Engineering Structures 25 (2003) 12791287 www.elsevier.

com/locate/engstruct

Sensitivity analysis and modication of structural dynamic characteristics using second order approximation
F. Aryana a, H. Bahai b,
b a Department of Mechanical Engineering, Bahonar University, Kerman, Iran Department of Systems Engineering, Brunel University, Uxbridge, Middlesex, UB8 3PH, UK

Received 9 September 2002; received in revised form 1 April 2003; accepted 1 April 2003

Abstract This paper presents a formulation in the form of an inverse eigen value problem for modication of vibration behavior of structures. The proposed method which is based on the second order approximation in Taylor expansion is expressed in terms of variables relating to stiffness or mass matrix parameters in a nite element formulation. An initial sensitivity analysis identies the regions within the structure where the modications would yield the required changes in the structures dynamic characteristics. An algorithm is developed which allows efcient modication of structural dynamics characteristics without iterations. These modications are conducted locally so that only elemental stiffness and matrices are affected. The algorithm is applied to four case studies and it is found that large modication of natural frequencies of up to 10% can be realized with an induced error of less than 5% for truss structures, and less than 3% for plane problems. 2003 Elsevier Science Ltd. All rights reserved.

1. Introduction The common industrial practice for optimising the vibration behavior of structures is to conduct a series of modications on the FEA simulations of the structure in order to achieve the required eigenfrequencies. This approach, known as the forward variation approach or design load analysis cycle is extremely time consuming, expensive and rarely yields to an optimum solution. The vibration optimisation problem can be dened as an inverse engineering problem. The inverse engineering refers to problems where the desired response of the system is known or decided but the physical system is unknown. These problems are difcult because a unique solution is rarely possible. The current state-of-the-art in the inverse approach to the vibration problem is only limited to structures modelled using simple linear springs, dampers and point masses. Very little attention has been given to formulating the inverse eigenvalue problem for two and

Corresponding author. Tel.: +44-1895-27400 x 2609; fax: +441895-812556. E-mail address: hamid.bahai@brunel.ac.uk (H. Bahai).

three dimensional and higher order nite elements which are most commonly used in simulation of real structures. Optimisation of vibration characteristics is dened as an inverse eigenvalue problem or problem of designing systems in order to produce the desired response. To eliminate the need to re-analyse the whole structure modelled by nite elements, an inverse approach is required in order to nd the exact modied parameters in various nite element formulations which yield the required natural frequencies. Early work in tackling the inverse eigenvalue problem by other researchers [1,2] utilised the 1st order terms of Taylors series expansion and is based on Rayleighs work. Others such as Chen and Garba [3] used the iterative method to modify structural systems. Recently Baldwin and Hutton [4] presented a detailed review of structural modication techniques. These were classied into categories of the techniques based on small modication, techniques based on localised modication and those based on modal approximation. Further research on structural modication was carried out by Tsuei et al. [57] who presented a method of shifting the desired eigenfrequencies using the forced response of the system. The method is based on modication of either the mass or stiffness matrix by treating

0141-0296/03/$ - see front matter 2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0141-0296(03)00078-6

1280

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

the modication of the system matrices as an external forced response. This external forced response is formulated in terms of the modication parameters, thus creating a modied eigenvalue problem. More recently Zhang and Kim [8] investigated the use of mass matrix modication to achieve desired natural frequencies. McMillan and Keane [9] investigated a method of shifting eigenfrequencies of a rectangular plate by adding concentrated mass elements. Sivan and Ram [10,11] extended further the research on structural modication by studying the construction of a mass spring system with prescribed natural frequencies. They obtained stiffness and mass matrices using the orthogonality principles. They [12] developed a new algorithm based on Josephs work [13] which involves the solution of the inverse eigen value problem. In the last few years the work on the inverse problem by Gladwell [14] started to be taken seriously by engineers and researchers interested in this eld of engineering. Mottershead [15] also considered the problem of resonance in the forced vibration of machines and structures by the design of physical modications to achieve targeted natural frequencies. His technique of achieving the required system include structural modications by adding a point mass, a grounded spring or by a spring connecting two co-ordinates. Li et al. [16] considered optimising dynamic behavior of a multi body system by conducting modications on its mass and stiffness matrices. The above techniques have predominantly been applicable to discrete systems made up of simple linear spring and mass elements. Even with these simple elements the problem of mapping of the physically viable stiffness and mass matrix to a real structure has not been fully resolved and the challenging problem of applying the inverse vibration problem to continuous nite elements has not yet been addressed. The method proposed in this paper is based on a matrix treatment procedure for modifying stiffness and mass matrices of the nite elements most commonly used in modelling structures as continuous systems. Our earlier work [17,18] was focused on bar and beam elements. This was then extended [20,21] to two dimensional elements. The new formulation signicantly improves the previous work by using the second order Taylor approximation in the inverse formulation. Moreover, the proposed technique conducts the modication on the mass and stiffness matrices at a local level thence reducing the computational effort considerably. 2. Theory 2.1. First derivative of eigenvalue Consider the rate of change of eigenvalue with respect to design variables [19]:

