Soft computing based multilevel strategy for bridge integrity monitoring

Soft computing based multilevel strategy for bridge integrity monitoring

• 1.
  Computer-Aided Civil andInfrastructure Engineering 25 (2010) 348–362 Soft Computing Based Multilevel Strategy for Bridge Integrity Monitoring S. Arangio & F. Bontempi∗ Department of Structural and Geotechnical Engineering, University of Rome “La Sapienza,” Via Eudossiana 18, Rome, Italy Abstract: In recent years, structural integrity monitor- ing has become increasingly important in structural en- gineering and construction management. It represents an important tool for the assessment of the dependability of existing complex structural systems as it integrates, in a unified perspective, advanced engineering analyses and experimental data processing. In the first part of this work the concepts of dependability and structural integrity are discussed and it is shown that an effective integrity assess- ment needs advanced computational methods. For this purpose, soft computing methods have shown to be very useful. In particular, in this work the neural networks model is chosen and successfully improved by apply- ing the Bayesian inference at four hierarchical levels: for training, optimization of the regularization terms, data- based model selection, and evaluation of the relative im- portance of different inputs. In the second part of the ar- ticle, Bayesian neural networks are used to formulate a multilevel strategy for the monitoring of the integrity of long span bridges subjected to environmental actions: in a first level the occurrence of damage is detected; in a fol- lowing level the specific damaged element is recognized and the intensity of damage is quantified. 1 INTRODUCTION The realization of high-cost and safety-critical construc- tions requires advanced approaches to take into ac- count their intrinsic complexity (Ciampoli, 2005). The complexity of this kind of structures can be related to several aspects, as for example, nonlinear dynamic be- havior (Adeli et al., 1978), various sources of uncertain- ties, both objective and cognitive, and strong interaction between components. ∗To whom correspondence should be addressed. E-mail: franco. [email protected]. Only by considering these aspects can a consistent evaluation of the structural performance be obtained. Therefore, it is necessary to evolve from the simplis- tic idealization of the structure as “device for channel- ing loads” to the analysis of the structural system as a whole, intended as “a set of interrelated components working together toward a common purpose” (NASA System Engineering Handbook, 2007). The correlation between different aspects can be taken into account by applying the principles and techniques of System Engi- neering, which is a robust approach to the creation, de- sign, realization, and operation of an engineered system (Bontempi et al., 2008). If the entire design process needs to be reviewed in the System Engineering framework, one includes re- quirements concerning the construction phase and the operation and maintenance during the whole life-cycle (Sarma and Adeli, 2002). To this aim, data collected on site are important both for checking the accomplish- ment of the expected performance during the service life and for validating the original design (Smith, 2001). This approach requires the definition of the quality of a complex structural system in a comprehensive way by an integrated concept, like dependability. The concept of dependability has been originally developed in the field of Computer Science and it is extended to struc- tural engineering as “the ability to deliver service that can justifiably be trusted” (Avižienis et al., 2004). This definition stresses the need for justification of trust. The alternate definition considers dependable “a system that has the capability to avoid service failures which are more frequent and more severe than acceptable.” All these factors are connected to the integrity of the structural systems, considered as the completeness and consistency of the structural configuration. Specif- ically, structural integrity refers “to the safe opera- tion of engineering components, structures and mate- rials, and addresses the science and technology which is C 2010 Computer-Aided Civil and Infrastructure Engineering. DOI: 10.1111/j.1467-8667.2009.00644.x
• 2.
  Multilevel strategy forbridge integrity monitoring 349 used to assess the margin between safe operation and failure.” During the service life the integrity, and conse- quently the overall dependability, can be lowered by deterioration and damage. The structural monitoring represents an essential tool to assess the evolution in time of the dependability of existing structural systems (Soyoz and Fukuda, 2009; Li et al., 2006). It includes issues like definition and analysis of the structural per- formance, from regular exercise to out-of-service and collapse, assessment of the environmental conditions, choice of the sensor systems and their optimal place- ment, use of data transmission systems and signal pro- cessing techniques, and methods for damage identifica- tion and model updating (Jiang and Adeli, 2005; Adeli and Jiang, 2006; Psimoulis and Stiros, 2008). In case of complex structural systems it can be diffi- cult to deal with the huge quantity of data coming from the monitoring process and various soft computing tech- niques have shown to be effective tools for data pro- cessing (Adeli and Jiang, 2006; Carden and Brownjohn, 2008; He et al., 2008; Jiang and Adeli, 2008a). In this article, a soft computing model, the Bayesian neural networks (Castillo et al., 2008; Adeli and Panakkat, 2009), is used to formulate a multilevel strat- egy for the assessment of the integrity of a long span suspension bridge subjected to wind actions and traffic loads. In the first step of the proposed strategy the oc- currence of damage is detected and the damaged por- tion of the bridge is identified; in the second step the specific damaged element is recognized and the inten- sity of damage evaluated. In the following, the concept of integrity monitoring for dependability is explained with reference to struc- tural systems and the multilevel strategy is illustrated. 2 STRUCTURAL INTEGRITY MONITORING FOR DEPENDABILITY For complex structural systems, where there are signif- icant dependencies among elements or subsystems, it is important to have a solid knowledge of both how the system works as a whole, and how the elements behave individually. In this contest, dependability is an inte- grated property that includes and describes the relevant aspects with reference to the system quality and its in- fluencing factors (Bentley, 1993). System dependability can then be thought of as being composed of three ele- ments (Figure 1): 1. the attributes, that is, the properties that quantify the dependability; 2. the threats, that is, the elements that can affect de- pendability; ATTRIBUTES THREATS MEANS MAINTAINABILITY RELIABILITY SAFETY AVAILABILITY FAILURE ERROR FAULT FAULT TOLERANT DESIGN FAULT DETECTION FAULT DIAGNOSIS FAULT MANAGING DEPENDABILITY Fig. 1. Dependability: attributes, threats, and means. 3. the means, that is, the tools that can be used to in- crease dependability. In structural engineering, relevant attributes are re- liability, safety, availability, and maintainability. These properties are essential to guarantee the safety of the system under relevant hazard scenarios, the survivabil- ity under accidental or exceptional scenarios, and the functionality under operative conditions. The threats for system dependability can be subdi- vided into faults, errors, and failures. According to the definition given in Avižienis et al. (2004), an active or dormant fault is a defect or an anomaly in the system behavior that represents a potential cause of error; an error is the cause for the system being in an incorrect state and it may or may not cause failure; failure is a per- manent interruption of the system ability to perform a required function under specified operating conditions. The problem of conceiving and building a dependable structural system can be considered at least by four dif- ferent points of view: 1. how to design a dependable system, that is a fault tolerant system; 2. how to detect faults, that is, anomalies in the sys- tem behavior; 3. how to localize and quantify (that is, diagnose) the effects of faults and errors; 4. how to manage faults and errors to avoid failures. This article is focused on points 2 and 3: fault de- tection and fault diagnosis. These aspects are strictly related to the integrity monitoring of the structural sys- tem: an efficient integrity monitoring system is expected
• 3.
