You are on page 1of 10

Engineering Structures 27 (2005) 14881497 www.elsevier.

com/locate/engstruct

On the optimal sensor placement techniques for a bridge structure


M. Meo, G. Zumpano
School of Engineering, Craneld University, Bedford, MK43 0AL, UK Received 13 May 2004; received in revised form 8 November 2004; accepted 30 March 2005 Available online 23 May 2005

Abstract Health monitoring systems set up to assure the safe operation of structures require linking sensors with computational tools able to interpret sensor data in terms of structural performance. Although intensive development continues on innovative sensor systems, there is still considerable uncertainty in deciding on the number of sensors required and their location in order to obtain adequate information on structural behaviour. This paper considers the problem of locating sensors on a bridge structure with the aim of maximizing the data information so that structural dynamic behaviour can be fully characterised. Six different optimal sensor placement techniques, three based on the maximisation of the Fisher information matrix (FIM), one on the properties of the covariance matrix coefcients, and two on energetic approaches, have been investigated. Mode shape displacements were taken as the measured data set and two comparison criteria were employed. The rst criterion was based on the mean square error between the FE model and the cubic spline interpolated mode shapes. The second criterion measured the information content of each sensor location to investigate on the strength of the acquired signals and their ability to withstand the noise pollution keeping intact the information relative to the structure properties. The results showed that the effective independence driving-point residue (EFI-DPR) method provides an effective method for optimal sensor placement to identify the vibration characteristics of the studied bridge. The variance method (VM) developed by the authors gave results very close to the EFI-DPR technique, in terms of the capability to capture the vibration mode shape and signal strength. However, the VM presented unique characteristics in the world of the OSP techniques, which is the indication of the optimal number of sensors (ONS). 2005 Elsevier Ltd. All rights reserved.
Keywords: Optimal sensor placement; Fisher information matrix; Covariance matrix

1. Introduction Deciding on an optimal sensor placement (OSP) is a common problem encountered in many engineering applications and is a critical issue in the construction and implementation of an effective structural health monitoring system. An optimal conguration can minimise the number of sensors required, increase accuracy and provide a robust system. OSP is important in cases where the properties of a system, described in terms of continuous functions, need to be identied using discrete sensor information. Numerous techniques have been advanced for solving the optimal
Corresponding author. Tel.: +44 1234 750111x5220; fax: +44 1234 752149. E-mail address: m.meo@craneld.ac.uk (M. Meo).

sensor placement problem and are widely reported in the literature. These have been developed using a number of approaches, some based on intuitive placement or heuristic approaches, others employing systematic optimisation methods. A comprehensive survey of sensor placement strategies for aerospace applications can be found in Padula and Kincaid [1], for the optimal sensor placement problem in the process industry in Naimimohasses et al. [2] and for the safe operation of nuclear reactors in Oh and No [3]. The characterisation of the dynamic behaviour of a real structure is possible only if a minimum amount of information is available. This, in turn, implies that a minimum number of sensors must be placed on the structure under assessment. Thus, the sensors must be judiciously placed in order to provide adequate information for the identication of the structural behaviour.

0141-0296/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.engstruct.2005.03.015

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

1489

Fig. 1. Nottingham suspension bridge.

The set of measured degrees of freedom for large scale civil structures, usually the displacements of low frequency modes, must give sufcient information to describe the dynamic behaviour of a structural system with sufcient accuracy to allow its health state and/or relevant modications to be determined effectively. Hence, the fundamental problem is how many and which degrees of freedom should be taken in the fault/damage identication process. In solving this problem due account has to be taken of economic factors, which may require a limited number of sensors being placed at accessible locations on the real structure. For this reason, it is crucial that those sensors which are employed are located in the most advantageous sites. Otherwise, incomplete modal properties will be measured and an accurate structural health monitoring (SHM) assessment will be impossible. Sensor locations must be individually determined for each structure to be estimated. Normally, the selection is based purely on the engineering judgement. In order to detect structural changes within the bridge, a more reliable method must be developed. The efciency of an SHM relies on the sensitivity of the acquired data to structural changes that may be obtained by an extended sensor network on the structures. The present study investigates techniques for selecting optimal sensor locations in a sensor network designed to monitor the health condition of a bridge by capturing sufcient information to identify structural dynamic behaviour. This work is part of a broader research project [46] aimed at undertaking the research necessary to set up a basic remote health monitoring system, using GPS sensors placed on an operational bridge, linked to new nite element/optimisation based health assessment software. 2. Bridge structure The different optimal sensor placement techniques were tested on a real large scale structure. The Nottingham Wilford suspension bridge (Fig. 1) was chosen as a test bed because it experiences deection in the decimetre range under normal environmental loading. The suspension bridge studied is a footbridge composed of two sets of suspension cables restrained by massive

