Study of Cross-Project Prediction of Rock Mass Classification Based on Feature Fusion

Abstract

For new tunnel boring machine (TBM) tunnel projects that lack specific data required for training prediction models of rock mass classification, it is essential to establish a cross-project prediction model based on historical projects. However, due to the uncertainty of geological conditions and the variability in data correlations, predicting rock mass classification for cross-project remains a significant challenge. This study proposes a cross-project prediction model of rock mass classification based on feature fusion of physics-driven and data-driven models. Basically, a physics-driven model was presented by creating a penetration grade matrix combined with statistical indicators. Simultaneously, a data-driven model was established by selecting contributing raw parameters based on the importance ranking provided by the LightGBM model. Based on the Yinchao project (YC), data from YC2-6 was used as a known database for model training and verification, while data from YC2-5 served as a test set to verify the cross-project prediction performance. Label denoising was performed to refine data quality, and data analysis was conducted between the mutual projects. The results indicate that there is a discrepancy between the data distributions of new tunnel and historical projects. The data-driven model attains good performance in single-tunnel prediction but suffers from overfitting in cross-project prediction, leading to a sharp decline in prediction accuracy. The proposed feature-fusion model attains an F1 score of 0.74 in cross-project prediction, outperforming the data-driven model which has an F1 score of 0.67. It can effectively improve the prediction performance for cross-project rock mass classification.