Data-driven estimation of surrounding rock areal fracture intensity with semi-supervised learning using measurement-while-drilling data

The areal fracture intensity (P21) of the surrounding rock is a critical indicator for evaluating tunnel excavation stability. However, conventional methods often fail to capture geological conditions ahead of the tunnel face. With the development of measurement-while-drilling (MWD) technology, real-time acquisition of surrounding rock information has become feasible. Therefore, establishing a relationship between MWD data and the P21 is of significant importance. This study proposes a semi-supervised deep learning framework that combines a multi-autoencoder unit training module (MAUT) with a multi-branch feature extraction network (MBEN). By integrating MWD, construction, and blasting data as dataset input, the proposed model is applied to the Yangjiawopu tunnel. The predicted P21 values exhibited a relative error of less than 10%, confirming the model’s superior predictive performance and engineering applicability. Ablation studies demonstrate that both the MAUT module and feature fusion enhance model accuracy. The shapley additive explanations (SHAP) analysis highlights excavation length, water inflow, and mechanical specific energy (MSE) as key predictors. The model effectively leverages unlabeled data and provides robust geological insights, offering a data-driven approach for early fracture detection and risk assessment ahead of the tunnel face.