Identication of Modal Damping Ratios of Four-Flue Chimney of a Thermo-Electrical Plant using Pseudo-inverse Matrix Method

M. Cavacece1, P.P. Valentini2, L. Vita2

1

Department of Structural Engineering - DIMSAT, University of Cassino, 03043 Cassino, Italy

2

Department of Mechanical Engineering, University of Rome, Tor Vergata, 00133 Rome, Italy
Abstract Some computational issues related to the identication of modal parameters of structures are presented in this paper. Optimal estimation of modal parameters often requires the solution of overdetermined linear system of equations. Hence, the computation of a pseudoinverse matrix is involved. In this paper the numerical performances of dierent algorithms for MoorePenrose pseudoinverse computation have been tested for modal analysis of four-ue chimney of a thermo-electrical plant. The computational scheme herein adopted for parameters identication is based on well known modal properties and has a fast rate of convergence to solution. The computation of the Rayleigh damping coecients and is an important step in the area of the modal superposition technique. The proposed approach can predict accurately damping ratios and all the eigenvectors without evaluating mass and stiness matrices.

Keywords: Modal analysis; Frequency response; pseudoinverse matrix method; Damped system. 1

Introduction

Recently analytical and experimental modal analysis, or vibration system identication, has become an increasingly popular technique. Developments in measuments have facilitated the acquisition of reliable experimental data. These procedures lead to extraction of the modal properties of a test structure [1, 2, 3, 5, 6, 7, 9, 10, 11, 12, 15, 16, 17]. The building response to dynamic excitation, such as wind or earthquake, can be evaluated using techniques, such as the direct integration of equations of motion to modal analysis or the use of response spectra. In each case the availability of the stiness, inertia and dissipative parameters is very important. The objective of this paper is to test dierent methods for the numerical computation of pseudoinverse matrix and to discuss their applicability in structural dynamics applications. The generalized inverse is a well known tool in modern linear matrix theory. It is usually used to solve a set of overdetermined or undetermined simultaneous equations [4] through the application of the least squares optimization criterion. As numerical test, by means of the proposed method the modal damping ratios and the modal shape of a chimney have been estimated assuming Rayleigh damping. In the rst two sections the basic equations used for the mathematical setting of the identication problems are briey deduced. The third section is dedicated to review dierent approaches for computing the MoorePenrose pseudoinverse matrix. In the fourth section the numerical results for dierent cases are discussed.

Fundamental equations of structural dynamics

The equations of motion [14] of a linear viscous damped multi degreeoffreedom system in matrix form are [m] {y } + [c] {y } + [k ] {y } = {F (t)} . (1)

By means of Laplace transformation of the response canonical equations of modal motion [8], assuming zero initial conditions, one obtains the transfer function
Nr Nr

H (s ) =
r =1

H r ( X i Xk ) r =
r =1

[s2

( X i Xk ) r 2] + 2 r r s + r

(2)

between the excitation at the k th location and response at the ith location. Assuming Rayleigh damping, the damping matrix could be expressed as a linear combination of the mass and stiness matrices [c] = [m] + [k ] (3) This type of damping is known as proportional damping. This condition is sucient to decouple the equations of motion (1), allowing the use of modal analysis. By means of equation (3) one obtains
2 + r = 2 r r .

(4) (5)

From dr = r equation (4) becomes

2 + r =2 2 2 r Er,q 2 1 r

(6)

where is dr = Er,q with Er,q experimental circular frequency of damped oscillation, obtained from the q th measurement and r th vibration mode.

Identication of modal parameters

The equations deduced in the previous sections constitute the main blocks of the method for the identication of modal parameters (Fig.2). In this section their use for such purpose is presented. Since the mathematical approach adopted is essentially a least squares solution of a redundant system of linear equations, from our treatment nonlinear equations are excluded. 3

3.1

Rayleigh damping

for q = 1, 2, . . . , Nq and mode r = 1, 2, . . . , Nr . By applying least-squares optimality criterion, the redundant Nq Nr equations system (7) is solved with respect to constants and . In this phase computation of the Moore-Penrose pseudoinverse matrix has its main role.

