• Like
  • Save
Neural network-based techniques for the damage identification of bridges: a review of recent advances, Arangio S.
Upcoming SlideShare
Loading in...5
×
 

Neural network-based techniques for the damage identification of bridges: a review of recent advances, Arangio S.

on

  • 344 views

Review Invited lecture at Third International Conference on Soft Computing Technology in Civil, Structural and Environmental Engineering (CC2013), 3-6 September 2013, Cagliari, Italy

Review Invited lecture at Third International Conference on Soft Computing Technology in Civil, Structural and Environmental Engineering (CC2013), 3-6 September 2013, Cagliari, Italy

Statistics

Views

Total Views
344
Views on SlideShare
344
Embed Views
0

Actions

Likes
0
Downloads
3
Comments
0

0 Embeds 0

No embeds

Accessibility

Categories

Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

    Neural network-based techniques for the damage identification of bridges: a review of recent advances, Arangio S. Neural network-based techniques for the damage identification of bridges: a review of recent advances, Arangio S. Presentation Transcript

    • Neural networks based techniques for damage identification of bridges: a review of recent advances Sapienza University of Rome – StroNGER s.r.l. S. Arangio stefania.arangio@uniroma1.it, stefania.arangio@stronger2012.com Cagliari, September 5th 2013
    • 2/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Introduction Part I Conclusions Part II Neural networks and Bayesian enhancements Outline Case study: Bayesian neural networks for the assessment of the bridge of the ANCRiSST benchmark problem Soft computing approaches for the structural assessment of bridges
    • 3/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Introduction Part I Conclusions Part II Neural networks and Bayesian enhancements Outline Case study: Bayesian neural networks for the assessment of the bridge of the ANCRiSST benchmark problem Soft computing approaches for the structural assessment of bridges
    • 4/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Methods for structural identification and damage detection Input – output techniques • The structure has to be artificially excited and in case of large structures it is not always possible • The operation of the structure has to be interrupted Only output techniques • The excitation is given by the ambient vibration • Measurements in real operational conditions • Suitable in case of continuous monitoring Traditional methods Soft computing methods • Time domain approaches • Frequency domain approaches • Neural networks • Genetic algorithms • Fuzzy Logic Introduction
    • 5/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Examples of structural assessment by using soft computing methods (2008-2013) Adeli H., Jiang X., Intelligent infrastructures – Neural Networks, wavelets, and Chaos Theory for Intelligent Transportation Systems and Smart Structures, CRC Press, Taylor & Francis, Boca Raton, Florida, 2009 Al-Rahmani A.H., Rasheed H.A., Najjar A.Y., A combined soft computing-mechanics approach to inversely predict damage in bridges, Procedia Computer Science, 8, 461 – 466, 2012 Arangio S., Beck J.L. Bayesian neural networks for bridges integrity assessment, Structural Control & Health Monitoring, 2012; 19(1), 3-21. Arangio S., Bontempi F. Soft Computing based Multilevel Strategy for Bridge Integrity Monitoring, Computer-Aided Civil and Infrastructure Engineering 2010; 25, 348-362. Bhattacharyya P., Banerji P., Improved Damage Classification and Detection on a Model Bridge using Fuzzy Neural Networks, 4th