  350 Arangio Bontempi to be able to preserve the structural dependability, diag- nosing deterioration and damage at their onset (Ou and Li, 2006). Even if there is no general consensus on its defini- tion, in analogy with biological systems, an intelligent monitoring system is expected to (Aktan et al., 1998; Isermann, 2006): 1. sense the loading environment as well as the struc- tural response; 2. reason by assessing the structural condition and health; even small faults should be detected and diagnosed; 3. communicate through proper interface with other components and systems; 4. learn from experience as well as by interfacing with humans for heuristic knowledge; 5. decide and take action for alerting controllers in case of accidental situations, or activate fault tol- erant configurations in case of reconfigurable sys- tems. Structural monitoring has a key role in the mainte- nance scheduling of the bridge structures and a great re- search effort has been devoted in the past 30 years to es- tablishing effective local and global methods for health monitoring in civil structures (Doebling et al., 1996; De Roeck 2003; Sohn et al., 2004, 2008; Jiang and Adeli, 2007; Li and Wu, 2008; Moaveni et al., 2008). Analyzing the problem in terms of the expected pay- off, it comes out that, in cases of complex structures, like long span bridges, for example, the monitoring process should be planned during the design phase and should be carried out during the entire life cycle to assess the structural health and performance under in-service and accidental conditions (Bontempi et al., 2008). This long-term monitoring of bridges, where long- term designates a period of time from 1 year to decades and desirably the entire life cycle, is a quite recent concept, enabled by recent advances in sensing, data acquisition, computing, communication, data, and infor- mation management (Ou and Li, 2006). Exploring long- term monitoring of structural responses was pioneered in China and in Japan (Abe and Amano, 1998; Lau et al., 1999; Wong et al., 2000). Nowadays several bridges are instrumented in Europe (Casciati, 2003), the United States (Aktan et al., 2002), Korea, and other countries, and the administration of the major coun- tries have developed guidelines to explain the advan- tages of long-term monitoring and to help the engineers in building effective monitoring systems (Aktan et al., 2002; Mufti, 2001; ISO, 2002; Task Group 5.3, 2002). In accord with the concepts reported in these guide- lines, long-term monitoring is based on the integration of different kinds of technologies (Figure 2): experimen- tal, analytical, and information technologies. Mathematical modeling Finite Element modeling ANALYTICAL TECHNOLOGIES INFORMATION TECHNOLOGIES EXPERIMENTAL TECHNOLOGIES Non-destructive evaluation Continuous monitoring Data acquisition Data processing Communication Interpretation Fig. 2. Issues in long-term monitoring implementation. Fig. 3. Steps of the information technology. Experimental technologies include nondestructive vi- sual inspection and continuous monitoring. Analytical technologies include mathematical and finite-element modeling. The last one, the information technology, assumes a key role: it covers the entire spectrum of efforts related to the acquisition, communication, pro- cessing, and interpretation of the data (Figure 3). The entire monitoring process needs a team of experts in civil, mechanical, and electrical engineering and com- puter scientists working together to take full advantage of the data. In fact, the desired outcome of structural monitoring is not data collection, but it is the generation of information and the creation of a base of knowledge about potential and existent system symptoms that will enhance decision making for fault management. 3 FAULTS-SYMPTOMS RELATIONSHIP As mentioned in the previous section, to detect and di- agnose a system fault, it is necessary to process the data
• 4.
  Multilevel strategy forbridge integrity monitoring 351 Fig. 4. Fault–symptoms relationship. coming from the monitoring process, that is, the sys- tem symptoms. However this is a complex task. The relationship between fault and symptoms can be repre- sented graphically by a pyramid (Figure 4). The vertex represents the fault, and the lower levels the possible events generated by the fault; the base corresponds to the symptoms. The propagation of the fault to the ob- servable symptoms follows a cause–effect relationship, and is a top–down forward process: a fault determines events that, as intermediate steps, influence the measur- able or observable symptoms (Isermann, 2006). On the other hand, the fault diagnosis proceeds in the reverse way (Figure 4); it is a bottom–up inverse process that relates the observed symptoms to the faults. This im- plies the inversion of the causality principle. However, one cannot expect to rebuild the chain only by measured data because usually the causality is not reversible or the reversibility is ambiguous (Füssel, 2002): the underlying physical laws are often not known in analytical form, or are too complicated for explicit numerical calcula- tion. Moreover, intermediate events between faults and symptoms are not always recognizable (Figure 4, right- hand side). The solving strategy requires integrating different procedures, either forward or inverse: this mixed solv- ing approach has been called total approach by Liu and Han (2004) and different computational techniques have been developed for this task (Adeli and Samant, 2000; Ghosh-Dastidar and Adeli, 2003). 4 KNOWLEDGE-BASED FAULT DETECTION AND DIAGNOSIS As shown in the previous section, fault diagnosis needs the integration of forward and inverse proce- dures with the heuristic knowledge coming from ex- perience or qualitative information. For this task, a knowledge-based analysis can be applied (Adeli and Fig. 5. Knowledge-based analysis for structural integrity monitoring. Balasubramanyam, 1988; Paek and Adeli, 1990; Adeli and Hawkins, 1991; Shwe and Adeli, 1993; Waheed and Adeli, 2000; Aktan et al., 1998) (Figure 5). The results obtained by visual inspection or instrumented monitor- ing (the inverse diagnosis system of Figure 4) are pro- cessed and combined with the results coming from the analytical model (the forward physical system of Fig- ure 4). Information technology provides the tool for such integration. The output of the information technol- ogy is then filtered by the available heuristic knowledge for decision making. An attractive aspect of the knowledge-based analysis is that it can cope with incomplete or uncertain data in- tegrating qualitative and quantitative information, com- ing from modeling and heuristics. To carry out the var- ious phases, different computational methods can be used. In several applications, inference models and soft computing techniques, like the Bayesian neural net- works used in this work, have shown their effectiveness (Adeli and Park, 1995; Pandey and Barai, 1995; Masri et al., 1996; Faravelli and Pisano, 1997; Hajela, 1999; Topping et al., 1999; Kim et al., 2000; Adeli, 2001; Ni et al., 2002; Kao and Hung, 2003; Waszczyszyn and Ziemianski, 2005; Ko and Ni, 2005; Xu and Humar, 2006; Lam et al. 2006; Jiang and Adeli, 2008a,b).
