A new method to predict rock fracture based on Gradient Boosting Decision Tree via multidirectional crack vibration monitoring

Precise prediction of rock fracture is essential for the accurate monitoring and early warning of dynamic disasters in underground engineering. To achieve this, a multidirectional crack vibration (MDCV) monitoring experiment was conducted during the uniaxial compression fracture of rock using vibration acceleration sensors with directional sensing capabilities. The crack vibration statistics in all rock fracture directions were calculated using rolling windows, and the Autoregressive Integrated Moving Average (ARIMA) model was applied for parameter extraction, yielding rolling statistical features. The results indicate that these features exhibit a clear response to rock fracture. The amplitude of rolling standard deviation (R_std) and rolling mean (R_m) exhibit a sudden amplitude increase preceding fractures. Additionally, the density of high-amplitude points for rolling skewness (R_sk) and rolling kurtosis (R_k) significantly increases before the fractures; Meanwhile, the rolling quantile 25 (R_25%) and rolling quantile 75 (R_75%) demonstrate greater sensitivity to the initial fracture stage. In addition, the characteristics and importance of rolling statistic features differs across different fracture directions, highlighting the necessity of MDCV monitoring for the rock's instability assessment and fracture dynamics. Ultimately, a rock fracture prediction model was established via Gradient Boosting Decision Tree (GBDT) method and validated through MDCV monitoring experiments on rocks subjected to different loading speeds. The results demonstrate that the proposed model effectively predicts fracture events with high warning accuracy and strong generalization ability, remaining unaffected by variations in loading speeds. This research offers a robust and scalable approach for early warning of dynamic disasters in underground engineering worldwide.