International Conference on Structural Health Monitoring of Intelligent Infrastructure (SHMII-4), 22-24 July 2009, Zurich, Switzerland, 2009. Cheng J., An artificial neural network based genetic algorithm for estimating the reliability of long span suspension bridges, Finite Elements in Analysis and Design, 46, 658–667, 2010. Cheng J., Li Q.S., Reliability analysis of structures using artificial neural network based genetic algorithms, Comput. Methods Appl. Mech. Engrg., 197, 3742–3750, 2008. Firouzi A., Rahai A., An integrated ANN-GA for reliability based inspection of concrete bridge decks considering extent of corrosion-induced cracks and life cycle costs, Scientia Iranica, 19 (4), 974–981, 2012. Flood I., Towards the next generation of artificial neural networks for civil engineering, Advanced Engineering Informatics 22, 4–14, 2008 Freitag S., Graf W., Kaliske M. Recurrent neural networks for fuzzy data, Integrated Computer-Aided Engineering - Data Mining in Engineering, 2011; 18(3), 265-280. Graf W.S., Freitag S., Sickert U., Kaliske M., Structural Analysis with Fuzzy Data and Neural Network Based Material Description, Computer-Aided Civil and Infrastructure Engineering 27, 640–654, 2012. Li S., Li H., Liu Y., Lan C., Zhou W., Ou J., SMC structural health monitoring benchmark problem using monitored data from an actual cable- stayed bridge, Structural Control and Health Monitoring, published online form March 26th 2013, DOI:10.1002/stc.1559 Mehrjoo M., Khaji N., Moharrami H., Bahreininejad A., Damage detection of truss bridge joints using Artificial Neural Networks, Expert Systems with Applications 35, 1122–1131, 2008. Park J.H., Kim J.T, Honga D.S., Hoa D.D., Yib J.H., Sequential damage detection approaches for beams using time-modal features and artificial neural networks, Journal of Sound and Vibration, 323, 451–474, 2009. Sgambi L., Gkoumas K., Bontempi F. Genetic Algorithms for the Dependability Assurance in the Design of a Long-Span Suspension Bridge, Computer-Aided Civil and Infrastructure Engineering 2012; 27(9), 655-675. Tsompanakis Y., Lagaros N.D., Stavroulakis G. Soft computing techniques in parameter identification and probabilistic seismic analysis of structures, Advances in Engineering Software 2008, 39(7), 612-624. Wang Y.M., Elhag T.M.S., An adaptive neuro-fuzzy inference system for bridge risk assessment, Expert Systems with Applications 34, 3099–3106, 2008. Zhou H.F., Ni Y.Q., Ko J.M., Constructing input to neural networks for modeling temperature-caused modal variability: Mean temperatures, effective temperatures Introduction
    • 6/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Introduction Part I Conclusions Part II Neural networks and Bayesian enhancements Outline Case study: Bayesian neural networks for the assessment of the bridge of the ANCRiSST benchmark problem Soft computing approaches for the structural assessment of bridges
    • 7/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Nonlinear feed-forward basis functions ( )         +         += ∑∑ == )2( 0 1 )1( 0 )1( 1 )2( , k D j jiji M j kjk bbxwgwfy wx ∑= += D i jijij bxwa 1 )1( 0 )1( ( )kk afy = ∑= += M j kjkjk bzwa 1 )2( 0 )2( ( )jj agz = NEURAL NETWORK MODEL ( ) ( )        = ∑= M j jjwfy 1 , xwx φ output units hidden units activations weights bias Neural network model
    • 8/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Traditional learning t ij t ij t ij www ∆+= − η1 ij ij W E W ∂ ∂ −=∆ η Learning rate Weights updating Minimization of a sum of squares error function Model fitting is obtained by modifications of the coefficients w t = correct value y = network value ( )         +         += ∑∑ == )2( 0 1 )1( 0 )1( 1 )2( , k D j jiji M j kjk bbxwgwfy wx Gradient descent algorithm [traingd] Conjugate gradient algorithm [traincg] Quasi – Newton algorithm [trainbfg] Levenberg – Marquardt algorithm [trainlm] ( ){ } ∑∑∑ == = +−= W i i N n oN t t n t n wxytE 1 2 1 1 2 2 ; 2 1 α w
