You are on page 1of 11

5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011, Cancún, México

RELIABILITY ISSUES IN DAMAGE IDENTIFICATION VIA STRUCTURAL HEALTH MONITORING
F. Lanata University of Nantes, France A. Del Grosso University of Genoa, Italy

Structural Health Monitoring (SHM) is becoming more and more used in practice, especially for large complex civil structures like bridges, offshore platforms, nuclear power plants, etc. In parallel with the diffusion of monitoring systems, the development of decision support systems (DSS) has been started, that utilize the results of the monitoring data interpretation process for supporting decisions concerning maintenance/rehabilitation planning. The use of life-cycle curves (LCC) has been recently proposed to be the core of DSS’s. The overall process, from data interpretation and damage identification to the definition of the life-cycle or fragility curves and finally to decision models is however affected by several uncertainties, both stochastic and epistemic. With reference to a real case of permanent static monitoring system, the paper is aimed at discussing the influence of the uncertainties affecting the damage identification process. After reviewing the probability models that better fit with the need of defining the reliability of the damage estimates, the paper discusses on the performance of the Proper Orthogonal Decomposition as a valid damage detection algorithm. The reliability of this latter is modeled in terms of probability of detection (PoD), probability of false alarms (PFA) and receiver operating characteristic (ROC) curve. Damage detection reliability on real data from a 2-year monitoring program on prestressed beams is considered for illustration and different ROC curves are deduced as a function of the threshold assumed for damage identification. The discussion is based on the availability of a reference period in which the structure can be considered as undamaged. Environmental conditions are indeed recognized to have a great influence on the reliability of damage identification procedures. Corresponding author’s email: delgrosso@dicat.unige.it

-1-

modelling and other factors. geometry. the actual state of degradation can also be determined by applying a damage identification process. The paper is aimed at discussing the influence of the uncertainties affecting the damage identification process. LUNAM University. The time-dependent reliability problem for ageing concrete structures was discussed by several authors. The process from data interpretation and damage identification to the definition of the life-cycle curves and finally to decision models is usually affected by several uncertainties. deterioration. Nantes. allowing updating -2- . at any required time. materials.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. both stochastic and epistemic. Andrea Del Grosso2 1 2 GeM. Environmental conditions are indeed recognized to have a great influence on the reliability of damage identification procedures. Genoa. 1 INTRODUCTION Safety and usability of existing concrete structures actually depend on the level of degradation and obsolescence due to ageing. if a Structural Health Monitoring system is installed on the structure. among which Mori & Ellingwood (1993). France DICAT. Italy ABSTRACT: Structural Health Monitoring (SHM) is becoming more and more used in practice. probability of false alarms (PFA) and receiver operating characteristic (ROC) curve. Damage detection reliability on real data from a 2-year monitoring program on pre-stressed beams is considered for illustration and different ROC curves are deduced as a function of the threshold assumed for damage identification. These are integrated into a method that estimates the probability of reaching the specified performance level during the service life of a structure (Aktan & Ellington 2007). especially for large complex civil structures. México RELIABILITY ISSUES IN DAMAGE IDENTIFICATION VIA STRUCTURAL HEALTH MONITORING Francesca Lanata1. Stewart & Rosowsky (1998) and Enright & Frangopol (1998) have been pioneering the subject. The reliability of the proposed damage detection method is modeled in terms of probability of detection (PoD). Reliability analysis in structural engineering has to enable quantification of uncertainties associated with loading. Cancún. Lifetime functions for concrete structures can be constructed from theoretical/experimental degradation models and. University of Genoa.

