DAMAGE DETECTION OF STRUCTURES BASED ON MINIMUM NUMBER OF

FREQENCY RESPONSE FUNCTIONS MEASUREMENTS

H. Y. HWANG AND J. Y. LEE

University of Ulsan, Ulsan, Korea 680-749

The identification method for stiffness constant parameters and damping coefficient

parameters of connections using test data is derived when a structure is attached to the

another structure via connections. Because of the inherent difficulties of deriving spatial

models from test data, the frequency response function (FRF) is used as a response model.

The identification method which used the FRFs works for each discrete frequency so that

the connection properties can be found for each frequency and can be averaged using

statistical methods for an accurate identification. If highly sensitive regions are excluded,

this identification method for connection using FRFs gives accurate estimations and can

be applied to a general structure in an easier manner than the modal model methods which

require a mathematical model of the mass, damping and stiffness matrices (or natural

frequencies and mode shapes).

1. INTRODUCTION

Coupling between one structure and a neighboring structure via the structural connections which

inevitably exist can have a major influence on the actual values of the natural frequencies and mode

shapes. Hence, accurate dynamic response of the complex structure is frequently restricted by the

capabilities to describe the connection part of the structure properly. For example, the popular finite

element method may not produce satisfactory dynamic behavior of a mechanical system due to the

uncertainty of the connection properties. The real structure usually includes many different kinds of

1
joints (bolted joints, sliding joints, socket joints, etc.) which contain at least some amount of

nonlinearity (e.g., clearance nonlinearity, dry friction, localized plasticity, etc.). Since boundary

conditions are determined by the connecting joints in a structure, the dynamic characteristics of the

structure are greatly influenced by the modeling of connections between structural components. This

leads to an increasing need for developing improved analytical models for connection parts.

However, it is extremely difficult to develope accurate anlytical models for connection parts by

just considering the geometry and properties of a structure. Analytical models for connections may

have modeling errors due to incorrect stiffness and mass properties of connections and nonlinearities

of connections and many other incorrect configurations of connections. Hence, we must have some

means of developing confidence in or of validating analytical models of connection parts. It is to use

vibration test data as a criterion, expecting the analytical modelings of connections to agree with test

data.

In this paper, the identification method for stiffness constant parameters and damping coefficient

parameters of connections using test data is derived when a structure is attached to the another

structure via connections. In this modified method, frequency response functions are used to find

physical connection parameters instead of natural frequencies and mode shapes. By using this new

method, connection parameters are directly obtained from the measured receptances and inertances

without introducing mathematical models of the mass, damping, and stiffness matrices from test

natural frequencies and mode shapes. This research greatly reduces the quantity of test data that

must be obtained from the coupled system because it requires only one numerical value of the

frequency response function among various frequency range values for an each connection. One

frequency response function contains infinite number of numerical values. This is particularly useful

when it is not realistic to obtain a complete set of vibration test data of natural frequencies and mode

2
shapes for a coupled structure, such as large space structures, large ocean structures, or complex

turbomachinary.

Many attempts at applying frequency-domain test analysis techniques have failed in the past

because of inaccuracies in the FRFs used. Recently, with the availability of better data acquisition

hardware and techniques, the use of measured FRFs for frequency-domain techniques has become

more practical.

First, in this paper, equations to find stiffness and damping constants using the test data of

frequency response functions are derived from the equations of motion for a structure with

connections. Next, for a better understanding of the derived equations, an example consisting of a

continuous beam system is examined.

A number of papers have been published which discuss theoretical structural dynamics, and a

considerable amount of experimental research has been performed which investigated coupled

structures [1-15]. However, the objective of most of these studies was to estimate the overall

behavior of the coupled structure. Only a few studies have tried to identify the connection

parameters themselves using experimental data [1, 2]. Huckelbridge and Lawrence [1, 2] developed

a procedure for identifying physical connection properties from free and forced response test data,

then verified it by utilizing a system having both a linear and nonlinear connection. Interface

connections in both the translational and rotational directions were addressed. Connection properties

were computed in terms of physical parameters so that the physical characteristics of the connections

could be better understood, in addition to providing improved input for the system model. They

used the modal model methods which require mass and stiffness matrices from test natural

frequencies and mode shapes so that identified results have an error and highly depend on the

amounts of test data and the mode selected.