• 5.
  352 Arangio Bontempi 5 BAYESIAN NEURAL NETWORK FOR FAULT DETECTION AND DIAGNOSIS 5.1 The neural network model and the probability logic framework The neural network concept has its origins in attempts to find mathematical representations of information processing in biological systems. Actually, there is a def- inite probability model behind it: a neural network is an efficient statistical model for nonlinear regression (Cheng and Titterington, 1994). It can be described by a series of functional transformation working in different correlated layers (Bishop, 2006). For example, for two layers yk(x, w) = h ⎛ ⎝ M j=1 w (2) kj g D i=1 w (1) ji xi + b (1) j0 + b (2) k0 ⎞ ⎠ (1) where yk is the kth output variable in the output layer; x is the vector of the D input variables in the 1input layer; w is the matrix including the adaptive weight pa- rameters w (1) ji and w (2) kj and the biases b (1) j0 and b (2) k0 (the superscript refer to the considered layer); M is the to- tal number of units in the hidden layer. The quantities within the brackets are the so-called activations; each of them is transformed using a nonlinear activation func- tion (h and g). The nonlinear activation functions are generally chosen to be sigmoidal or tanh functions be- cause of the so-called universality property (Cybenko, 1989). In the traditional learning approach, the values of the parameters w are obtained during the training phase by minimizing an error function (Adeli and Hung, 1994), for example, the sum of squared errors with weight de- cay (Bishop, 1995) E = 1 2 N n=1 No k=1 yk (xn ; w) − tn k 2 + α 2 W i=1 |wi |2 (2) where yk is the kth neural network output correspond- ing to the n-th realization of x, tn k is the relevant target value, N is the size of the considered data set, N0 is the number of output variables, W is the number of param- eters in w, and α is a regularization parameter. The sec- ond term in the right-hand side is a decay regularization that penalizes large weights. Neural network learning can be framed as Bayesian inference, where probability is treated as a multival- ued logic that may be used to perform plausible infer- ence (Jaynes, 2003). The roots of this probability logic approach are in the work by Bayes published in 1763 (Bayes, 1763). He presented a method for updating probability distributions based on available data that would come to be known as Bayes’ theorem, and that forms the foundation of a framework for probabilis- tic inference. The power of this theorem was shown by Laplace (1812) and Jeffreys (1939) who applied it to the analysis of real data set. Although this frame- work had its origin in the 18th century, the practical application of Bayesian methods was for a long time severely limited by the difficulties in carrying through the full Bayesian procedure. The developments of ap- proximation theories and stochastic sampling methods, along with dramatic improvements in the power of com- puters, have recently opened the door to the practi- cal use of Bayesian techniques in an impressive range of applications across all disciplines. In recent years in civil engineering, for example, the probability logic ap- proach has been successfully applied to system identifi- cation problems and structural health monitoring (Beck and Katafygiotis, 1998; Beck and Yuen, 2004; Muto and Beck, 2008). Starting from the early works of MacKay (1992) and Buntine and Weigend (1991), there has been a growing interest for the application of this framework in the field of neural networks methods (MacKay, 1994; Neal, 1996; Lampinen and Vethari, 2001; Barber, 2002; Lee, 2004; Nabney, 2004). To pose the neural network model within the Bayesian framework, the learning process needs to be interpreted probabilistically: the network output can be considered as the conditional average of the target data (Bishop, 1995). Because the model does not reproduce the data set exactly, the error ε = t – y(x; w) between the target value t and the network output y needs to be interpreted probabilistically using a prediction-error probability model: a Gaussian distribution with mean zero and constant inverse variance β = 1/σD 2 is a model supported by the principle of maximum differential en- tropy (Jaynes, 2003). Thus, modeling the predictions as independent and identically distributed (i.i.d.), the like- lihood function for a data set D = xn , ,tn is given by p(D | w, β, M) = β 2π
• 6.
  N·N0 exp − β 2 N n=1 No k=1 yk (xn ; w)− tn k 2 (3) where M denotes the Bayesian model class that specifies the forms of the likelihood function and the prior prob- ability distribution discussed next. Although the like- lihood function does take into account the uncertain prediction error, it does not quantify the uncertainty in the values of the parameters w. In the Bayesian frame- work, this can be represented by a prior PDF p(w | M) over the parameters w, which expresses the relative
• 7.
  Multilevel strategy forbridge integrity monitoring 353 Fig. 6. Learning as inference. plausibility of each value. Because generally there is a little idea of what the values should be, it is usual to se- lect the prior as a rather broad distribution. Using once again the principle of maximum differential entropy, this requirement suggests a Gaussian prior distribution with zero mean of the form p(w | α, M) = α 2π W/2 exp − α 2 |w|2 (4) where α = 1/σ2 W represents the inverse variance of the distribution. Using available data, Bayes’ theorem up- dates the prior probability distribution over the parame- ters p(w | α, M) to give the posterior PDF p(w | D, α, β, M): p(w | D, α, β, M) = p(D | w, β, M) p(w | α, M) p(D | α, β, M) . (5) This posterior distribution is always more compact than the prior distribution if the data informs the model, as indicated schematically in Figure 6, expressing the fact that something has been learned. Therefore, by maximizing the posterior, the most plausible values of the parameters wMAP can be found. Instead of finding a maximum of the posterior prob- ability in Equation (5), it is usually more convenient to seek instead a minimum of its negative logarithm. As shown in Figure 6, for the chosen prior distribution and likelihood function, the negative log probability is just the usual sum of squares function in Equation (2). Therefore, the conventional learning approach can be derived as a particular approximation of the Bayesian framework where only the MAP (maximum a posteri- ori) parameter values are utilized. 5.2 Bayesian enhancements for neural networks The optimization of the parameters w, that is, the so- called model fitting, is only one level of inference where Bayesian approach can be applied to neural networks. The potential enhancements that can be obtained by ap- plying this framework at further levels in a hierarchical fashion are often not appreciated. The various levels can be summarized as follows (Arangio, 2008): 1. Level 1: Model fitting: task of inferring appropriate values for the model parameters, given the model and the data. 2. Level 2: Optimization of the regularization terms α and β that make level 1 a better conditioned in- verse problem. 3. Level 3: Model class selection: the Bayesian ap- proach allows an objective comparison between models using alternative network architectures. 4. Level 4: Automatic relevance determination (ARD): the relative importance of different inputs can be determined using separate regularization coefficients.
