Quantification of the Significance of M&C variables on Pavement Performance Using Neural Network.BackgroundProblem statementObjectiveNeural Network Modeling (Top-down approach)Input and Output Data to MLPBinary Representation of IRI for data outputThree-fold cross-validationAverage Correct Classification RateConnection Weight Analysis using 9-6-5 network structureInput Node Share of Output ConnectionsDiscussionQuantification of the Significance of M&C variables on Pavement Performance Using Neural Network.Jae-ho Choi CS539Civil & Environmental EngineeringBackgroundPavement Performance – Single variable to represent overall condition of pavement surface. Specification development1. Method-type specification2. End-result specification3. Performance-related specification(PRS)Problem statementThe payment schedules in end-result specification – based on historical performance of construction industry not on the loss in performance.Pay factor items & pay adjustment schedules are largely based on engineering judgment.Objective Identify Material & Construction (M&C) variables.Find the relative importance of the different variables to the development of any PRS.FCPDSNACStiffACAVStabil.RVLTC….….IRINeural Network Modeling (Top-down approach)Input and Output Data to MLPAge AADT ThicknessLayer_No PI LLP200 AC AVIRI Input & Output variables Output variableExterior factors Structural factors Materialtesting factorsBinary Representation of IRI for data output IRI Binary Representation0.00 ~ 1.25 1 0 0 0 0 1.25 ~ 1.5 0 1 0 0 0 1.5 ~ 2.0 0 0 1 0 02.0 ~ 2.5 0 0 0 1 02.5 ~ 3.0 0 0 0 0 1Three-fold cross-validation Trial 1Trial 2Trial 3Used where there is a scarcity of labeled examples compared with the complexity of the problem Train data set Test data setAverage Correct Classification Rate Network StructureValidation oneValidation twoValidation three9-3-5 51.67 63.79 70.119-6-5 62.22 66.67 75.869-9-5 53.55 60.34 75.869-18-5 58.33 62.07 74.13Fixed parameters( Learning Rate – 0.1, 0.01 , Momentum – 0.1, 0.5, Epoch size - 15000)Connection Weight Analysis using 9-6-5 network structureHidden NodeV1 V2 V3 V9 OUT1 8.36 7.47 14.06 …. 7.93 -3.522 -1.64 1.36 -4.45 …. 2.04 -10.983 0.91 -5.15 -10.14…. -8.10 -10.724 5.18 1.54 2.34 …. -0.35 -5.475 3.06 -10.01 5.37 …. 2.32 1.296 -3.70 3.12 6.95 …. 9.77 -11.81Input Node Share of Output Connections Layer_NoMeanThick. Age P200 ….. AC7.08(%) 7.60(%) 14.26(%) 19.02(%) …..7.32(%)DiscussionP200 Age AV AADT LL PI Mean Thickness Asphalt Content Layer_noMaterial characteristics are more significant than pavement structural factors. Ex) P200, AV, LL, PI > Mean Thickness, Layer_no This result can be used to develop new components for PRS. The relative importance of the different variables is important to the design process and is important to the contractor in determining which factors have the largest effect on the bid
View Full Document