From equation (6), the following system could be obtained 2 2 E 2 1 1,1 2 2 2 1 E1,2 2 1 1 . 2 . 1 1 . . . 2 2 2 1 E . 1,Nq 1 2 . 1 . . . . . 2 2 2 2 E 1 2 r,1 r 1 2 2 2 2 2 Er,2 r = . . . . . . 2 1 r 2 2 2 2 E r,Nq . . . . . 2 . 1 N r 2 2 2 1 N E 2 3 r Nr ,1 . . 2 2 . 2 3 ENr ,2 2 1 N . r . . 2 2 2 3 E Nr ,Nq

(7)

3.2

Transmissibility function

From equation (2), being 1 = (2f1 )2 , 2 = (2f2 )2 , . . . , r = (2fr )2 , . . . , Nr = (2fNr )2 the square of the eigenvalues, the real component [Hik (f )] of the trans4

missibility function at response location and excitation location k of chimney is ( Xi Xk ) 1 f 1 1 fd1 f 1 fd1
2 2 2 2

[Hik (f )] =

f + 2 1 fd1

( Xi Xk ) 2 f 1 2 fd2 f 1 fd2
2 2

2 2

f + 2 2 fd2 f fdNr f fdNr

2 2

...+

( Xi Xk ) r f 1 r fdr f 1 fdr
2 2

2 2

( X i Xk ) N r 1 Nr
2

+ ...+ 1 f fdNr

f + 2 r fdr

2 2

(8)

+ 2 Nr

the imaginary component [Hik (f )] of the transmissibility function at response location i and excitation location k of chimney is [Hik (f )] = ( X i Xk ) 1 f 21 1 fd1 f 1 fd1
2 2

f + 2 1 fd1

( Xi Xk ) 2 f 22 2 fd2 f 1 fd2
2 2

f + 2 2 fd2

+ ...+

( X i Xk ) r f 2Nr Nr fdr f 1 fdr

2 2

f + 2 r fdr

+ ...+

( Xi Xk ) N r Nr 1 f fdNr

2Nr

f fdNr f fdNr (9)

2

2 2

+ 2 Nr

Then, the modulus of the transmissibility function is Hik (f ) = [Hik (f )]2 + [Hik (f )]2 (10)

From equations (8) and (9) with f = f 1 , f2 , ..., fNp , the following system could be obtained [A] {X } = {b} 5 (11)

with 1 1 1 2 2 2 f1 f 1 + 2 1 1 f fd1 d 1 f 2 1 2 1 f 1 d 1 2 2 2 f f 2 2 + 2 1 1 fd1 fd1 . . . [A] = f p 1 2 1 1 fd1 2 2 2 f f p p + 2 1 1 fd f d1 1 . . . f Np 1 2 1 1 fd1 2 2 2 f f Np Np + 2 1 1 fd1 fd1

f 1 fd1

2 2

...

1 1 Nr f 1 fdNr

f 1 fdNr

...

f 2 1 2 Nr Nr fdNr

2 2 + 2 N r

2 2

fdNr

. . .

f 2 fdNr

2 2 + 2 N r

. . .

2 f 2 fdNr

...

f p 1 2 Nr Nr fdNr

. . .

f Np fdNr

2 2 + 2 N r

. . .

2 f p fdNr

...

f Np 1 2 Nr Nr fdNr

f Np fdNr

2 2 + 2 N r

2 f Np fdNr

(12)

By applying least-squares optimality criterion, the redundant Np Nr equations system (11) is solved with respect to constants (Xi Xk )1 , (Xi Xk )2 , . . ., (Xi Xk )r , . . ., (Xi Xk )Nr . In this phase computation of the Moore-Penrose pseudoinverse matrix has its main role.

Hik Hik . . . {b} = Hik . . . H ik

f 1 f 2 f p f Np

( Xi Xk ) 1 ( Xi Xk ) 2 . . . {X } = ( X i Xk ) r . . . ( X i Xk ) N r

(13)

The Moore-Penrose pseudoinverse matrix

This section, for completeness, summarizes the main properties of the MoorePenrose pseudoinverse matrix and the steps of the dierent algorithms tested during the dynamic simulations [18].

