Research On Back Balculation Method Of Asphalt Pavement Modulus Based On Neural Network

Resilience modulus of structural layer is one of the important indexes to evaluate pavement strength.There is a highly nonlinear mapping relationship between resilience modulus and deflection basin.Theoretically,the modulus of resilience can be calculated by inversion of deflection basin data.However,the inverse calculation of modulus has the characteristics of multivariable and nonlinearity,which is a difficult task.Because of the strong ability of neural network to deal with nonlinear dynamic data,and more and more scholars believe that the geometric parameters of deflection basin play an indispensable complementary role in pavement strength evaluation,this paper compares some commonly used neural network structures,selects the most suitable structure and optimizes it,and determines the iterative direction of training parameters by Bayesian regularization algorithm.Taking various geometric parameters of deflection basin as characteristic variables,these parameters are transformed into comprehensive principal component characteristics.In order to make features more recognizable,function transformation is carried out according to the data features of feature variables and target variables,and the spatial distribution of input parameters and output parameters is changed to optimize the modeling effect.In view of the different difficulty in calculating the resilience modulus of each structural layer,a layer-by-layer inverse calculation model is designed,which firstly calculates the modulus of soil base which is most sensitive to the input variables,and then takes the modulus of soil base as one of the input parameters,and then calculates the resilience modulus of surface layer and base layer.The modulus value predicted by neural network is taken as the initial theoretical value,and the theoretical value is adjusted and iterated until the calculated theoretical deflection value is close to the measured value.Experimental results show that BP network has good fitting ability,and Bayesian regularization can improve the generalization ability of neural network.The effect of deflection basin shape parameters as input variables is better than deflection values,the feature transformation method can significantly reduce the back calculation error,the step-by-step back calculation mode can optimize the back calculation effect to a certain extent,and the forward calculation iteration method can effectively adjust the theoretical modulus of neural network prediction.The test method provides a new idea for improving the accuracy of modulus back calculation.