Anda di halaman 1dari 11

Engineering Structures 30 (2008) 30343044

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Dynamic monitoring of a long span arch bridge


Filipe Magalhes , lvaro Cunha, Elsa Caetano
Faculty of Engineering of the University of Porto (FEUP), R. Dr. Roberto Frias, 4200-465 Porto, Portugal

article

info

a b s t r a c t
A new multi-channel dynamic monitoring system was recently installed in a long span concrete arch bridge that crosses the Douro River in the city of Porto, Portugal: the Infante D. Henrique bridge. This paper describes the experimental and numerical studies developed shortly after construction of the bridge, characterizes the installed monitoring system and presents the results achieved with the software developed to process the data that is continuously received through the Internet. Preliminary studies included the development of an ambient vibration test and the construction of a numerical model of the bridge that was tuned to fit the bridge dynamic properties identified by the ambient vibration test. The routines implemented include the on-line automatic identification of the bridges natural frequencies with the Frequency Domain Decomposition method, enabling the tracking of the bridges first 12 natural frequencies. This unique feature is only possible due to the combination of high-quality acquisition equipment with state of the art processing algorithms. 2008 Elsevier Ltd. All rights reserved.

Article history: Received 29 January 2008 Received in revised form 15 April 2008 Accepted 16 April 2008 Available online 22 May 2008 Keywords: Ambient vibration test Dynamic monitoring system Automatic modal parameters identification Frequency domain decomposition

1. Introduction An increasing interest in permanent observation of the dynamic behaviour of bridges has been observed during the last years, not only owing to the ageing of a huge number of structures, but also because of the increasing complexity of new bridges. In addition, recent technological advances have contributed to make the installation and operation of permanent dynamic monitoring systems more practical and economical and to permit an almost immediate analysis of the bridges condition. System remote control and real-time data retrieving are allowed by nowadaysstandard technology. According to Ko and Ni [1], all over the world, there are so far about 40 long-span bridges (with spans longer than 100 m) instrumented with health monitoring systems. In order to profit from the latest technological developments and to have a system that is really useful for evaluation of the structures condition, a continuous online processing of the collected data is required. The outputs produced should then serve as indicators of the structures health. Modal parameters and especially natural frequencies can be used for that purpose, as demonstrated in [2]. It is however of utmost importance that subsequent to the identification of natural frequencies, numerical models are applied in order to extract the effect on natural frequencies of environmental variables (e.g. air temperature and humidity) and, if significant, the effect of the extra mass associated

Corresponding author. Tel.: +351 225081854; fax: +351 225081835.


E-mail address: filipema@fe.up.pt (F. Magalhes). URL: http://www.fe.up.pt/vibest (F. Magalhes). 0141-0296/$ see front matter 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.engstruct.2008.04.020

with the traffic over the bridge. After elimination of the influence of these factors, frequency changes can only be due to stiffness reductions associated with damage. In this context, the development and validation of tools for automatic identification of natural frequencies based on the measurement of bridge responses during its normal operation is fundamental, as the success of subsequent damage detection algorithms depends on the accuracy of these natural frequency estimates. Furthermore, it is essential that these routines are sufficiently robust to run on an online basis, in order to provide in almost real-time parameters that characterize the structures condition. This feature is especially helpful for rapid assessment of critical infrastructures after the occurrence of natural or manmade disasters. At present, it is opportune to perform practical applications on full-scale bridges with advanced commercially available dynamic monitoring hardware combined with processing routines based on the latest theoretical developments. Under these circumstances, a multi-channel dynamic monitoring system was recently installed in a long span concrete arch bridge to continuously evaluate the variation of its dynamic modal parameters. The final goal of the application is to show the feasibility of the use of damage detection methodologies based on modal parameters shifts. This paper describes experimental and numerical studies developed before equipment installation, characterizes the monitoring system used and presents the results achieved with Matlab routines developed to process the data received through the Internet. Preliminary studies included the development of an ambient vibration test that provided important information for the design

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

3035

Fig. 1. Aerial view of Maria Pia, Infante D. Henrique and Luiz I bridges (from bottom to top).

of the monitoring system and the construction of a numerical model of the bridge that was tuned to fit the bridge dynamic properties identified by the ambient vibration test. This numerical approach was essential to improve the understanding of the structure dynamic behaviour and to create a baseline model for future damage detection studies. The routines implemented perform on-line automatic identification of the bridges natural frequencies based on the Frequency Domain Decomposition method, enabling tracking of the bridges first 12 natural frequencies. This unique feature is possible because high-quality acquisition equipment was combined with state of the art processing algorithms. 2. Description of the bridge The Douro River is crossed by several outstanding bridges linking the cities of Porto and Gaia, located at the north of Portugal. The 19th Century metallic Maria Pia and Luiz I Bridges and the 20th