4.1

Denitions

The main properties of the MoorePenrose pseudoinverse matrix [A]+ of a matrix [A] are: [A] [A]+ [A]+ [A]
T

= [A] [A]+ = [A]+ [A]

[A] [A]+ [A] = [A] [A]+ [A] [A]+ = [A]+ 7

When [A] is a square matrix with full rank, then its pseudoinverse coincide with the inverse. The MoorePenrose pseudoinverse matrix is associated with the least squares solution of the linear system of equations [A] {x} = {b} (14)

where the number m of equations is not equal to the number n of unknowns and [A] does not necessarily has full rank. In particular, the following cases are distinguished: Overdetermined system of equations (m > n) By requiring that [B ] is h [A] {x} = {b} is a minimum, one obtains [A]T [A] {x} = [A]T {b} Therefore, the solution of (1) can be stated as {x} = [A]T {b} (17) (16)
2 2

(15)

is the right pseudoinverse matrix. Undetermined system of equations (m < n) The solution is obtained imposing the minimum of the Euclidean norm g x
2 2

(18)

with {x} subjected to (14). Thus, introduced the new objective function, g g + {x}T ([A] {x} {b}) the solution is achieved solving the system I AT AT 0 8 x (20) (19)

or {x} = [A]+ {b} where [A]+ = [A]T [A] [A]T is the left pseudoinverse matrix.
1

(21)

(22)

4.2

The least squares method

Since there is abundance of software procedures for computing the least squares solution of a system of algebraic equations [A] {x} = {b} (23)

the computation of the pseudoinverse can be reduced to such solution. These procedures are often based on QR decomposition by means of Householder reections or GS orthogonalization. Let {b1 } = {b2 } = {b3 } = . . . {bm } = 1 0 0 ... 0 0 1 0 ... 0 0 0 1 ... 0
T T T

, , , (24)

0 0 0 ... 1

(25)

The procedure diers according to the dimensions of [A]. Case m > n 1. Solve m times the system [A]T [A] {x} = [A]T {b} 9 (26)

2. From the m solutions {x1 }, {x2 } ,. . . , {xm } one can form the pseudoinverse matrix [A]+ = Case n < m 1. Solve m times the system I AT AT 0 x (28) {x1 } {x2 } {x2 } . . . {xm } (27)

Also in this case the pseudoinverse is given by equation (27). The square matrices in (26) and (28) are singular or ill conditioned. Thus their solution requires special care. As mentioned, an appropriate way to solve least squares problems is by means of GS orthogonalization or Householder QR factorization.

4.3

Grevilles method
[A]+ =

1. Decompose the matrix [A]mn into row vectors {ai }, (i = 1, 2, . . . , m) aT aT aT . . . aT 1 2 3 m (29)

2. Let matrix [Ai ]in = with [A1 ]1m = {a1 }1m . 3. For i = 2, . . . , m compute the matrices [A]+ as
+ T T [Ai ]+ ni = [Ai1 ] {bi } {di } | {bi }

Ai1 ai

(30)

(31)

10

where
+ {d i }+ 1(i1) = {ai } [Ai1 ]

{ci }1n = {ai } {di} [Ai1 ] { ci } {bi }1n = ( ci = 0) {ci } {ci }T {bi }1n = ( ci = 0) 1 + {d i } {d i }T {a1 } [A1 ]+ = ( a1 = 0) {a1 } {a1 }T [A1 ]+ = ai ( a1 = 0) (32) 4. After m repetitions [Am ]+ gives the pseudoinverse [Am ]+ mn of matrix [A] {di } A+ i1
T

Modal testing of chimney

The height of the studied chimney, w.r.t. ground placed at 4 m above sea level, is 200 m (Fig.1). The building is composed of the following parts: An external shell, made of reinforced concrete, with an height of 195.8 m and with diameter of section variable from 23.0 m (at the bottom) to 19.1 m. The foundation is a cylindrical structure (with a diameter of 55.0 m and 3.0 m thickness) and a series of poles (with a diameter of 1 m each) drilled in the ground to a depth of 56.0 m above sea level. Four internal ues, with a diameter of 5.0 m each, made of acid-repellent and insulated bricks. Each ue is divided into 10 portions of 17.0 m, except the highest one, which is substantially shorter, and each one is placed on a horizontal platform made of reinforced concrete that distributes the weight to external shell. At the top of chimney there is a cover platform at a height of 197.0 m. 11

