Enhancing tunnel safety with machine learning models for ground behavior prediction
Jian Zhou, Yulin Zhang, Yingui Qiu, Kang Peng, Manoj Khandelwal
Abstract
Accurate prediction of ground behavior is pivotal in ensuring safety and reliability during tunnel construction. This paper addresses the limitations of conventional models by leveraging machine learning techniques for predicting ground behavior in rock tunnelling. The study focused on enhancing safety in complex systems, evaluates eight classification models—bagging meta-estimator (Bagging), random forest (RF), k-nearest neighbors (KNN), multilayer perceptron (MLP), support vector machines (SVM), stochastic gradient descent (SGD), logistic regression (Logistic), and decision trees (DT)—using a dataset of 207 cases categorized into behavior types: non-squeezing, rockburst, squeezing, and self-supporting. Six key input parameters are considered, encompassing joint factors, tunnel diameter, and overburden height. Evaluation metrics including accuracy, kappa, Matthew’s correlation coefficient, and Hamming loss, along with ROC analysis and confusion matrix, are employed. The results demonstrated that RF exhibited the highest performance in classifying ground behavior for rock tunnelling, with an accuracy of approximately 0.9762, surpassing other classifiers. Moreover, the Shapley Additive Explanations (SHAP) approach was utilized to evaluate the predominant features and their respective contributions towards predicting ground behavior. This study offers valuable insights into the prediction of ground behavior, which can inform future endeavors aimed at enhancing safety and efficiency in tunnel construction projects.