1

3 School of Civil, Environmental

4 and Mining Engineering
5

7 JOURNAL PAPER SUBMISSION

8 CONDITION MONITORING OF STRUCTURES

9 THROUGH VIBRATION TESTINGS

10 By

11 F17S23

12

13 Hongzhi Wang, HW: 1674371

14 Jiamou Peng, JP: 1654874
15 Man Chun Martin Lee, ML: 1604637
16 Yong Wang, YW: 1642840
17

18

19 Supervisor: Dr. Ching-Tai Ng

20

21 October 2017
22
23 CONDITION MONITORING OF STRUCTURES THROUGH VIBRATION TESTINGS

24 Hongzhi Wang, Jiamou Peng, Man Chun Martin Lee, Yong Wang, Ching-Tai Ng
25 Campus, The University of Adelaide, SA 5005, Australia
26 ABSTRACT

27 Nowadays, the issue of infrastructure aging in Australia has been recognized as one of the

28 greatest challenges facing civil and structural engineers. It has significant impacts on the

29 structure vibration characteristics which has important influences on the serviceability, design life

30 and safety of structures. This paper presents the application of Least Square modal identification

31 method to obtain structure vibration characteristics. The proposed methodology consists of

32 conducting a field vibration test under ambient environment. Besides, multi-degree of freedom

33 numerical data has been also used to carry out the verification process during modal identification

34 development. The bridge structural parameters, such as Natural frequency (f), Damping ratio(𝜁)

35 and Mode shape (Φ) have been estimated successfully and the results were promising and

36 encouraging.

37 1. Introduction

38 The popularity of describing structure vibration characteristics came from rapidly grown of

39 engineering industries in the last few decades. System identification was emerged as a research

40 area which deals with developing mathematical models to describe and characterise an unknown

41 system with aid of experimental data. It has been successfully adopted and implemented in

42 aerospace and mechanical engineering in the last few decades.

43 It has also attracted significant attention in civil engineering, one important application of the

44 system identification is to describe structure vibration characteristics. This can be done by obtaining

1
45 the most fundamental modal parameters, namely natural frequencies, mode shapes and damping

46 ratios through structural vibration testing. Moreover, modal parameters are adopted for structural

47 health monitoring (SHM) and structural damage detection. All of these applications of the modal

48 parameters could lead to improvement of structure serviceability, safety and minimize or avoid

49 catastrophic structure failures.

50 There are essentially two techniques which can be employed to obtain the three modal parameters,

51 forced vibration and ambient vibration test. The forced vibration test, also referred as known input

52 method, assesses structures by means of large inertial shakers or drop weights. A short-coming of

53 this method is it requires extensive instrumentation and traffic interruptions. Furthermore, it is

54 costly to obtain significant levels of response for large structures. Due to these disadvantages, the

55 other method which referred as output-only method, has attracted increasing research attention in

56 recent years. It uses traffic and natural excitation such as winds and earthquakes to vibrate the

57 structure. The structural responses by the unknown input forces can be recorded and used for further

58 analysis. It has been shown in past years that this method could provide reliable and accurate

59 estimation of modal parameters (Brincker, Zhang and Andersen, 2001).

60 Once test data has been gathered, modal identification is employed to analyse the gathered data and

61 obtain modal parameters. There are many modal identification algorithms existing in the literature

62 but most of them can be grouped into two strategies. One is the estimation of range of graphs and

63 use interpretation techniques to estimate modal parameters. The other one is to construct

64 mathematical models and determine a set of model parameters that best describe the system’s

65 behaviour.

66 The overall objective of this paper was defined, which is to investigate, develop and verify a

67 computer based modal identification method which could be used to determine the three
2
68 fundamental modal parameters mentioned and hence enable structural vibration characterization in

69 civil engineering.

70 The procedure of the methodology to estimate the structural parameters are demonstrated by the

71 results of numerical and field vibration test presented in Chapter 3. Chapter 4 is related to

72 verification of the computer program developed which is based on the Least Square modal

73 identification method. Chapter 6 contains discussion and interpretation of results obtained from real

74 vibration field tests.

75 2. Methodology

76 2.1 Fourier Transformation (FT) and Fast Fourier Transformation (FFT)

77 It is important to convert data from time domain into frequency domain in order to ascertain the

78 different frequency contents within a signal. This helps to understand the physical behaviour of the

79 measuring parameters within a system. The FT is the way to achieve it.

80 However, the measured data are usually discrete and often contain noises. This makes difficult to

81 construct FT for experimentally collected data. Hence, alternative method for computing FT is

82 required and it is generally done by artificially select number of data points from the time domain

83 data. This number would help define artificial time period and frequency resolution for calculated

84 spectrum. The alternative method is often referred as Fast Fourier Transformation (FFT). The

85 extracted data based on the selected number of data points can be expressed as the following

[ ]
N
−i 2 π ( k −1 ) ( i−1 )
y ( k )=∑ x ( i ) exp
86 i=1 N
( 1)

87 where x(i) is the original time sequence with i=1 to N, N is number of samples, y(k) is the FFT

88 sequence at frequency fk=(k-1)/NΔt (k=1 to N).