    • 9/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Probabilistic interpretation ( ){ }     −−∝ 2 ; 2 exp),,,( ww nn xytMxtp β β 1) Probabilistic interpretation of the network output 2) Probability model for the prediction error );( wxyt −=ε Gaussian µ = 0 σD 2 = 1/β 3) Predictive PDF The output approximates the conditional average of the target data hyperparameter 4) Prior PDF ( )       −= 2 2 exp 1 ),( w Z Mwp W α α α Gaussian µ = 0 σw 2 = 1/α hyperparameter
    • 10/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Network learning as inference 10 =),,( MwDp β ( ) ( ){ }         −− ∑∑= N n N t n t t n D o xyt Z 1 2 ; 2 exp 1 w β β Likelihood ( ) ( ) ( )MDp MwpMwDp Dwp ,, ,,, )M,,,( βα αβ βα =Bayes theorem evidence priorxlikelihood posterior = ( ) ( ){ } ∑∑∑ == +−= W i i N n N t n t t n wxytwE o 1 2 1 2 2 ; 2 αβ w ( ) ( ){ }∑∑= −=− N n N t n t t n o xytMwDp 1 2 ; 2 ,,log w β β ( ) ∑= =− W i iwMwp 1 2 2 ,log α α max (posterior) = min (negative log posterior) =− )M,,,(log βαDwp ( ) ( )=−− MwpMwDp ,log,,log αβ ( )       −= 2 2 exp 1 ),( w Z Mwp W α α α Prior ( )Mwp ( )MDwp ,
    • ( ) ( ) ( ){ }         −−= ∑∑= N n N t n t t n D o xyt Z MwDp 1 2 ; 2 exp 1 ,, w β β β ( )       −= 2 2 exp 1 ),( w Z Mwp W α α α ( ){ }∑ ∑ ∑= = +− =− N n N t W i i n t t n o wxyt Dwp 1 1 2 2 ; 2 ),,,(log αβ βα w M DATA PRE- PROCESSING OUTPUT NETWORK ARCHITECTURE n°INPUT n°UNIT IN THE HIDDEN LAYERS POSTERIOR: BAYES’ THEOREM ( ) ( ) ( )MDp MwpMwDp Dwp ,, ,,, ),,,( βα αβ βα =M w = wMAP? yes INFERENCE OF NEW DATA DATA POST PROCESSING PROBABILISTIC MODEL • NOISE MODEL • PREDICTIVE PDF • LIKELIHOOD • PRIOR ),,,( Mxtp βw ( )MwDp ,, β ),( Mwp α OPTIMIZATION (MINIMUM OF )),,,(log MβαDwp− no INPUT ED EW ( )Mwp ( )MDwp , 1) Model fitting
    • 12/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Bayesian techniques for neural networks • Level 1 Model fitting: inferring the model parameters given the model and the data • Level 2 Optimization of the hyperparameters α and β • Level 3 Model class selection: optimal model complexity • Level 4 Automatic relevance determination (ARD): evaluation of the relative importance of different inputs Network learning as inference (model fitting) is only one level in which Bayesian inference can be applied in the neural network field Hierarchical multi-level approach PartI
    • POSTERIOR FOR α, β TRAINING: OPTIMIZATION w = wMAP? ?( ) ( )DMEVDMEV ii 1−> INFERENCE OF NEW DATA CHOOSE MODEL Mi-1 ? POSTERIOR FOR Mi α, β = αMP, βMP DATA PRE- PROCESSING OUTPUT NETWORK MODEL Mi N HIDDEN = i N INPUT = k POSTERIOR FOR w yes DATA POST PROCESSING PROBABILISTIC MODEL no INPUT CHOOSE INITIAL α, β INITIALIZE WEIGHTS w RE-ESTIMATION OF α, β yes no Wγ ≈ yes no i= i+1 is α1,…,αk ‘very large’? k= k-1 yes no ( ) ( ) ( )MDp MwpMwDp Dwp ,, ,,, ),,,( βα αβ βα =M 1st level Model fitting
    • POSTERIOR FOR α, β TRAINING: OPTIMIZATION w = wMAP? ?( ) ( )DMEVDMEV ii 1−> INFERENCE OF NEW DATA CHOOSE MODEL Mi-1 ? POSTERIOR FOR Mi α, β = αMP, βMP DATA PRE- PROCESSING OUTPUT NETWORK MODEL Mi N HIDDEN = i N INPUT = k POSTERIOR FOR w yes DATA POST PROCESSING PROBABILISTIC MODEL no INPUT CHOOSE INITIAL α, β INITIALIZE WEIGHTS w RE-ESTIMATION OF α, β yes no Wγ ≈ yes no i= i+1 is α1,…,αk ‘very large’? k= k-1 yes no ( ) ( ) ( )MDp MwpMwDp Dwp ,, ,,, ),,,( βα αβ βα =M 1st level Model fitting 2nd level Evaluating the hyperparameters α, β ( ) ( ) ( )MDp MpMDp Dp βαβα βα ,,, ),,( =M
