Rapid pavement aggregate gradation estimation based on 3D data using a multi-feature fusion network

Abstract

Non-destructive pavement aggregate gradation estimation during construction is required because the aggregate distribution is difficult to control but impacts pavement performance significantly. This paper proposes a multi-feature fusion network based on Residual Convolutional Neural Network (ResNet) to estimate the aggregate gradation using 3D data collected by laser scanners. Ten geometric parameters and eighteen 2D-wavelet parameters are merged to the fully connected layer of the ResNet. Eight categories of mixtures are divided into 800 samples as a dataset (720 for training and 80 for testing). The proposed model (ResNet+MLP) performs better (F1-score = 0.96) in 8-types classification than other classifiers and is applied to estimate the aggregate gradation, where the R-square is 0.86. A field test validation for the proposed model on the newly constructed pavement is conducted and proves its practicability. This paper establishes the association between 3D pavement texture and the aggregate gradation distribution, paving a new way for pavement quality evaluation.