3
 89 The multiplication of y(k) and non-zero sampling time interval Δt is FFT and it equals FT when the

90 artificial time segment becomes infinitely large.

91 2.2 Power Spectral Density (PSD)

92 The PSD could help obtain the basic modal information of the measured response data. It can show

93 the properties of the stationary stochastic process in the frequency domain which help provide

94 reasonable initial guesses for natural frequencies and can estimate the damping level. It can also be

95 used for determining reasonable exciting frequency range in the forced vibration test. The definition

96 of the PSD is shown in Equation (2).

¿
1
s ( f )= lim X T fX T ( f )
97 T→∞ T
( 2)

98 where T is the duration of time series, * denotes the complex conjugate transpose, f denotes the

99 frequency abscissa and

T/2

100
X T ( f )= ∫ x ( t ) exp (−i 2 π ft ) dt
−T /2
( 3)

101 Note the expression above tends to become FT as T becomes infinitely large.

102 In discrete series, PSD could be reformulated as (3)

lim 1
T→∞
103 S ( f k )= ¿¿¿
T
¿

104 where y (k) is the FFT sequence, N is number of samples and Δt is time interval of samples.

4
105 Note that Equation (4) describes a two-sided spectrum, symmetric about the y-axis. Normally, only

106 the positive frequencies are concerned. One-sided spectrum 2×S (fk) is used to represent all the

107 information in the positive axis.

108 2.3 Singular Value Decomposition (SVD)

109 Experimental/Operational noises often contained within measured data, the noise level needs to be

110 reduced in order to minimize its impact on modal identification process. The first and most

111 important way to reduce noise levels is to pass the data through the SVD algorithm.

112 The output from the SVD consists of three matrices U, V, S. The U and V matrices are unitary, that

113 is UUH=1, where the superscript H indicates a Hermitian transpose. The S matrix contains the

114 singular values on its diagonal and the values are real. The three matrices from a decomposition of

115 the original matrix as (5)

H
A=USV
116
( 5)

117 Note that the SVD is performed for each of the matrices at each frequency and for each

118 measurement.

119 2.4 Least Square Identification (LS) for Unknown Inputs (Ambient Vibration Tests)

120 The FFT for the collected acceleration data is expressed in the same way with Equation (6) for LS

121 with unknown inputs.

√
N
2 Δt ( k−1 ) ( j−1 )
122 ^
F k= ∑
N j=1
^
ẍ j exp[−2 π i¿
N
] ( k =1 ,… , N ) ¿ ( 6 )

123 The measured acceleration is modelled as

5
 ^ẍ j= ẍ j ( θ ) +ε j
124
( 7 /13 )

125 where ẍ j (θ) is the model acceleration response characterized by the set of modal parameter u. The

126 ε j is the prediction error which equals the difference between the model response and measure data.

127 Assume linear classically damped dynamics with m contributing modes, the model response is
m
ẍ j ( θ )=∑ Φ ( i ) η̈i ( t j )
128 given by i=1
( 8/14 )

129 where Φ(i)∈ R n (i=1 , … , m) is the ith mode shape confined to the measured DOFs and η̈i (t j) is the

130 ith modal response which satisfying the equation of motion:

2
η̈i ( t )+ 2ξ i ωi η̇i ( t ) +ω i ηi ( t )= pi ( t )
131
( 9/15 )

132 where ω i=2 π f i ; f i∧ζ iare the natural frequency (Hz) and damping ratio of the ith mode and pi is the

133 ith modal force.

134 It is often assumed that the prediction error in Equation (7) is white noise. This leads to assumption

135 of the prediction error being Gaussian where the noise is independent and identically distributed

136 (Au, 2011). Assume stationary modal forces with a constant spectral density matrix in ambient

137 vibration test, the joint probability density function (PDF) for the FFT described in Equation 3.18

138 can be written in the following vector form:

139
Z k=
[ ]
ℜ Fk
ℑ Fk
∈R
2n

( 10 /16 )

6
140 It is known that Zk are asymptotically independent for large sample size. Additionally, each Zk has a

141 Gaussian distribution with covariance matrix given in terms of the FFT of model correlation

142 functions.

143 If sampling rate is high, the covariance matrix asymptotically equals

[ ]
T T
1 Φ ℜ H k Φ −ΦIm H k Φ S
C k= + e I 2n
144 2 Φ ℑ H k Φ ΦRe H k Φ
T T
2
( 11/17 )

145 for large N. Note Φ=¿, Se is the spectral density of the prediction error. I2n denotes the 2n×n

146 identity matrix and Hk∈C m × m is the theoretical spectral density matrix of the modal acceleration

147 response and is equal to

H k =diag ( hk ) S diag ( h¿k )

148
( 12/18 )

149 The S∈C m × m and it is the spectral density matrix of modal forces; hk∈C m is a vector of modal

150 transfer functions and its ith element equals to

[ ]
−1
hik = ( β ik−1 ) +i ( 2 ξ i βik )
2
151
( 13/19 )

152 and diag(hk)∈C m × m means diagonal matrix with the ith diagonal element equal to hik. The asterisk

153 denotes conjugate transpose and βik=fi/fk is frequency ratio. The (i,j) element of Hk is given by

¿
H k ( i, j )=S ij hik h jk
154
( 14 /20 )

155 where Sij denotes the (i,j) element of S.

