Bridge Maintenance, Safety, Management and Life-Cycle Optimization Frangopol, Sause & Kusko (eds) 2010 Taylor &

p; Francis Group, London, ISBN 978-0-415-87786-2

Observed dynamic characteristics of an overpass bridge during destructive testing

D.M. Siringoringo, T. Nagayama, Y. Fujino, D.Su & C.Tandian
Department of Civil Engineering, The University of Tokyo, Tokyo, Japan

ABSTRACT: Vibration measurement and analysis of dynamic characteristics of an overpass bridge during a full-scale destructive test are described. Damage is introduced systematically by cutting one of bridge piers at the footing level and inducing initial settlements. This type of damage is expected to simulate a condition where a bridge suffers from non-uniform pier settlement. By applying time and frequency domain vibration analysis, as well as system identification, evolution of dynamic characteristics caused by damage is quantified. The results show that changes of natural frequencies are clearly visible, thus can be used as indicator of damage presence, while the change in mode shapes can be used as the local damage indicator. In addition, application of novelty detection based on multivariate outlier analysis of the auto-spectra function is also discussed.

1 INTRODUCTION Vibration measurement during progressive damage test of a full-scale structure is a very rare and important learning opportunity. From such measurement, one can observe evolution of dynamic characteristics, validate damage detection method, and formulate the baseline criteria for typical structural deteriorations. This paper describes such a study. It includes vibration measurement and destructive test, analyses the data and presents the results of vibration analysis. Ambient vibration measurements were conducted before, during and after introduction of damage. Sensitive features that can be utilized as indicators of damage were extracted from vibration characteristics. 2 TESTED BRIDGE The tested bridge is the S101 Overpass Bridge located in Reibersdorf, Upper Austria, west side of Vienna, Austria. The bridge crossed over the national highway A1 Westautobahn Austria. It is a post-tensioned concrete bridge with the main span of 32 m, side spans of 12 m, and the width of 6.6 m (Figure 1). The deck is continuous over the piers and is built into abutment. The bridge, built in 1960, was a typical overpass in the national highway. Although there were no known significant structural problems,

the bridge had to be demolished to allow a space for additional lane of highway underneath. Before demolition, series of vibration test was carried out by the Vienna Consulting Engineers (VCE 2009). The authors participated in two-day measurement from December 10 until 11, 2008.

1200

3200

1200

2:

120

360

470

damaged pier

60

120

damaged pier Cutting the pier

Figure 1. Destructive Test of S101 Bridge

Measurement system consists of six triaxial accelerometers. During two days measurement, ambient vibration of bridge was measured and six sensor

1664

configurations were employed. Two sensors (i.e. node A and B) were kept at the same place throughout measurement to provide reference for timesynchronization. Four other sensors were the roving sensors that moved from one end to the other end of the bridge (Figure 2). To measure the bridge in undamaged condition, three sensor arrangements were utilized (sensor arrangement 1,2, and 3 in Figure 2). For damage cases sensor arrangement 4,5, and 6 were employed. During the two-day measurement, temperature condition was relatively equal with average temperature of -2oC.

At the last stage, a steel plate was inserted to close the gap between pier and the footing. In this condition the pier rested on the plate and the stage is named Retrofitted stage.

Figure 2. Sensor Arrangement

3 DAMAGE SCENARIOS Damage was introduced to the structure by cutting the pier column just above the pier footing. The cut was made twice each was approximately 5 cm layer of column. Damage 1 and Damage 2 is defined as a state where the first and second cut is made, respectively (Figure 3). During cutting process, a steel column was placed alongside the pier and tightened to the pier with steel rods. A hydraulic jack was placed on bottom of the steel column to provide a temporary support. Immediately after the cutting process was completed, the temporary steel column was lowered gradually by releasing the pressure in hydraulic jack. This caused the vertical settlement of the bridge at the location of pier column. To introduce the initial pier settlement, pressure in hydraulic jack was released and the temporary steel column was lowered 1 cm. This was followed by 1 cm vertical settlement of pier (Damage 3). The settlement was further increased to 2 cm by further lowering the temporary steel column 1 cm (Damage 4). Finally, at Damage 5, the steel column was lowered until 3 cm but the total settlement of the bridge is only 2.7 cm. From this point, no further vertical settlement was observed. The pier column was suspended completely; hence the hydraulic jack and temporary steel column did not function anymore. During the process of damage and pier lowering, vertical settlement of the bridge at the pier location was recorded by geodetic leveling and laser system.