• 8.
  354 Arangio Bontempi Regarding the first two levels, the traditional and the Bayesian framework usually give equivalent results (MacKay, 1992). The addition of the third level, the model class selection, has shown to be very effective. In fact, the number of adaptive parameters of the net- work model, that is, the model class, has to be fixed in advance, and this choice has a fundamental importance. It is not correct to choose simply the model that fits the data better: more complex models will always fit the data better but they may be over-parameterized and so they make poor predictions for new cases. The problem of finding the optimal number of param- eters provides an example of Ockham’s razor, which is the principle that one should prefer simpler models to more complex models, and that this preference should be traded off against the extent to which the models fit the data (Sivia, 1996). The best generalization perfor- mance is achieved by the model whose complexity is neither too small nor too large. The third level of inference mentioned above deals with this task: the Bayesian framework provides an objective and structured framework for dealing with the issue of model complexity, and allows an objec- tive comparison between models with alternative net- work architectures (Beck and Yuen, 2004). The most plausible model class among a set M of NM candi- date ones is obtained by applying Bayes’ Theorem as follows: p(Mj | D, M) ∝ p(D | Mj ) p(Mj | M) . (6) The factor p(D | Mj) is known as the evidence for the model class Mj provided by the data D. Equation (6) shows that the most plausible model class is the one that maximizes p(D | Mj)p(Mj) with respect to j. If there is no particular reason a priori to prefer one model over another, they can be treated as equally plausible a priori and a noninformative prior, that is, p(Mj) = 1/NM, can be assigned; then different models can be compared just by evaluating their evidence (MacKay, 1992). Once the optimal architecture has been determined, the last issue that should be considered is the relative importance of each input variable. If the available data comes from real systems it could be difficult to separate the relevant variables from the redundant ones. In the Bayesian framework, this problem can be addressed by the ARD method, proposed by Mackay (1994) and Neal (1996). To use this technique, a separate hyperparame- ter αi is associated with each input variable: this value represents the inverse variance of the prior distribution of the parameters related to that input. In this way, ev- ery hyperparameter explicitly represents the relevance of one input: a small value means that large parameters are allowed and the corresponding input is important; on the contrary, a large value means that the parameters are constrained near zero, and hence the corresponding input is less important. The ARD allows a fourth level of inference to be ap- plied to the neural networks model. Once the architec- ture of the model is defined, the importance of every in- put is evaluated: if some hyperparameter is very large, the related input will be dropped from the model and the optimal architecture for the new model will be re- estimated. The four levels of inference are summarized in the flowchart in Figure 7. Starting from the simple process of model fitting, further steps have been added to in- clude the other three levels of inference: evaluation of the hyperparameters, model class selection, and ARD. More details can be found in Arangio (2008). The improvements that can be obtained by applying the first three levels are well documented in the exist- ing literature (MacKay, 1992, 1994). On the contrary, the fourth level is usually applied independently and in this way the benefits of an integrated approach are not fully exploited. In this work the evaluation of the relative importance of each input is included in the it- erative process. In this way, once the optimal architec- ture of the model is defined, it is possible to recognize eventual redundant parameters and drop them from the model. 6 MULTILEVEL STRATEGY FOR BRIDGE INTEGRITY ASSESSMENT The Bayesian neural networks discussed in the previous section is applied in a multi-step strategy for the assess- ment of the integrity of the long suspension bridge in Figure 8 (Arangio, 2008). The considered bridge has a main span of 3,300 m and it carries six road lanes in the external box girders and two railway tracks in the cen- tral one; detailed information on the bridge project and its history can be found in Bontempi (2006). A multi-step approach has been followed because it has been shown that is more effective to consider inde- pendently the tasks of damage detection, location, and quantification (Ceravolo et al., 1995; Ko et al., 2002). In the first step of the strategy the occurrence of damage or anomalies in the bridge is detected, and the damaged portion of the structure is identified. If some damage is detected, the second step of the procedure is initiated: using a pattern recognition approach, the specific dam- aged member within the whole area is identified, and the extent of damage is evaluated. The entire procedure has been carried out working on a finite-element model of the bridge but it could be applied in the same way to an existent structure.
• 9.
  Multilevel strategy forbridge integrity monitoring 355 Fig. 7. Hierarchical Bayesian framework for neural networks. Fig. 8. Steps of the damage identification strategy. 6.1 Step 1: Damage detection In the first step of the proposed strategy, the response of the structure is monitored at various measurement points, located at groups of three (A, B, and C) every 30 m along the bridge deck. One neural network for each intermediate point (B) is built and trained using the time-histories of the response of the structure sub- jected to wind actions and traffic loads (due to the pas- sage of a train) in the undamaged situation. The time- histories of selected structural response parameters are sampled at regular intervals, thus generating series of discrete values. A set of such values from the instant t – k to t is used as input for the network models, and the value at the instant t + 1 is used as the target output (left-hand side of Figure 9). Then, the trained models are tested on new input pat- terns, corresponding to different time intervals and to
• 10.