7
156 The set of modal parameters u includes natural frequencies { f i}, damping ratios {ξ i }, mode shape

157 matrix Φ , spectral density matrix S for the modal forces and spectral density S e of the prediction

158 error.

159 The objective function of LS algorithm can be written in the following form:

1 1
160
L ( θ )= ∑
2 k
ln det C k ( θ )+ ∑ Z k C k ( θ ) Z k
2 k
T −1

( 15 /21 )

161 Provided with large sufficient data, Equation 3.27 can be well approximated by the second order

162 Taylor series Equation (16).

1
L ( θ ) ≈ L ( θ^ ) + ( θ−θ^ ) ^ H L ( θ−θ^ )
T

163 2
( 16/22 )

164 where θ^ is the most probable value (MPV) which minimises Equation (16) and ^
H L is the Hessian of

165 L at the MPV.

166  Computational Difficulties

167 The MVP of θ^ can be determined by minimizing Equation (16), this requires repeated calculations

168 of the determinant and inverse of Ck for different k and different trials of u. However, the

169 optimization process if often ill-conditioned. This is because Ck is close to being singular, especially

170 for data with high signal to noise ratio (Au, 2011). To see this, ReHk, ImHk∈ R m ×m have a rank of at

171 most m and thus the first term in Equation (11) has a rank of at most 2m. From the resonant

172 frequency band, the prediction error term is relatively insignificant so Ck∈ R 2 n× 2 nis dominated by

173 the first term, indicating it is close to singular under the usual case where m<n.

8
174 If numerical optimization is used to find all the MVPs, then the computational cost is expressed by

175 two factors: the number of measured DOFs n and the number of modes m (Au, 2011). The number

176 of modal parameters to be optimized is shown in Equation (17) when the Hermitian nature of S is

177 taken into account.

2 2
178 n p =2 m+ m +1+mn=( m+1 ) +mn=O( m¿¿ 2+ nm)(17 /23)¿

179 The growth dimension of n is a major concern in here because the number of DOFs can be

180 significant for large structures. As consequences, it increases number of parameters to be identified

181 and dimension of matrix during optimization process.

182 On the other hand, the growth of dimension with m is a secondary concern even if it is quadratic.

183 This is because frequency band dominated by small number of modes can be selected for modal

184 identification. The number of closely spaced modes equals to m in general case which often less

185 than four. For well separated modes, mode shapes can be identified separately and m=1 in such

186 case.

187 Due to the mathematical structure of the objective function, it is difficult or almost impossible in

188 some cases to efficiently optimize all parameters simultaneously. However, it has been found that

189 the MVP for mode shape under well separated modes can be obtained analytically which

190 significantly reduces computational effort. For general case (close spaced modes where m>1), it is

191 still possible to mathematically derivate the objective function such that MPV of some modal

192 parameters can be calculated efficiently calculate, hence providing partial optimal solution that

193 could be used as initial guess to determine all the MPVs on next stage through numerical iteration

194 process (Au, 2011).

9
195 Because the mathematical derivation is significantly different for well separated and close spaced

196 modes. The following writing and derivation for the LS algorithm is therefore divided into two

197 parts, modal identification with single mode and multiple (cross) modes.

198 2.5 Single Mode Least Square Identification (LS) using Unknown Inputs (Ambient Vibration Tests)

199 Assuming data is dominated by a single mode under a selected frequency band and only the FFTs

200 on this band are used to calculate the mode, m=1 and u consists of the natural frequency f, damping

201 ratio ξ and mode shape Φ ∈ R n, spectral density S of the modal excitation and prediction error Se. It

202 is also assumed that the mode shape is normalized to have unit norm.

203 From Equation (11), ImHk is a scalar factor which equals to zero since (ImHk)T=-ImHk. On the other

204 side,

[ ]
2 −1
ℜ H k =S ( β k −1 ) + 2ξ β k
2 2
=S Dk
205
( 18 /24 )

206 where

[ ]
2 −1
Dk = ( β k −1 ) + ( 2 ξ β k )
2 2
207
( 19/25 )

208 behaves like dynamic amplification factor, the frequency ratio is defined as β k =f /f k.

209 Combining Equation (18) and (19) into Equation (11)

[ ] [ ]
T
ΦΦ 0 E 0
+ ( S e /2 ) Ι 2 n=( 1/2 ) k
2
C k =( S Dk /2 )
210 0 Φ ΦT 0 Ek
( 20 /26 )

211 Where

10
 T 2
E k =S D k Φ Φ + S e Ι n
212
( 21/27 )

[ ]
−1
Ek 0
C k−1=2 −1
213 0 Ek
( 22/28 )

−2 n 2
det Ck =2 (det E k )
214 and
( 23/29 )

215 next step is to express Ek through its eigenspace decomposition. To do this, we define an

216 orthonormal basis B= {a j ∈ R n : j=1 , … , n } and set a 1=Φ ∈ R n∧let {a 1 , a 2 , … , an } be an orthonormal

217 basis in the orthogonal complement of the subspace spanned by a1. From this basis, the identity
n
matrix Ι n=∑ a j a j such that
T
218
j=1

n n

219
E k =S D k a1 a1 + Se T 2
∑ a j a j =( S Dk =S ) a 1 a +∑ S e 2 a j a jT
T
e
2
1
T

j=2 j=2
( 24/30 )