    • 15/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Issues in neural network design: selection of the optimal model RULES OF THUMBS -…between the input layer size and the output layer size (Blum, 1992) - (Software Neuroshell, 2000) - (Berry and Lynoff, 1997) - n = dimension needed to capture 70-80% of the variance (Boger and Guterman, 1997) OPTIMAL NUMBER OF UNITS (“OCKHAM’S RAZOR”) )( 3 2 oI NNn += INn ⋅<2 examplesNn ⋅< 30 1 They aren’t rigorous methods INPUT LAYER OUTPUT LAYER HIDDEN LAYERS PartI
    • POSTERIOR FOR α, β TRAINING: OPTIMIZATION w = wMAP? ?( ) ( )DMEVDMEV ii 1−> INFERENCE OF NEW DATA CHOOSE MODEL Mi-1 ? POSTERIOR FOR Mi α, β = αMP, βMP DATA PRE- PROCESSING OUTPUT NETWORK MODEL Mi N HIDDEN = i N INPUT = k POSTERIOR FOR w yes DATA POST PROCESSING PROBABILISTIC MODEL no INPUT CHOOSE INITIAL α, β INITIALIZE WEIGHTS w RE-ESTIMATION OF α, β yes n o Wγ ≈ yes no i= i+1 is α1,…,αk ‘very large’? k= k-1 yes no ( ) ( ) ( )MDp MwpMwDp Dwp ,, ,,, ),,,( βα αβ βα =M 1st level Model fitting 2nd level Evaluating the hyperparameters α, β 3rd level Model class selection ( ) ( )MpMDpDMp ∝)( prior = constantevidence ( ) ( ) ( )MDp MpMDp Dp βαβα βα ,,, ),,( =M
    • POSTERIOR FOR α, β TRAINING: OPTIMIZATION w = wMAP? ?( ) ( )DMEVDMEV ii 1−> INFERENCE OF NEW DATA CHOOSE MODEL Mi-1 ? POSTERIOR FOR Mi α, β = αMP, βMP DATA PRE- PROCESSING OUTPUT NETWORK MODEL Mi N HIDDEN = i N INPUT = k POSTERIOR FOR w yes DATA POST PROCESSING PROBABILISTIC MODEL no INPUT CHOOSE INITIAL α, β INITIALIZE WEIGHTS w RE-ESTIMATION OF α, β yes n o Wγ ≈ yes no i= i+1 is α1,…,αk ‘very large’? k= k-1 yes no ( ) ( ) ( )MDp MwpMwDp Dwp ,, ,,, ),,,( βα αβ βα =M 1st level Model fitting 2nd level Evaluating the hyperparameters α, β 3rd level Model class selection ( ) ( )MpMDpDMp ∝)( prior = constantevidence is α1,…,αk ‘very large’? 4th level Automatic Relevance Determination ( ) ( ) ( )MDp MpMDp Dp βαβα βα ,,, ),,( =M
    • 18/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Introduction Part I Conclusions Part II Neural networks and Bayesian enhancements Outline Case study: Bayesian neural networks for the assessment of the bridge of the ANCRiSST benchmark problem Soft computing approaches for the structural assessment of bridges
    • 19/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio The ANCRiSST benchmark problem • Consortium of 20 research institutions • Established in 2002 with the purpose of: • assessing current progresses on smart materials and structures technology • Developing synergies that facilitate joint research projects that cannot easily carried out by individual centers In October 2011 they opened for researchers in the SHM community a benchmark problem based on a real bridge: the TianjinYonghe bridge http://smc.hit.edu.cn/ PartII
    • 20/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Description of the Tianjin Yonghe bridge Tianjin Hangu 25.15 99.85 260 99.85 25.15 • Cable-stayed bridge • Opened to traffic since December 1987 • After 19 years of operation damages were detected and the bridge was retrofitted • A sophisticated SHM system has been designed and implemented by the Research Center of Structural Health Monitoring and Control of the Harbin Institute of Technology PartII
    • 21/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Structural Health Monitoring System Tianjin Hangu 2515 5600 5885 5900 5600 5600 5900 5885 5600 2515 1 (3) 2 (4) 3 (5) 7 (9) 9 (10) 11 (12) 13 (14) Uniaxial/biaxial accelerometers Hygrothermograph Anemometer 1, 3, 5, 7, 9 11, 13 2, 4, 6, 8, 10, 12, 14 During 2008: • Continuous monitoring system • 14 uniaxial accelerometers on the bridge deck (downward and upward) • On the top of the tower: 1 biaxial accelerometer; 1 anemometer; 1 temperature sensor downward and upward PartII
