| Summary: | Rockbursts occur in many deep underground excavations and have caused non-negligible casualties or property losses in deep underground building activities over the past hundreds of years. Effective early warning approaches to judge the practical situation of a rock mass during excavation are one of the best ways to avoid rockbursts, while proposing high demands for monitoring data and computational methods. In this study, a data-driven method based on spectral clustering to predict rockburst intensity was proposed. Considering the fact that the original spectral clustering has some defects, an improvement strategy that selects K-medoids, or an improved variant of K-medoids to replace the original K-means clustering as the latter clustering process, was executed. First, the hyperparameters and selections of the latter clustering algorithms were determined, and improved K-medoids with related hyperparameters were determined by 65 rockburst samples collected in underground engineering cases. Based on the previous configurations of flow and hyperparameters, the remaining 17 samples were labeled using a concise labeling flow, which was also based on spectral processes in spectral clustering. The results of the control experiments show that the proposed method has certain feasibility and superiority (82.40% accuracy performance) in rockburst intensity prediction for underground construction.
|
|---|