Rutting prediction using deep learning for time series modeling and K-means clustering based on RIOHTrack data

Abstract

Considering that traditional statistical regression and machine learning rutting prediction models hardly capture time series characteristics of rutting, this study employed deep learning (DL) techniques to develop multivariate time series models to predict the rutting progression curve for asphalt pavements commonly constructed in China from the data relating to the rutting depth of 19 asphalt pavements from RIOHTrack. 32 features related to climate, traffic, pavement structure, and materials, were selected from initial 50 features using variance threshold and feature importance analysis. k-means clustering was used to classify these pavement sections according to their rutting progression pattern. Three DL models (RNN, LSTM, and GRU) with different architectures and sequence lengths (SLs) (24 models produced) were trained from historical rutting depths and the selected exogenous variables, and their performances were then compared with ARIMAX, Gaussian process, and mechanistic-empirical (M−E) models. The results show that the architecture 2 based GRU model with an SL of 11 is found to have the best prediction performance (R2 = 0.90). The DL model can capture the pattern of the observed rutting progression curve which contains linear and seasonal trends. Moreover, the optimal DL model is more accurate than the statistical time series models and the M−E model. Permutation feature importance analysis shows the impacting features mainly include historical rutting depth, temperature, age, as well as binder penetration and air void of the second asphalt concrete layer.