Figure 1: Chimney

The material of the chimney is isotropic with Youngs modulus E = 36 GPa, density = 1500 kg/m3 and Poisson coecient = 0.2. Experimental tests for measurement of the dynamic response of the chimney have been performed using a shacker. The shaker used is a lowfrequencyrange machine (ISMES BF50/5), with the following features: maximum amplitude of generated force: 50 KN; maximum frequency of the generated force: 5 Hz; structural mass (without inertial masses): 1840 kg. The shaker has been placed at the top of chimney. At a position of 197.0 m a.s.l. the shaker excited the chimney by means of a sinusoidal load with variable 12

Experimental response in time domain Discrete Fourier Transform Experimental transmissibility function Circular frequencies of damped oscillation

Finite Element Method

Natural circular frequencies

Moore - Penrose pseudoinverse matrix

Rayleigh coefficients damping

Moore - Penrose pseudoinverse matrix

Modal shapes Transmissibility function by proposed method

Figure 2: Flow chart of the method for the identication of modal parameters

frequency, along two orthogonal directions [13]. The response of the structure has been monitored at 30 dierent locations along the chimney. In particular, the 22 piezoelectric accelerometers and 8 eightdynamical seismometers have the following features: Accelerometer type: Endeveco mod. 2262C-25, frequency response: linear from 0 up to 500 Hz, sensitivity: 9.81 cm/s/V; Seismometer type: Teledyne mod.Geotech S-13, frequency response: linear from 1 up to 50 Hz, sensitivity: 0.00170 cm/s/V. Experimental setup allows us to measure circular frequencies of damped oscillation and transmissibility functions. The measured circular frequencies of damped oscillation are summarized in table 1. Natural circular frequencies r are obtained 13

by nite element method [13] (Table 2). The model of chimney has been meshed using four-node standard shell elements with 5 d.o.f. per node. Referring to table 2 3, we can calculate r Er,q , right side of system (7) with Nr = Nq = 3. The transmissibility functions have been computed for the nodal displacement at measurement locations. The experimental analysis leads to the values of the real part [H (f )] and of the imaginary part [H (f )] of the frequency response at the top of chimney. Experimental results of the frequency response at the top of chimney have been evaluated referring to the values of frequency of Table 4. Mode r th 1 2 3 Er,1 [rad/s] 1.2 10.04 21.960 Er,2 [rad/s] 1.213 10.034 21.953 Er,3 [rad/s] 1.238 10.022 21.972

Table 1: Experimental circular frequencies of damped oscillation

Mode r th 1 2 3

r [rad/s] 1.257 10.053 22.054

Table 2: Natural circular frequencies

14

Mode r th 1 2 3

2 2 r Er,q

2 2 2 r Er,q

2 2 2 r Er,q

[rad/s] 0.748 1.022 4.068

[rad/s] 0.659 1.235 4.216

[rad/s] 0.435 1.577 3.800

Table 3: Right side of system Measure q th 1 2 3 4 5 Frequency [Hz] 0.01 0.1597 0.194 3.0 3.495 Frequency response [ms2 N 1 ] [H ] = 0.01 [H ] = 0.25 [H ] = 0.72 [H ] = 0.0001 [H ] = 0.1

Table 4: Experimental values of frequency response

Complete modal data

Lets suppose that the modal data are complete in terms of the number of measured modes. For this application we have an overdetermined system of equations, where the number of equations m = 9 is greater than the number of unknowns n = 2. Referring to Table 3, system (7), with Nr = Nq = 3, provides the following solutions = 0.583 , = 7.065 103 (33) Damped oscillation frequencies and modal damping ratios are summarized in table 5. 15

