Artificial Neural Network and Multilinear Regression Models. Journal of Hydrologic Engineering. ASCE, Vol.18. HENNING U. V., JENS T., and JRGEN K. Nonlinear dynamical system identification from uncertain and indirect measurements.International Journal of Bifurcation and Chaos. Vol 14, 06, 2004. Nash. M., 1998. Use of the Multivariate Mode Indicator Function for Normal Mode Identification, IMAC, pp. 740-744. Balm&S, E., 1993. Integration of Existing Methods and User Knowledge in a MIMO Identification Algorithm for Structures with High Modal Densities, IMAC, pp. 613-619. Van der Auweraer, H., Otte, D., Debille, J., Leuridan, J., 1992. Optimum Force Vector Conditioning in Incremental Sine Testing, IMAC, pp.299-306. Balm&S, E., Chapelier, C., and Lubrina, P., An Evaluation of Modal Testing Results Based on the Force Appropriation Method, IMAC-XIII, pp. 47-53.