Deep Learning-Based Inversion Analysis of Interlayer Contact Parameters in Pavement Structures

Accurate evaluation of interlayer contact conditions is essential for understanding pavement performance, yet conventional inversion methods often fail to capture these parameters effectively. This study presents a deep learning–based framework that integrates a Spectral Element Method (SEM) forward model with an enhanced one-dimensional LeNet-5 Convolutional Neural Network (CNN) for simultaneous inversion of horizontal and vertical contact parameters, elastic modulus, and layer thickness using Falling Weight Deflectometer (FWD) data. The supervised CNN achieved high accuracy, with average relative errors of 2.63% and 1.17% for horizontal and vertical contact parameters, respectively. To address limited abnormal data, a convolutional autoencoder was employed for unsupervised pavement state evaluation. Field validation confirmed the model’s ability to distinguish pre- and post-grouting conditions through reconstruction error and structural similarity analysis. The proposed approach offers a robust, non-destructive, and intelligent solution for quantitative inversion and qualitative assessment of pavement structures.