Analysis Of Vertical Dynamic Deflection Of Flexible Pavement Profile Based On Neural Network

This paper analyses the dynamic deflection response of individual layers onthe cross section of the pavement along with increase of load repetitions, by aradial basic function (RBF) neural network model, comparing with ABAQUS’sresult.It is built the constitution of the RBF model, considering the effect of thelocal deflection of the pavement surface on the corresponding points of otherlayers. The training and testing data of the RBF model use the last profiledeflections on the tops of four layers, P401AC Surface, P209Base, P154Subbaseand Medium Strength Subgrade, measured by the National Airport Pavement TestFacility (NAPTF), located at the Federal Aviation Administration’s (FAA) WilliamJ. Hughes Technical Centre. The surface deflection curve of Layer P401is used forthe input data, because the developing process of the curve could be recorded andobserved in the test. The top deflection curves of the other three layers are set asthe output data, respectively, because their dynamic process in the test wasinvisible unknown.First of all, in order to seek the deflection response of individual layers of thepavement profile, it simulates to predict in different situations, such as thedeformation of intermediate points of the three adjacent points on the top surface.There are raised points in the deformation curve of the fourth layer (the mediumstrength subgrade), so it chooses two methods: the actual selection and the thesmooth handling for the deformation curve in the data selection. Moreover, thereare three types from the initial deformation to the final one in each methods:divided into50stages equally, the1/i step length and divided randomly. Then itcompares the error between predicted results and real deformation, and the error ofthickness between the predicted results and the initial value.Secondly, it studies the deformation developing process of the pavementprofile, considering the ratio between the initial thickness and the average thickness ofthe lth layer in the jth stage, the the ratio between the average deflection and the averagethickness of the lth layer in the jth stage, the ratio between the maximum deflection andthe average thickness of the lth layer in the jth stage, the ratio of the average thicknessvalues of any two layers in the jth stage, the ratio of the maximum deflection values ofany two layers in the jth stage.Finally, the finite element software ABAQUS is used to model and analyzethe pavement, establishing the axisymmetric model and the3D model respectively.And it compares the finite element method with the RBF neural network, by the correlation coefficient, similarity level and average error with measured bymulti-depth deflectometers. The results show that the RBF neural network isfesible and accurate to predict the deflection response of individual layers of thepavement profile.