Century concrete Arrbida and S. Joo Bridges got, at the beginning of the 21st Century, the company of a new concrete arch bridge: the Infante D. Henrique bridge (Fig. 1). The Infante D. Henrique Bridge is composed of two mutually interacting fundamental elements: a very rigid prestressed reinforced concrete box beam, 4.50 m deep, supported by an extremely shallow and thin reinforced concrete arch, 1.50 m thick, as shown in the elevation and cross-sections represented in Fig. 2. The arch spans 280 m between abutments and rises 25 m until the crown, thus exhibiting a shallowness ratio greater than 11/1. In the 70 m central segment, arch and deck meet to define a box-beam 6 m deep. The arch has constant thickness and its width increases linearly from 10 m in the central span up to 20 m at the springs [3]. Owing to the high stiffness of the deck with respect to the slenderness of the arch, the structure behaves as a beam bridge defined between abutments and with intermediate elastic supports 35 m apart. The extreme shallowness and flexibility of the arch implied significant complexity of construction and required very accurate control of geometry, deformations and forces. This control was performed by three separate instrumentation systems: one for concrete elements (comprising strain gages, clinometers and temperature sensors); another for temporary stay cables; and another one for granite rocky slopes. The Laboratory of Vibrations and Structural Monitoring (ViBest, http://www.fe.up.pt/vibest) of FEUP was contracted to measure on a regular basis the tension forces installed in the provisional stay cables using vibration chord theory [4]. Those three instrumentation systems provided an accurate understanding of the structural behaviour, and supported important decisions of the designer and the constructor at several critical moments of construction. Owing to the evident usefulness of the monitoring activity during this phase, the owner decided to keep the static monitoring system working after the bridge opening to traffic, and agreed to install a new long term dynamic monitoring system, suggested by ViBest with the support of the designer. It is relevant to point out that the bridge is one of the main entrances into the Porto city centre and so it is crossed by high volumes of traffic. During rush hours, the deck is frequently congested with cars and buses.

Fig. 2. Elevation and cross-sections of the bridge.

3036

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

Fig. 3. Instrumented cross-sections (the ellipse indicates the reference cross-section) and image of one tri-axial strong motion recorder with the external GPS sensor.

3. Ambient vibration test and numerical modelling 3.1. Ambient vibration test procedure The design of a dynamic monitoring system requires previous knowledge of the structures natural frequencies and mode shape configurations. The most convenient way to estimate the modal parameters experimentally is with an ambient vibration test. Furthermore, the data collected during these tests is of the same type of the data recorded by the dynamic monitoring systems, so the same identification algorithms can be used. The ambient vibration test of the Infante D. Henrique bridge was developed without disturbing its normal use, taking profit from the vibrations induced by traffic and wind. To measure very low amplitude accelerations, 4 tri-axial 18-bit strong motion recorders were used (Fig. 3, http://www.geosig.com). These devices are constituted by high sensitivity internal force balance accelerometers, 18 bit analog-to-digital converters, batteries that enable autonomy for one day of tests, internal memories to store the acquired data and external GPS sensors to ensure accurate time, so that synchronous measurements are recorded by the various units independently, avoiding the use of cables. During the ambient vibration test, two recorders served as reference permanently located at the reference cross-section (section 8 in Fig. 3) of the deck, at both sides of the deck (upstream and downstream). The other two recorders scanned the bridge deck measuring the acceleration along three orthogonal directions at both sides of the other 15 cross-sections represented in Fig. 3. For each sensor layout, time series of 16 min were collected. The sampling frequency was 100 Hz, a value that is imposed by filters of the acquisition equipment [5] and which is much higher than that required for this bridge, as the most relevant natural frequencies of the bridge are below 10 Hz. Therefore, a decimation (reduction of the sampling rate after the application of a digital low-pass Chebyshev filter with a cutting frequency equal to 0.4 times the new sampling frequency) was applied before the use of the identification tools, reducing the sampling frequency from 100 Hz to 20 Hz. 3.2. Identification of modal parameters Identification of modal parameters from data collected during the ambient vibration test was achieved with two separate outputonly identification methods with different theoretical bases: the Frequency Domain Decomposition (developed in the frequency domain) and the Data-driven Stochastic Subspace Identification (developed in the time domain). A detailed description of the results provided by both methods is presented in [6], where it is also shown that estimates provided by the two approaches are very consistent. In this paper, only the results of the FDD method are presented, as an adaptation of this method is going to be used in processing of data collected by the monitoring system.