Figure 3. Damage scenarios

4 VIBRATION ANALYSIS 4.1 Spectrogram Analysis The frequency content of response was observed in the frequency range 1-50 Hz. The dominant frequencies at the girder in vertical direction appear in the range of 1-15 Hz. Figure 4 shows the spectrogram of Reference sensor A and B throughout the measurement. The ordinate consists of two parts: the undamaged part (frame number 1-10), and damaged part (frame number 11 onward). In the undamaged part, one can see four distinct vertical lines representing four natural frequencies within the range of 3 to 14 Hz. The first line is around 4 Hz, second line is around 6 Hz, the third and the forth line is around 9 Hz and 13 Hz, respectively. Despite the fact that amplitude of ambient vibration was small, the fre-

1665

quency spectra show very clear peaks indicating well-separated modes and suggesting that the records have high signal-to-noise ratio. The four vertical straight lines indicate constant peaks in frequencies plot and small variation of natural frequency estimates. Unlike the undamaged part, the damaged part shows distinct variation of frequencies. Starting from frame number 11 one can observe the leftward shift of natural frequencies especially the forth mode (13 Hz). The other modes show apparent shift starting from frame 20 onward, which correspond to the time when the bridge experienced 2cm of vertical settlement. Leftward frequency shift of the first, second and third modes continue until frame number 34. Largest shifts were observed at the time when the pier was completely suspended indicating the significant reduction of stiffness. Starting from frame number 35 onward, we can observe rightward shift of the natural frequencies. For the 1st mode, the frequency shifted back almost to its original position, while some residual frequency shifts were observed for the 2nd, 3rd, and 4th modes. Note that frame 35 corresponds to the time when the steelplate was inserted and the structure was in the Retrofitted stage. This result indicates that the steel-plate insertion reduces the vertical flexibility of the structure as evident by the rightward shift of the 1st and 3rd mode (all are bending), but not in the same degree as it reduces the torsional flexibility, as evident by residual frequency shift in 2nd and 4th mode (torsional modes). Results of spectrogram analysis clearly reveal the evolution of natural frequencies during damage stages and can be used as indicators of structural damage.

(James et al. 1993) and Eigensystem Realization Algorithm (Juang and Pappa 1985) are employed in this study. In the NExT, the cross-correlation function (CCF) between reference and roving nodes are computed and treated as the free-vibration responses. Considering sensor arrangement, the global system identification is performed in two parts, one is for undamaged stage where fourteen measurement nodes are used, and the other is for damaged stage, where only six measurement nodes are used. To investigate the effect of variability and to estimate the confidence bounds of identified modal parameters, a statistical analysis is essential. By quantifying statistical properties we can interpret the results with confidence. For this purpose, the Bootstrap method (Efron and Tibshirani 1993) is employed in this study. The Bootstrap method randomly selects and replicates the response from limited number of record to create ensemble average of larger population of response. Statistical properties of the ensemble average are computed to determine the bounds of uncertainty. To implement Bootstrap analysis, large number of CCF data set was randomly selected from the available time-normalized CCF data, and the CCF ensembles were formed. On each ensemble, the CCF ensemble average was computed, and then treated as Markov parameter in the Hankel matrix of ERA. This procedure was repeated for large number of times to form a histogram of the identified modal parameters. The confidence bounds of modal parameters are calculated by the percentile interval method that computes the 95% confidence limit by sorting the modal parameters in an ordered list and defining the value of upper and lower 2.5% percentile. 4.3 Result of Modal Analysis on Damage Stages
Table 1. Identified natural frequencies and damping ratios (value in brackets denotes 95% confidence interval by Bootstrap) _______________________________________________ Mode Frequencies (Hz) ____________________________________ Damage 3-4 Damage 5 Retrofitted _______________________________________________ 1st Bending 3.90 (0.13) 3.65 (0.05) 3.94 (0.07) 1st Torsion 5.84 (0.08) 5.22 (0.20) 5.76 (0.05) 2nd Bending 9.21 (0.13) 8.16 (0.39) 9.04 (0.06) 11.76 (0.31) 10.28 (0.16) 11.06 (0.44) 2nd Torsion __________________________________________________ _______________________________________________ _____________________________________________ Mode Damping Ratio (%) ____________________________________ Damage 3-4 Damage 5 Retrofitted _______________________________________________ 1st Bending 1.98 (3.19) 2.10 (1.31) 2.76 (2.62) 1st Torsion 2.14 (1.23) 2.72 (2.54) 1.93 (1.05) 2nd Bending 2.12 (1.25) 1.93 (3.91) 2.13 (0.63) nd 2 Torsion 1.49 (3.07) 1.23 (2.29) 1.48 (3.65) __________________________________________________ ______________________________________________