    • 22/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio damaged area Damage situation 1 Cracks at the closure segment at both side spans August 2008: 2 damages are detected PartII
    • 23/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Damage situation 2 Damaged piers PartII
    • 24/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Available data set Health condition Damaged condition • Time histories of the accelerations recorded at the 14 deck sensors on January 1st and January 17th 2008 (registrations of 1 h carried out for 24 h ) • Environmental information (wind, temperature) • Biaxial accelerations at the top of the tower • Time histories of the accelerations recorded at the same 14 deck sensors on July 30th 2008 (registrations of 1 h carried out for 24 h) • Accelerations collected by field testing August 7th to 10th 2008 (not used) PartII
    • 25/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Procedure for neural network training time history of the acceleration recorded at sensor # Structural system Ambient excitation 1+−dtf 2−tf tf1−tf 1+tfTraining of the neural network model in undamaged condition 2+tf Test of the trained neural network model on a new time history
    • 26/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Neural network based damage detection strategy 14 groups of networks have been created (one for each measurement point e one for each hour of measurements) 14 (points) x24 (hours) = 336 neural network models Tianjin Hangu 1 (3) 2 (4) 3 (5) 7 (9) 9 (10) 11 (12) 13 (14) accelerometers
    • 27/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Detection of anomalies If ∆e ≈ 0 the structure is considered as undamaged If ∆e is large an anomaly is detected
    • 28/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Damaged area Error in the approximation of the accelerations in the undamaged sections Training Undamaged Damage detection Tianjin Hangu
    • 29/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Bayesian model class selection The most plausible class can be obtained applying Bayes’ Theorem: ( ) ( )( , ) |j jj p M D p D M p M∝M M prior = cost evidence provided by D The various model can be compared by evaluating their evidence       − +      +−− γγ α N E MP W 2 ln 2 12 ln 2 1 ln 2 1 A ++++− jjMP MP D HH N E ln2!lnln 2 ββ( )=iMDpln Data fit term Penalizing term “Ockham factor”
    • 30/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Bayesian model class selection The chosen model has 3 hidden units Model 1 2 3 4 5 N parameters 7 13 19 25 31 gamma 2,00 3,03 4,02 5,00 6,00 MP j MP D MP j β N Eβ ln 2 +− 20770 22682 25078 22153 23500 ( ) MP j MP j HH ln2!ln + 2,08 3,99 5,95 8,01 10,16 data fit term 20772 22686 25084 22161 23510 MP j MP W MP j α W Eα ln 2 ln 2 1 ++− A -13,08 -79,32 -158 -213 -266       − +      γNγ 2 ln 2 12 ln 2 1 -3,31 -3,51 -3,66 -3,8 -3,86 penalizing term -16 -83 -162 -217 -270 log evidence 20756 22603 24922 21944 23240
    • 31/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII BAYESIANMODELSELECTION
    • 32/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Error in the approximation of the undamaged conditions downriver upriver ∆e at the various locations Data for training: January 1st 2008 (H1 to H24) Data for testing: January 17th 2008 (H1 to H24) Undamaged conditions
