Probabilistic machine learning approach to predict incompetent rock masses in TBM construction

The prediction of incompetent rock masses susceptible to tunnel wall collapse is a common concern in any construction project using a tunnel boring machine (TBM). In China, to identify these weak locations, field geological surveys are traditionally conducted in accordance with the rock mass rating system specified by the national classification standard GB50487-2008, which categorizes the rock masses into five grades. Grade V designates the poorest rock masses, which are very vulnerable to a tunnel collapse. With the advent of machine learning methods based on big data, many researchers have attempted to apply such methods to classify rock mass quality based on the 5-grade system. However, the results of those previous studies have shown that predictions of grade V rock masses were not sufficiently accurate. This study proposes a new approach that combines a Bayesian boosting probabilistic model with two parallel machine learning methods, namely, a convolutional neural network (CNN) and a random forest (RF). To improve the accuracy of the proposed methodology, we simplified our predictions by merging grades IV and V into one category corresponding to incompetent rock masses, and the remaining grades into another category defining competent rocks. Our research is based on the big field database collected from the Songhua River water diversion tunnel project and the features extracted from it, namely, the torque and field penetration indices (TPI and FPI, respectively). The results demonstrate that by using the binary classifier and Bayes boosting model, most of the incompetent rock mass predictions achieve a probability level greater than 0.9. The model and calculated results were subjected to rigorous verification using the tenfold cross-validation method, Precision-Recall-F-1 tests, and receiver operating characteristic (ROC) curve.

Automated summary

This study proposes a new approach that combines a Bayesian boosting probabilistic model with two parallel machine learning methods for classifying the quality of rock masses. The results show that by using the binary classifier and Bayes boosting model, most of the predictions achieve a probability level greater than 0.9.