Mode r th 1 2 3

dr [rad/s] 1.219 10.032 21.962

r 2.366 101 6.453 102 9.113 102

Table 5: Circular frequencies of damped oscillation and modal damping ratio

6.1

Modal shapes

Substituting modal damping ratio r (Table 5) and the ones of frequency response (Table 4) into system (11), we obtain the following solutions:
2 XT 1

= 0.432 ,

2 XT

= 2.746 ,

2 XT

= 3.744

(34)

at the top of chimney (i = k = T ) by means of the pseudoinverse matrix (Fig.1). Method Least Squares Householder Least Squares Gram-Schmidt Grevilles Method CPU (time ratios) 1.0 2.0 3.0

Table 6: CPU time In Fig.3 there is the comparison among the transmissibility functions obtained by proposed methodology, experimental results and FEM ones. The Fig.4 shows the real and the imaginary part of the transmissibility fucntion deduced by the pseudoinverse matrix. Referring to Table 6, there is the comparison among the elapsed CPU time obtained by proposed methods.

16

0,6

Experimental
Transmissibility (m/s /N) 10
0,5

FEM Proposed Method

-5

0,4

0,3

0,2

0,1

0,0 0,25 0,50 0,75 1,00 1,25 1,50 1,75 2,00 2,25 2,50 2,75 3,00

Frequency [Hz]

Figure 3: Transmissibility function

Real Part
0,4

Imaginary Part

0,3
-5

Transmissibility (m/s /N) 10

0,2

0,1

0,0

-0,1

-0,2

-0,3

-0,4

-0,5

-0,6

0,0

0,5

1,0

1,5

2,0

2,5

3,0

Frequency [Hz]

Figure 4: Real and imaginary part of the transmissibility function 17

Incomplete modal data

Lets suppose that the modal data are incomplete only in terms of the number of measured modes. If there is noise, structural identication will be ill-posed. In this context, the word noise has wide implications, as for example: any causal or random factors which should not or cannot be modelled; further information is not available; some information of the system has been lost; some vibration data cannot be analyzed; the modal data are incomplete in terms of the number of measured modes. When the second mode of the eigensolution is omitted, we have an overdetermined system of equations, where the number of equations m = 6 is greater than the number of unknowns n = 2. Referring to table 3, we obtain the following solutions = 0.602 , = 7.037 103 (35)

Damped oscillation frequencies and modal damping ratios are summarized in Table 7.

Mode r th 1 2 3

dr [rad/s] 1.219 10.032 21.962

r 2.438 101 6.529 102 9.124 102

Table 7: Circular frequencies of damped oscillation and modal damping ratio for incomplete modal data

18

Mode r th 1 2 3

Complete modal data r 2.366 101 6.453 102 9.113 102

Incomplete modal data error r % 2.438 101 3.0 6.529 102 1.1 9.124 102 0.1

Table 8: Modal damping ratio

Conclusions

The generalized inverse, an elegant mathematical technique, provides only one solution to solve an overdetermined problem with or without noise. In this paper, the application of the generalized inverse allows us to determine modal damping ratios and modal shapes of chimney. The investigation is divided in two parts. In the rst part, we determine the modal damping ratios assuming Rayleigh damping. In the second part, the modal shapes and the frequency response of displacement are computed. It is shown that Rayleigh damping can predict accurately the response of chimney in a large frequency range, because the chimney has a symmetrical structure and the material of chimney has linear elastic properties. The generalized inverse, very useful tool in linear matrix theory, has the following features: 1. By using an overdetermined system of equations with complete modal data, we obtain modal damping ratios. In addition, the acquired signals and the computed results of frequency response of displacement show a good agreement. 2. When the second mode of the eigensolution is omitted, the physical parameters, obtained by using the generalized inverse method, are quite similar to physical ones, obtained by using overdetermined system of equations with complete modal data (Table 8). Therefore, generalized inverse minimizes the eect of noise in this mechanical system. 19

3. Application of the generalized inverse in structural identication is not completely successful by using a set of simultaneous equations which is undetermined. Finally, the generalized inverse is a means for determining the least squares solution of a set of simultaneous equations which is overdetermined. We demonstrate that this method leads to reliable results for modal damping ratios of a chimney with negligible damping. In fact, the response of chimney obtained by this method show good agreement with experimental response of the chimney subjected to a generic wind periodic excitation and the response obtained by nite elements method.