2011). Cancún. First. PFA). A bayesian modeling of monitoring results is finally introduced by defining the probability of detecting damage conditional to an actual existing damage (Probability of Detection. theoretical aspects from detection theory are reminded. Okasha & Frangopol 2010). With reference to a real case of permanent static health monitoring systems (Del Grosso & Lanata 2010). in particular uncertainties due to environmental changes. the paper is aimed at discussing the performance of the Proper Orthogonal Decomposition as a damage detection algorithm (Lanata & Del Grosso 2006). and the validation and updating of existing prediction models (Frangopol et al. 2004) and dynamic data (Kim et al. Looking for the best detection performances. and on laboratory (Park et al. The SHM efficiency depends on the ability to detect the structural changes under environmental and operational conditions (Sohn 2007). 2006) and in-field experiments (Cheung et al. PFA). The uncertainties affecting the damage identification process on the assessment of the structural health state will be taken into account. Robust and reliable methods capable of detecting. For example. On the contrary. Despite of the above interest that has been largely expressed by means of research papers in conferences and scientific journals. special emphasis is given to the influence of the environmental conditions on the probability of false alarm. Schoefs et al. The difficulties in obtaining a good performance curve in presence of strong temperature variations (noise) and small levels of damage is also discussed -3- . 2008. from data interpretation and damage identification to the definition of the lifecycle curves and finally to decision models is affected by several uncertainties. 2010). Benedetti et al. locating and estimating damage whilst being insensitive to changes in environmental and operating conditions have yet to be agreed upon. the experimental approach is presented and the performance estimate of damage detection algorithms (DDA) for a given detection threshold is discussed through the Receiver Operating Characteristic (ROC) curve (Schoefs et al. the use of monitored extreme data allows the reduction of uncertainties associated with numerical models. The damage indicators extracted from the proposed algorithm are used to set a reference period and a threshold value associated to damage presence. PoD) and the probability of detecting damage conditional to non-existing actual damage (Probability of False Alarm. as soilstructure interaction. because this latter depends on the noise level and the chosen threshold only. 2011). The overall process. For complex structures in particular. the main reasons being the lack of extensive field evidence in the disclosure of structural defects and the inability to interpret the data obtained from hundreds of sensors in a timely manner. based on the characteristics of real monitoring data. the probability of detection should always take larger values than the probability of false alarm (low noise sensitivity). México of the lifetime functions (Orcesi & Frangopol 2010.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. and in-service structural behaviour influenced by building/loading conditions and settlements that are expected to separate the current behaviour from the theoretical one (Vallone & Giammarinaro 2007. on static (Omenzetter et al. 2009). A lot of damage detection methods have been tested in the last years. The process performance can be completely described by the couple (PoD. Then. both stochastic and epistemic. the probability of detection depends on the damage extension. very few practical examples can be reported to date. In the paper. SHM has been proved to be the only way to understand both complex interaction mechanisms. 2008).

1 PROBABILISTIC MODELING OF DDA RESULTS BASED ON DETECTION THEORY Theoretical background and basic concepts for PoD and PFA Despite of the great number of damage identification approaches that have been proposed in the literature. Cancún. 2 2. monotonically increasing -4- . The proposed strategy uses the whole information from the damage identification process to optimize the results. From a theoretical point of view. a first way to visualize the performance of a selected algorithm and to compare the performances of different approaches is to plot on the same graph. Most of the time. their capabilities and abilities to well perform are different from those given by numerical simulations and laboratories sets (Del Grosso et al. Let ad be the minimal damage level. In a probabilistic scheme. Nevertheless. even partially. the results of a DDA. a method usually used for evaluating the performances of non-destructive testing (NDT) is proposed and extended to the damage identification process (Rouhan & Schoefs 2003). the probability of detection is defined as: (1) where a is the damage extension. the use of the PoD alone is not suitable. performance evaluations only deal with the probability of detection. this is the receiver operating characteristic (abbreviated ROC) curve. This curve is basically the PoD plotted as function of the PFA (Arques 1982). which is the probability to detect an existing damage. The probability of detection (PoD) and the probability of false alarm (PFA) are both needed to characterize. Thus. in particular in complex environment and harsh conditions. All algorithms have limitations and. environmental conditions and so on. PFA). For a given detection threshold there is a way to define method performance by the couple (PoD. México and the possibility of improvement of the detection process through post-processing of acquired data is illustrated. this is a convex. and not with the size of the damage. It is induced especially by noise with several possible sources: human. Another information which is called probability of false alarm (abbreviated PFA) is to be considered to take into account the role of uncertainties. In harsh insitu conditions. false detections increase dramatically and non-existing large damages can be detected (Del Grosso & Lanata 2010). their ROC curves. under which it is assumed that no detection is done.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. 2008). a and ad are stochastic variables. In the following. Data from a 2-year monitoring program on pre-stressed beams is considered for illustration and different ROC curves are deduced as a function of the threshold assumed for damage identification. The optimal damage identification approach should be able to detect the smaller possible damage at the right place and at the right time. False alarms correspond to the detection of nonexisting damage. The discussion is based on the availability of a reference period in which the structure can be considered as undamaged. nature of phenomenon to be measured. This leads to lower performances than theoretically expected. So. there is an increasing need to assess the performance and the validity of the selected algorithms with a comparative tool.

