Life Cycle Cost Analysis of Ageing Structural Components Based on Non-destructive Condition Assessment

Abstract Inspection-based maintenance strategies can provide an efficient alternative for ageing civil engineering components subjected to ageing and degradation. The technical and/or economic efficiency of such strategies depends on many factors, such as the mechanisms involved in the loss of performance; the availability, cost and efficiency of inspection techniques; the relation between what can be measured through inspections and the level of performance of the structure; the level of required serviceability of the structure; and the direct and indirect economic losses due to a reduction in the performance of a structure. On this basis, it is studied here, using Monte Carlo simulations, the benefits and limitations of an inspection-based maintenance strategy. The quality of the inspection technique is analysed in terms of its sensitivity to defects in their initial stage of development, and on its discriminant ability (detection of a real defect, while avoiding false alarms). This study is carried out with ageing characteristics, inspection models and cost assumptions that can cover a wide field of ageing assets. The influence of several factors is highlighted to see how they influence optimal strategies; a focus is given on the quality of inspections and on the allowable probability of failure.


INTRODUCTION
Many of the structures that were built around the 1960s are coming to the end of their service lives and are showing signifi cant signs of deterioration. Consequently, bridge maintenance costs are already very high and are increasing. Currently the biggest challenge that structure owners/managers face is fi nding the optimum balance between the increasing number of deteriorating structures, and the limited funds available for their upkeep. The demolition and replacement of large engineering structures results in high economic and environmental costs, further increasing the need for effi cient management plans to maintain these structures.
As a result, a lot of research has been conducted in this area over the last decade to develop methods of maintenance management that optimise maintenance budgets (Estes & Frangopol, 1999;Faber & Sorensen, 2002;Kong & Frangopol, 2003;2005;Lauridsen et al, 2006;Radojicic et al, 2001: Stewart, 20012005;Stewart et al, 2004;Stewart & Mullard, 2006). The main objective is to fi nd the optimal maintenance management plan, thereby optimising the lifecycle cost of the structure. Many of these methods rely on quantitative data from inspections, rather than qualitative and subjective data. Therefore, monitoring and inspections are key aspects in this process. The information from these tests can be The main focus of this paper is on inspection-based maintenance, and how the quality of inspection and its ability to provide useful information infl uence risks and costs, and change the optimal time between inspections. The maintenance strategy chosen depends on many factors, such as the rate of deterioration, the mode of failure, the correlation between the measured parameters, the performance of the structure, and the consequence of failure. The aim of this paper is to use a cost-based analysis to determine under what conditions inspection-based maintenance is an effi cient strategy.
Through Monte Carlo simulations, this paper studies the benefi ts and limitations of inspection based maintenance strategies. The study is based upon data and models developed and used in the European Union-funded MEDACHS research project. MEDACHS is dedicated to the optimisation of the service life of structures in marine environments. Both analytical analysis and Monte Carlo simulations use deterioration models, and information supplied from inspections, to determine the most suitable inspection method along with the optimum inspection period, for different deterioration rates. The probability of detection (PoD) and probability of false alarm (PFA) are used in this study to quantify the quality of an non-destructive testing (NDT) method.
In addition, the results of this analysis are then used to compute annual total cost of a structure using inspection-based maintenance and to see, for a given set of parameters, what factors determine the optimum maintenance strategy.