220 The determinant of Ek can now be written in a form shown in Equation (25)

221 [( ) ]
det Ek =( S D k + S e 2) S e2 (n−1 )=
S Dk
S e
2
+1 S e2 n
and the inverse of Ek equals
( 25 /31 )

n
−1
E k−1=( S Dk + Se 2 ) a1 a1T + ∑ S e−2 a j a jT
222 j=2
( 26 /32 )

223 Using a1=Φ and

224
∑ a j a jT =Ιn −a1 a1T =Ι n−Φ ΦT
j=2
( 27 /33 )
11
225 we can obtain

[ ( )]
−1
2
−1 −2 −2 S T
E k =S e Ι −S
n e 1+ e ΦΦ
226 S Dk
( 28 /34 )

227 after simplifications.

228 Substituting Equation (25) and (28) into Equation (22) and (23), the objective function Equation

229 (15) becomes

L ( f , ξ , S , σ 2 , Φ ) =−n N f ln2+ ( n−1 ) N f ln S e2 + ∑ ln ( S D k + S e2 ) + S e2 ( d−Φ T A Φ )

230 k where
( 29/35 )

[ ( )]
−1
2
S
A=∑ 1+ e Dk
231 k S Dk
( 30/36 )

T T
D k =ℜ F k ℜ F k + ℑ F k ℑ F k
232
( 31 /37 )

d=∑ ( ℜ F k T ℜ F k ℑ F k T ℑ F k ) =trace ( A0 )
233 k
( 32/38 )

A0 =∑ Dk
234 k
( 33 /39 )

235 Based on matrix theory, the MPV of the mode shape corresponds to the eigenvector of A with the

236 largest eigenvalue while minimizing L with respect to the mode shape under the norm constraint.

237 Since Dk is positive, so does the largest eigenvalue of A. After minimizing with respect to mode

238 shape, Equation (29) becomes

12
 L ( f , ξ , S , σ 2 ) =−n N f ln2+ ( n−1 ) N f ln S e2 + ∑ ln ( S D k + S e2 ) + S e−2 ( d− ^λ )
239 k where ^λ is the largest
( 34/ 40 )

240 eigenvalue of A in Equation (30).

241 It can be seen that L in Equation (34) only depends on four remaining parameters and their optimal

242 values can be obtained through unconstrained numerical optimization. It has been found that

243 computational effort has been reduced significantly before transformation.

244  High Signal to Noise Asymptotic

245 Define signal to noise ratio as the model spectral density of structural response with respect to the

246 prediction error at frequency fk, it can be expressed mathematically as

S Dk
γ k= 2
247 Se
(35 /41 )

248 This factor has significant influences on the expression of A in Equation (30) and L in Equation

249 (34). This is because in the cases of data with good quality (γ k >>1)

[ ( )] ( )
−1
2 2
−1 −1 S S
250
(1+ γ )k = 1+ e
S Dk
∼ 1− e ∼ 1
S Dk
( 36/42 )

251 Using the zeroth order approximation, the matrix A in Equation (30) becomes

A ∼ ∑ D k =A 0
252 k
(37 /43 )

253 which is a constant matrix. This implies the MPV of mode shape can be calculated directly with

254 known data.

13
255 The MPV of S and Se can also be obtained analytically. Using the first order approximation in

256 Equation (36)

A ∼ ∑ [1−(Se ¿¿ S D k )] Dk =∑ D k −S e S ∑ D k−1 D k =A 0−S e−1 S−1 ∑ Dk −1 Dk (38)¿

2 2 −1
257
k k k k

258 ^ , we can obtain the following

Using Equation (38) and evaluating Equation (39) at Φ=Φ

259
L ∼−n N f ln 2+¿ ∑
k
{
ln Dk + ( n−1 ) N f ln S e + S e
2 −2
[ (d − ^λ0 )
( n−1 ) N f ]}
(N ∑ Dk−1 D^ k ) ]
−1 −1
+ N f [ln S +S f
k
(39/45)

260 where

^λ =Φ ^T A Φ ^
0 0
261
( 40/46 )

262 is the maximum eigenvalue of A0 and

^
Dk = Φ^ T Dk Φ
^
263
( 41/47 )

264 Equation (39) shows that S and Se affect L through the third and fourth term. Notice both of the

265 terms are in form of lnx+c/x which has a unique minimum of (1+lnc) at x=c. The MPV of Se is thus

266 equals

^ 2 d− λ^ 0
Se =
267 (n−1)N f
( 42/ 48 )

268 and the MPV of S is given by

14
 ^S ( f , ξ )=N f−1 ∑ Dk ( f , ξ )−1 D
^k
269 k
( 43/49 )

270 Because Dk depends on {f , ξ }, thus ^

Se can be directly calculated once data is obtained but S still

271 depends on the remaining parameters { f ,ξ } . Hence, L is now only dependent on f and ξ

[
L ( f , ξ ) ∼ ∑ ln D k ( f , ξ ) + N f ln N f ∑ D k ( f ,ξ )−1 ^Dk ] +const
−1

272 k k
( 44 /50 )

273 Only the MPV of {f , ξ } is required to be determined by numerical optimization, this process is

274 extremely fast as it does not involve solving eigenvalue problem during iteration.