Acknowledgements
The authors wish to acknowledge Prof. Ettore Pennestr` for his helpful advice during the development of this work.

References
[1] Ewins, D. J. and Gleeson, P. T. 1982. A Method for Modal Identication of Lightly Damped Structures. Journal of Sound and Vibration. 84(1):57-79. [2] Luk, W.L. 1987. Identication of physical mass, stiness and damping matrices using pseudo-inverse. 5th IMAC. 679-685. [3] Lagomarsino, S. 1993. Forecast models for damping and vibration periods for duildings. Journal of Wind Engineering and Industrial Aerodynamics. 48: 221 - 239. [4] To, W. M. and Ewins, D.J. 1995. The Role of the Generalized Inverse in Structural Dynamics. Journal of Sound and Vibration. 186(2): 185-195. [5] Watanabe, Y., Isyumov, N., Davemport, A.G. 1997. Empirical aerodynamic damping function for tall building. Journal of Wind Engineering and Industrial Aerodynamics. 72: 313-321. 20

[6] Mohammad, D. R., Khan, N.U. and Ramamurti, V. 1995. On the Role of Rayleigh Damping. Journal of Sound and Vibration. 185(2): 207-218. [7] Dovstam, K. 1997. Recpetance Model based on Isotropic Damping Function and Elastic Displacement Model. Int. J. Solids Structures. 34 (21):2733-2754. [8] Clarence W. de Silva, 2000. Vibration Fundamentals and Practice. CRC Press: New York. [9] Prells, U. and Friswell, M.I. 2000. A Measure of non - proportional Damping. Mechanical Systems and Signal Processing. 14(2): 125-137. [10] Adhikari, S. and Woodhouse, J. 2001. Identication of Damping: Part 1, Viscous Damping. Journal of Sound and Vibration. 243(1): 43-61. [11] Gaylard, M. E. 2001. Identication of Proportional and other sorts of Damping Matrices using a Weighted Response - Integral Method. Mechanical Systems and Signal Processing. 15(2): 245-256. [12] Liu Kefu, Kujath Marek R., Zheng Wanping 2001. Evaluation of Damping Non-Proportional Using Identied Modal Information. Mechanical Systems and Signal Processing. 15(1): 227-242. [13] Cavacece M., Valentini P.P. 2003. An Experimental And Numerical Approach to Investigate the Dynamic Response Of A Four-Flue Chimney Of A Thermoelectrical-Electrical Plant. The Structural Design of Tall and Special Buildings. 12: 283 - 291. [14] Singiresu S. Rao 2004. Mechanical Vibrations. Pearson Education: Singapore. [15] Sang-Hyun Lee, Kyung-Won Min, Jae-Seung Hwang, Jinkoo Kim 2004. Evaluation of equivalent damping ratio of a structure with added dampers. Engineering Structures. 26: 335-346. [16] He, J. and Fu, Z.-F. 2004. Modal Analysis. Butterworth Heinemann: Oxford. 21

[17] Trombetti T., Silvestri, S. 2006. On the modal damping ratios of shear-type structures equipped with Rayleigh damping systems. Journal of Sound and Vibration. 292: 21-58. [18] Pennestri, E. and Cheli F. 2006. Cinematica e Dinamica dei Sistemi Multibody. 1. Casa Editrice Ambrosiana: Milano. Symbol [m] [c] [k ] , {y } {F (t)} q = 1, 2, ..., Nq r = 1, 2, ..., Nr ( X i Xk ) r f r fr dr fdr Er,q i [H (s)] s Description mass (inertia) matrix damping matrix stiness matrix Rayleigh damping coecients displacement response vector forcing excitation vector measure modal shape r th modal shape at response location i and excitation location k frequency r th natural circular frequency r th natural frequency r th circular frequency of damped oscillation r th frequency of damped oscillation experimental circular frequency, acquired for measure q and mode r modal damping ratio for the ith normal mode transfer function matrix Laplace variable: = 2f in frequency domain Table 9: Symbols

22