Fig. 4. Average normalized singular values.

The first step of the FDD method is to construct a spectrum matrix of ambient responses for each test setup, with one row for each measurement degree of freedom and with one column for each degree of freedom elected as reference. Therefore, columns contain cross spectra relating the structural response at all measurement points with the corresponding response at each reference point. In the present application, four time series were considered for each instrumented section: upstream and downstream vertical accelerations, mean values of the two measured lateral accelerations and mean values of the two measured longitudinal accelerations. Hence, spectra matrices with 8 rows and 4 columns were organized for each setup. The elements of these matrices were estimated using the Welch procedure [7] by dividing the available time series in segments of 102.4 s (2048 points), considering an overlap of 66% between segments and adopting a Hanning window to reduce the leakage. The selected parameters allowed the realization of averages over 26 time segments and produced spectra with a frequency resolution of 0.00977 Hz. It can be shown [8] that, under some assumptions (white noise excitation, low damping and orthogonal mode shapes for close modes), the singular values of the spectrum matrix in the vicinity of each natural frequency form auto-spectrum density functions of single degree of freedom systems with the same frequency and damping as the structure vibration modes. Fig. 4 presents the average of normalized singular values covering all setups. This graphic is a synthesis of the frequency content present in each setup and allowed the identification of thirteen resonant frequencies (marked with the dashed vertical lines) in the frequency range of analysis (05 Hz). Mode shapes are estimated from singular vectors of spectrum matrices evaluated at the identified natural frequencies and associated with singular values that contain the peaks. In each setup, eight mode shape components are calculated (number of time series processed in each setup). These are then grouped together using the components at reference sections (estimated in all setups) to correctly scale the segments identified in each setup. Figs. 57 show some of the identified modes of vibration of the bridge deck plotted with Artemis software [9]. These are compared with the ones provided by the numerical model that is described in the next sub-section.

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

3037

Fig. 5. Top view of the first, second and third lateral bending modes.

Fig. 6. Lateral view of the first 5 vertical bending modes.

3.3. Numerical modelling Structural behaviour of the bridge was modelled with ANSYS software using 3D bar finite elements. Cross section properties (area, moments of inertia, torsion moment of inertia and shear deflection constants) were defined according to the geometry of the deck, arch and columns. The same elasticity modulus of 37 GPa (value provided by tests performed during the bridge construction) was adopted for the deck and arch concrete and an elasticity modulus of 34 GPa (value defined by the Eurocodes for a C35/45 concrete) was considered for the concrete of the columns. Connections between deck and the highest columns (M1 and M6, Fig. 2) are monolithic, whereas connections with other columns and abutments are provided by two unidirectional sliding pot bearings. These bearings allow movements along the longitudinal direction of the bridge and rotation in all directions.

However, for low levels of excitation, as is the case during ambient vibration tests, the behaviour of these connections can be different. For low levels of vibration, friction forces can prevent displacement or rotation. To analyse the influence of the behaviour of these connections on the modal parameters, three alternative models were developed: - Model 1 (M1): longitudinal displacement and rotation free in all pot bearings; - Model 2 (M2): longitudinal displacement and rotation fixed in all pot bearings; - Model 3 (M3): longitudinal displacement and rotation fixed in the pot bearings of the columns but free in the pot bearings of the abutments.

3038

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

Fig. 7. Lateral view of the first and second torsion modes. Table 1 Modal parameters obtained with the numerical models that were developed Modea Exp. Freq. (Hz) V1 V2 V3 V4 L1 L2 L3 T1 T2 0.810 1.135 1.405 1.993 0.770 1.734 3.309 2.212 3.734 Model 1 Freq. (Hz) 0.541 1.081 1.244 1.766 0.794 1.767 3.353 2.169 3.632 Error (%) Model 2 Freq. (Hz) 1.054 1.164 1.473 2.088 0.799 1.788 3.394 2.211 3.724 Error (%) 29.96 2.56 4.84 4.82 3.77 3.11 2.54 0.00 Model 3 Freq. (Hz) 0.701 1.144 1.465 2.086 0.794 1.768 3.357 2.185 3.641 Error (%) Model 4 Freq. (Hz) 0.810 1.149 1.466 2.086 0.794 1.768 3.357 2.185 3.641 Error (%) MAC 0.995 0.994 0.992 0.994 0.996 0.989 0.965 0.956 0.952

33.29 4.76 11.46 11.35


3.12 1.90 1.30

13.56
0.79 4.27 4.72 3.12 1.96 1.42

0.12
1.23 4.34 4.72 3.12 1.96 1.42

1.90 2.73

0.27

1.18 2.49

1.18 2.49

a Mode type, defined considering the most relevant components (L lateral bending; V vertical bending; T torsion).