3
 Ewins ([3], [10]) reviewed some of the methods available for making vibration analyses of

complex structures where one or more of the component substructures is modeled from

experimental, rather than theoretical data. The study concentrated on methods of analysis which

were based on response properties.

Time-domain and frequency-domain methods for coupling substructures with general linear

damping were considered in [4]. The time-domain method used state variable notation for each

substructure. The frequency-domain method utilized the discrete Fourier transform and fast Fourier

transform to get transient response solutions. Suarez and Singh [5] presented a method for

synthesizing the real modes of substructures to obtain the complex modes of the combined structure

including nonproportional damping effects which may exist in the combined structure.

Wang and Liou [6] proposed a method to synthesize the frequency response functions of a

structure which is composed of two beams and linear joint springs and dampers. They introduced a

simple method based on statistical criteria to overcome the problem caused by measurement noise.

In the synthesis process, it was assumed that one substructure was fixed completely, however, in an

actual measurement situation, grounding a substructure is a very difficult task. In [7], Wang

identified the stiffness, mass and damping matrices of a linear mechanical system using a weighted

frequency response function combined with the instrumental variable method. The effect of

measurement noise was of particular concern.

Yoshimura [8-9] performed experimental measurements to obtain vibratory characteristics and

quantitative values of the rigidity and damping at a bolted joint, welded joint and slide way which

were the representative joints in the machine tool structure under study. He described methods and

procedures for adapting equivalent spring stiffness and damping coefficients to bolted joints and

slide ways in computer aided machine tool design.

4
 Chen and Cherng [11] presented a simple and effective modal synthesis method via combined

experimental and finite element techniques in which the "constraint modes method" was used to

determine the dynamic properties of complex structures. To satisfy the rotational compatibility at

the common boundary, an experimental procedure was proposed to measure the generalized dynamic

compliance.

In [12], a unique methodology was proposed to identify the joint structural parameters of a

machine tool by combining the dynamic data system methodology with the finite element method.

The structural parameter identification of a simple system with a complete modal matrix was

introduced together with modal analysis by the dynamic data system method.

In [13], both time-domain and frequency-domain Component Mode Synthesis methods were

reviewed and according to the types of component modes used in synthesis process, the methods

were further subdivided into cases such as free-interface component normal modes, constraint

modes, inertia-relief attachment modes, etc. A large number of references about Component Mode

Synthesis methods were also given in this paper.

Ghlaim and Martin [14] solved a matrix set involving eigenvalues and eigenvectors of the

substructure, together with a connection matrix to give the complex roots of the system. Reduction

was applied by approximating the substructures by a reduced set of eigenvalues and an equally

reduced set of displacements in the eigenvector.

Ewins and Henry [15] provide the necessary introduction and grounding for a study of vibration

characteristics of individual turbomachine blades including joint characteristics.

Most of these experimental studies of connection parameters involved comparing the frequency

response functions for a range of different cases instead of identifying numerical values of

connections [3-15].

5
 In this paper, instead of comparing the frequency response functions with connections, methods to

find numerical values of the connection parameters for complicated real structures using frequency

response function data were developed.

2. FORMULATION OF THE EQUATIONS OF MOTION

Assume that a hypothetical structural component is attached to fixed structure via springs and

dampers as shown in Figure 1.

For structure with connections, equations of motions are given by [6],

x A   x A   x A    f A  
 M     D      K      (1)
 x a   x a   
 x a   f a   f ja 

where "a" represents the region on the connection of structure and "A" represents the region except

region "a" on structure and {fja} represents the interface force acting on the region "a" , {f A} and

{fa} are applied forces.