    • 33/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Error in the approximation of the damaged conditions ∆e at the various locations Data for training: January 1st 2008 (H1 to H24) Data for testing: July 30th 2008 (H1 to H24 Damaged conditions!
    • 34/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio PartII Difference of the errors The difference of error in the approximation suggests the presence of structural anomalies around sensor #10
    • 35/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Validation of the results: Structural assessment by applying the Enhanced Frequency Domain Decomposition • Data collection and signal preprocessing • Construction of the the Power Spectral Density matrix (PSD) • Whelch averaged modified periodgram method • 50 % overlapping and periodic Hamming windowing • Singular Value Decomposition (SVD) of the PSD • Identification of modal frequencies and mode shapes • Evaluation of the damping PartII
    • 36/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio 0 0,2 0,4 0,6 0 0,5 1 1,5 2 2,5 3 H6 H11 H15 H17 H19 H21 SingularValues(health) f [Hz] 0 0,1 0,2 0,3 0 0,5 1 1,5 2 AverageSingularValues(health) f [Hz] EFDD: Singular Values DecompositionPartII 0 0,5 1 1,5 2 0 0,5 1 1,5 2 AverageSingularValues(damaged) f [Hz] 0 0,5 1 1,5 2 0 0,5 1 1,5 2 2,5 3 H6 H9 H12 H15 H18 H20 H22 H23 H24 Undamaged conditions Damaged conditions Average Singular values Average Singular values
    • 37/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Comparison of the mode shapesPartII The decrease of the frequencies suggests the presence of damage f=0.4075 Hz FEM (“AS BUILT” CONDITION) FEM Mode 1 - f=0.452 Hz FEM Mode 2 - f=0.632 Hz FEM Mode 3 - f=0.937 Hz Mode 1 - Mode 2 - f=0.594 Hz Mode 3 - f=0.896 Hz Mode 1 - f=0.262 Hz Mode 2 - f=0.388 Hz Mode 3 - f=0.664 Hz UNDAMAGED CONDITION DAMAGED CONDITION
    • 38/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Introduction Part I Conclusions Part II The ANCRIiST benchmark problem Description of the bridge and available monitoring data Outline Neural network based damage detection strategy Results
    • 39/39PartIPartIIConclusionsIntroduction NEURAL NETWORKS BASED TECHNIQUES FOR DAMAGE IDENTIFICATION OF BRIDGES: A REVIEW OF RECENT ADVANCES S. Arangio Conclusions Soft computing approaches, like the neural networks model, have proven to be effective for dealing with large quantities of data and, recently, have been widely used for the structural assessment of Civil structures and infrastructures. Neural networks can be significantly improved by applying Bayesian inference at different levels in a hierarchical way: Bayesian Neural Networks (BNN) The BNNs have been applied for processing the monitoring data coming from the bridge of the ANCRiSST SHM benchmark problem and have shown to be able to detect the presence of an anomaly. The current work is focused on the development of methods for the localization of the detected damage Conclusions
    • POSTERIOR FOR α, β TRAINING: OPTIMIZATION w = wMAP? ?( ) ( )DMEVDMEV ii 1−> INFERENCE OF NEW DATA CHOOSE MODEL Mi-1 ? POSTERIOR FOR Mi α, β = αMP, βMP DATA PRE- PROCESSING OUTPUT NETWORK MODEL Mi N HIDDEN = i N INPUT = k POSTERIOR FOR w yes DATA POST PROCESSING PROBABILISTIC MODEL no INPUT CHOOSE INITIAL α, β INITIALIZE WEIGHTS w RE-ESTIMATION OF α, β yes n o Wγ ≈ yes no i= i+1 is α1,…,αk ‘very large’? k= k-1 yes no email: stefania.arangio@uniroma1.it stefania.arangio@stronger2012.com Prof. Bontempi and his research team www.francobontempi.org of Sapienza University of Rome are gratefully acknowledged. This research was partially supported by StroNGER s.r.l. from the fund “FILAS - POR FESR LAZIO 2007/2013 - Support for the research spin off”.