{y}T[K]{y}{y}T[M]{y}

(1)

where z is the eigenvalue, {y} is the eigenvector and [k], [m] are stiffness and mass matrices of the system respectively. If we choose bj as the nite elements physical or geometrical property, where j is the element number, then the derivatives with respect to bj of the terms in the elements of matrices not containing bj become zero. Using this reasoning, we showed in our previous work [20] that the vectors and matrices in the above equation can be changed to a reduced form and so the rate of change of eigenvalue with respect to bj becomes: Krs(i) (i) (i)Mrs(i) (i) y y yy bj bj r s bj r s
(i)

(2)

The index i denotes the ith eigenvalue and eigenvector and the sufces r,s are summed over the numbers of the rows and columns of the stiffness and mass matrices of the particular element under consideration. They can be found by using the connectivity matrix [B]. The global numbers of nodes of the element j will be located in the row j of the matrix [B].As an example, for the case of plane truss with linear elements the r and s indices are determined from: r,s (2B(j1)1),(2B(j2)),(2B(j2)1),(2B(j2)). 2.2. Second derivative of eigenvalue The second derivative of eigenvalue can be derived by differentiating Eq. (2) with respect to bk, where bk is another geometrical or physical property of the system (as a special case, it can be the same as bj): yr Krs (i)Mrs (i) (i) 2Krs 2 2 y yr bjbk bk bj bj s bjbk 2Mrs Mrs (i) y. bjbk bk bj s
(i) (i) (i) (i)

(3)

(4)

All the terms in the above equation are known except y(i) , the derivative of eigenvector. bk 2.3. First derivative of eigenvector We start with the equation of motion of the system subjected to undamped free vibrations: (Kmn Mmn)yn 0 (m,n 1,2,N)
(i) (i)

(5)

where N is the total degree of freedom of the system. Differentiating with respect to bk gives: y (Kmn Mmn)(i) y. (Kmn Mmn) bk bk
(i) (i) (i)

(6)

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

1281

Given the fact that the eigenvectors of the original system are linearly independent and span the space they occupy, any linear combination of the eigenvectors can be expressed as:
(l) yn aiklyn bk (i)

(i) (l) (i) M (i) mn 2ymMmnaiklyn ym yn. bk

(14)

The left hand side of the above equation is zero except for l = i and therefore: 1 (i) Mmn (i) y. aiki ym 2 bk n (15)

(7)

We next compute the coefcients aikl. We begin by substituting the above notation in Eq. (6). We therefore have: (Kmn Mmn) (i) yn. (Kmn Mmn)aiklyn bk
(i) (l) (i)

As before, this can be expressed in the reduced form: 1(i)Mtu (i) y. aiki yt 2 bk u (16)

(8)

Multiplying two side of the above equation by ym(p) with the assumption that p i, we get: aikl(ymKmnyn ymMmnyn) (Kmn Mmn) (i) yn (p ym bk
(p) (i) (p) (l) (i) (p) (l)

(9) i)

3. Sensitivity analysis The developed algorithm involves the computation of the rate of change of each eigenvalue (i)with respect to each property bj and then ranking the rates by the order (i) will show of their magnitudes. The largest value of bj th the most sensitive element for the i eigenvalue. It is assumed that the original eigenvalues and eigenvectors of the system are calculated from an initial nite element run. The following algorithm will then determine the required changes in the systems properties to achieve the structures desired natural frequencies. 1. The elemental stiffness and mass matrices in the global coordinates system are determined. 2. The rate of change of the stiffness and mass matrices with respect to various geometrical or physical properties, bj are computed. 3. From the connectivity matrix, the global node numbering of the captured elements and the global numbers of the rows and columns of their position in the stiffness and mass matrices are determined. 4. The eigenvectors are converted to their reduced form. 5. The rate of change of eigenvalues with respect to the property bj is calculated from Eq. (2). 6. Steps 15 are repeated for all the elements. 7. The largest value for the rate of change of the ith eigenvalue is determined. 8. The coefcients aikl in (12) and (16) are computed and the rate of change of eigenvectors with respect to bj is calculated. 9. The second derivative of eigenvalue with respect to bj is computed from (4). 10. The variation of bj is determined from (17). In the case of modication of a single eigenvalue with respect to a property b, Eq. (18) should be used.

