Rockburst prediction in kimberlite with unsupervised learning method and support vector classifier

One of the most serious types of mining disasters in many countries, rockburst leads to injuries, deaths, and damages to facilities, which explains the need to study its prediction. However, due to highly non-linear relationship between the occurrence of rockburst and burst impact factors, traditional mechanism-based prediction methods cannot generate precise results. This paper employed a support vector classifier (SVC) to predict rockburst in kimberlite pipes at a diamond mine. We collected 246 groups of data based on real rockburst cases from all over the world as a supportive database. A novel dimensionality reduction method, t-SNE, helped to reduce relevance of original data attributes, and then, an unsupervised learning method (clustering) was adopted to relabel original data to determine relative intensity of these rockburst cases. After the processed prediction data was fed into the trained SVC model, the prediction results were obtained, which matched real rockburst cases that recently occurred at this mine. This data-driven prediction method can be easily conducted and does not rely on the discussion of rockburst mechanism, which has wide potential applications in rockburst prediction in engineering.