Fig. 8. Normalized auto-spectra of longitudinal accelerations measured simultaneously at the deck and at the Gaia abutment.

existence of friction forces. Therefore, the appropriate modelling of the behaviour of the bridge requires the inclusion in Model 3 of horizontal springs to simulate the additional stiffness provided by the abutments. This last conclusion led to the development of a more correct model of the bridge (Model 4). This model is similar to Model 3 but includes a horizontal spring at each abutment with a stiffness constant that was adjusted in order to obtain good matching between numerical and experimental frequencies. Table 1 shows that the correlation between modal parameters of the final numerical model and the experimental ones is very good, with relative errors of natural frequencies lower than 5% and MAC values always greater than 0.95 [10] (ratio that measures the correlation between mode shapes:1 means that the mode shapes only differ on a scale factor; 0 means that the modes shapes are orthogonal). 4. The monitoring system Traditional monitoring systems are based on one central acquisition system to which all the sensors are connected. However, it is more convenient to have the digitizers distributed along the monitored structure, in order to reduce the length of the sensor cables, as these are sensitive to electrical interferences that can corrupt the sensors electrical signals. Additionally, after digitization, the information collected by several sensors can be transmitted by a single Ethernet cable. Therefore, with this alternative solution, the installation becomes simpler and the signal noise is reduced. These advantages are enhanced when large civil structures are involved. Whenever the distance between digitizers is longer than 90 m, it is necessary to replace the Ethernet cable by a fibre optical cable, that can span long distances without signal interference, making this arrangement even more flexible. As high signal synchronization is essential for modal analysis, this type of architecture of the monitoring system requires good synchronization of the digitizers clocks. This can be achieved with a synchronization cable connecting the distributed digitizers or with one GPS antenna and receiver for each digitizer to

Natural frequencies of the most relevant modes provided by the three models are presented in Table 1. Models 1 and 2 establish the lower and upper bounds of the numerical natural frequencies calculated using the material properties previously defined. A significant variation of the natural frequencies of vertical modes is noted (especially the associated with the first mode). Modes identified experimentally show that the deck undergoes longitudinal movements. Therefore, the hypothesis of fixing longitudinal movements in the abutments is not correct. On the other hand, model M3 shows that, even fixing the relative movements and rotation in all columns, the numerical frequency of the first mode is considerably lower than the experimental one. In order to better understand the longitudinal behaviour of the bridge, two additional measurements were performed: longitudinal acceleration at the deck near the expansion joint in the Gaia bank and longitudinal acceleration of the Gaia abutment. Spectra of the collected time series are represented in Fig. 8. These graphics show the presence of some of the natural frequencies of the bridge in the response of the abutments and, in particular, the presence of the natural frequency of the first vertical mode. This means that the abutment is mobilized in the movements associated with the first vertical bending mode, due to the

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

3039

Fig. 9. Scheme of the monitoring system

synchronize their internal clocks using the information provided by the in view satellites. The dynamic monitoring system of the Infante D. Henrique bridge is essentially composed by 12 force balance accelerometers, 2 digitizers and recoding units and an internet router, which are installed inside the deck box girder and distributed along the bridge according to the scheme presented in Fig. 9. Since the structure is almost symmetric and the previously performed ambient vibration test has proven that the mode shapes are also approximately symmetric, it was decided to instrument just one half of the bridge. Therefore, the 12 available accelerometers were distributed along four sections. Three sensors equip each section: one to measure the lateral acceleration and two for the vertical acceleration at the downstream and upstream sides (the ambient test showed the existence of torsion modes in the analysed frequency range). The force balance accelerometers used (FBA ES-U2 from Kinemetrics) have a dynamic range of 145 dB, are sensitive in the frequency range DC to 200 Hz and their measuring range can go up to 4 g. In the present application a measuring range between 0.25 and +0.25 g was fixed, in order to optimize the sensitivity of the sensors and so reduce the effect of noise, while keeping a conservative acceleration range (the maximum observed acceleration is lower than 10 mg). Each digitizer (www.Q330.com) allows the connection of six dynamic channels, is equipped with a 24-bit analog-to-digital converter and permits simultaneous telemetry of the acquired data to a central site and a link to a local recording unit. In the present installation, the digitizers were placed at sections S2 and S4 (Fig. 9) together with local backup disks. These recording units guarantee that there are no data losses in case of failure of the Internet connection. Digitizers located at sections S2 and S4 are connected to each other by an Ethernet cable. Another Ethernet cable links section S4 and the router that is the interface between the local network and the Internet.