Fig. 2. Bridge nite element model.

masonry anchorage. The span sidewalk is 3.65 m wide and 68.58 m long, and it is composed of a steel deck covered by a oor of wooden slats. A suspension bridge is a complex structural system in which each member plays a different role. It is inherently a exible structure with some form of stiffness incorporated in the design. The stiffness is usually obtained with a properly designed cable system or with a stiffening truss. Stringers and oor beams transfer the weight of the deck to the suspenders and trusses and then to the suspension cables. Suspension cables then transfer the loads to the anchorage and the towers. Suspension rods pull down on the chain causing it to sag and to pull inwards at its ends. As the main span is loaded, the load causes the cables to pull on the support tower and hence pull the other span up. The tension in the main cables and the consequent movement of the towers gives rise to uplift in the main span. In order to provide input data for the OSP methods a three-dimensional nite element model of the bridge was built. The construction of a nite element model, capable of accurately replicating the behavior of the real world structure, was undertaken using the SAFESA Method [8]. The three-dimensional nite element model developed is shown in Fig. 2. The vibration properties were calculated

1490

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

Fig. 3. GPS positioning.

by performing a modal analysis using the nite element analysis code and pre/post-processor system ANSYS. The reference frame has its X axis across the bridge width, Y axis along the bridge span and Z axis orthogonal to the plane XY . The GPS sensors were used to measure the vibration properties of the bridge. Recent advances in GPS receiver technology and data-processing software have made GPS a much more cost-effective tool to monitor the structural movement of large scale civil structures, such as bridges, tall buildings and offshore platforms. Experimental results highlight GPS as a viable means of measuring thermally induced displacements, shorter-term transient motion and long-term settlements of the foundations of large civil structures with an accuracy of the order of a few millimetres [4,5]. The global positioning system provided by Leica Corporation was capable of acquiring the real time absolute three-dimensional positions at the rate of 10 Hz. GPS dimensions and allocation requirements were such that the only suitable locations were the bridge handrails (Fig. 3). The possibility of disturbances due to the handrails modal characteristics during the acquisition of bridge global mode shapes (less than 10 Hz), was checked out by comparing their respective frequencies. The modal analysis results conrmed that the handrails eigenfrequencies were by far higher than those of interest (larger than 60 Hz). The sampling theory (Nyquist theorem) states that the maximum measurable frequency has to be less than half of the GPS sampling rate (5 Hz). Therefore, according to the modal frequencies estimated with the FE analysis, the rst three global modal properties (Table 1, Fig. 4) were used as input data to nd the optimal sensor locations. Hence, the OSP methodologies considered and described, in the next paragraph, employed only the rst three mode shapes and the candidate sensor positions were those dened by the bridge handrails. The objective of the OSP selection was to identify the best locations of the available sensors (10 GPs) in such a way to capture the dynamic response of the structure. Six different OSP methods were investigated. The rst method uses the effective independence method (EFI, Section 3) developed for in-orbit modal identication of large space structures [9], based on the maximization of

