Compaction quality evaluation of subgrade based on soil characteristics assessment using machine learning

Abstract

The reliable correlation model for intelligent compaction (IC) is to be developed by integrating characteristics of the filling material and control parameters of the vibratory roller. Key characteristics of the backfill soil from the construction site of Rongwu Highway are tested through the modified Proctor test and the direct shear test. A well-documented dataset is built by summarizing 4000 shear strengths from open literature and 246 data from laboratory tests in this study. The PSO-BP-NN model is developed based on this dataset to predict the shear strength and compactness of the subgrade soil in the scope of mechanical properties and compaction powers. The importance analysis of the input variables is performed with the Random Forest algorithm. The influence mechanism is analyzed in sequence. The plain soil and the lime soil demonstrate typical compaction curves, and the lime soil is less sensitive to the influence factors. An optimal compaction power exists for determining the optimal moisture content utilizing the shear strength, tending to be less conservative. The moisture content is the most important factor for the compactness, followed by the compaction power; the particle size is suggested to be considered for real-time evaluations. The compaction mechanism is mainly attributed to the water film theory and the electrochemical property of the filling soil. This study aims to provide a reliable model to estimate the compactness in the aspect of material properties so as to enhance the accuracy of the IC model.