THEORETICAL BACKGROUND OF PROBABILISTIC MODELLING OF INSPECTION RESULTS
The models that are used to determine the optimum maintenance management strategy are just estimations that predict how the structure will behave over time.
Since these models include uncertainty, it can be useful to carry out regular inspections to ensure that the structure is behaving as predicted, or to detect possible problems. The information from the tests can be used to update these models and come up with a more economical maintenance strategy over the remaining lifetime of the structure (Faber & Sorensen, 2002).
Given the size of the defect, and the inspection method being used, there is a certain PoD (Faber & Sorensen, 2002;Onoufriou & Frangopol, 2002;Straub & Faber, 2003). The PoD can be evaluated using: where a d is the detection threshold, under which it is assumed no crack can be detected (for a particular method), and a is the measured crack length (Schoefs & Clement, 2004). For example, Onoufriou & Frangopol (2002) developed equation (2) to calculate the PoD that corresponds to an NDT method that has a 90% probability of detecting a 40 mm long crack.
The PoD and PFA can be used to quantify the quality of an NDT method. A variety of NDT methods, each with different costs and quality, can be used to assess the condition of a structure over its lifetime (Onoufriou & Frangopol, 2002). For a given crack size, ROC curves (a plot of PoD versus PFA) of different NDT methods can be used to compare the quality of different methods (Schoefs & Clement, 2004;Rouhan & Schoefs, 2003). For a given test, the PoD depends on the crack size, the detection threshold and noise. The PFA, however, depends only on the detection threshold and noise. Therefore, the PoD is the probability that the quantity "signal + noise" is greater than the detection threshold, and the PFA is the probability that the quantity "noise" is greater than the detection threshold (Rouhan & Schoefs, 2003). Noise can depend on environmental conditions, human interference and the nature of what is being measured. Rouhan & Schoefs (2003) further developed this methodology by focusing on the probability that a defect exists after an inspection has been carried out. The PoD is the probability that an existing defect is detected, whereas a more useful parameter is the probability that a defect exists, given the results of an inspection. A decision scheme is introduced that considers four inspection events, which are represented in fi gure 1 and by equations (3) to (6).

Deterioration growth and failure models
The simple model describing the random growth of a structural defect is: where the growth rate k is a Gaussian random variable, assumed to be constant for each component before its failure or repair. The regular evolution of the defect size randomly varies for two consecutive cycles, from an initial defect size d 0 equal to 0.25.
Each year, the annual probability of failure, p F , is calculated using the Weibull cumulative distribution function: where the parameter m is the Weibull exponent, which determines the spread of the curve. The purpose of the limit defect size, d 1 , is described by equation (9) and fi gure 2.
In this context, failure must not be understood as structural failure (even if Weibull law has often been used to describe the statistical distribution of concrete or timber strength), but from a more general point of view it is the fact that a limit state is reached, as it will be discussed at section 4.3.

Models for decision and inspection
The aim of optimised maintenance is to detect and repair soon enough, before the defect size reaches values for which the probability of failure (annual probability or probability between two inspections) becomes non-acceptable. The level of acceptability depends on the induced consequences of failure; since some failures have few consequences, one can accept a relatively high probability (this corresponds to the "serviceability limit states") when in other cases (eg. if lives are at stake), the acceptable level is very low. Thus the allowable annual probability of failure, p FA , will be chosen after a consideration of . It is thus possible to deduce the critical defect size d c , corresponding to this probability, using equation (10), obtained directly from equation (8): Scheduled inspections are simulated every T i years, with a technique of known quality Q'. The measurement induces a random noise, which is assumed to be a random error on the estimate of the d value, following a Gaussian distribution, whose standard deviation is inversely proportional with Q'. It comes the measured value d meas .
where d is the true value and n the measurement noise. Depending on the respective values of d, d meas and d c , one follows the event tree scheme where the question "exists" is replaced "with larger than critical". It is also assumed that it exist a minimum defect size d min that can be detected with a technique of quality Q'; d min increasing if Q' increases. Figure 3 shows what is the time evolution of the defect size, combining effects of deterioration, inspection and repair. Squares correspond to the measured values assessed during inspection (here every fi ve years). The non-perfect quality of the technique implies that the assessed values differ from the true ones. As soon as the assessed value is larger than the critical value d c (here taken at 0.41), the component is repaired and assumed to return to its initial minimal defect size d 0 . A new cycle, with its own growth rate k, begins.
When the quality of the technique is good (thus the technique is expensive), the assessed value is near the true one. In this case, the decision for repair is generally a good decision (events E1 or E4 on fi gure 1). With a poor quality technique, the measurement noise is large and the assessed value can differ signifi cantly from the true one. This often results in a premature decision for repair, since the assessed value is larger than the threshold when the true one is not (event E2 on fi gure 1). In some cases, a defect whose size exceeds the threshold size may be missed, implying a non-detection (event E3 on fi gure 1).