  356 Arangio Bontempi Fig. 9. Flowchart of the chart of the Step 1 procedure for damage detection. both undamaged and damaged situations. For each pat- tern, the set of values from ft+n−k to ft+n−1 is used as input, and the value ft+n is predicted and compared with the target one. If the error in the prediction is negligible, the struc- ture is considered as undamaged; if the error is higher than a threshold value (eventually defined according to expert opinion), the presence of an anomaly is detected (Figure 10). The anomaly may correspond to a damage state or simply to a change of the characteristics of the exci- tation. To distinguish the changes in the structural re- sponse due to variations in the excitation from those due to damage, the prediction errors are checked in all mea- surement points, according to the procedure schemati- cally represented in the flowchart of Figure 9. If the prediction is wrong in several locations, that is the difference e between the mean value of the errors in training and testing is different from zero in different measurement points, it can be concluded that the char- acteristics of the excitation are probably different from those assumed, and the trained neural network models are unable to represent the actual time-history of the re- sponse parameters. In this case, the models need to be updated according to the new excitation. On the other hand, if the difference e is large only at one or a few points and generally decreases with the distance from those points, it can be concluded that the considered portion is damaged. To illustrate the proposed approach, data is simu- lated using a dynamic model of the suspension bridge where damage is implemented as a reduction of stiff- ness of a structural element. The following scenarios are considered: 1. Hangers: reduction of stiffness from 5% to 80%; 2. Cables: reduction of stiffness from 1% to 10%; 3. Transverse beam: reduction of stiffness from 5% to 30%. The training data set for every network model in- cludes 1,000 samples of the time-history of the response parameters that were found to be the most sensitive to a stiffness reduction (Arangio and Petrini, 2007), that is the rotation of the deck around the longitudinal axis in case of wind actions, and the vertical displacements of the deck in case of traffic loads. 6.2 Step 2: Identification of damage location and severity Having recognized that a portion of the structure is damaged, the second step of the procedure is initiated; it is aimed at identifying the specific damaged element
• 11.
  Multilevel strategy forbridge integrity monitoring 357 Fig. 10. Location of the measurement points on the bridge deck and identification of the damaged portion by considering the errors in the approximation; also shown the potentially damaged elements of each portion. Fig. 11. Neural network for the identification of damage location and intensity. (a hanger, the cable, or a transverse beam), and at eval- uating the damage intensity. A pattern recognition ap- proach is used. To create the training data set, the errors in Step 1 obtained by the neural network approximation of the response time-histories at three different points of the damaged portion (A, B, and C in Figure 11) are col- lected, by considering different damage scenarios. For each damage scenario, the training data set has as input the mean values of the errors in A, B, and C, and, as output, a vector including the five possible locations of damage and its intensity (Figure 11). 7 RESULTS OF THE INTEGRITY ASSESSMENT PROCEDURE 7.1 Results of step 1: Damage detection The different network models were trained using the time-histories of the response of the bridge in undam- aged conditions. The network architecture has been de- termined by the Bayesian approach discussed in Sec- tion 5: the optimal network models consist of 2, 2 and 1 nodes in the input, hidden and output layers, respec- tively. An example of the evolution in time of the differences between the predicted and the target values in the sets of training and test data is reported in Figures 12 and 13; both undamaged and damaged conditions are con- sidered. It is possible to notice that when time-histories related to various damage scenarios are proposed to the trained networks the errors in the approximation in- crease. There is a difference e between the mean val- ues of the error in undamaged and damaged conditions. In Figures 14 to 16 the increments e of the mean values of the error with respect to the undamaged situ- ation are shown for different levels of damage in the ca- bles, the hangers, and the transverse beam. Both wind actions and traffic loads are considered and the results are compared.
• 12.
  358 Arangio Bontempi 0.0 0.3 0.6 0.9 0 20 40 60 80 Training error Test error (undamaged) err t [s] Fig. 12. Differences between the network values and the correct value in case of undamaged structure. 0.0 0.3 0.6 0.9 0 20 40 60 80 Training error Test error (damaged) err t [s] Δe mean -damaged mean -undamaged Fig. 13. Differences between the network values y and the correct value t in damaged conditions in a case example (considered damage: 5% reduction of stiffness in one cable). (a) Damage intensity (%) – cable (pos 1/5) 0.0 0.3 0.6 0.9 1.0% 3.0% 5.0% 10% train wind Δe Fig. 14. Increment of the error in the approximation of the response time-history of the cable under wind actions and traffic load. 0.00 0.03 0.06 0.09 20% 40% 50% 80% train wind (c) Damage intensity (%) – hanger (pos 2/4) Δe Fig. 15. Increment of error in the approximation of the response time-history of the hanger under wind actions and traffic load. 0.00 0.03 0.06 0.09 5% 10% 30% 50% train wind (b) Damage intensity (%) – transverse beam (pos 3) Δe Fig. 16. Increment of error in the approximation of the response time-history of the transverse beam under wind actions and traffic load. Looking at the results shown in Figures 14 to 16, it is possible to note that the proposed method is more ef- fective when responses from high speed excitation (like traffic) are considered instead of responses due to slow speed excitation (like wind). Thus, in the following step, only the structural response due to the passage of one train is considered. 7.2 Results of step 2: Identification of damage location and intensity Once the damaged portion of the whole structure is rec- ognized, the specific damaged element and the intensity of damage are identified using a pattern recognition ap- proach. Various damage scenarios, corresponding to the reduction of the stiffness in the hangers, the cables, and the transverse beam in the identified damaged portion is simulated, and a training set consisting of 370 exam- ples is created. The network architecture is always de- termined by the Bayesian approach discussed in Section 5. The optimal network model has 11 units in the hidden layers. After the training phase the network is tested with 30 new input vectors that are not included in the training set, and the related damage scenarios are obtained and compared with the target ones. To give a global and in- tuitive representation of the results, two quantities are defined: 1. The position, which gives a measure of the error in the location: pos(i) = t × y |t| · |y| (7) 2. The intensity, which gives a measure of the error in the quantification: int(i) = |t| t|y| . (8) If these quantities are equal to one, the damage is well localized and its intensity is correctly estimated. These
• 13.