Interface force {fja} at connections can be expressed as

 k 1 x a1  d 1 
x a1 

 k x a 2 
 
d 
x 

 f   
2
  
2 a 2

 
ja
 

 k n x an

 
d n 
x an



 x a 1  
 x a 1 

(2)

 x 
  
 x 

  K  
a 2
   D  
a 2

 
c c
 

 x an

 
 
x an



where

 k1   d1 
 k2   d2 
 Kc   
 ,  D c  
  


(3)

 
kn  
 d n 

Then, equation (1) can be expressed as

6
 x     0  0   x A     0  0   xA   fA  
 M  A     D        K      
 xa  

  0  D c    x a     0  K    x
c a    fa  
(4)

Let

  0  0    0  0 
 q c   xA  , xa  , fc   fA  , fa  ,  D c4    0  D   ,  K c4    0  K   (5)
T T

 c   c 

Now, equation (4) can be expressed as

 M q c    D W  q c    K W  q c   f c   (6)

where [D]w = [D] + [Dc4], [K]w = [K] + [Kc4], and subscript w represents with connections.

Similarly, for structure without connections, equations of motions are given by

 M  q c    D  q c    K q c   f c  (7)

Let {fc}={f}ejt, {qc}={x}ejt, so that equations (6) and (7) can be written:

([K]w + j[D]w - 2 [M]){x}ejt = {f}ejt (8)

([K] + j[D] - 2 [M]){x}ejt = {f}ejt (9)

Then, using notations in equation (6), connection parameters, j[Dc4] + [Kc4], can be found

using the following equations.

j[Dc4] + [Kc4] = j[D]w - j[D] + [K]w - [K] (10)

By adding and subtracting - 2 [M] to the left hand side, equation (10) becomes

7
 j[Dc4] + [Kc4] = ([K]w + j[D]w - 2 [M]) - ([K] + j[D] - 2 [M]) (11)

where subscript w represents structure with connections.

From equation (9), frequency response function, [H()] can be represented by

[H()] = {x}/{f} = ([K] + j[D] - 2 [M])-1 (12)

A frequency response function is the relation between the Fourier transform of the system output

(response) and the Fourier transform of the system input (applied force).

Hence, equation (11) becomes

j[Dc4] + [Kc4] = [H]-1w – [H]-1w/o (13)

Alternatively,

  0  0    0  0 
  d1    k1 
       = [H]-1w – [H]-1w/o (14)
j D c 4    K c 4   j d2 k2
 0      0  
       
  d n     k n  
 

Hence, damping coefficient di and spring constant ki can be determined by measuring FRF of the

structure with and without connections.

3. EXAMPLE: IDENTIFICATION OF CONNECTIONS USING FRFs

In this section, the identification of connections using frequency response functions (FRFs) will

be carried out for a continuous beam system. For a system model shown in Figure 2.(a), frequency

response functions for the connected structure and unconnected structure could be obtained from an

experiment with the physical model. However, representative FRFs that could be obtained from the

test will be synthesized numerically from the continuous system model given in Figure 2.(b).

8
 The finite element model of a cylinder type beam in Figure 2.(b) was used to create natural

frequencies and modes shapes. The outer diameter and the inner diameter of the cylinder type beam

are 2.00 inchs and 1.75 inchs each and the lenth of a beam is 75 inches. Young’s Modulus is 30106

psi and mass density is 5.282710-4 (lb-s2/in2). First four natural frequencies of the beam with free-

free boundary conditions from experiments are given by 84.0 (Hz), 232.0 (Hz), 450.0 (Hz), 734.0

(Hz) excluding rigid body modes. Analytical natural frequencies from the finite element model are

given by 85.3 (Hz), 234.0 (Hz), 456.4 (Hz), 750.6 (Hz), and so on.

Frequency response functions and inverse frequency response functions corresponding to the

given models with connection stiffness 103 (lb/in) can be numerically simulated using the following

frequency response functions and are plotted in Figure 3.

