| description abstract | The classification of surrounding rock mass is essential for characterizing rock properties and geological conditions in tunneling engineering. While numerous empirical rock mass classification systems have been proposed (e.g., rock mass rating system, rock structure rating system), they tend to heavily rely on engineers’ experience, which is unfavorable for tunnel construction, particularly in deep-buried and ultralong tunnels. Alternatively, machine learning, i.e., artificial intelligence (AI), methods estimate the classification of the surrounding rock mass using certain readily available rock indices (e.g., volumetric joint count). However, most machine learning models are considered black box models, leading to unexplainable predictions. In addition, employing all measurements of readily available rock indices as input may lead to excessive model complexity and a reduction in generalization performance. In this case, a Gaussian process classification (GPC) approach combined with the synthetic minority oversampling technique (SMOTE), Bayesian framework, and SHapley Additive exPlanations is proposed in this study for the probabilistic classification of the surrounding rock mass and selection of the optimal GPC model based on imbalanced and sparse measurement data. It is worth noting that the proposed method can also provide physics-informed explanations for the prediction and model class selection results and determine the significant input variables for each grade of the surrounding rock mass. A real-life example is employed to illustrate and validate the proposed approach. The results show that the F1 score of the optimal GPC model reaches 0.93, which is comparable with those of the GPC model with all input variables. | |