Table 1 Modal frequencies calculated using FE model Number 1 2 3 4 5 Frequency (Hz) 1.44 2.79 4.66 6.75 8.9

the Fisher information matrix determinant. The second technique used, EFI-DPR [10], is a compromise between the EFI and an energetic approach, the driving-point residue (DPR [11]), and was developed for large mechanical structures such as engine casings and nozzles. A second OSP technique that to some extent has been derived from the EFI method is the kinetic energy method (KEM) [12], which maximises the kinetic energy content of the signals acquired. The KEM was used for large scale civil structures. The third OSP technique employed is the variance method (VM), which was developed by the authors and based on the most informative subset (MIS) [13]. This technique was designed to reduce the data amount, with the aim to study their dependences and characteristics, in meteorology and economic studies. The last two approaches, the eigenvalue vector product (EVP) [14,15] and the non-optimal driving point (NODP) [10] are energy based methods and they tend to select the sensor locations that maximise the vibration energy content of the signal acquired. A comprehensive description of the OSP methods utilised is given in the next paragraphs. 3. The effective independence method The problem being addressed is that of placing M sensors on a structure at locations which will allow the computational software to give the best t to a set of N targeted mode shapes. The EFI sensor placement method [9] was developed to maximise both the spatial independence and signal strength of the N targeted mode

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

1491

Fig. 4. First three mode shapes.

shapes by maximising the determinant of the associated Fisher information matrix. The vector of the measured structural responses denoted by ys can be estimated as a combination of N mode shapes through the expression:
N

ys = q + w =
i=1

q i i + w

(1)

where: is the matrix of FEM target mode shapes; q is the coefcient response vector; w is a sensor noise vector, assumed stationary random with a mean value zero; N is the column number of (n by N matrix, n being the number of the candidate sensor positions); i is the i th column of that is the i th target mode shape selected. This representation of structural response is based on the concept [7] that, the response in any point of an elastic structure can be obtained in the time or frequency domain as a linear combination of mode shape values. In this way, the i th coefcient of ys is a linear combination of i th mode shape vectors, where qi is a multiplier coefcient that is a function either of time or frequency. Evaluating the coefcient response vector using an efcient unbiased estimator and then, estimating the covariance of the error results in: J = E[(q q)(q q)T ] = 1 T 2
1

Fig. 5. EFI sensor location.

distribution (EID) vector given by: E D = []2 1 {1}k where: is the eigenvector matrix of Q. is the associated eigenvalue matrix. {1}k is the sum of all coefcients belonging to row k. The square is simply the square of any single element of the matrix product. Hence, to ensure the maximisation of the Q determinant, at each iteration, the sensor position having the smallest E D coefcient is eliminated, from the initial set of candidate sensors until the number of candidate sensor n equals the xed sensor number M being used. The major drawback with EFI, common to all OSP methodologies based on the Fisher information matrix, is that in order to have a non-singular matrix Q the minimum sensor number employable must be equal to the target mode shape number N. The results obtained from this OSP technique give rise to a fairly uniform spacing for the sensor locations along the length of the beam as shown in Fig. 5. 4. The EFI-DPR method A limitation of the EFI method is that sensor locations with low energy content can be selected with a consequent possible loss of information. The EFI-DPR (driving-point (4)

= Q 1

(2)

where Q is the Fisher information matrix [16]. GT =


T ys (q) q

Q(q) =

1 T G G 2

(3)

and G is the sensitivity matrix. Hence, the best estimation of q occurs when Q is maximised, therefore the procedure for selecting the best sensor placements is to unselect candidate sensor positions such that the determinant of the Fisher information matrix is maximised. In order to achieve this, an iterative algorithm was developed [9], which evaluates the candidate location sensor contributions employing the effective independence

1492

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

Fig. 6. EFI-DPR sensor selection.

Fig. 7. KE method sensor selected.

residue [11]) method [10] eliminates this problem by multiplying the candidate sensor contribution of the EFI by the corresponding driving point residue (DPR) coefcient:
N

DPRi =
j =1

i2j j

(5)

resulting in the following change of the EID vector expression [10]: E Di = []2 1 {1}i DPRi (6) where j is the j th target mode frequency. In essence the DPR is a weighting factor of the E D vector. This methodology concentrates sensor positions in the high energy content regions resulting in sensors quasi-uniformly spaced and symmetrically deployed as shown in Fig. 6. 5. The kinetic energy method The kinetic energy method (KEM) [12] sensor location procedure is similar to the EFI method. The main difference is that the KEM objective is to nd a reduced conguration of sensor placements which maximizes the measure of the kinetic energy of the structure rather than the determinant of the FIM Q. The Fisher information matrix Q is weighted with the FE model mass matrix M generating the kinetic energy matrix KE dened as follows: KE = T M. (7) Decomposing the mass matrix into lower (L) and upper (U ) triangular Choloskey factor matrices = U M = LU (8) (9)