Synchronization between digitizers is achieved with two GPS antennas and receivers that allow continuous update of the internal clocks of both units. The data produced by the two digitizers is received at FEUP (Faculty of Engineering of the University of Porto), where data integrator software generates ASCII files with as many columns as the number of sensors, corresponding to the acceleration time series sampled with a predefined rate and length. For monitoring this bridge, a sampling frequency of 50 Hz and a length of 30 min were selected. Consequently, each half an hour a new ASCII file with 12 columns and 30 60 50 = 90 000 lines is created at a PC located at FEUP. These files are then processed by the software described in the next section. In the selection of the time series length, priority was given to the quality of modal parameters estimates in order to obtain a system that is able to detect small stiffness changes. If the purpose was to have a rapid alarm at the occurrence of an anomaly, the length of the files could be reduced, but the accuracy of the estimates would decrease. A good compromise could be then to present less accurate natural frequency estimates using shorter time periods, to rapidly detect important and sudden anomalies, and obtain more accurate estimates at the end of each 30 min. Finally, it is important to state that the configuration of the system (e.g. sampling frequency and length of the time segments) and the analysis of parameters used to check the system condition (e.g. quality of the GPS signals, digitizer internal temperature and input voltage) are performed remotely. This dynamic monitoring system is complemented by an independent static monitoring system (performing one or two acquisitions per hour) that was installed in the bridge during construction [11], and comprises strain gages, clinometers and temperature sensors. Measurements of the temperature sensors embedded in the concrete are essential for the development of numerical models that extract the effect of temperature from identified natural frequencies. Furthermore, a weather

3040

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

- pre-processing of the data to eliminate the offset and to reduce sampling frequency from 50 to 12.5 Hz (the first 12 modes are below 5 Hz); - processing of the data, with automatic identification of modal parameters; - creation of a database with processing results; - display of plots with the most relevant results. Analysis of modal parameter changes with environmental conditions (e.g. temperature and humidity) has not yet been done, because that study requires several months of measurements and the system has been working for just two months. The definition of thresholds for damage detection is dependent on the previous analysis. Routines already developed for the automatic identification of modal parameters are described in Section 5.1. Fig. 11 presents the plots created by the software that has been developed for two groups of pre-processed time series associated with two distinct periods: during the night under little traffic and during a rush hour. 5.1. Automatic identification of the modal parameters Automatic identification of natural frequencies is one of the most important features of a continuous dynamic monitoring system, since the success of damage detection algorithms based on these dynamic parameters is strongly dependent on the accuracy of the estimates. Consequently, much research is being developed on this subject following different approaches. Lau et al. [12] present a literature review on the last developments of the methods that fit a numerical model to data. Brincker et al. [13] present an alternative non-parametric method in the frequency domain that is an automatic implementation of the Frequency Domain Decomposition method (FDD the method applied to the data collected during the ambient vibration test). As a first approach to interpret data collected by the monitoring system installed, the FDD method was used. Therefore, this method was implemented in Matlab to automatically process the time series that are created each half an hour. In this paper, the results of the first two months of the system operation (from 2007/10/17 to 2007/12/17) are presented. As already mentioned, the first step of the FDD method consists of the calculation of a spectrum matrix for each data set. In the present application, all the sensors were elected as references and so, this is a 12 by 12 matrix (the number of acceleration channels is equal to 12). The auto and cross spectra of the matrix were calculated with a resolution of 0.0061 Hz, using segments from the time series with a total length of 30 min and adopting a Hanning window and an overlap of 50%. Then, singular value decomposition of this matrix was performed, producing, for each frequency, 12 singular values and vectors. The graphic of the first singular values as a function of the frequency allows visual identification of the natural frequencies, which are associated with the peaks. Fig. 12 presents a colour map that consists of a top view (the colour is a function of the

Fig. 10. Dataflow inside the monitoring system.

station exists close to the bridge recording all the important environmental variables (air temperature, humidity and wind velocity and direction) whose measures can also be used to investigate the possible effect of these variables on the identified modal parameters. 5. Routines for data processing In the context of a monitoring program, it is very important to have good tools for continuous processing of data arriving, in order to permanently extract parameters with physical meaning that can then be used to evaluate the structure health. Fig. 10 presents the dataflow inside the monitoring system, highlighting the Matlab routines that are being developed in order to execute several tasks every 30 min (after the creation of a new 30 min duration ASCII file). Routines already implemented include online execution of following tasks: - creation of a database with the original data (sampled at 50 Hz) that can be used later to test alternative processing methodologies;

Fig. 11. Time series after pre-processing, collected during the night and during a rush hour.