By using the orthogonality property of eigenvectors with respect to stiffness matrix and their normalization with respect to mass matrix, the left hand side of the above equation must be zero except for l = p. We therefore have: (Kmn Mmn) (i) yn ym bk . (i) (p) ( )
(p) (p) (i)

aikp

(10)

After simplication we get: ym aikp

Kmn (i)Mmn (i) y bk bk n . (i) (p) ( )

(11)

Now by using the same reasoning as before, the sufces m,n can be changed to the reduced forms:
(p)

yt aikp

Ktu (i)Mtu (i) y bk bk u . (i) (p) ( )

(12)

The sufces t,u must be summed over the number of the rows and columns of element k which contains property bk. In the situation where p = i, we can use the following procedure for computing aikl. Starting with the normalization of eigenvectors with respect to the mass matrix we have: ymMmnyn 1.
(i) (i)

(13)

Differentiating with respect to bk and also using the notation of Eq. (7) we get:

1282

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

Fig. 1.

Plane truss.

4. Modication of natural frequency Our previous algorithm [21] was based on the rst order approximation for modication of natural frequencies. The present algorithms utilizes the second term of the Taylor expansion: 1 2 b b bj bj 2 bjbk j k
(i) (i) (i)

Fig. 3. Error in results versus changes of rst frequency of the truss structure.

(17)

Where (i) is the components of the known vector {} and bj is the components of the unknown vector {b}. In the general case where the modications are to be conducted by changing the properties of a group of elements, the above equation is converted to a system of nonlinear equations. However, in the special case that a single eigenvalue is modied by changing a single property, the above equation will be a scalar equation in the form of: 12 b (b)2 b 2b2 (18)

Fig. 4.

Error versus the change of cross sectional area of element 3.

5. Numerical application For comparison purposes the four case studies conducted using the proposed algorithm use the same examples as given in our previous work [21]. These comprise of a truss structure (Fig. 1), a plane stress model of a cantilever (Fig. 5), a plane stress model of

Fig. 5.

Plane stress cantilever.

a bracket (Fig. 9) and a plane strain model of a dam (Fig. 12). 5.1. Plane truss The 12 plane truss cantilever shown in Fig. 1 is considered as the rst case study. The physical properties of the elements in the truss structure are: Modulus of elasticity, E = 2e11(pa) Density, r = 7860(kg/m3) Elemental Cross sectional area, A = 0.0005(m2)

Fig. 2. The behavior of the rst frequency of the truss structure versus the change of cross sectional area of element 3.

The sensitivity analysis revealed element 3 as the most

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

1283

sensitive element with respect to a change in the bars cross-sectional area. In order to determine the required modication of cross sectional area for this element to achieve the desired rst natural frequency, the rate of change of the rst eigenvalue [21] with respect to the cross sectional area of element 3 is calculated: 4.69e7(rad / s2m2) A3
(1)

(19)

The coefcients aikl are computed by using Eqs. (12) and (16) and it is interesting to note that sufcient accuracy can be achieved in the calculation of the rate of the change of eigenvectors by using only six eigenvectors. In fact this procedure is similar to a Fourier expansion where only a selected number of the terms in the series are retained without signicant loss of accuracy. The coefcients aikl are shown Table 1. The second derivative of the rst eigenvector with respect to the cross sectional area of element 3 is computed by using Eq. (4): 2 7.6208e10(rad / s2m4) A2 3
(1)

Fig. 6.

The behavior of rst frequency in plane stress cantilever.

6. Plane stress cantilever A rectangular cantilever is modeled using six plane stress, constant strain elements as shown in Fig. 5. This example is chosen to illustrate the required modication in the thickness of element 1 (the most sensitive element) in order to achieve the desired rst natural frequency. The physical properties of the elements are:Modulus of elasticityE = 2e11(pa)Density r = 7860 (kg/m3)Poisons ratio n = 0.28Thickness of elements t = 0.01(m2) The rate of change of the rst eigenvalue with respect to the thickness of element 1 [21] is equal to: 0.499e8(rad / s2m2) t1
(1)

(20)

Fig. 2 shows the behavior of the rst frequency against the change of cross sectional area of element 3. The nearly straight line is based on rst order theory, the full curve is generated from the present algorithm which is based on a second order theory and the dashed curve is the exact behavior, obtained by calculating the natural frequencies iteratively for each modied structure using the ANSYS program. The rst order theory is useful for sensitivity analysis and optimization of structures which require small modications in their natural frequencies. The second order theory is a much better candidate for large modication of frequencies up to 15%. Fig. 3 shows the error in the results against the changes of frequency and Fig. 4 shows the error against the change of cross sectional area in element 3. It is seen from these results that the second order approach yields much better accuracy compared to the linear approximation illustrated by the curve obtained from the rst order approach.