Therefore after evaluating , the optimal sensor placement process is identical to that described for the EFI. Differently from Heo et al. [12], the FE model used for the bridge is very accurate and the candidate sensor locations are restricted to the bridge handrails. In order to calculate the mass matrix reecting the global mass distribution of the bridge a dynamic reduction of the bridge FE model was undertaken. This was accomplished using substructuring [17], which partitions the FE model DOFs into master (m, retained DOFs) and slaves (s, removed DOFs). The mass and the stiffness matrices were partitioned into four submatrices as described below: K = K mm K ms K sm K ss M= Mmm Mms . Msm Mss (11)

In accordance with Guyan [18] the reduced mass and stiffness matrices at the master DOFs were computed as follows:
1 K = K mm K ms K ss K sm 1 1 M = Mmm K ms K ss Msm Mms K ss K sm 1 1 + K ms K ss Mss K ss K sm .

(12) (13)

Finally, considering as master DOFs of the bridge FE model those belonging to the bridge banisters, a reduced mass matrix was evaluated and then Eq. (7) was used for the computation of the kinetic energy matrix (KE). The sensor locations identied by the KE method tend to be concentrated only on one side of the bridge handrails, as shown in Fig. 7 in contrast with the EFI and the EFI-DPR that distributed an equal number of sensors on both sides of the bridge.

the KE matrix can be represented as the product of the matrix and its transpose similarly as for the FIM Q, as described by the expression below: KE = T . (10)

6. Eigenvalue vector product The eigenvalue vector product (EVP) is an energy based technique [14,15] consisting in the evaluation of the vector

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

1493

Fig. 8. EVP sensor selection.

Fig. 9. NODP sensor selection.

EVP calculated using the following expression:


N

EVPi =
j =1

|i j |.

(14)

This technique selects the sensors with the largest EVP values in order to prevent the choice of sensors placed on nodal lines of a vibration mode and to maximise their vibration energy. As it can be seen in Fig. 8, the sensor distribution estimated by the EVP method is mainly concentrated in a small area of the bridge. It is clear that this method will not be able to capture accurately the vibration mode shapes of the bridge. 7. Non-optimal driving point based method The non-optimal drive point (NODP) [10] is an energy based method and it is generally used to nd the optimal excitation point. It is based on the concept that the amount of vibration energy of any mode shapes depends on the relative positions of the excitation source and the mode shape nodal lines. However, the amount of the vibration energy measured by the sensors is a function of the relative positions of the sensors and mode shape nodal lines. Therefore, the NODP based method was used to identify the optimal sensor placements as well. The methodology consists in an iterative algorithm that unselects the candidate sensor position having the smallest target mode shape displacement that is: NODPi = min |i j |.
j

physical and mathematical considerations on the coefcients of the covariance matrix [C] [19]. The most informative subset (MIS) technique is a statistics method that relies on a covariance analysis approach. This method was developed for problems related to evaluation and approximation of a phenomenon by a nite number of observations. Therefore, for our aims, the phenomenon investigated by the MIS technique is the evaluation of the mode shapes in non-instrumented points of the structure. Therefore, supposing that each target mode shape is evaluated (observed) in M(y M = {y1 , y2 , . . . , y M }) of the possible n locations, the mode shapes need to be estimated in the remaining nM positions y p = {y M+1 , y M+2 , . . . , y M+ p }(n = M + p). This problem can be solved using the best linear unbiased estimator of y p [13] given by the following equation:
1 y p = Cpp CMM y M

(16)

where if the matrix Y is assumed equal to {y M , y p } = [FE ]T , the matrices Cpp and CMM are the diagonal matrices of its covariance matrix [13]: cov(Y ) = cov(y M ) CMp CMM CMp = . CpM Cpp CpM cov(y p ) (17)

Evaluating the covariance of the estimation error as follows:


1 Dpp = cov( y p y p ) = Cpp CpM CMM CMp .

(18)

(15)

As shown in Fig. 9, the NODP selection technique places the sensors asymmetrically and mainly located at bridge mid-span. 8. Variance method The variance method (VM) is an evolution of the most informative subset (MIS) technique [13] according to

Then, any kind of monotonic function Q(Dpp ) could be used to measure the efciency of the estimator measuring the dynamic response in the location x i for i from 1 to M. Taking as efcient measure of the estimation, the Dpp determinant (D-criterion), the best M sensor positions will be those that minimise the Dpp determinant. Therefore, considering that: cov(Y ) = CMM CMp 1 = |CMM ||Cpp CpM CMM CMp | CpM Cpp = |CMM ||Dpp |.