  Multilevel strategy forbridge integrity monitoring 359 0.0 0.2 0.4 0.6 0.8 1.0 1.2 0 10 20 30 Test number pos Fig. 17. Identification of the damage position in the test examples. 0.0 0.3 0.5 0.8 1.0 1.3 1.5 0 10 20 30 Test number int Fig. 18. Identification of the intensity in the test examples. quantities are evaluated for each of the 30 test samples and the results are shown in Figures 17 and 18. The lo- cation can be detected in almost 90% of the considered cases and the intensity is correctly estimated in 66% of the cases. 8 CONCLUSIONS In this work the concept of dependability has been dis- cussed and its original meaning has been extended to the structural engineering field. It has been shown that this term describes the overall quality performance of a complex structural system and its influencing factors in an integrated way. The different aspects related to dependability are strictly connected with the concept of structural in- tegrity. During the service life the integrity, and conse- quently the dependability, can be lowered by damages. The structural monitoring represents an essential tool to assess the evolution in time of the dependability of existing structural systems. Fundamental tasks of integrity monitoring are fault detection and diagnosis. Fault diagnosis from experi- mental data is an inverse problem and the reconstruc- tion of the fault-symptom chain can be very difficult. A solution can be achieved by applying a knowledge- based procedure that integrates the solving procedures with the heuristic knowledge coming from experience or qualitative information. For this task, different soft computing methods can be suitable. In particular, in this work, the Bayesian neural network model has been used to formulate a hierarchical integrity assessment strategy. The proposed approach has been applied for the anal- ysis of the time-history of the response of a long span suspension bridge subjected to ambient excitations. The strategy could be useful for damage identification of large structural systems instrumented with on-line mon- itoring systems. The presented example case has been developed on a numerical model of the structure but the strategy can be applied on real structural systems as well: various neural networks models could be selected and trained in a continuous way using the time-histories of the structural response; in this way the occurrence of anomalies can be detected almost in real time. When an anomaly is recognized, numerical simulations can be carried out to create the data set to develop the second step of the strategy. In this way experimental data are used for damage detection and the results of the numer- ical analyses can help to identify the damaged element and to quantify the intensity of damage. ACKNOWLEDGMENTS The authors wish to thank Professors H. Li (Harbin In- stitute of Technology), J.L. Beck (California Institute of Technology), F. Casciati, and L. Faravelli (Univer- sity of Pavia) for discussions related to this study. The reviewers of the article are acknowledged for the care- ful reading and the very useful suggestions. The sup- port of Prof. H. Adeli is also recognized. The financial support of University of Rome “La Sapienza” is also acknowledged. The opinions and the results presented here are however the responsibility only of the authors and cannot be assumed to reflect the ones of University of Rome “La Sapienza.” REFERENCES Abe, K. Amano, K. (1998), Monitoring system of the Akashi Kaikyo Bridge, Honshi Technical Report, 22(86), 29–34. Adeli, H. (2001), Neural networks in civil engineering: 1989– 2000, Computer-Aided Civil and Infrastructure Engineering, 16(2), 126–42. Adeli, H. Balasubramanyam, K. V. (1988), A novel ap- proach to expert systems for design of large structures, AI Magazine, pp. 54–63. Adeli, H., Gere, J. Weaver, W., Jr. (1978), Algorithms for nonlinear structural dynamics, Journal of Structural Divi- sion, ASCE, 104(ST2), 263–80.
• 14.
  360 Arangio Bontempi Adeli, H. Hawkins, D. (1991), A hierarchical expert sys- tem for design of floors in highrise buildings, Computers and Structures, 41(4), 773–88. Adeli, H. Hung, S. L. (1994), An adaptive conjugate gradi- ent learning algorithm for effective training of multilayer neural networks, Applied Mathematics and Computation, 62(1), 81–102. Adeli, H. Jiang, X. (2006), Dynamic fuzzy wavelet neural network model for structural system identification, Journal of Structural Engineering, ASCE, 132(1), 102–11. Adeli, H. Panakkat, A. (2009), A probabilistic neural net- work for earthquake magnitude prediction, Neural Net- works, 22, 1018–24. Adeli, H. Park, H.S. (1995), Optimization of space struc- tures by neural dynamics, Neural Networks, 8(5), 769– 81. Adeli, H. Samant, A. (2000), An adaptive conjugate gradi- ent neural network—wavelet model for traffic incident de- tection, Computer-Aided Civil and Infrastructure Engineer- ing, 15(4), 251–60. Aktan, A. E., Catbas, F. N., Grimmelsman, K. A. Pervizpour, M. (2002), Development of a Model Health Monitoring Guide for Major Bridges, Report DTFH61- 01-P-00347, Federal Highway Administration Research and Development, Drexel Intelligent Infrastructure and Transportation Safety Institute, Federal Highway Admin- istration, US Department of Transportation, available at: http://www.ishmii.org/Download%20Folder/FHWA%20 Guide%209-8%20-%20SHM%20Guidelines.pdf, accessed January 26, 2010. Aktan, A. E., Helmicki, A. J. Hunt, V. J. (1998), Issues in health monitoring for intelligent infrastructure, Smart Ma- terials and Structures, 7, 674–92. Arangio, S. (2008), Inference model for structural systems integrity monitoring: Neural networks and Bayesian en- hancements. Ph.D. thesis on Structural Engineering, Uni- versity of Rome “La Sapienza.” Arangio, S. Petrini, F. (2007), Application of neural net- works for predicting the structural response of a long sus- pension bridge subjected to wind actions, in Proceedings of the Third International Conference on Structural Engineer- ing, Mechanics and Computation, SEMC 2007, Cape Town, South Africa. Avižienis, I., Laprie, J. C. Randell, B. (2004), Dependabil- ity and its threats: A taxonomy, 18th IFIP World Com- puter Congress, Toulouse (France), Building the Infor- mation Society, Kluwer Academic Publishers, 2004, ISBN 1-4020-8156-1, 12, 91-120. Barber, D. (2002), Bayesian methods for supervised neural networks, in Handbook of Brain Theory and Neural Net- works, MIT Press, Cambridge, MA. Bayes, T. (1763), An essay towards solving a problem in the doctrine of chances, Philosophical Transactions of the Royal Society of London, 53, 370–418. Beck, J. L. Katafygiotis, L. S. (1998), Updating models and their uncertainties-Bayesian statistical framework, Journal of Engineering Mechanics, 124, 455–61. Beck, J. L. Yuen, K.