several numbers can be defined as a function of the vector c of the fixed threshold values: . Thus. one must remind that the probability of false alarm depends on the noise and detection threshold only.nf(c): number of non-existing damages that are detected by the algorithm. It does not depend on damage size. the damage size.nn(c): number of existing damages that are not detected by the algorithm. 3 3. but the probability of detection is an increasing function of the damage size. When assessing the ROC curve using experimental data. The best threshold value has to be set for each tested DDA.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. no matter what the observation is. 2010) on numerical simulation and experimental data. (2008). . More generally in this framework. The threshold can be for example defined using an adequate confidence interval assessed from the observations of the reference period. The probability of detection is a function of the detection threshold. As previously mentioned. and the noise. As the definition of the threshold value in the case of damage detection algorithm is not easy to achieve. always lying above 45° diagonal of the ROC space. for a given detection threshold. the threshold value ad should be defined observing the variation of the feature/s for a reference period during which the structure is supposed to be undamaged. the probability of false alarm is a constant. 3. PFA) = (1. -5- . Cancún. the variables a and ad in Eq. PFA) plane are: (PoD. the probability of detection should always take larger values than the probability of false alarm (low noise sensitivity). discrete values for PFA and PoD are obtained for given conditions. México function. . A parallel approach is used to evaluate the performance of a damage detection algorithm already applied by the Authors (Lanata & Del Grosso 2006. (1) can represent the observed current value and the threshold value of one or more features extracted from the measurements by the particular damage identification algorithm. In this case. since in that case the detection result is the same. The diagonal line running from lower left to upper right (curve "PoD = PFA") is the line of no performance.nr(c): number of non-existing damages that are not detected by the algorithm.1 PRACTICAL ASSESSMENT USING MONITORING DATA From the inspection results to the damage detection algorithms In the following.nb(c): number of existing damages that are detected by the algorithm.2 Basic definitions from experimental data Using experimental data. Looking for the best detection performances. definitions taken from the assessment of non-destructive testing performance are used. The perfect ideal algorithm is represented by a ROC curve reduced to a single point whose coordinates in the (PoD. so that PoD ≥ PFA. the inspection results from a NDT are herein substituted by the damage identification results from the Proper Orthogonal Decomposition. The length of the reference period depends on the available information and on the periodicity of the time histories as shown in Posenato et al. . 0). When reading ROC curves. a parametric study using a vector c of threshold values is suggested. Del Grosso et al.

For this reason. Cancún. conditional to damage detection. 2003). the detection result could be: no damage. and .E3: presence of damage. Considering the binary random variable "presence of damage" X. while the second one deals with the probability of wrong assessment (PWA) and has to be considered as an indicator of the PFA. whose value is 1 if damage is detected. Then four events are defined: . By definition. 3.E1 = [d(X) = 0 | X = 0].5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. as for in-service structures the state of the structure is not known. . conditional to no damage detection. a Bayesian approach from decision theory can be used (Rouhan & Schoefs.3 Bayesian modeling With the same logic. these events are not independent and some of them are complementary. . conditional to an actual existing damage. . 0 otherwise.E4 = [d(X) = 1 | X = 1]. whose value is 1 if damage is present. (3) (4) -6- .E1: no presence of damage. according to Bayes' rule: PoD(X) = P(E4) = P(d(X) = 1 | X = 1) PFA(X) = P(E3) = P(d(X) = 1 | X = 0) This gives the right definitions of the PoD and the PFA: . México As the number of the detected damages can exceed the number of existing damages. it is possible to note as d(X) the random damage decision function. Anyway. Thus.the PoD is the probability to decide damage presence (damage detection).E3 = [d(X) = 1 | X = 0]. the corresponding ratio is not a probability. conditional to damage detection. .E2 = [d(X) = 0 | X = 1]. conditional to no damage detection. After data analysis. 0 otherwise. two sets of probabilities are defined: (2) (3) The first probability concerns the probability of good assessment (PGA) and is somehow related to POD. or presence of damage.E2: no presence of damage. . the probability of false alarm PFA and the probability of detection PoD could be written. it is necessary to consider four events: .E4: presence of damage.