Figure 4. Spectrogram showing the evolution of natural frequencies with respect to damage level

4.2 Modal Analysis and Bootstrapping Global modal parameters are derived from ambient acceleration response under the assumption of stationary random excitation. To extract modal parameters, the Natural Excitation Technique (NExT)

The first four mode shapes are shown in Figure 5. Results of identification for damage stages are listed in Table 1. In this table the results are divided into

1666

three groups: 1) Damage 3 and 4 (frame number 1629), 2) Damage 5 (frame 30-33) and 3) Retrofitted stage (frame 34-36).
Table 2. Changes in frequencies due to damage ______________________________________________ Mode Frequency Changes (%) ____________________________________ Damage 3-4 Damage 5 Retrofitted _______________________________________________ 1st Bending -3.01 -9.23 -2.01 1st Torsion -7.49 -17.31 -8.76 2nd Bending -4.51 -15.40 -6.27 2nd Torsion -12.02 -23.09 -17.26 __________________________________________________ ______________________________________________

Figure 6 shows the comparisons of 95% confidence bound estimated by Bootstrap method and Table 2 lists the identified change of frequencies as a result of damage. It can be seen in the figure and table that for Damage 3 and 4, natural frequencies of the second, third and forth modes experience significant changes as denoted by frequency changes that are larger than the 95% confidence bound. These changes despite small can be considered statistically significant and be used with confidence as damage indicators. On contrary, frequency change of the first mode is statistically insignificant because its value is smaller than the 95% confidence bounds and thus cannot be used as damage indicator. During Damage 5, the changes in natural frequencies of all modes become more significant. All frequency changes are now larger than the 95% confidence bounds.

Figure 6. Confidence bound of identified natural frequency

Figure 5. Identified mode shapes for undamaged case

For damping ratio (Table 1), there is slight increase in the mean value as a result of damage. The averaged values of damping for all four modes were between 1.2 to 1.5 % for undamaged structure with small bound of 95% confidence. These values increase slightly up to 2% for Damage 3, 4, and 5; and up to 2.7% for Retrofitted condition. Note, however, that in damage condition the 95% confidence bound were significantly larger than that during undamaged. These large bounds indicate large variation or scatternes in damping estimates. Therefore, even though damage created damping change (i.e. increased the damping), they are statistically insignificant and cannot be used with confidence as damage indicator. 4.4 Mode Shapes Local Change Another important aspect is the effect of damage on the mode shape. Simulation using FEM suggested that when damage altered the support condition, significant change in mode shapes were resulted. One of the two stationary points on the mode shape that initially located at the pier-girder connections moves vertically as the results of pier damage. This change is understandable because vertical movement that was restrained by pier on footing level is now released due to damage. Observation on identified mode shapes reveals similar outcome. Unfortunately, due to limited num-