Expected costs
When a scheduled inspection is carried out, every ΔT years, there are various costs involved. As well as the cost of inspection, there may also be an associated repair cost, or failure cost, both of which depend on the results from the inspection. For each inspection, there can be fi ve possible outcomes (events E(C 01 ), E(C 023 ), E(C 045 ), and two variants of the latter two), with varying associated costs (see fi gure 4).
The mean annual total cost of the structure is used to determine the optimum maintenance management plan. The total cost is the sum over the period of study (or its annual average) of the building cost, inspection cost, repair cost and failure cost. The three latter costs are expressed in terms of their relative ratio to the building cost C 0 . Figure 4 reproduces the event tree for costs after inspection (when there is a year without inspection, the only possible costly event is failure).

Inspection cost
The cost of a single inspection is computed using: where C 0 is the initial construction cost, Q' is the technique quality coefficient, k i is the inspection coeffi cient, and a is a weighting exponent that links the quality and the cost.

Repair cost
The cost of repair is calculated using: where k r is the repair coeffi cient and where it is assumed that the cost increases with the defect size.

d d min
In case of detection When a repair has been carried out, the size of the defect is assumed to be returned to the original size, d 0 . Repair may occur due to the detection of a defect (good decision; GD), or due to a false alarm (FA).
In the computations, the cost of repair is calculated separately for these two instances, so that the relative costs can be compared and the consequences of false alarms can be analysed.

Failure cost
The cost of failure of the system is also calculated relatively to the cost of construction: where k f is the failure impact coeffi cient.
For the consistency of simulations, one can wait for some link between the impact of failure and the "Life cycle cost analysis of ageing structural components . annual probability of failure p FA defi ned above, which, knowing the delay ΔT between two inspections, will induce the value of the critical defect size. The reason is that when failure has important consequences, one needs to be more severe and to repair before the p f (d) is too large. This question will be addressed below. A discounting factor r can also be introduced to account for the advantage one can have in delaying his expenses.

ANALYSIS OF OPTIMAL INSPECTION STRATEGIES
As all models are defined, it is thus possible to perform simulations. Since it has been chosen to model both the material deterioration process, the material random failure, the detection process and the associated costs, many parametric studies can be performed. It is thus necessary to focus on some specifi c points, such as to highlight some interesting effects.

Optimum delay between two inspections
In a fi rst series of simulations, we will address the optimality of delay between inspections, the defect growth law being given, and the quality of the technique being given. We will show how the optimal solution depends on assumptions on induced costs (here through k i , k r and k f coeffi cients values).  Figure 6 synthesises results when the delay ΔT between inspections is varied between 1 and 10 years. Some comments arise: • The number of failures increases with the delay, since the failure probability increases and two inspections with a too large delay cannot prevent it (the defect is below the critical threshold at the fi rst inspection).
• The number of repairs after a false alarm is larger than that of repair undertaken on the basis of a good decision, and this difference is larger when the delay decreases. This comes from the error measurements, whose consequence is (with a 50% probability) the overestimation of the defect size, which leads to many repairs before needed (see fi gure 4). A direct consequence is that the average service life jumps from less than 40 years (38.1 = 10 000/262.6 if ΔT = 1 year) to more than 50 years (53.0 = 10 000/188.6 if ΔT = 10 years).  Costs calculated for varying ΔT delay are compared on fi gure 6. Inspection costs, repair costs (distinguishing those due to false alarm and those due to good decision) and failure induced costs are considered separately. With the chosen set of parameters, the repair cost remains small and the optimum delay value is determined by the fact that, if the delay increases, the inspection costs decrease and the failure costs increase. The resulting optimum is here equal to four years. Figure 7 gives the results for the same simulations for a variant in which only cost weights have been changed (table 2). In this case, one has given a larger weight to repair costs. The result is an increase of the optimum interval, with a best result for six or seven years. The shape of the function shows that it is very sensitive to economical assumptions.