p. but gives also a spatial information following each eigenvector at each sensor location. POD is based on an orthogonal decomposition of the covariance matrix of the process variables along the direction that explain the maximum variation of the data. the covariance matrix has to be computed and the relative proper orthogonal modes can be obtained.2 Assessment of ROC points The eigenvectors extracted by the Proper Orthogonal Decomposition are used as damage features in the analysis of PoD and PFA. 64 time trajectories are available to assess the ROC point. the time variation of the relative eigenvector at each sensor location is available. (2010). For each structural static mode. conditional to no actual existing damage. EPFL and SMARTEC SA. In this study POD has been proposed to extract a spatial and temporal correlation between the installed sensors. 4. previously validated on numerically simulated experiments.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. Cancún. placed outdoor under environmental conditions and one of which has been subjected to increasing artificial damages. In the present paper the performance of the Proper Orthogonal Decomposition as a damage detection method is investigated. The measurements have to be arranged to form the measurement matrix: each column represents a displacement history at a particular sensor location and each row represents the spatial distribution of the response at a given time instant. 4 4. As damages have been introduced too early during the -7- . The experiment has been conducted for more than two years and consists in the continuous monitoring of strains and temperatures in two pre-stressed concrete beam specimens. The algorithm is a data reduction tool that is capable of transforming a number of related process variables into a smaller set of uncorrelated variables (Lanata & Del Grosso 2006). the method allows to follow the time variations of the eight extracted eigenvectors.A.. the results of some applied damage detection algorithms and the method proposed to remove the temperature effects are presented in Del Grosso et al. four on the undamaged one). The details of the experimental program. an intrinsic advantage of strain measurements consists of an assessment of static modes which are much more sensitive to damage. After subtracting the mean displacements of each sensor location. The extracted eigenvectors need a certain period of time to get a stationary behavior to be assumed as reference state. México . In this way. related in a certain way to the normal modes contained in the static structural response. has been performed to validate different damage identification algorithms. The eigenvalues and the orthogonal eigenvectors are calculated from the covariance matrix using a window containing at each step all previous measurements.the PFA is the probability to decide damage presence (damage detection). carried out by the Authors in cooperation with Autostrade per l’Italia S. While in usual dynamic applications modal parameter modifications are not considered very sensitive to damage.1 CASE STUDY Presentation of the experiment and data analysis An experimental monitoring program. As a consequence. The strain time histories have been detected by means of fibre optic SOFO sensors (eight installed on the damage beam. that can be applied to the SHM of reinforced and pre-stressed concrete beams under static continuous monitoring programs.

Figure 2 shows again that the higher damage threshold leads to a shorter distance δ and that a better efficiency of the DDA is obtained.3*standard deviation). As confirmed from the previous data interpretation.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. 4. The same procedure is performed on pre-processed filtered data to remove the effect of temperature variations. Considering the distance δ between the experimental points and the best performance point with coordinate (0. Three thresholds have been defined using the standard deviation of measurements of the reference state: the first one corresponds to a confidence interval of 67% (mean +/standard deviation). México monitoring. A time window of 5 and 15 days respectively is used to discretize the monitoring period in N intervals. The PGA and the PWA are evaluated as described at 3.1 Results on noisy and pre-processed data The computation is initially performed on rough data. Cancún. No significant differences are noticed in results obtained with the two time windows. 1). Figure 1 shows that the higher damage threshold leads to a shorter distance δ. The extracted eigenvectors contain the structural response and also all the environmental information. three months at the end of the monitoring have been assumed as the reference undamaged period to define the damage detectable threshold.2. -8- . as presented in Del Grosso & Lanata (2010). the PGA and the PWA give a maximum value for the PGA of 0. for this reason. The ROC points are finally obtained for each eigenvectors and plotted for the three investigated thresholds. they are mixed to the non-stationary initial trend of the eigenvectors. each one containing or not the artificial damage.82 and a minimum value for the PWA of 0.08. the second one to the 95% (mean +/. In this case. the performance of the Proper Orthogonal Decomposition improves after the data pre-processing to remove the noise due to temperature effects. A better performance of the DDA on pre-processed data is evident.18 for the first mode. The ROC points are then obtained for each eigenvectors and plotted for the three investigated thresholds.2 giving a maximum value for the PGA of 0.92 and a minimum value for the PWA of 0.2*standard deviation) and the third one to the 99% (mean +/. it is possible to have an evaluation of the optimal efficiency of the DDA under specific conditions. Deeper analyses are necessary to improve the ROC curves and to compare the performance of different damage detection algorithm on the same available database. A vector of damage threshold is defined for each structural mode at each sensor location.

1 0 CI 68% CI 95% CI 99% 0 0. The application of statistical pattern recognition methods for damage detection to field data. Kiremidjian. Decisions en traitement du signal. M.1 0 CI 68% CI 95% CI 99% 0 0. Benedetti.3 0. PY. and Kehoe.4 0. B. V. Sarabandi. Ellingwood. Mode 2 at each sensor location (from 1 to 8) corresponding to different damage thresholds.5 0.4 0.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. KK.9 0. Structural health monitoring of wind towers: remote damage detection using strain sensors. A. 5 REFERENCES Aktan.2 0.5 0.3 0.4 0.9 1 PFA Figure 2.9 1 PFA Figure 1.9 0.8 0. Smart Materials and Structures. A. and Zonta. P. BEAM A .5 0.1 0. 20 (13pp). Paris: Editions Masson.6 PoD 0. Cancún. -9- .3 0. and Wenzel. Smart Materials and Structures. AE.4 0.6 0.MODE 1 1 0. 1982. ROC points from experimental data. Mode 1 at each sensor location (from 1 to 8) corresponding to different damage thresholds.7 0.2 0. BR.8 0. ROC points from experimental data. D. 2007.MODE 2 1 0. 2011. 133(3): 311-323. Journal of Structural Engineering.8 0.1 0. 2008. Cabrera.7 0.6 0. Arques. C. H.7 0. Cheung. Performance-based engineering of constructed systems. 17 (12pp).6 PoD 0. Nair.7 0. México BEAM A .2 0. Fontanari.3 0.8 0.5 0.2 0.