275 It is clear that Taylor series approximation of the term (1+gk-1)-1 in A for small gk-1 is used in the

276 asymptotic approximation. If value of gk-1 is small in the selected frequency band then a good

277 asymptotic approximation can be guaranteed. However, the converse of this is not necessary. A

278 good asymptotic approximation can still be obtained even if gk-1 at the spectral tail of the mode is

279 not small. This means although Taylor approximation is poorest near the spectral tail of the mode,

280 spectral density matrix ^

Dk contributes little to A as ^
Dk is insignificant compared to those at the

281 peak. Thus, the asymptotic approximation is overall more robust than the Taylor approximation

282 (Au, 2011).

283 Since Dk~1/4ξ 2 at the spectral peak, one could use spectral peak to define signal to noise ratio in

284 real practice. The peak signal to noise ratio is given by

S
γ= 2 2
285 4 Se ξ
( 45/51 )

15
286 Equation (45) describes the degree of excitement level of the target mode with respect to the

287 background prediction error which consists measurement noise and response from other non-

288 resonant modes. It is not hard to see from Equation (45)that modes with frequency significantly

289 higher than the target mode have little effects on the prediction error as their dynamic amplification

290 is second-order small in the selected frequency band. Vibration modes have significant low

291 frequency compared to the target mode influence the prediction error pseudo-statically because their

292 dynamic amplification is O(1). The resultant prediction error is likely to be dominated by

293 measurement noise for the first few interested modes. Thus, g is large if the measurement noise is

294 small compared to the resonant response. If the modes of interest are high and the prediction error is

295 dominated by the pseudo-static contribution from lower non-resonant modes, then S e=O(S) and

296 g~1/4ξ 2>>1. It should be noted that higher modes can be difficult to identify in the real field as

297 unknown coloured environmental or human excitations are often present (Au, 2011).

298 3. Numerical Case Study

299 The purpose of numerical study is to verify the workability of modal identification method with

300 considering a multiple DOFs system which consists of 10 DOFs systems. It is mainly based on the

301 Least square identification of single modes as mention in Chapter 2.5.

16
302 Initially, the Root SV and PSD graph were generated and then an appropriate initial guess of natural

303 frequency could be made to evaluate modal parameters. All the test data length were divided into

304 2000 samples set to generate Root PSD and SV (Figure 1 and Figure 2) in order to obtaining an

305 optimal view of various peaks. Besides, it also could demonstrate whether there are cross modes

306 between or not based on the overlap of different lines.

307 The eight mode shapes out of ten were chosen which demonstrate in Table 1, and It represents

308 biases between modal answer which comes with data and estimation by modal identification, this is

309 because the mode 9 and 10 are difficult to determine the accurate range based on the Figure 2

310 shown (SVD).

311 In general, this verification of ambient vibration tests indicates the procedure of FFT method for

312 structure modal identification. The most probable parameters were estimated by the known range

313 of FFT on PSD graph, which presents consistent behaviour of the MPVs under multi degree of

314 freedom structure. As the consequence, most of mode shapes and damping ratios were well-

315 identified comparing with the modal answers, so the development of modal identification could be

316 recognized as having an ideal workability to estimate modal parameters of real-life infrastructures

317 in field vibration tests which will be addressed in the following chapter.

318 4. Experimental Studies

319 The field vibration test was carried out to gather the vibrations of the real-life structure, Mcdonnell

320 Avenue Pedestrian and Cyclist Bridge as shown in the Figure 5-1 and Figure 5-2, which shows the

321 location of the structure and a side view of the structure, and It is a single span bridge across

322 Karrawirra Parri, in the Linear Park, Adelaide, South Australia. This structure has not been

323 monitored since the construction processes completed.

17
324 The general facts of bridge are 30 meters in length and 3 meters in width, and also bridge railing are

325 assembled along span. It was constructed by concrete with additional brick blocks reinforcing the

326 two supports. It is noticed that, the Figure 5-3 demonstrates the information of setups and there are

327 4 rigid supports altogether on the both side of bridge, which indicates that giving 4×2×3=24 DOFs

328 was measured by seismic recorders triaxially. However, only vertical direction will be considered

329 as the major factor in this vibration test. Moreover, the reference points were located at the 2/5

330 span, primarily because it would not be affected by vibration to evaluate the modal parameters, such

331 as mode shapes.

332 The bridge were covered by seven proposed setups and Table 4-1 shows the setup plan, which was

333 arranged in chronological order of the test. As verification process was completed on site, other

334 data was processed after field vibration test. Due to 7 setups were conducted during test, 70 minutes

335 data were extracted from data acquisition and 42 minutes data of 7 setups were eventually applied

336 for outputs due to the consideration of 4 minutes transition in each setups. The whole field

337 vibration test took about 4 hours approximately (from 10.30am – 14.20 pm 5/10/2017).

338 The initial setup and calibration procedure (Figure 5-4) was completed before the field test

339 conducting, this is due to ensure a certain amount of vibration responses could be acquired and all

340 seismic recorders were synchronized with GPS clock, thus, estimating modal parameters.