H()conn = 0.0000002/(0.000012-2) + 6.3427342/(257.71242-2) + 11.511122/(638.28832-2) +

10.572032/(1505.4022-2) + 9.9883512/(2884.472-2) + 9.6281592/(4725.7422-2) +

9.2926012/(7015.0262-2) + 8.9216282/(9743.0362-2) + 8.4925232/(12901.362-2) +

7.996772/(16480.762-2) + 7.4343942/(20469.232-2) + ............

H()free= 3.9802422/(0.000012-2) + 9.2054382/(0.000012-2) + 9.9640132/(536.21352-2) +

9.8515792/(1470.3082-2) + 9.6940612/(2867.6312-2) + 9.4827512/(4716.0722-2) +

9.2108762/(7008.9232-2) + 8.8716942/(9738.9742-2) + 8.4603282/(12898.582-2) +

7.9753092/(16478.822-2) + 7.4198172/(20467.882-2)+............

In simulating FRF of a beam with free-free boundary condition, rigid body modes were added

that are the first and the second terms in the above second equation. To avoid numerical difficulties

in the inverse process, natural frequency of the rigid body mode was assumed to be 0.00001 (rad/s)

and arbitrary rigid body mode shape can be chosen.

9
 If it is now assumed that the above computed frequency response functions are those obtained

from a test, using equation (14), the connection spring constant, k c, can be found using the following

equation.

(15)

It should be also noted that this expression is not constant but is frequency-dependent unlike

methods involving eigenvalues and modal matrices.

Using equation (15), the identified connection spring constant, k c, is plotted in Figure 4 as a

function of frequency. Except for small peaks, the identified connection spring constant was

uniformly 103 (lb/in) (the assumed value). The small peaks occurred at the frequency where the FRF

becomes zero and it is the result of finite numerical precision in the computation. In practice it may

be necessary to use some type of the statistically averaging process of excluding these kinds of

peaks over a range of frequencies to accurately estimate kc using this approach. Namely, small

numerical values of FRFs closed to zero should be excluded to avoid infinite values in the inverse

process.

In a practical application the measurements will include errors arising from many potential

sources, and these will affect the estimation of kc. Two different situations will be considered: (a)

errors in measurement of the resonant frequencies, (b) errors in the amplitude measurement of the

FRFs.

In Figure 5.(a), kc is plotted for the frequency range of 0 through 4000 (rad/s) when a one percent

error exists in the first non rigid body resonant frequency. In Figure 5.(b), k c is plotted for the

frequency range of 0 through 4000 (rad/s) when a one percent error exists in the second non rigid

body resonant frequency. In Figure 5.(c), kc is plotted for the frequency range of 0 through 4000

10
(rad/s) when a one percent error exists in the third non rigid body resonant frequency. In Figure 5.

(d), kc is plotted for the frequency range of 1100 through 1300 (rad/s) when a one percent error

exists in the third resonant frequency to check double peaks that exist in Figure 5.(c). This kind of

double peaks can be explained using Figure 3.(c) and Figure 3.(d). Since, k c is the difference

between these two inverse FRF graphs, small shift in one peak create double peaks in kc graph.

From Figure 5, it can be said that the identification of k c using equation (15) give accurate results

except values near the frequencies where inverse FRFs become infinity.

Next, uniformly distributed random numerical noise was used to simulate measurement errors in

the amplitude of the FRFs. At each computed point of the simulated FRF, the amplitude was

assumed to be "contaminated" by measurement noise. The noise was simulated by adding a ±10%

uniformly distributed random number. Specifically, it was assumed that:

(16)

where A1= (1+ ) and A2= (1+ ) and  is a uniformly distributed random number with || 10%.

The resulting plot of kc as a function of frequency is shown in Figure 6.

If the result in Figure 6 is considered with FRFs or inverse FRFs in Figure 3, it can be shown that

a greater error in k1 will be created if the values of FRFs are close to zero (or if the inverse values of

FRFs are close to infinity).

For very high connection stiffness cases of k=10 6(lb/in), identified spring constants are plotted in

Figure 7. As expected, it shows slightly higher estimation but still gives useful values.

