Prediction of tunnel rockbursts based on data preprocessing technology considering influences of stress gradient of surrounding rock

As the current rockburst prediction investigation frequently ignores outliers, missing values, sample imbalance in the rockburst dataset and the influences of surrounding rock stress gradient, a complete preprocessing process of rockburst data is proposed, and the hole diameter index that indirectly represents the stress gradient of surrounding rock of tunnel is employed to establish the multi-factor comprehensive prediction model for tunnel rockbursts. At the stage of the data collection, considering the variation in stress conditions between the tunnel, stope and tunnel group, 306 samples of rockbursts in tunnels are isolated from the rockburst database. At the stage of determining prediction index, five indices are selected including the hole diameter (D₀), the maximum tangential stress ( \sigma _\theta \max ), the uniaxial compressive strength ( \sigma _\textc ), the uniaxial tensile strength of the rock (σ_t) and the elastic energy deformation index (W_et). At the stage of the data preprocessing, the multiple imputation method of random forest (MI-RF) is introduced to fill in the missing values. Three unsupervised algorithms including the K-nearest neighbor (KNN), the isolation forest (IForest) and the local outlier factor (LOF) are introduced to comprehensively evaluate the rockburst dataset and removed outliers. The adaptive comprehensive oversampling (ADASYN) algorithm is introduced to expand the number of minority samples. At the stage of the model validation, five types of models including the support vector machine (SVM), the random forest (RF), the gradient boosted decision trees (GBDT), the adaptive boosting algorithm (AdaBoost) and the extreme gradient boosting algorithm (XGBoost) are adopted for comparison. The results demonstrate that the aforementioned models based on the data preprocessing and the hole diameter index are all the best among similar algorithm models. Without the data preprocessing, the model considering the hole diameter index is better than those without considering the hole diameter.