Effect of the inspection quality
The inspection quality Q' has an infl uence on the ability of detecting sooner (via the d min value), of taking good decisions (because a higher quality reduces the noise measurement n, equation (11)) and on inspection cost (equation (12)). Figure 8 shows what is the infl uence of the quality of the technique, when it is varied from Q' = 5 to Q' = 512. Unit cost data are those of Case C (table 2) and a constant delay ΔT = 7 years is assumed between inspections.
The simulation shows that the optimum is obtained for a technique of an average quality. A too sophisticated technique is too expensive without decreasing significantly the repair and failure costs (for this ΔT value and this defect growth kinetics), while a too rough technique, even cheap, leads to too many repairs and thus increases the corresponding costs.

Effect of the allowable probability of failure
In the previous simulations, the allowable annual probability of failure was fi xed (p FA = 10 -2 ). If one considers that this parameter is itself a degree of freedom during the design process, this enriches the optimisation process, since the "optimal solution" (in terms of minimum cost) will correspond to an optimal set of ΔT, Q' and p FA . The allowable annual probability of failure p FA is varied in the following from 10 -4 to 10 -1 . This range of values mainly corresponds to a Serviceability Limit State, since it gives 0.5% probability to near certainty for a 50-year period.
A second series of simulations has been undertaken such as to better understand this point. The number of simulations corresponds to about 2500 to 4500 life cycles, depending on the data set. This number being larger than in the fi rst series (section 4.1), it improves the quality of convergence and reduces the computational noise that could be seen on fi gures 6 to 8. Figures 9 to 11 show how, as a function of ΔT

Reference simulation (Case A) Variant (Case B) Variant (Case C)
Inspection The quality Q' has been replaced by Q, with: such as to see better the infl uence of the technique quality. Thus the value Q' = 8 used above corresponds to Q = 3. The same patterns can be seen for the two probability values and the shape of the surfaces is compatible with what has been observed previously on fi gure 5. For instance, the number of false alarms decreases (since the techniques induces some noise)   when the ΔT delay increases. However, for a good or very good technique (Q > 5 or Q' > 32), the number of false alarms is very low. Figure 11 shows that the total number of failures normally increases when ΔT delay increases, but it increases less when a good technique is used.
It can also be seen that changing the p FA value has two adverse consequences in terms of total cost: a more severe probability of failure decreases the number of failures when it increases the number of repairs (the average service life is reduced since the threshold value of the defect size for repair is lowered). This proves that this value of the allowable probability is of worth to be considered for optimisation.
Since all results regarding optimal costs also depend on what is assumed on single costs (k i , k r and k f coeffi cients introduced in equations (12) to (14)), it has been decided to choose some sets of values for these coeffi cients and to look at optimality for the {ΔT, Q, p FA } set. For practical applications, it will of course be necessary to identify, from fi eld data, the real relative ratio between inspection, repair and failure cost corresponding to any real asset management project. Figure 12 shows what has been obtained with two slightly different sets: • Set A: k i = 0.0007, k r = 0.1, a = 0.06, k f = 1 • Set B: k i = 0.0007, k r = 0.1, a = 1.5, k f = 1.
The main difference between the two sets is the a coeffi cient (cost exponent defi ned in equation (12)), which quantifi es how the inspection cost increases with the quality of the technique. A high value of this coeffi cient (Set B) penalises very good techniques whose technical effi ciency for detection of critical defects is not counterbalanced against their cost. The surfaces have been built by considering four values of p FA = {10 -1 , 10 -2 , 10 -3 , 10 -4 } and by keeping that corresponding, for each {ΔT, Q} set, to the lower cost.
The optimal solution, respectively, corresponds to the range {p FA , ΔT, Q} = {10 -3 , 5, 6-7} with the Set A and to {p FA , ΔT, Q} = {10 -3 , 6, 3-4} with the Set B, confi rming the high sensitivity of the optimal strategy to the input values. In fact, this shows that, if the cost of inspections quickly increases with their quality (Set B), it is better to use a rough technique and to have a large delay between inspections, which can cause a signifi cant number of failures. On the contrary, if the quality of the inspection does not imply a high additional cost, this justifi es to use more frequently a more accurate technique, thus reducing the number of failures (Set A). The observed number of failures decreases from about 50 (Set B) to less than 20 (Set A) for 100,000 years of simulation, which comes to an annual probability of failure of about 0.5 10 -3 (Set B) to 0.2 10 -3 (Set A). These values are well below the p FA , values since they include the effects of preventive maintenance (when the defects would lead to a larger probability of failure, inspection and maintenance tend to renew the component).