(21)

By using a similar procedure as in the previous case study, the second derivative of the rst eigenvalue is calculated: 2 4.0454e9(rad / s2m4) t2 1
(1)

(22)

Fig. 6 shows the variation of the rst frequency with the change in the thickness of element 1. The nearly straight line is based on rst order theory. The full curve is generated from the present algorithm which is based on a second order theory and the dashed curve is the exact behavior of the rst frequency. Fig. 7 shows the error in results against the changes

Table 1 Coefcients a13L a131 2.626 a132 144.095 a133 96.66 a134 24.55 a135 12.418 a136 7.78 Fig. 7. Error in results versus the change in the rst frequency.

1284

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

of the rst frequency and the Fig. 8 shows error against the change of thickness of element 1. This gure again illustrates the improved accuracy obtained from the second order method in comparison to the exact results obtained from running the modied structure and obtaining the new eigen frequencies.

7. Plane stress bracket A real bracket is modeled with 114 plane stress, constant strain elements as shown in Fig. 9. This case study is selected to illustrate the required modication of thickness of the highlighted region in the model in order to achieve the desired rst natural frequency. In the larger models, the sensitivity analysis is conducted by considering blocks of elements each pertaining to a distinct region of the model. The highlighted region in the Figs. 3 and 4 are the most sensitive regions. The physical properties of the elements are: Modulus of elasticity, E = 2e11(pa) Density, r = 7860(kg/m3) Poisons ratio, n = 0.28 Elemental thickness, t = 0.05(m) The rate of change of rst eigenvalue with respect to the thickness of the highlighted region [21] is rst calculated: () t
(1)

The double summation is carried out over the elements which are located within highlighted region of the bracket. Fig. 10 shows the behavior of the rst frequency against the change of thickness of the bracket. The nearly straight line is based on the rst order theory, the full curve is generated from the present algorithm which is based on a second order theory and the dashed curve shows the exact behavior of the rst frequency. Fig. 11 shows the error in the results against the changes of the rst frequency. 8. Plane strain dam A real dam is modeled with 705 plane strain, constant strain elements as shown in Fig. 12. This problem is considered as the fourth case study to illustrate the required modication of density of the highlighted region in order to achieve the desired rst natural frequency. The physical properties of the elements are: Modulus of elasticity, E = 2e10(pa) Density, r = 2410(kg/m3) Poisons ratio, n = 0.11 The rate of change of the rst eigenvalue with respect to the density of the highlighted region [21] is: 0.3185 () ri r
(1)

28.258e8 ti

(1)

(23)

(1)

(25)

Where is the highlighted region in the model where a group of elements are to be modied. By using a similar procedure as in the previous case study, the second derivative of the rst eigenvalue is computed: 2 t2
(1)

By using the procedure like the previous case study, the second derivative of the rst eigenvalue is equal to: 2
(1)

r2

( )

2 1.2989e4 rirj

(1)

(26)

()

2 2.4844e10 titj

(1)

(24)

Fig. 8.

Error in results versus change of thickness of element 1.

Fig. 9. Plane stress bracket (highlighted region is chosen for modication).

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

1285

Fig. 10.

The behavior of the rst frequency in plane stress bracket versus the change of thickness of region .

Fig. 11.

Error in the results versus a change in the rst frequency.

The double summation is carried out over the elements which are located in the highlighted region of the dam. Fig. 13 shows the behavior of the rst frequency against the change of density. The nearly straight line is based on rst order theory, the full curve is generated from the present algorithm which is based on a second order theory and the dashed curve is the exact behavior of the rst frequency. The comparison between the rst order and second order approaches follow the same pattern as the previous case studies. Fig. 14 shows the error in the results against the changes of the rst frequency. 9. Conclusions and discussion An inverse eigenvalue formulation based on the second order approximation has been developed in order to determine the required geometrical and material modi-

Fig. 12. Plane strain dam (highlighted region is chosen for modication).

1286

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

Fig. 13. The behavior of the rst frequency of the plane strain dam versus a change in the density of the highlighted region.

Fig. 14.