341 Furthermore, the original sampling rates were 2000 Hz because of the consideration of results

342 accuracy and each test conducted 10 minutes in order to minimize the effect on transition between

343 setups.

344 Before the implementation of test, four seismic recorders were allocated together to collect the

345 similar amplitude vibration data in order to inspect the natural frequency of structure based on FFT

18
346 as shown in Figure 5-4 and ensure all the seismic sensors had been synchronized as indicated

347 before.

348 Then, the root PSD and root SV spectrum diagrams were generated to inspect the data allocation as

349 shown in Figure 5-5.

350 It can be seen that data acquisition process was successful due to the amplitudes are similar in

351 accordance with the root PSD graph. Also, the data has been synchronized ideally because of the

352 different spread on the root SV (Figure 5-5) since it could be recognised as an ideal indication of

353 synchronous data (Zhu and Au, 2017).

354 4.1 Field Vibration Test Results

355 It should be noted that natural frequencies under 20 Hz were considered in this research project.

356 This is because we are only interested in modal parameters that characterise the whole structural

357 vibrations. Natural frequencies that higher than 20 Hz for the tested foot bridge are likely to be

358 associated with vibrations from local structural elements. This is evidenced by Figure 6-1 which

359 shows there are four significant peaks for all sensors blow 20 Hz, indicating four sets of modal

360 parameters exist for the whole structural vibration characteristics. The significant peak between 40

361 Hz and 50 Hz are likely due to local vibration hence was omitted. PSD plots for all setups have

362 similar trend with respect to Figure 6-1, hence only two PSD plots are shown in here for illustration

363 purpose.

364 The peaks from the PSD plots were used to provide initial guess for natural frequency and select

365 frequency range where data can be used for modal identification. The initial guess was estimated

366 from the locations where all four sensors have the similar frequency peaks. The frequency band was

19
367 selected to surround the initial guess, It does not need to be excessively wide otherwise it increases

368 the risk of modelling error as the tail contains little information.

369 It can be seen from Table 6-2 that four sets of calculated natural frequencies are highly consistent

370 whereas damping ratios vary considerably large for different setup. However, this pattern is not

371 surprising. In general, these variations could because experimental errors, existing structural

372 damages and variations of environmental conditions during the entire field vibration test.

373 In addition, It is known that accurate experimental estimation of damping ratio is still an open

374 problem in the literature. This is because damping ratio can be affected significantly by factors such

375 as material type/density, pressure and temperature (Stephens and Scavullo, 1865).

376 Because of this, the definition of a well-established formulation to represent damping forces is still

377 debated. Hence, calculation of damping ratio is usually characterised by large error bounds. On the

378 other hand, natural frequencies and mode shapes can be relatively easily and accurately calculated

379 (Carlo Rainieri and Giovanni Fabbrocino, 2014).

380 Despite this explanation, there may still be some other reasons for the large errors in the calculated

381 damping ratios. One could be due to insufficient data length, that is, recording time for each setup is

382 not long enough to determine damping ratio to a desired accuracy. If this is true, then modal

383 parameter precision and sampling time are generally governed by damping ratio.

384 Four mode shapes corresponding the four sets of modal parameters were also determined which are

385 shown in Figure 6-2 to Figure 6-5.

20
386 5. Conclusion

387 The aim of this paper is to investigate and estimate the structural parameters of the real-life

388 structure by using Least Square modal identification. The acquired data served as input to the

389 verified single mode ambient program. It was found that the first calculated natural frequency align

390 with vibration frequency introduced by pedestrians which post safety concern for the tested bridge

391 and local community. The damping ratio and natural frequencies were found within acceptable

392 limits. The calculated mode shapes were physically sound and reasonable. Overall, it can be said the

393 results from the single mode ambient program were promising and encouraging.

394 However, as Table 6-3 summarising footbridge design codes with respect to natural frequencies in

395 different region, the natural frequency should be taken into account in footbridge design to

396 minimise the likelihood of resonance caused by pedestrians. This is because Occhiuzzi et al. (2007)

397 have concluded that pedestrians would induce lateral and vertical direction vibration to the

398 footbridge when crossing. If this introduced vibrations are in similar range compared to the natural

399 frequency of the footbridge, resonance could be resultant which would have negative impacts on the

400 safety of the structure and pedestrians.

401 It should be pointed out that there has not been a detailed guideline for considering and calculating

402 lateral natural frequency on footbridge design in Australia. Dallard et al. in 2001 suggested that

403 lateral natural frequency could be estimated as half of the vertical natural frequency.

404 One can see that the calculated vertical natural frequency from the first set of modal parameters

405 Table 6-2 falls into similar range with the second column in Table 6-3. In addition, this calculated

406 vertical natural frequency is roughly twice in magnitude with respect to the value in the third

407 column of Table 6-3. Dallard et al. conducted studies in 2001 about failure reasons for the

408 Millennium Bridge in London in 2000. He concluded one of the reasons was the introduced
21
409 vibrations by pedestrians were in similar frequency range compared to Table 6-3 which caused

410 resonance of the footbridge. He also found out that the designers of the bridge did not aware of the

411 minimum criterion of lateral bridge frequency during design process of the bridge.