(19)

1494

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

The minimisation of the Dpp determinant is equivalent to the maximisation of a measure of the CMM determinant or the maximisation of the determinant of the [FE ]T covariance matrix for a set of M sensor positions: max cov([FE ]T ) .
M set

(20)

However, the maximization of the determinant of [FE ]T guarantees the rank of the target mode shape matrix restricted to the M locations selected to be equal to N, but it does not assure that the signal strength is maximised. Moreover, the results produced are identical to those yielded by the EVP technique (Fig. 8), which are not capable of capturing the vibration mode shapes. Therefore, taking into account the following properties of the covariance matrix coefcients ci j [19]: For i = j if ci j = 0, the i th and the j th row of [FE ] are independent. For i = j ci j = i2 that is the vector {FE }i variance. Consequently, the diagonal coefcients can be considered as a measure of the signal strength acquired by the i th candidate sensor position. Hence, the simultaneously satisfaction of the maximisation of the determinant of CMM and the maximisation of the signal strength can be met for a CMM matrix, of which the off-diagonal coefcients tend to zero, while its diagonal terms tend to 1 (the maximum for the standard deviation is 1constant distribution). This is equivalent to the maximisation of the function Vr:
M

Fig. 10. VM sensor selection.

Vr =
i=1

cii Depi

Depi =
i= j

ci j .

Fig. 11. MSE comparison.

(21) respect the bridge symmetry, since the covariance matrix coefcients are sensible to the signs of the mode shape displacements. In some extent VM match more closely the energy distribution along the structure due to mode shapes. 9. OSP comparison In order to compare the capability of the different OSP techniques analysed to capture the vibration behaviour of the bridge two different criteria were used. Conceptually, a criterion of optimality should also be related to maximum damage information, i.e. sensor locations should be chosen in such a way that they produce reliable and sensitive information on the potential damages of the structural system. However, this aspect was not considered in this paper. The rst criterion assesses the capability of each OSP technique to capture the dynamic response of a structure by measuring the mean square error (MSE) between the FE model mode shape (FE) and the mode shape obtained by a cubic spline interpolation (CS) of the displacement measured at the sensor locations selected. The MSE of each vibration mode i th was normalised with respect to

For this reason is sufcient to select M sensor locations (k) having the highest ratio between the diagonal covariance coefcient ckk and the sum of its off-diagonal terms Depk . However, this way to maximise the function Vr, common to most of the energetic OSP techniques (EVP, NODP), leads to a concentration of the sensor locations selected in few regions of the bridge handrails similar to those provided by the MIS and EVP technique. The problem was solved by selecting the local maxima of the function Vr sorted in decreasing order. This solution has the advantage to: Distribute the sensors selected throughout the structure, assuring an improved capture of the vibration mode shapes. Suggest the number of sensors employable through the number of local maxima. This point is an important breakthrough for OSP, since none of the techniques described in the literature gives any information about the optimal number of sensors to use. Consistently reduce the computation time of the MIS, since the evaluation of the determinant of the CMM is avoided. In Fig. 10, the sensor locations selected having the largest Vr values are plotted. The sensor locations selected do not

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

1495

Fig. 12. EFI-DPR: (a) 1st mode shape; (b) 2nd mode shape; (c) 3rd mode shape.

its standard deviation i with the aim to evaluate the total MSE (TMSE, Eq. (22)) and it is given by the following expression:
1 i2 n j =1 2