1667

ber of sensor, only half-span of mode shapes can be analyzed. To compare these half-span mode shapes of damaged bridge with the complete span mode shape of undamaged bridge, the first and second modes (i.e. bending and torsion) are normalized to the maximum value that occurred in the midpoint of the span. Mode shapes comparisons are shown in Figure 7.a and b. In both modes, we can observe larger relative modal displacement at the pier-girder connection during damage stages. The largest modal displacement was observed in Damage 5, when the pier is completely suspended. In addition, during Damage 5, node that has the highest modal displacement in the torsion mode shifts toward damaged pier as suggested by FEM.

and second mode, respectively. And for Damage 5 case, the value becomes 0.47 and 0.93, respectively. A comparison of the change in modal displacement due to damage and the 95% confidence bounds is presented in Figure 8. For mode 1, the changes in mean values are 0.03, 0.26 and 0.07 for Damage 3 and 4, Damage 5, and retrofitted stage respectively. These changes are significantly larger than the 95% confidence bounds. For mode 2, the changes in mean values are even larger: 0.2, 0.5 and 0.29 for Damage 3 and 4, Damage 5 and retrofitted stage respectively. These changes are also significantly larger than the 95% confidence bounds indicating that in both modes, the changes of modal displacement on the pier-girder connection are statistically significant and can be used as damage indicator.

Figure 8. Quantification of modal displacement change caused by pier settlement

5 DAMAGE DETECTION USING MULTIVARIATE OUTLIER ANALYSIS The use of multivariate outlier analysis in detecting damage has gained an increased attention recently. The main idea of this approach is to combine the use of damage-sensitive features and statistical novelty detection such as outlier detection. Outlier is an element of data set that appears inconsistent with the rest of data and thus perceived to be governed by other mechanisms. The advantage of this approach over modal-based damage detection is that it requires only data from undamaged structure. Based on this data, the statistical pattern of undamaged structure is formulated and utilized as a baseline for evaluation of future data. When the statistical property of future data is inconsistent with that of baseline, they will be considered as possible outliers. In this study we investigate the feasibility of auto-spectra of acceleration as damage sensitive feature. The auto-spectra contain at least two basic information of structure that is sensitive to damage: natural frequency indicated by spectra peaks, and damping indicated by the sharpness or the width of

Figure 7.a. Effect of damage on (a) 1st Bending mode, (b) 1st Torsional Mode

To quantify the significance of the change in modal displacement, the 95% confidence bounds of identified mode shapes were computed for all damage stages. For comparison, the modal displacements at pier-girder connection in the first and second mode generated by FEM were also computed for undamaged and damaged stages. For undamaged stage, their values are 0.017 and 0.03 for the first

1668

are beneath the frequency peaks. It is a simple feature that can be derived directly from ambient measurement and it captures changes in natural frequency and damping simultaneously. To implement the method, a number of data point (m) within the frequency range of interest is selected from spectra plot and treated as a multivariate feature vector. The procedure is repeated for several data sets under the same undamaged condition. In order to provide a clear distinction between inliers and outliers there needs to be some threshold value. In this study we follow closely the method by (Worden et.al. 2000), in which the threshold value is set by Monte Carlo simulation using the following steps: 1. Create a bank of data for undamaged condtion. For this purpose, 100 equally-sampled data points (m = 100) are selected as feature vector from auto-spectra of reference A and B during undamaged stage. Data in the feature vector are within the frequency range of 3-7 Hz. Since there are only 10 data frame for undamaged condition, only 20 number of observation (n) are available. This was considered insufficient. To provide appropriate mean and covariance matrix for undamaged condition larger number of observation is required. For this purpose, the feature vectors were randomly copied 500 times, and for each copy, Gaussian random vectors with the rms of 1% of the maximum value were added to simulate noise. 2. Compute the largest Mahalanobis squared distances. The largest values of Mahalanobis distance were calculated exclusively for each observation. This procedure was repeated for large number of trials (e.g. 10.000 trials in this study) and all the largest values are ordered. The threshold is then defined as 99.99 percentile value of the largest Mahalanobis squared distance in all trials. After defining the threshold, Mahalanobis distances for all feature vectors in each damage stage were computed and their status as outliers or inliers were conformed. Figure 9 shows the values of Mahalanobis distance for four damage cases. The figure consists of two parts, one is training set, which represent undamaged stage, and the other is the testing sets, obtained from auto-spectra of reference sensors in damage stages. In the training set, the mean vector and covariance matrix are estimated by following the procedure in Section 4.6 for 1000 trial. One can see from this figure that all points in the undamaged stage fall well below the threshold line. There is no false negatives observed, indicating that the threshold line clearly separates the condition of damage and undamaged. For Damage 1and 2, and Damage 3 most of the points are detected as outliers. False positives, however, are obtained from two points in each damage stages. These two points are from data sets with the frequencies of 3.8 and 3.9 for mode 1, and frequen-