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

3041

- repetition of these steps in the reduced search domain, until a specified number of identified natural frequencies has been achieved, the search set is empty or the maximum spectrum amplitude is lower than a eventually predefined noise level. One of the advantages of this method is the relatively small number of user-defined parameters. These include frequency resolution, the frequency interval under analysis, the MAC value limit of the points in the same modal domain and the minimum number of points of the modal domains. The number of expected natural frequencies and the noise level are optional parameters that can be used to shorten the number of searches for maxima. In the present application, the analysis was performed over the frequency range from 0.5 to 4.5 Hz, with a frequency resolution of 0.0061 Hz (0.8% of the lowest natural frequency). A number of 12 natural frequencies was specified for automatic identification, a limit value of 0.4 was selected for the MAC and only modal domains with more than 7 points were considered. The specification of frequency resolution is critical. If the value is too small (the segments used are longer, the number of averages is lower), the amount of noise is higher, and so automatic identification becomes more difficult. On the other hand, higher values limit the accuracy on the identification of natural frequencies. Experience gained in the analysis of data from this bridge showed that it is better to select low values for the MAC (0.4). In this way, the number of points in the search domain is reduced more rapidly and the number of false identifications is minimized. However, it should be pointed out that this strategy is not adequate if the number of sensors is small (the ability of the MAC to distinguish mode shapes increases with the number of sensors) and similar modes shapes for adjacent natural frequencies are expected. The minimum number of points in the modal domain is obviously related to the MAC limit. To sum up, it is important to stress that the correct selection of all the parameters for automatic identification of modal parameters should result from sensitivity tests performed during an evaluation period. Fig. 13 illustrates the output of the implemented automatic identification algorithm applied to one of the time series sets collected during the observation period. The graphic shows the first two singular values and the modal domain associated to each peak. The colour of the modal domains is dependent on the order of the natural frequencies identification and the corresponding ordinates are the MAC values. It can be noticed that in the vicinity the 2.2 Hz natural frequency, there are small peaks that were included in the respective modal domain and not identified as

Fig. 12. Colour map with the variation of the signals frequency content over time (from 2007/10/17 to 2007/12/17).

amplitude) of all the first singular values spectra associated to all 30 min intervals of the observation period (from 2007/10/17 to 2007/12/17, a total of 48 61 = 2928 setups). Thirteen vertical alignments are evident in this figure. They characterize the evolution of the bridge natural frequencies during the period under study. It is also shown that a single picture can be used to provide an overview of the structure dynamic behaviour during a long observation period. The spectra of the first singular values are the starting point for the automatic identification of the natural frequencies. The methodology adopted was based on the one described in [13] and consists on the repeated application of the following steps: - identification of the maximum of the spectra in the search domain; at the beginning this is equal to a user predefined frequency range; - selection of the points around the identified peak that lead to similar mode shapes (identified from the first singular vectors), using the MAC coefficient [10] as similitude measure; - if the number of selected of points is larger than a predefined value and if there are points in both sides of the peak, the identified peak is elected as a natural frequency and the selected group of points is called the modal domain; otherwise, the peak is assumed to be related to data noise (this simple criterion for the selection of peaks with physical meaning worked very well in this application, but for the processing of lower quality data a more robust criterion may be required, implying analysis of the bell shape); - redefinition of the frequency search domain by removal of the previously selected group of points;

Fig. 13. Automatic identification of natural frequencies in the setup: 2007/11/18 12:00.

3042

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

Fig. 14. Identified natural frequencies of the first 12 modes during the period from 2007/10/17 to 2007/12/17.