Table 2 MSE comparison OSP method EFI EFI-DPR EVP NODP KEM VM Mean square error Mode shape 1st 7.66E09 1.57E08 1 3.03E06 5.30E04 4.50E08 TMSE 2nd 1.31E07 5.57E08 1 7.51E06 4.25E03 6.12E08 3rd 2.68E09 1.09E08 1 1.63E06 2.52E04 1.18E8 1.42E07 8.23E08 3 1.22E05 5.03E03 1.18E07

iCS j n

iFE j . (22)

MSE =
i=1

The cubic spline interpolation was carried out using a MATLAB coded function, separately for each handrail. The MSE criterion results are summarized in Table 2 and shown in Fig. 11. As expected the MSE of the EVP method was the highest due to the concentration of sensors only on the left banister of the bridge (Fig. 8). In contrast, the MSE for the EFI-DPR method gave results slightly better than, respectively, VM and the EFI method and about 3 and 5 orders of magnitude lower than the NODP and KE method, respectively. The comparison between the FE and the spline interpolated mode shapes using the EFI-DPR technique for all three vibration modes investigated is shown in Fig. 12. The second comparison criterion was based on the concept that the strength of the signals acquired associated

with modal characteristics should be as high as possible in order to reduce the noise. In order to comply with this important aspect, the Fisher information matrix (FIM) determinant behaviour was recorded during the sensor selection and displayed in terms of percentages of its initial value against the progressive number of the dropped candidate sensors (Figs. 13 and 14). The number of possible locations was 2042. This procedure allowed us to have a dynamic comparison between the number of sensors employed and the signal

1496

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

Fig. 13. Fisher information matrix determinant.

Fig. 14. Zoom of the FIM determinant.

strength. In Fig. 13 the overall behavior of the FIM determinant is shown. Since the number of available GPS sensors was 10 the area of interest was between 2042 and 2032 sensors. Therefore, in order to highlight the area of concern in Fig. 13 a second diagram was plotted in a logarithmic scale (Fig. 14). Two vertical lines were drawn; the rst (the dash dot line) highlights the FIM determinant value using a network of sensor made of 10 GPS, while, the dotted line represents a three sensor monitoring network which is the smallest number of sensors which can be used according to the FIM based techniques and KEM (see Section 3) since the FIM would be singular for a smaller number of sensors. Therefore, beyond this line the FIM determinant was kept constant. Moreover, in this diagram, the VM is also reported, but only for a number of sensors installed smaller than the optimal sensor number (the number of local maxima of the Vr function was 14).

The diagram shows that the trend and the values of the signal strength of the EFI, EFI-DPR and NODP method were very close, while the signal strength of the EVP and KEM were very small. The VM underperforms the FIM and NODP techniques, even though it has the signal strength of the same order for 10 sensors installed. Decreasing the number of the sensors of the sensing network, the VM behavior improves with respect to the FIM techniques and becomes better than the NODP method for a number of sensors smaller than 8. Finally, analyzing the results of both comparison criteria the EFI-DPR proved to be a better technique among those analyzed because of its: Superior capability of capturing the dynamic response of the structure by reproducing, with the smallest MSE, the shapes of the bridge modal properties (Fig. 11 and Table 2).

M. Meo, G. Zumpano / Engineering Structures 27 (2005) 14881497

1497

Symmetric distribution of the sensors on the bridge banisters respecting the bridge symmetry (Fig. 12). Marginally smaller information content in the signals acquired by its sensor distribution than the EFI sensor distributed network (Fig. 14). A better vibration energy content in the signals acquired than the EFI technique since the EFI-DPR selection procedure tends to select higher vibration energy sensor location (see Section 4). 10. Conclusions In this paper, several OSP methods for a suspended bridge were studied. The sensor locations were chosen in such a way that they should be able to give relevant information in the reconstruction of modal and dynamic characteristics of the bridge under investigation. Moreover, the strength of the signal acquired had to be able to withstand the noise presence without introducing any appreciable change in the bridge properties of interest i.e. the rst three target mode shapes. To assess the efciency of the six OSP techniques two comparison criteria were used. The rst investigated the ability of a set of measures to reproduce the shapes of the modal properties by estimating the MSE between their FE representation and the cubic spline interpolation of their values at the sensor locations selected. The second criterion measured the signal strength to the noise pollution by evaluating the information content of each sensor set by calculating the determinant of the FIM. The comparison criterion results showed the better performance of the EFI based techniques compared to the KEM, VM and energetic approach based methods. The results highlight that the effective independence drivingpoint residue (EFI-DPR) method provides an effective method for optimal sensor placement to identify the low frequency vibration characteristics of the studied bridge because of its smaller MSE, the large signal strength and its capability of respecting the bridge symmetry. The OSP technique developed by the author gave results very close to the EFI-DPR technique, in terms of the capability to capture the vibration mode shape and in terms of signal strength. However, the VM presented a unique characteristic in the world of the OSP techniques, which is the indication of the optimal number of sensors (ONS). This characteristic was not exploited since the ONS turned out to be larger than the number of GPS sensors available. However, the authors are condent of the efciency and the reliability of this new feature (ONS) introduced by the VM, for future applications.