-V. (2004), Model selection using response measurements: Bayesian probabilistic approach, Journal of Engineering Mechanics, 130, 192–203. Bentley, J. P. (1993), An Introduction to Reliability and Qual- ity Engineering, Longman, Essex, U.K. Bishop, C. M. (1995), Neural Networks for Pattern Recogni- tion, Oxford University Press, London, U.K. Bishop, C. M. (2006), Pattern Recognition and Machine Learning, Springer, New York. Bontempi, F. (2006), Basis of design and expected per- formances for the Messina Strait Bridge, Proceedings of BRIDGE 2006, Hong Kong [CD-ROM]. Bontempi, F., Gkoumas K. Arangio S. (2008), Systemic approach for the maintenance of complex structural sys- tems, Structure and Infrastructure Engineering, 4, 77–94, doi: 10.1080/15732470601155235. Buntine, W. Weigend, A. (1991), Bayesian back- propagation, Complex Systems, 5, 603–43. Carden, E. P. Brownjohn, J. M. W. (2008), Fuzzy clustering of stability diagrams for vibration-based structural health monitoring, Computer-Aided Civil and Infrastructure Engi- neering, 23(5), 360–72. Casciati, F. (2003), An overview of structural health monitor- ing expertise within the European Union, in Z. S. Wu and M. Abe (eds.), Structural Health Monitoring and Intelligent Infrastructure, Lisse, Balkema., pp. 31–7. Castillo, E., Menéndez, J. M. Sánchez-Cambronero, S. (2008), Traffic estimation and optimal counting lo- cation without path enumeration using Bayesian net- works, Computer-Aided Civil and Infrastructure Engineer- ing, 23(3), 189–207. Ceravolo, R., de Stefano, A. Sabia, D. (1995), Hierarchical use of neural techniques in structural damage recognition, Smart Material and Structures, 4, 270–80. Cheng, B. Titterington, D. M. (1994), Neural networks: A review from a statistical perspective, Statistical Science, 9(1), 2–54. Ciampoli, M. (2005), Reliability assessment of complex struc- tural systems, Proceedings of AIMETA 2005, Firenze Uni- versity Press, Florence, Italy, ISBN 88-8453-248-5. Cybenko, G. (1989), Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals, and Systems, 2, 303–14. De Roeck, G. (2003), The state-of-the-art of damage detection by vibration monitoring: The SIMCES experience, Journal of Structural Control, 10, 127–43. Doebling, S. W., Farrar, C. R., Prime, M. B. Shevitz, D. W. (1996), Damage identification and health monitoring of structural and mechanical systems from changes in their vibration characteristics: A literature review, Los Alamos National Laboratory Report LA-13070-MS, available at: http://www.lanl.gov/projects/ei/shm/pubs/lit review.pdf, ac- cessed January 26, 2010. Faravelli, L. Pisano, A. A. (1997), A neural network ap- proach to structure damage assessment, Proceedings of In- telligent Information Systems (IIS), Grand Bahama Island, Bahamas, December 1997. Füssel, D. (2002), Fault diagnosis with tree-structured neuro- fuzzy systems, Fortschr-Ber, VDI Reihe 8, 957, VDI Verlag, Düsseldorf. Ghosh-Dastidar, S. Adeli, H. (2003), Wavelet-clustering- neural network model for freeway incident detection, Computer-Aided Civil and Infrastructure Engineering, 18(5), 325–38. Hajela, P. (1999), Application of neural networks in modelling and design of structural systems, in Z. Waszczyszyn (ed.), Neural Networks in the Analysis and Design of Structures, Springer-Verlag, New York. He, X., Moaveni, B., Conte, J. P. Elgamal, A. (2008), Modal identification study of Vincent Thomas Bridge using simulated wind-induced ambient vibration data,
• 15.
  Multilevel strategy forbridge integrity monitoring 361 Computer-Aided Civil and Infrastructure Engineering, 23(5), 373–88. Isermann, R. (2006), Fault-Diagnosis Systems. An Introduc- tion from Fault Detection to Fault Tolerance, Springer, New York. ISO (2002), Mechanical vibrations—Evaluation of measure- ment results from dynamic tests and investigations on bridges, ISO /TC 108/SC 2, DIS 18649:2002(E). Jaynes, E. T. (2003), Probability Theory: The Logic of Science, Cambridge University Press, Cambridge, UK. Jeffreys, H. (1939), Theory of Probability, Clarendon Press, Oxford, UK. Jiang, X. Adeli, H. (2005), Dynamic wavelet neural network for nonlinear identification of highrise buildings, Computer- Aided Civil and Infrastructure Engineering, 20(5), 316–30. Jiang, X. Adeli, H. (2007), Pseudospectra, MUSIC, and dynamic wavelet neural network for damage detection of highrise buildings, International Journal for Numerical Methods in Engineering, 71(5), 606–29. Jiang, X. Adeli, H. (2008a), Dynamic fuzzy wavelet neu- roemulator for nonlinear control of irregular highrise build- ing structures, International Journal for Numerical Methods in Engineering, 74(7), 1045–66. Jiang, X. Adeli, H. (2008b), Neuro-genetic algorithm for nonlinear active control of structures, International Journal for Numerical Methods in Engineering, 75(7), 770–86. Kao, C. Y. Hung, S. L. (2003), Detection of structural dam- age via free vibration responses generated by approximat- ing artificial neural networks, Computers and Structures, 81, 2631–44. Kim, S. H., Yoon, C. Kim, B. J. (2000), Structural monitor- ing system based on sensitivity analysis and a neural net- work, Computer-Aided Civil and Infrastructure Engineer- ing, 15(4), 309–18. Ko J. M. Ni, Y. Q (2005), Technology developments in structural health monitoring of large-scale bridges, Engi- neering Structures, 27, 1715–25. Ko, J. M., Sun, Z. G. Ni, Y. Q. (2002), Multi-stage iden- tification scheme for detecting damage in cable-stayed Kap Shui Mun Bridge, Engineering