CONCLUSIONS

11
 When an analytical model does not exist, experimentally obtained modal parameters must be used

for the identification of connection parameters and for the explanation of dynamic properties that are

not obvious from the analytical models. However, the cost to measure all possible modes will be

outside of budget constraints, and it is likely that modes will be missed or that some of the identified

modes are not true modes of the structure. Hence the modal mode method which used the natural

frequencies and mode shapes can not be used when it is difficult to create an accurate modal data

from the test. The identification method of connections in this paper which used the FRFs directly

has a better technical merit in this aspect than methods using modal data of natural frequencies and

mode shapes [1, 2] to create system matrices M, C, K because FRFs are easily measured dynamic

properties of the structure. Besides, many difficulties and restrictions in the calculation of these

matrices from test data can be avoided.

The identification method which used the FRFs, as discussed in the previous section, works for

each discrete frequency so that the connection properties can be found for each frequency and can be

averaged using statistical methods to overcome the problems caused by measurement noise. The

identified connection properties near the frequency ranges where the FRFs are close to zero should

be excluded in the averaging process. If those highly sensitive regions are excluded, the

identification method for connection using FRFs can be applied to a general structure in an easier

manner than the modal model methods which require a full mathematical model of the mass,

damping and stiffness matrices (or natural frequencies and mode shapes).

Since highly sensitive regions can be excluded during the avering process, identified results with

only one sets of FRFs are more accurate than the modal model method. Identification of

connections using modal model methods was accurate only when full modal data was available [16].

Also, if FRFs are directly used to find the connection parameters, curve fitting processes that are

required to find natural frequencies and mode shapes can be omitted in obtaining the test data.

12
 This connection identification method of using FRFs can be also applied to the model update

methods and damage detections of structures using vibration test data.

REFERENCES

1. S.W. DOEBLING 1996 AIAA Journal, 34(12), 2615-2621. Minimum-Rank Optimal

Update of Elemental Stiffness Parameters for Structural Damage

Identification.

2. J.S. LEW 1995 AIAA Journal 33(11), 2189-2193. Using Transfer Function Parameter Changes

for Damage Detection of Structures.

3. M. ABDALLA 1998 Proceedings of the 16th International Modal Analysis

Conference (Santa Barbara, California) Feb., 144-150. Enhanced Damage

Detection Using Linear Matrix Inequalities.

4. K.C. YAP and D.C. ZIMMERMAN 1998 Proceedings of the 16th International

Modal Analysis Conference (Santa Barbara, California) Feb., 165-171. The

Effect of Coding On Genetic Algorithm Based Structural Damage Detection.

5.

(5) Sheinman, I., "Damage Detection and Updating of Stiffness and Mass

matrices Using Mode Data," Computers & Structures, Vol. 59, No.1, 1996.3,

pp. 149-156.

(6) Pandey, A.K., Biswas, M., "Damage Detection in Structures Using Changes in

Flexibility," Journal of Sound and Vibration, Vol. 169, No. 1, 1994, pp. 3-17.

(7) Li, D., Zheng, Z., He, K., Wang, B., "Damage Detection in Offshore

Structures By the FRF Method," Proceedings of the International Offshore

Mechanics and Arctic Engineering (Calgary), 1992, Vol. 1, Part B, pp. 601-604.

13
(8) Lim, T.W., Bosse, A., Fisher, S., "Structural Damage Detection Using Real-

Time Modal Parameter Identification Algorithm," AIAA Journal, Vol. 34, No. 11,

1996, pp. 2370-2376.

(9) Lim, T. W., "Structural Damage Detection of a Planar Truss Structure Using a

Constrained Eigenstructure Assignment," Proceedings of the 1994 AIAA

Dynamics Specialists Conference (Hilton Head, SC), AIAA, Washington, DC,

April 1994, pp.336-346.