412 Based on findings from Dallard et al. in 2001 and the calculated first natural frequency, one can

413 conclude that resonance is likely to be resultant on the tested footbridge due to pedestrian crossing,

414 especially with large number of people crossing simultaneously. This posts a safety hazard for the

415 serviceability of the bridge and safety of the local community.

416 6. Acknowledgements

417 We would like to express our sincere appreciation and gratitude to Doctor Ching-Tai Ng (Alex)

418 who has supervised us through the research this year. His vast knowledge, encouragement and

419 guidance throughout the year have been instrumental to the progress and completion of this

420 research.

421 7. References

422 Arakawa, T & Yamamoto, K 2004, "Frequencies and Damping Ratios of A High Rise Building
423 Based on Micro-tremor Measurement", viewed 8 October 2016.

424 Au, S, Zhang, F & Ni, Y 2013, "Bayesian operational modal analysis: Theory, computation,
425 practice", Computers & Structures, vol. 126, pp. 3-14.

426 Au, S. 2011, Fast Bayesian FFT Method for Ambient Modal Identification with Separated Modes.
427 Journal of Engineering Mechanics, 137(3), pp.214-226.

428 Au, S. 2012, Fast Bayesian ambient modal identification in the frequency domain, Part I: Posterior
429 most probable value. Mechanical Systems and Signal Processing, 26, pp.60-75.

430 Au, S., 2017, Operational Modal Analysis. Springer.

22
431 Beskhyroun, S, L. Wotherspoon, Q. T. Ma & [Link], 2013, “Ambient and Forced Vibration
432 Testing of a 13-Story Reinforced Concrete Building”, 2013 NZSEE Conference, Department of
433 Civil and Environmental Engineering, The University of Auckland, Auckland.

434 Brincker, R., Zhang, L. and Andersen, P. (2001). Modal identification of output-only systems using
435 frequency domain decomposition. Smart Materials and Structures, 10(3), pp.441-445.

436 Carlo Rainieri & Giovanni Fabbrocino, 2014. Operational Modal Analysis of Civil Engineering
437 Structures. Springer.

438 Celik, O, Sucuoğlu, H & Akyuz, U 2015, "Forced Vibration Testing and Finite Element Modeling
439 of a Nine-Story Reinforced Concrete Flat Plate-Wall Building", Earthquake Spectra, vol. 31,
440 no.2, pp. 1069-1081.

441 Celik, O. C., 2002. Forced Vibration Testing of Existing Reinforced Concrete Buildings,
442 [Link], Civil Engineering Department, Middle East Technical University, Ankara, Turkey.

443 Dallard, P, Fitzpatrick, T, Flint, A, Low, A, Smith, R, Willford, M & Roche, M 2001, "London
444 Millennium Bridge: Pedestrian-Induced Lateral Vibration", Journal of Bridge Engineering, vol. 6,
445 no. 6, pp. 412-417.

446 David G. Stepbens and Maurice A. Scavullo, 1865, “Investigation of air damping of circular and
447 rectangular plates, a cylinder, and a sphere”, National Aeronautics and Space Administration,
448 Washington, D.C., pp 1-24.

449 Freitas, T, 2007 “Modal Parameters Identification Using Only Response Data – Stochastic

450 Subspace Identification and Frequency Domain Decomposition –”, 19th International Congress
451 of Mechanical Engineering, Brasilia, Brazil.

452 Gentile, C & Saisi, A 2011, “Ambient vibration Testing and condition assessment of the
453 Paderno iron arch bridge (1889)”, Construction and Building Materials, pp 3709-3720.

454 Magalhaes, F. et al 2012, “Ambient and free vibration tests of the Mallau Viaduc: Evaluation of
455 the alternative progress strategies”, Engineering Structures, pp 372-384.

456 Maia, N.M.M. and Silva, J.M.M, 2001, “Modal analysis identification techniques”, The Royal
457 Society, 359(1778), pp. 29–40.

23
458 Nakata, S. et al 2012, “Influence of vibration methods, structural components and excitation
459 amplitude on modal parameters of low-rise building”, 5th world Conference on Structural and
460 Monitoring, pp 1-15.

461 Peeters, B & De Roeck, G 2001, "Stochastic System Identification for Operational Modal
462 Analysis: A Review", Journal of Dynamic Systems, Measurement, and Control, vol. 123, no. 4, p.
463 659.

464 Peeters, B & Ventura, C 2003, “Comparative Study of Modal Analysis Techniques for Bridge
465 Dynamic Characteristics ", Mechanical Systems and Signal Processing, vol. 17, no. 5, pp. 965-
466 988.

467 Reynders, Edwin, 2012 "System Identification Methods for (Operational) Modal Analysis: Review
468 and Comparison". Archives of Computational Methods in Engineering, vol. 19, no. 1, pp. 51-124.

469 Salawu, O & Williams, C 1995, "Bridge Assessment Using Forced-Vibration Testing", Journal of
470 Structural Engineering, vol. 121, no. 2, pp. 161-173.

471 Sinha, Jyoti Kumar. Vibration Analysis, Instruments and Signal Processing. 1st ed. Boca Raton:
472 CRC Press, 2015.

473 Trifunac, M 1972, "Comparisons between ambient and forced vibration experiments",
474 Earthquake Engineering Structure Dynamic, vol. 1, no. 2, pp. 133-150.

