AN INSIGHT OF ROCKBURST INTENSITY PREDICTION USING MACHINE LEARNING

Abstract

Rockburst is a violent explosion of rock, which happens mostly in high geo-stress conditions in underground excavations. This phenomenon occurs due to of the sudden release of strain energy stored in rock mass. It is of one the most serious hazard in underground mining as it can lead to economic losses (damage of equipment and underground structures), and most importantly to the injury and causality of workers. Therefore, many researchers have extensively investigated rockburst phenomenon in order to develop more reliable prediction models. Many predictive models based on Machine Learning have been proposed. However, as all these models are different, they show a wide range of capability depending on many factors, such as number of input variables, data quality among others. The aim of this thesis is to predict rockburst severity through machine learning techniques the predictive models for rockburst intensity. To this end, two different databases were employed. The first one consists of 344 cases containing info of rockburst occurrence, it is intensity and mechanical rock properties. The second encompasses 254 rockburst cases induced by seismic events and containing multiseismical data. Several machine learning algorithms were implemented. The results indicated that common machine learning developed models classify events with near 60% prediction accuracy. After the reduction in the number of categories of rockburst severity classes prediction accuracy increased up to 86%. The results obtained are far from existing works, where the prediction accuracy approaches 100%. However, after using their test sets, the accuracy increased significantly up to 80 percent, given that this thesis used conventional methods without hyperparameters and algorithms. Therefore, it can be assumed that the used test set is biased. It was concluded that after reduction the groups and increase of prediction accuracy with individual developed models can be applied in site-specific with suitable severity events. Otherwise, it is necessary to increase the databases to further improve the models for wider use in mines with rock bursts of varying severity. This study allows us having a new insight of rockburst prediction on existing models and it’s application for machine algorithm based rockburst severity prediction models.