Abstract:

This paper aims to accurately predict rockburst levels and ensure the safety of construction personnel and equipment. First, from the perspective of rock burst mechanism, eight indicators of burial depth (D), uniaxial compressive strength (UCS), uniaxial tensile strength (UTS), rock brittleness index (B₁、B₂), maximum tangential stress (MTS), stress concentration coefficient (SCF) and elastic deformation energy index (W_et) were analyzed, and a rock burst prediction index system was established. Secondly, to address the problem of data imbalance in rockburst samples, the Tomek Link of under sampling method was introduced to improve the (Smote) for mixed oversampling of rockburst training samples. Finally, the SmoteTomek-GBDT rockburst prediction model was constructed, and the validity of the model was verified with 38 sets of data and compared with other models. The results show that the accuracy of SmoteTomek-GBDT is 92.1%, and it is a 5.3% improvement over unsampled and 10.5% improvement over Smote sampled, which is better than the random oversampling model, and avoids cross-grade rockburst misclassification, which is of some significance for accurate rockburst prediction.

Key words: rockburst prediction, gradient boosting decision tree (GBDT), synthetic minority oversampling technique (Smote), rockburst index, Tomek Link