24
475 Xia, Y, Chen, B, Weng, S, Ni, Y & Xu, Y 2012, "Temperature effect on vibration properties of
476 civil structures: a literature review and case studies", Journal of Civil Structural Health
477 Monitoring,vol. 2, no. 1, pp. 29-40.

25
478 Tables

479 Table 1 Setup Plan in Chronological Order, Mcdonnell Avenue Pedestrian and Cyclist Bridge

Sensor
Setup Ref. 1 Ref. 2 Rover 1 Rover 2 Remark
1 104 204 101 201 Left End
2 104 204 102 202
3 104 204 103 203
4 104 204 105 205
5 104 204 106 206
6 104 204 107 207
7 104 204 108 208 Right End
480

Mode 1 2 3 4 5 6 7 8
Identified
Natural
1.001 2.987 4.892 6.703 8.351 9.805 11.068 12.067
Frequency
Damping Ratio 0.011 0.009 0.010 0.011 0.010 0.010 0.009 0.009
Modal Answer
Natural
1.001 2.980 4.893 6.696 8.350 9.817 11.065 12.066
Frequency
Damping Ratio 0.010 0.010 0.010 0.010 0.010 0.010 0.010 0.010
481 Table 2 Resluts of 10 DOFs building (1st-8th Mode)
482

483 Table 3 Modal Parameters of Mcdonnell Avenue Pedestrian and Cyclist Bridge

Mode 1(V) Mode 2(V) Mode 3(T) Mode 4(V)

Setup f (Hz) ζ (%) f (Hz) ζ (%) f (Hz) ζ (%) f (Hz) ζ (%)
1 2.476 0.043 8.806 0.015 11.666 0.012 18.212 0.017
2 2.152 0.054 8.558 0.032 11.705 0.003 18.172 0.023
3 2.198 0.089 9.010 0.025 11.452 0.011 18.517 0.033
4 2.306 0.069 8.755 0.011 11.708 0.013 18.145 0.017
5 2.434 0.092 8.804 0.022 11.695 0.016 18.147 0.013
6 2.520 0.042 8.765 0.008 11.693 0.013 18.059 0.016
7 2.482 0.030 8.487 0.022 11.694 0.006 18.183 0.012
Mean 2.367 0.060 8.741 0.019 11.659 0.011 18.205 0.019
COV 6.26% 40.32% 1.98% 44.55% 0.79% 43.54% 0.80% 38.60%
484 * Coefficient of variation (COV) = standard deviation; V and T means vertical and torsional mode
485 shapes respectively

1
486 Table 4 Summary of natural frequencies in international codes (Guidelines for the design of
487 footbridges, p.1891, 2005)
Limit values
Code/Standard Horizontal
Vertical
American Guide Spec. < 3 Hz
Eurocode 2 (ENV 1992-2) 1.6 Hz - 2.4 Hz 0.8 Hz - 1.2 Hz
DIN-Fachbericht 102 1.6 Hz - 2.4 Hz, 3.5 Hz - 4.5 Hz
Eurocode 5 (ENV 1995-2) < 5 Hz
SBA (former East Germany) 1.0 Hz - 3 Hz
< 1.3 transverse
SIA 260 (Switzerland) 1.6 Hz - 4.5 Hz
< 2.5 longitudinal
BS 5400 (Great Britain) < 5 Hz
Austroads (Australia) 1.5 Hz - 3 Hz
Japanese Footbridge Design Code (1979) 1.5 Hz - 2.3 Hz
488

2
489 Figures

Mode 2
Mode 3 Mode 5

Mode 1 Mode 4
Mode 6
Mode 7

Mode 8

Mode 9

490

491 Figure 1 Root PSD 10 DOFs Structure

492

493

494

Mode 1 Mode 6 Mode 7

Mode 2 Mode 8
Mode 4 Mode 5
Mode 3 Mode 9

495

496 Figure 2 Root SV for 10 DOFs Structure

497

498

499
3
500

N
McDonnell Avenue
Pedestrian and Cyclist
Bridge

The University of
Adelaide
Adelaide CBD
501

502 Figure 3 The McDonnell Avenue Pedestrian and Cyclist Bridge on Google Map
503

504

505

North side of bridge South side of bridge

506

507 Figure 4 Side View of McDonnell Avenue Pedestrian and Cyclist Bridge
508

509

4
510

511

512 Figure 5 Setup locations, Mcdonnell Avenue Pedestrian and Cyclist Bridge
513

514

515

516

517 Figure 6 The process of data validation and synchronization

518

5
519

520 Figure 7 The plots of data Verification and Synchronization (PSD and SV)
521

522

523 Figure 8 PSD plots for setup 4 and 5

524

Y
X
525

526 Figure 9 1st Mode Shape 2.37 Hz, 5.98%

6
 Z

Y
X
527

528 Figure 10 2nd Mode Shape 8.74 Hz, 2.43%

Y
X
529

530 Figure 11 3rd Mode Shape 11.66 Hz, 1.06%

Y
X

531

532 Figure 12 4th Mode Shape 18.21 Hz, 1.88%

533