CONCLUSIONS
This study demonstrates that the cost and the quality of inspections have a signifi cant infl uence on the optimum delay between inspections, and hence the minimum annual total cost of the structure over its lifetime.
In general, inspection-based maintenance provides a more economical alternative to systematic maintenance, as repairs are only carried out based on the results obtained from inspections. Also, the extent of a repair is based on the size of the defect that was detected by the inspection. However, it has to be accounted for that the inspection-based maintenance requires that the repair decision is taken on fi rm basis, assuming correct information provided by accurate techniques. If it is not the case, over-cost can be induced by too conservative decisions. Of course, since many parameters describe the whole system (defect growth, detection process, manager's choices, economics), it is not possible to draw universal conclusions. It is possible to study the infl uence of few parameters while considering the others as constant, but many coupling effects exist and the quantitative results (regarding optimality) depend on these assumptions. It is the reason why it can be interesting to devote more specifi c studies to the relations between inspection quality and inspection cost. The simulation, however, offers, in its present state, a very powerful tool for a better understanding of the role of each parameter in a such complex context. Therefore, to build the optimum maintenance management plan, it is necessary to have accurate models for deterioration, inspections and repair, and to have a good understanding of the infl uencing parameters involved in each of these stages. Further work is being carried out as part of the French SENSO project to investigate the quality of information provided by a variety of NDT techniques on several parameters quantifying the condition of concrete.

DENYS BREYSSE
Denys Breysse is a civil engineering professor. His fi elds of interest are variability in materials, soils and structures, and the consequences on uncertainties on the overall response of a system (structure or asset). He is president of the French Association of Civil Engineering Universities (AUGC).

SIDI MOHAMMED ELACHACHI
Sidi Mohammed Elachachi joined University of Bordeaux 1 in 2006 and is currently an associate professor in Civil Engineering. His fi elds of interest are soil-structure interaction (including spatial variability), management systems, structural reliability and structural optimisation. He has supervised three PhDs and 12 Masters students in the areas mentioned above.

EMMA SHEILS
Emma Sheils is in her third year of studying for a PhD in the Department of Civil Structural and Environmental Engineering in Trinity College Dublin. Her PhD is on "The Development of a Two Stage Inspection Process for the Optimisation of Maintenance Planning for Infrastructure Elements & Networks".

FRANCK SCHOEFS
Dr Franck Schoefs is an assistant professor at the Institute for Research in Civil and Mechanical Engineering (GeM) within the Nantes Atlantic University in France since 1998. He graduated in 1992 from Ecole Normale Superieure de Cachan with the best student in civil engineering award. He received his PhD in 1996. Franck's interests include stochastic fi nite element of random geometries (X-SFEM); matrix response surfaces for the modelling of wave and wind extreme loading on framed structures (storms); probabilistic deterioration modelling of steel and concrete in coastal areas; probabilistic modelling of inspection results; risk analysis of deteriorated structures; and probabilistic modelling of loading from structural monitoring.

ALAN O'CONNOR
Dr Alan O'Connor is Senior Lecturer in Structural and Bridge Engineering at the University of Dublin, Trinity College. He is a Chartered Engineer with considerable experience in safety assessment of existing structures. He has authored over 80 technical papers, has spoken at numerous international conferences, and is the recipient of numerous awards and research contracts. His research interests include probability-based assessment and optimised maintenance management for structures, structural robustness, and modelling of stochastic systems.