Fig. 15. Two zooms of Fig. 14 to show the successful identification of closely spaced natural frequencies and variation of the 11th mode natural frequency with daily temperature oscillations.

new natural frequencies, as it would have happened if a simple algorithm of peaks selection were used. Automatic identification of the natural frequencies of all the 2928 collected setups allowed the construction of the graphic presented in Fig. 14, which contains the estimates provided by the independent analysis of all the datasets. This shows the variation of the first natural frequencies of the bridge during the observation period (some blank spaces exist due to maintenance tasks that were needed during the first months of the systems operation). It is clear that the algorithm for automatic modal identification provided very good results, as the number of points associated with wrong estimates (the ones that are clearly apart from the horizontal alignments) is quite small. Moreover, it is possible to eliminate the wrong estimates, considering these points as outliers. In the present application, the points outside the interval defined by the average values +/1.5 times the standard deviation were classified as false estimates (outliers). Using this statistic classification, the percentage of false natural frequencies estimates is equal to 1.9%. Fig. 15 presents two zooms of the graphic presented in Fig. 14 after elimination of the outliers. It is relevant to observe that the methodology was able to identify the first two closely spaced natural frequencies and the very small changes of the natural frequencies of the 11th mode with the oscillation of ambient temperature. During the monitoring period associated with the presented zoom, the daily average temperature was about 16 C and the amplitude of daily temperature variation was around 10 C. The 11th mode is one of the modes that exhibit higher frequency dependency with temperature and, in this case, daily variations of about 1% (aprox. 0.04 Hz) were observed. Table 2 presents the mean values and standard deviations of natural frequencies estimates provided by the monitoring system and compares these values with the ones previously obtained with the ambient vibration test. Mean values associated with the ambient vibration test are consistently lower. This difference can be explained by the temperature effect, because the

Table 2 Natural frequencies identified by the Ambient Vibration Test and by the Monitoring System Modea Ambient Vibration Test Frequency (Hz) L1 V1 V2 V3 L2 V4 T1 V5 L3 V6 T2 V7 0.770 0.810 1.135 1.405 1.734 1.993 2.212 3.013 3.309 3.490 3.734 4.339 Std. F. (Hz) 0.0011 0.0029 0.0014 0.0013 0.0012 0.0029 0.0025 0.0051 0.0050 0.0052 0.0049 0.0051 Monitoring System Frequency (Hz) 0.780 0.825 1.147 1.419 1.755 2.017 2.230 3.049 3.345 3.525 3.769 4.391 Std. F. (Hz) 0.0027 0.0057 0.0020 0.0033 0.0039 0.0051 0.0046 0.0074 0.0069 0.0069 0.0074 0.0108

a Mode type, defined considering the most relevant components (L lateral bending; V vertical bending; T torsion).

ambient vibration test was performed during the summer (natural frequencies decrease with temperature increase), and also by hardening of the concrete during the last 2 years (last pouringJune 2002; ambient vibration test June 2005; monitoring - October December 2007). On the other hand, the standard deviations of the estimates provided by the monitoring system for the natural frequencies are higher than the ones resulting from the ambient vibration test, owing to the effect of temperature, which is obviously more significant during a long observation period. The mode shapes associated with natural frequencies identified by the monitoring system are estimated by the first singular vectors associated with the selected peaks. The average modal ordinates of the first 12 modes estimated from all the setups collected during 2007/12/17 are represented in the complex domain at Fig. 16 (the vertical and lateral modal ordinates are represented in different colours in order to allow the distinction between the different types of modes). All the modal ordinates are over the horizontal axis, meaning that all the mode shapes are real.

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

3043

Fig. 16. First 12 mode shapes identified with the continuous monitoring system using the time series collected during 2007/12/17, and means and minima of the MAC ratios between the plotted modes and all the modes identified in the period between 2007/10/17 and 2007/12/17. Lateral modal ordinates represented in red; vertical modal ordinates represented in blue.(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 17. Comparison between the mode shapes identified with the ambient vibration test and with the continuous monitoring system using the time series collected during 2007/12/17.

These results prove the quality of the identified modal parameters and show good synchronization of the digitizers provided by the GPS sensors (an imperfect synchronization would lead to phase angles different from 0 or 180 between the degrees of freedom measured by different digitizers). The mode shapes plotted were compared with the mode shapes identified from all the data sets collected during the monitoring period under analysis, using the MAC ratio. Fig. 16 also presents the average MAC values and the minima MAC values that resulted from this comparison. The mean values are always greater than 0.99 and the minima values are always above 0.8, showing that the mode shapes estimated during the first two months of monitoring are very consistent. Direct comparison between the mode shapes provided by the ambient vibration test (represented in Figs. 57) and the ones estimated by the monitoring system is not possible, because the measured degrees of freedom are not exactly the same. However, the overlay of both estimates shows that they are very similar. Fig. 17 presents such comparison for the first four mode shapes.

