Developing a real-time compaction quality assessment methodology for subgrade based on semi-supervised co-training

Abstract

The current intelligent methodologies for real-time evaluation of subgrade compaction quality predominantly rely on supervised machine learning algorithms. However, the scarcity of sand cone test significantly impedes model performance, thereby severely limiting the applicability of intelligent subgrade compaction. This paper proposes a semi-supervised co-training algorithm for real-time evaluation of subgrade compactness during the construction procedure, leveraging unlabeled data to enhance the model performance. Based on the compaction datasets from various subgrade scenarios, the proposed PSO-XGB-Co-training KNN-PLS (PSO-XGB-CoKP) algorithm is utilized to train the unlabeled data, boasting a 20.4% reduction in Mean Squared Error (MSE). The semi-supervised co-training algorithm is modified by employing different regressors as co-trained sub-models and increasing the number of regressors. The model is optimized by levering the accuracy and computational cost, and an optimal data augmentation volume is recommended through the sensitivity study. This study provides an alternative approach for leveraging unbalanced and small-sample datasets to develop a reliable intelligent methodology for evaluating subgrade compaction quality in engineering practice.