Intelligent analysis of subbase strain based on a long-term comprehensive monitoring

Abstract

The irreversible development of road subbase strain under long-term loading may result in emergency in the foundation of transportation infrastructure systems, and thus the corresponding analysis is of great importance to the safe of society and economy. To improve the resilience of transportation infrastructure systems, the analysis of subbase strain, aiming for an accurate and reliable prediction, is of necessity for transportation engineers. Traditional methods mainly include mathematical and statistical regression analysis solely based on the monitoring stress/strain data. To analyze the monitoring data more comprehensively and more accurately, we conducted an intelligent analysis of subbase strain based on a long-term monitoring and deep learning approaches, which comprehensively considers the environment conditions (temperature, water content, pressure, etc.), the mechanical responses of other structural layers, etc. The comprehensive monitoring system was installed on the section from Nancun to Shiyingmen of 108 National Highway in Beijing in 2012, including asphalt strain sensors, embedded three-dimensional strain sensors, temperature sensors, osmotic pressure sensors, and others. In this study, the long-term road monitoring data of the eight years from 2012 to 2020 was used for analysis. The traditional Random Forest (RF) method and three kinds of deep learning models, including Long-Short Term Memory neural network (LSTM) model, Bidirectional LSTM-Convolution Neural Network (BiLSTM-CNN) model and Temporal Convolution Network (TCN) model, were employed to analyze the development of subbase strain. Test results showed that deep learning methods for the analysis of long-term monitoring data is better than that uses the traditional machine learning algorithm, in which the prediction accuracy of RF is 70.15%, the prediction accuracy of LSTM method is 76.94%, the prediction accuracy of TCN is 90.46%, and the predication accuracy of BiLSTM-CNN is 94.57%. Finally, sensitivity analysis of the TCN model has been performed to examine the impact of timesteps on evaluation metrics. The study could serve reference to predict the development of subbase strain in engineering projects.