The main limitations of the described procedure for the automatic identification of modal parameters are the accuracy of the natural frequencies estimates dependency on the frequency resolution and the inadequacy to estimate modal damping ratios. It is possible to obtain estimates for these coefficients, from auto-correlation function resulting from the inverse Fast Fourier Transforms of the points of the selected modal domains, by the fitting of exponential decays to the envelopes of those functions. However, the quality of this fitting relies on the adequacy of the selected interval and this task is not easy to automate. Limited frequency resolution prevents the observation of temperature effects on lower frequencies, as very low variations are expected. 6. Conclusions This paper presents the experimental and numerical work developed before setting up a continuous dynamic monitoring system in a long span concrete arch bridge, characterises the installed monitoring equipment and shows the most relevant results provided by the application of the routines implemented, to the data collected during the first 2 months of monitoring.

3044

F. Magalhes et al. / Engineering Structures 30 (2008) 30343044

A preliminary ambient vibration test was essential for the design of the monitoring system, especially for the selection of measuring points, and for calibration of the developed numerical model. The latter is an important basis to simulate damage scenarios in order to test the feasibility of application of vibration based damage detection techniques. The numerical model was also fundamental to understand the strong effect of the bridge supports on its dynamic properties. The Matlab routines developed to perform on-line processing of the data continuously transmitted through the Internet proved to be very efficient and robust. In fact, they allowed fully automatic identification of the bridges first 12 resonant frequencies with a very small number of false identifications. This unique feature was only possible because high-quality acquisition equipment was combined with state of the art processing algorithms, after an evaluation period necessary to specify the parameters required by the automatic identification procedure. The automatic version of the Frequency Domain Decomposition method showed also ability to identify closely spaced modes (as it is the case of the bridge first two modes) and to detect very small frequency variations (aprox. 0.04 Hz) motivated by the oscillation of air temperature. As a future system enhancement, implementation of an alternative procedure for automatic identification of natural frequencies based on the SSICOV method is now under development. These new routines will certainly allow an even more accurate identification of the natural frequencies and also identification of modal damping ratios. The results presented are related with a relatively short period of observation; the next months of operation will be surely very useful to confirm the efficiency of the routines developed and to create a database for the development of numerical models to extract the influence of environmental variables and of traffic intensity (if relevant) on the identified modal parameters.

Acknowledgments The authors would like to acknowledge: (1) all the supports provided by the Portuguese Foundation for Science and Technology (FCT) to ViBest/FEUP for the development of research in the area of Long-Term Dynamic Monitoring; (2) the Ph.D. Scholarship (SFRH/BD/24423/2005) provided by FCT to the first author; (3) the support provided by the bridge designer, Prof. Ado da Fonseca, and the bridge owner, Metro do Porto. References
[1] Ko JM, Ni YQ. Technology developments in structural health monitoring of large-scale bridges. Eng Struct 2005;27:171525. [2] Peeters B, De roeck G. One-year monitoring of the z24-bridge: Environmental effects versus damage events. Earthquake Eng Struct Dyn 2001;30:14971. [3] Ado Da Fonseca A, Millanes Mato F. The infante henrique bridge over the river douro, in porto, portugal. Struct Eng Internat 2005;15. [4] Ado Da Fonseca A, Bastos R, Cunha , Caetano E. Monitoring of temporary cables in Infante D. Henrique Bridge. Struct Health Monitoring 2002. [5] GeoSIG, GSR - 12/16/18 operation manual; 2000. [6] Magalhes F, Cunha , Caetano E, Ado Da Fonseca A, Bastos R. Evaluation of the dynamic properties of the infante dom henrique Bridge. In: IABMAS; 2006. [7] Welch PD. The use of fast fourier transform for the estimation of power spectra: A method based on time averaging over short modified periodograms. IEEE Trans Audio Electro-acoust 1967;AU-15. [8] Brincker R, Zhang L, Andersen P. Modal identification from ambient responses using frequency domain decomposition. In: IMAC XVIII; 2000. [9] SVS, ARTeMIS Extractor Pro, Release 3.41. Structural vibration solutions, Aalborg, (Denmark); 19992004. [10] Allemang RJ, Brown DL. A correlation coefficient for modal vector analysis. In: IMAC I; 1982. [11] Ado Da Fonseca A, BASTOS RO. Monitorizao em fase de servio do comportamento estrutural da Ponte Infante D. Henrique. Encontro Nacional de Beto Estrutural, Porto, Potugal; 2004 [in Portuguese]. [12] Lau J, Lanslots J, Peeters B, Van der Auweraer H. Automatic modal analysis: Reality or myth? In: IMAC XXV; 2007. [13] Brincker R, Andersen P, Jacobsen N-J. Automated frequency domain decomposition for operational modal analysis. In: IMAC XXV; 2007.

Anda mungkin juga menyukai