cies of 6.1 and 5.9 for mode 2, respectively. The two false positives data sets that are from frame 11 and 12 that correspond to the time when damage is still in progress. Therefore, it is understandable that the damage has not changed the structure significantly. When damage has significantly changed the characteristics of structure such as the case of Damage 4 and Damage 5, all points are unambiguously detected as outliers. Note that the distance between undamaged points and that of damage points of Damage 4 and Damage 5 increase as damage increases. This is rather expected result since the larger the damage is the more auto-spectra deviate from the undamaged pattern. In these two cases, the distance between outliers and threshold line can be used as indicator of damage severity.

Figure 9. Mahalanobis distance plot for each damage stage

6 CONCLUSIONS The study has described the process of vibration measurement and presented the results of vibration analysis. Important results of the study are summarized as follow: 1. A non-uniform pier settlement simulated as damage in this study, affects global stiffness of structure significantly. This is evident by the significant change in frequency of loworder modes. The effects are more obvious in torsional modes than in bending modes, as indicated by larger changes in frequencies of torsional modes than that of bending modes. This finding can be used as an indicator of the presence of a non-uniform pier settlement. 2. Damage in the form of pier settlement also alters the mode shapes locally. Modal displacements at the pier-girder node for damage cases increase significantly suggesting immediate effect of constraint-losing at the boundary condition. The changes are evident

1669

from bending and torsional low-order modes and are well predicted by FEM. Effects of damage on mode shapes are more obvious in torsional modes than in bending modes as indicated by larger changes in modal displacement of pier-girder node of torsional modes than that of bending modes. 3. In general damping increases as the damage level increases. Estimations from ERA indicate that damping in damage stages increase up to 2.5-3% from previously 1.5% in undamaged stage. 4. Feasibility of multivariate outlier detection using auto-spectra as damage features has been investigated in this study. The results show that the use of Mahalanobis distance can detect the presence of damage at the earliest stage (i.e. Damage 1). When damage has significantly changed the characteristics of structure such as the case of Damage 4 and Damage 5, all points are unambiguously detected as outliers indicating the clear presence of damage. The distance between threshold line and damage points in outlier detection are increasing as damage becomes larger. This distance can be used further as indicator of damage severity. 7 ACKNOWLEDGEMENT The authors wish to express their sincere gratitude to Dr. Helmut Wenzel, Robert Veit-Egerer, Monika Widmann from Vienna Consulting Engineers (VCE) for this precious test opportunity, and acknowledge their fruitful discussions and assistance during the authors technical visits. REFERENCES:
Efron B., and Tibshirani R.J. 1993. An Introduction to the Bootstrap, volume 57 of Monographs on Statistics and Applied Probability, New York: Chapmann & Hall. James G III, Carne T.G.,, Lauffer J.P. 1993. The Natural Excitation Technique (NExT) for Modal Parameter Extraction from Operation Wind Turbines, Sandia Report SAND921666.UC-261. Juang JN, Pappa RS,1985. An Eigensystem Realization Algorithm For Modal Parameter Identification And Model Reduction, Journal of Guidance, Control, and Dynamics, Vol. 8(5): 620-627 VCE 2009. Progressive damage test S101 Flyover Reibesdorf, Vienna Consulting Engineers (VCE) Internal Report, Report Nr.08/2308, Vienna Austria, May 2009 Worden K., Manson G., and Fieller N.R.J. 2000. Damage Detection Using Outlier Analysis, Journal Sound and Vibration. Vol. 229 (3):647-667

1670