References
[1] Padula SL, Palumbo DL, Kincaid RK. Optimal sensor/actuator locations for active structural acoustic control, AIAA Paper 981865. [2] Naimimohasses DM, Barnett DM, Smith PR. Sensor optimization using neural network sensitivity measures. Measurement Science Technology 1995;6:1291300. [3] Oh DY, No HC. Determination of the minimal number and optimal sensor location in a nuclear system with xed incore detectors. Nuclear Engineering and Design 1994;152:197212. [4] Roberts GW, Cosser E, Meng X, Dodson AH, Morris A, Meo M. A remote bridge health monitoring system using computational simulation and single frequency GPS data. In: Proceedings of the 16th international technical meeting of the satellite division of the institute of navigation. 2003. [5] Roberts GW, Meng X, Cosser E, Dodson AH, Morris A, Meo M. A remote bridge health monitoring system using computational simulation and GPS sensor data. Presented at the deformation measurements and analysis, 11th international symposium on deformation measurements, international federation of surveyors (FIG), Commission 6Engineering surveys, Working Group 6.1. May 2003. [6] Zumpano G, Meo M. Corrosion damage identication on plate by optimisation algorithm. In: Proc. of the rst European workshop structural health monitoring. 2002, p. 13742. [7] Ewins DJ. Modal testing: theory, practice and application. 2nd ed. Research Studies Press LTD; 2000. [8] Morris AJ, Vignjevic R. Consistent nite element structural analysis and error control. Computer Methods in Applied Mechanics and Engineering 1997;140(12):87108. [9] Kammer DC, Brillhart RD. Optimal sensor placement for modal identication using system-realization methods. Journal of Guidance, Control and Dynamics 1996;19:72931. [10] Imamovic N. Model validation of large nite element model using test data. Ph.D. dissertation. Imperial College London; 1998. [11] Worden K, Burrows AP. Optimal sensor placement for fault detection. Engineering Structures 2001;23:885901. [12] Heo G, Wang ML, Satpathi D. Optimal transducer placement for health monitoring of long span bridge. Soil Dynamics and Earthquake Engineering 1997;16:495502. [13] Fedorov V, Hackl P. Optimal experimental design: spatial sampling. Calcutta Statistical Association Bulletin 1994;44(MarchJune): 1734. [14] Doebling SW. Measurement of structural exibility matrices for experiments with incomplete reciprocity. Ph.D. dissertation. Colorado University; 1996. http://sdcl.colorado.edu/Publications/1995/Theses/Doebling_PhD.pdf. [15] Larson CB, Zimmerman DC, Marek EL. A comparison of modal test planning techniques: excitation and sensor placement using the NASA 8-bay truss. In: Proceedings of the 12th international modal analysis conference. 1994. p. 20511. [16] Dowski E. Fisher information and Cramer-Rao bound. Colorado University; 2002, http://www.colorado.edu/isl/papers/info/node2.html. [17] Kohnke P. ANSYS theory manual. ANSYS Inc.; 2001. [18] Guyan RJ. Reduction of stiffness and mass matrices. American Institute of Aeronautics and Astronautics Journal 1965;3(2):380. [19] Duda RO. Covariance. Department of Electrical Engineering, San Jose State University; 1997. http://www.engr.sjsu.edu/~knapp/HCIRODPR/PR_Mahal/cov.htm.