Compaction quality assessment of road subgrades using explainable deep graph learning framework

Abstract

Compaction-quality assessment based on machine learning is an attractive topic in road construction research. However, existing methods do not consider the structural information of data when predicting the compaction degree. Thus, an explainable deep graph learning framework is proposed for the intelligent compaction quality assessment of road subgrades. In this method, a multi-domain analysis is first used to extract different indicators from the vibration signals of a vibratory roller. Second, the indicators for the different sampling points are constructed as graph structure data. Finally, an alternating graph-regularized regression network (AGRN) is developed to learn features from the graph data and aggregate the features using a regressor to predict the compaction degree. Through experimental verification, the proposed method displays an improved generalization ability and a high prediction accuracy when compared with other methods. Moreover, Shapley additive explanations (SHAP) are introduced to measure the marginal contributions of indicators for predicting the compaction degree in compaction quality assessments.