(10) 강현규, 박중완, 류치영, 홍창선, 김천곤, “광섬유 센서 신호의 실시간 디지털 처리에 의한 구조물의 동적 변형률 측정,”

한국항공우주학회지, 제26권, 제2호, 1998, pp.47-59.

(11) Hwang, H.Y., "Identification Techniques of Structure Connection

Parameters Using Frequency Response Functions," Journal of Sound and

Vibration, Vol. 212, No. 3, 1998, pp. 469-479.

(12) Hwang, Ho-Yon, "Identification Methods for Physical Parameters in

Structures," Doctoral Thesis, Georgia Institute of Technology, May 1993.

(13) Daryl L. Logan “A First Course in the Finite Element Method,” Second Edition, PWS Publishing

Company, Boston, 1992

12. J.X. YUAN and X.M. WU 1985 ASME Journal of Engineering for Industry 107, 64-69.

Identification of the Joint Structural Parameters of Machine Tool by DDS and FEM.

13. R.R. JR. CRAIG 1985 Proceedings of the Joint ASCE/ASME Mechanics conference, June 24-

26, 1985 Albuquerque, New Mexico. A Review of Time-Domain and Frequency-Domain

Component Mode Synthesis Method.

14. K.H. GHLAIM and K.F. MARTIN 1984 Proceedings of the 2nd International Modal Analysis

Conference, 1984. 683-689. Reduced Component Modes in a Damped System.

14
15. D.J. EWINS and R. HENRY AGARDograph No. 298, AGARD Manual on Aeroelasticity in

Axial-Flow Turbomachines 2, 14-1-14-27. Structural Dynamic Characteristics of Individual

Blades.

16. HO-YON HWANG 1993 Doctoral Thesis, Georgia Institute of Technology, May, 1993, Atlanta,

GA. Identification Methods for Physical Parameters in Structure.

15
 k1

Structure d1
kn

dn
Figure 1. Fixed structure with connections

kc
system

connection

(a) System model

(b) Beam model for frequency response functions

16
 H(w)conn H(w)free

0.002 0.002
0.0015 0.0015
0.001 0.001
0.0005 0.0005
rad/s rad/s
-0.0005 1000 2000 3000 4000 -0.0005 1000 2000 3000 4000
-0.001 -0.001
-0.0015 -0.0015
-0.002 -0.002

(a) With connection (H()conn) (b) Without connection (H()free)

1/H(w)conn 1/H(w)free

40000 40000
20000 20000
rad/s rad/s
-20000 1000 2000 3000 4000 -20000 1000 2000 3000 4000

-40000 -40000

(c) With connection (H-1()conn) (d) Without connection (H-1()free)

kc

2000
1500
1000
500
rad/s
1000 2000 3000 4000
-500
-1000

Figure 4 : Identified connection spring constant, k c, as frequency varies

17
 kc

2000
1500
1000
500
rad/s
1000 2000 3000 4000
-500
-1000
(a) One percent error in the first non rigid boby resonant frequency of H()conn

kc

2000
1500
1000
500
rad/s
1000 2000 3000 4000
-500
-1000
(b) One percent error in the second non rigid boby resonant frequency of H()conn

kc

2000
1500
1000
500
rad/s
1000 2000 3000 4000
-500
-1000
(c) One percent error in the third non rigid boby resonant frequency of H()conn

18
 kc
6
2. 10
6
1. 10

rad/s
1175 1200 1225 1250 1275 1300
6
-1. 10
6
-2. 10
(d) Expanded plot of data in part (c)

Figure 5 : Identified connection spring constant, k c, when measurement errors exist in the resonance
frequencies

kc

5000
4000
3000
2000
1000
rad/s
500 1000 1500 2000
-1000
-2000
Figure 6 : Identified connection spring constant, k c, for uniformly distributed 10 % random noises in
magnitudes of H()conn and H()free

kc
6
1.2 106
1.1 106
1. 10
900000
800000
700000
rad/s
0 1000 2000 3000 4000

Figure 7 : Identified connection spring constant, kc, for assumed value of kc =106

19