Error in the results versus a change of the rst frequency.

cations for pre-dened natural frequency values of structures. The method has been validated by applying it to four case studies. The results are compared against the previous results from a rst order formulation and exact solutions and it is shown that large modications of frequencies can be conducted very efciently with an acceptable level of accuracy. In the case of the truss structure the natural frequencies can be shifted by up to 10% with a maximum error of 6%. In the case of the two dimensional elastic problem a natural frequency shift of 10% was obtained within an error bound of 3 to 4% in plane stress and plane strain applications. The results obtained in these case studies indicate that using the second order approach the model can be modied up to about 10% with much improved accuracy in comparison with the linearised rst order approach. Moreover, it is found that better accuracy can be obtained in the cases involving planner problems compared to models made up of truss elements. The great advantage of the proposed formulation is that the modication is conducted only on small parts of

the global stiffness and mass matrices, requiring minimal computational processing time and memory. The modications can therefore be conducted on the stiffness and mass matrices in their assembled form. This is a very important feature of the proposed technique which allows it to be used as an add-on tool to most commercial FEA codes which do not necessarily make the assembled stiffness and mass matrices accessible. When compared with the classical approach to optimization techniques used in nite elements, it is found the proposed model always yields the exact required eigen frequencies with minimal numerical effort. This is a distinct advantage over the classical techniques where the design space is iteratively searched until a near optimum solution is realized. Moreover, since the proposed algorithm always works with the components of stiffness and mass matrix of individual elements it is much more computationally economical. The other important feature of the proposed technique is that the optimisation is carried in one step with no iterations.

F. Aryana, H. Bahai / Engineering Structures 25 (2003) 12791287

1287

References
[1] Belle Van. Higher order sensitivities in structural design. AIAA J 1982;20:2868. [2] Vanhonacker P. Differential and difference sensitivities of natural frequencies and mode shapes of natural structures. AIAA J 1980;18:1511. [3] Chen JA, Garba JA. Analytical model improvement using modal test results. AIAA J 1980;18. [4] Baldwin J, Hutton S. Natural modes of modied structures. AIAA J 1985;23(11):173743. [5] Tsuei YG, Yee E. A method for modifying dynamic properties of undamped mechanical systems. Journal of dynamic systems. Measurement and Control 1989;111:4038. [6] Yee E, Tsuei YG. Method for shifting natural frequencies of damped mechanical systems. AIAA J 1991;29(11):19737. [7] Yee E, Tsuei YG. Modication of stiffness for shifting natural frequencies of damped mechanical systems. DE-Vol. 38. Modal Analysis, Modelling, Diagnostics and ControlAnalytical and Experimental ASME 1991;10:16. [8] Zhang Q. Kim Ki-ooK. A review of mass matrices for eigenproblems. Journal of Computers and Structures 1993;46:10418. [9] McMillan AJ, Keane AJ. Shifting reasonances from a frequency band by applying concentrated masses to a thin rectangular plate. Journal of Sound and Vibration 1996;192(2):54962. [10] Sivan D, Ram YM. Mass and stiffness modication to achieve desired natural frequencies. Communication in Numerical Methods in Engineering 1996;12:53142. [11] Ram YM. Enlarging a spectral gap by structural modication. Journal of Sound and Vibration 1994;176(2):22534. [12] Sivan D, Ram YM. Optimal construction of mass-spring system [13] [14] [15]

[16]

[17]

[18]

[19] [20]

[21]

with prescribed model and spectral data. Journal of Sound and Vibration 1997;201(3):32334. Joseph KT. Inverse eigenvalue problem in structural design. AIAA J 1992;30(12):28906. Gladwell G.M.L. Inverse Problems in Vibration. Martinus Nijhoff Publishers. Mottershead JE. On the zeros of structural frequency response functions and their sensitivities. Mech. Sys. Sig. Proc. 1998;12(5):5917. He J., Li T. Optimal relocation of resonances and anti-resonances of a dynamic system by local modications. Proc. 15th IMAC, Orlando, FL, 1997. Djoudi M.S., Bahai H. Relocation of natural frequencies using physical parameter modications. Proc. of the rst MIT conference on Computational Fluid and Solid Mechanics, June 1115, 2001, MIT, vol. 1. Djoudi H, Esat Bahai I. An inverse eigenvalue formulation for optimisaing the dynamic behaviour of pin-jointed structures. Journal of Sounds and Vibration 2002;253(5):103950. Edward J, Haue K, Chol K. Design sensitivity analysis of stractural systems. Academic Press Inc, 1986. Aryana F., Bahai, H. Design Optimisation of Structures Vibration Behaviour. Proc. of the rst MIT conference on Computational Fluid and Solid Mechanics, June 11-15, MIT, (abstracts). Aryana F., Bahai H. Design optimisation of structures vibration behaviour using rst order approximation and local modication, accepted for publication in the international journal of computers an structures.