Structures, 24, 857– 68. Lam, H. F., Yuen, K. V. Beck, J. L. 2006, Structural health monitoring via measured Ritz vectors utilizing artificial neural networks, Computer-Aided Civil and Infrastructure Engineering, 21, 232–41. Lampinen, J. Vethari, A. (2001), Bayesian approach for neural networks-review and case studies, Neural Networks, 14(3), 7–24. Laplace, P. S. (1812), Théorie Analytique des Probabilités, Courcier. Lau, C. K., Mark, W. P. N., Wong, K. Y., Chan, W. Y. K. Man, K. L. D. (1999), Structural health monitoring of three cable-supported bridges in Hong Kong, in F. K. Chang (ed.), Structural Health Monitoring 2000, Lancaster, Tech- nomic, pp. 450–60. Lee, H. K. H. (2004), Bayesian Nonparametrics via Neural Networks, ASA-SIAM Series on Statistics and Applied Probability. Li, H., Ou, J., Zhao, X., Zhou, W., Li, H. Zhou, Z. (2006), Structural Health Monitoring System for the Sahndong Binzhou Yellow River Highway Bridge, Computer-Aided Civil and Infrastructure Engineering, 21(4). Li, S. Wu, Z. (2008), A non-baseline method for damage locating and quantifying in beam-like structure based on dynamic distributed strain measurements, Computer-Aided Civil and Infrastructure Engineering, 23(5), 404–13. Liu, G. R. Han, X. (2004), Computational Inverse Tech- niques in Nondestructive Evaluation, CRC Press, Boca Ra- ton, FL. MacKay, D. J. C. (1992), A practical Bayesian framework for back-propagation networks, Neural Computation, 4(3), 448–72. MacKay, D. J. C. (1994), Bayesian methods for backpropa- gation networks, Model of Neural Networks III, chapter 6, Springer, New York, pp. 211–54. Masri, S. F., Nakamura, M., Chassiakos A. G. Caughey, T. K. (1996), Neural network approach to detection of changes in structural parameters, ASCE Journal of Engi- neering Mechanics, 122, 350–60. Moaveni, B. He, X. Conte, J. P. (2008), Damage identi- fication of a composite beam based on changes of modal parameters, Computer-Aided Civil and Infrastructure Engi- neering, 23(5), 339–59. Mufti, A. (2001), ISIS Canada Corporation 2002, Guidelines for Structural Health Monitoring, Design Manual no. 2, available online on 02.2008 at http://www.isiscanada.com/ manuals.html/, accessed January 26, 2010. Muto, M. Beck, J. L. (2008), Bayesian updating of hysteretic structural models using stochastic simulation, Journal of Vi- bration and Control, 14(1–2), 7–34. Nabney, I. T. (2004), NETLAB—Algorithms for Pattern Recognition, Springer, New York. National Aeronautics and Space Administration (NASA). (2007), Systems Engineering Handbook, Available on- line at http://education.ksc.nasa.gov/esmdspacegrant/ Documents/NASA%20SP-2007-6105%20Rev%201%20 Final%2031Dec2007.pdf, accessed January 26, 2010. Neal, R. M. (1996), Bayesian Learning for Neural Networks, Springer, Lecture Notes in Statistics 118, Springer, New York. Ni, Y. Q., Wong, B. S. Ko, J. M. (2002), Constructing input vectors to neural networks for structural damage identifica- tion, Smart Materials and Structures, 11, 825–33. Ou, J. Li, H.,(2006), The state of the art and the state of the practice of structural control and monitoring in the Main- land of China, Proceedings of the 4th Conference on Struc- tural Control and Monitoring, San Diego, USA, July 13–15, (Regional Panel Report). Paek, Y. Adeli, H. (1990), Structural design language for coupled knowledge-based systems, Advances in Engineer- ing Software and Workstations, 12(4), 154–66. Pandey, P. C. Barai, S. V. (1995), Multilayer perceptron in damage detection of bridge structures, Computer and Struc- tures, 54(4), 597–608. Psimoulis, P. A. Stiros, S. C. (2008), Experimental as- sessment of the accuracy of GPS and RTS for the de- termination of the parameters of oscillation of major structures, Computer-Aided Civil and Infrastructure Engi- neering, 23(5), 389–403. Sarma, K. C. Adeli, H. (2002), Life-cycle cost optimiza- tion of steel structures, International Journal for Numerical Methods in Engineering, 55(12), 1451–62. Shwe, T. T. Adeli, H. (1993), AI and CAD for earthquake damage evaluation, Engineering Structures, 15(5), 315–19. Sivia, D. S. (1996), Data Analysis: A Bayesian Tutorial, Oxford Science, Oxford University Press, New York. Smith, I. (2001), Increasing knowledge of structural perfor- mance, Structural Engineering International, 12(3), 191–5.
• 16.
  362 Arangio Bontempi Sohn, H., Farrar, C. R., Hemez, F. M., Shunk, D. D., Stinemates, D. W., Nadler, B. R. Czarnecki, J. J. (2004), A review of structural health monitoring literature: 1996– 2001, Report LA-13976-MS, Los Alamos National Labora- tory, New Mexico. Sohn, H., Kim, S. D. Harries, K. (2008), Reference-free damage classification based on cluster analysis, Computer- Aided Civil and Infrastructure Engineering, 23(5), 324–38. Soyoz, S. Fukuda M. Q. (2009), Long-term monitoring and identification of bridge structural parameters, Computer- Aided Civil and Infrastructure Engineering, 24(2), 82– 92. Task Group 5.3 (2002), Monitoring and safety evaluation of existing concrete structures, State of the Art Report, fib. Topping, B. H. V., Khan, A. I., Sziveri, J., A. Bahreininejad, A., Leite, J. P. B., Cheng, B. Ivanyi, P. (1999), Neural networks in advanced computational problems, in Neural Networks in the Analysis and Design of Structures, Springer, New York. Waheed, A. Adeli, H. (2000), A knowledge-based system for evaluation of superload permit applications, Expert Sys- tems with Applications, 18(1), 51–8. Waszczyszyn, Z. Ziemianski, L. (2005), Neural networks in the identification analysis of structural mechanics problems, in Z. Mroz and G. E. Stavroulakis (eds.), Parameter Identi- fication of Materials and Structures, Springer, New York. Wong, K. Y., Lau, C. K. Flint, A. R. (2000), Planning and implementation of the structural health monitoring sys- tem for cable-supported bridges in Hong Kong, Volume 3995 — Nondestructive Evaluation of Highways, Utilities, and Pipelines IV, A. Emin Aktan, S. R. Gosselin (eds.) June 2000, Newport Beach, CA. Xu, H. Humar, J. (2006), Damage detection in a girder bridge by artificial neural network technique, Computer- Aided Civil and Infrastructure Engineering, 21, 450–64.