2010. H. Posenato D. Use of monitoring extreme data for the performance prediction of structures: General approach. 2010. Lanata. and Frangopol. Optimization of Bridge Maintenance Strategies based on Structural Health Monitoring Information. IABMAS Proc. 33: 26-41. 46(5): 609–618. Time-dependent reliability of Deteriorating Reinforced Concrete Bridge Decks. Okasha. 30: 3644-3653. and Nouy. and Park. Schoefs. Taipei. 2003. D. Inaudi. Novel Approach for Multicriteria Optimization of Life-Cycle Preventive and Essential Maintenance of Deteriorating Structure. Smart Materials and Structures. Model free data interpretation for continuous monitoring of complex structures. 5th Int. Ahmad. H. A. A. Omenzetter. F. Hybrid health monitoring of prestressed concrete girder bridges by sequential vibration-impedance approaches. Park. and Lanata. D. Y. DV. 119(5): 1600-1621. Trans. and Frangopol. and Kim. and Moyo. F. and Frangopol. and Rosowski. Experimental Mechanics. A Long-Term Static Monitoring Experiment on R. R. Identification of unusual events in multi-channel bridge monitoring data. Life-Cycle of Civil Engineering Systems. JT. MP. Journal of Structural Engineering. D. Rouhan. Pardi. A. and Smith. Enright. 15(6): 1811-1829. S. Beams using Fiber Optic Sensors. F. México Del Grosso. Structural Safety. Lanata. 2010. Safety and Managemenent. F. BR. and Schoefs. and Ellingwood. A. A. Mechanical System and Signal Processing. Strauss. Polynomial chaos representation for identification of mechanical characteristics of instrumented structures. 20(1): 91-109. Bridge Maintenance. Taiwan Tech. Schoefs. Damage detection and localization for continuous static monitoring of structures using a Proper Orthogonal Decomposition of signals. and Mercalli. Structural Safety. Proc. Health Monitoring for Corrosion Detection in Reinforced Concrete Bridges. Clément. Park. Yàñez-Godoy. A. Engineering Structures. Advanced Engineering Informatics. D. 31(5): 409-419. Assessment of ROC Curves for Inspection of Random Fields. Multiple crack detection of concrete structures using impedance-based structural health monitoring techniques. Frangopol. P. F. ComputerAided Civil and Infrastructure Engineering. S. Brownjohn. Conf. 2010. A. Conf. WS. A. and Mercalli. Honga. Del Grosso. 2006. F. Probabilistic modeling of inspection results for offshore structures. Mori. and Del Grosso. I. S. Taylor and Francis. F. Engineering Structures. 11-15 July 2010. 2008. Effects of environmental and operational variability on structural health monitoring. MG. 365: 539–560. Cancún.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. Phil. 2004. L.. A. F. 1998. Philadelphia. A. P. 26: 173-189. Stewart. 18: 409-430. ASCE Journal of Structural Engineering. 2009. 2010. Validation of an SHM procedure for concrete bridges based on static strain. Y. Del Grosso. Structural Safety. Posenato. A. DS. Structural Safety. 4th Int. 2006. A. 22: 135-144. 2011. J. Reliability-based Service-Life Assessment of Aging Concrete Structures.10 - . D. NM.C. 1993. 1998. 20(11): 960-971. Orcesi. Lanata. 2008. JH. Lanata. F. Sohn. Probabilistic Analysis of Resistance Degradation of Reinforced Concrete Bridge Beams under Corrosion. CB. 25: 379-399. . 2008. D. Engineering Structures. 136(8): 1009-1022. Yun. 32(1): 115-128. and Lanata. and Roh. 2007. Kim. Soc.

intern. P. Cancún.. MS. México Vallone. conf. . Monitoring of Termini Imerese harbour (Sicily. and Giammarinaro. Applied Geophysics for Engineering: Proc.11 - . 17-18 December 2007.5th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-5) 2011 11-15 December 2011. Italy) using satellite DInSAR data. Messina. 2007.