Machine learning based inversion for earth rock dam compaction density

Abstract

In water resources and hydropower engineering and highway engineering, the compactness test of earth-rock mixed materials of the earth-rock dam and embankment plays an important role for construction quality assessment. The key is how to accurately and quickly evaluate the compaction density, which requires advanced methods. Meanwhile, machine learning has become an important and powerful tool for parameter prediction. In this work, a method to obtain the compactness of an earth-rock dam using machine learning algorithms is developed and verified. In so doing, first, by analyzing and processing the measured time history signal from a new fast Portable Falling Weight Deflectometer (PFWD), including time domain, frequency domain and mechanical impedance analyses, six optimized characteristic parameters are obtained to invert for the compactness, including the stiffness coefficient, the maximum time difference, the maximum amplitude ratio, the central frequency difference, the maximum stiffness impedance and the resonance frequency. Pearson coefficient was used to analyze the correlation among the six parameters and Gini index was used to analyze the influence degree of the parameters on the relative wet density. Then, the Decision Tree (DT) model, Random Forest (RF) model and Gradient Boosting Decision Tree (GBDT) model were used to predict the wet density and dry density of materials, respectively, and compared. PFWD was applied on a dam to acquire the time history signal, which was then preprocessed following a series of procedures including integrity check, outlier handling and normalization etc, and machine learning algorithms were used for training the data and parameter prediction. Examples show that, considering the generalization and optimization ability, for the GBDT model, the root-mean-square error (RMSE) of the predicted wet density of the earth-rock material is within 0.041 g/cm3 and that of the predicted dry density is within 0.032 g/cm3, which meet the requirements for construction detection and quality control. The proposed method based on PFWD signal analysis and machine learning algorithm models help better solve the nonlinear mapping problem between the PFWD signal and material density, and can accurately invert for the wet density and dry density of materials. It is an effective way for rapid detection and control of earth-rock dam construction quality.