Machine Learning Approaches for Long-term Rock Burst Prediction

• Author / Creator

  Pu,Yuanyuan

• A rock burst refers to a sudden collapse and outburst of rocks mostly from the surface of underground excavations. As one of the most serious geological disasters, rock bursts have killed hundreds of miners and injured more. A problem that remains to be solved in rock engineering is to predict an impending rock burst accurately, and more specifically, to determine the rock burst’s location, time, and severity. The problem of developing an accurate prediction of rock bursts can be further refined by the following two aspects: long-term rock burst prediction and short-term rock burst pre-warning. Long-term prediction serves for the design stage of a project, such as the pre-excavation of a tunnel or the pre-mining of a workface, which can be regarded as an integral assessment for the rock burst potential for the engineering field. By contrast, rock burst pre-warning serves the project construction stage, such as in the excavation of a tunnel and in the proceeding of a mining workface and focuses on the specific location and time of an impending rock burst. In this research, we mainly focus on long-term rock burst predictions, although in the literature review chapter, we also discuss some research processes for short-term rock burst pre-warning.
  Long-term rock burst prediction can be abstracted as a classification problem mathematically, in which we can introduce machine learning to solve it. Further, using machine learning in rock burst prediction can overcome drawbacks such as subjectivity and inconsistency brought from previous traditional approaches. Both supervised learning and unsupervised learning classification models are employed in this research to predict rock bursts over the long term in a diamond mine in Northern Canada.
  In the literature review chapter of this thesis, several research terms were introduced and further explanation was given some potentially confusing terms such as the long-term prediction and short-term pre-earning. With main goal of achieving the development of a more accurate form of rock burst prediction method, we have investigated the traditional approaches and machine learning approaches. Meanwhile, the thesis points out the pitfalls of the current use of machine learning for this task as well as potential solutions. The third chapter presents a novel strategy to build the initial ground stress field, which is the main premise when starting new underground excavations. A Decision Tree model is used to back analyze the initial ground stress based on data collected from the field and the previously built finite element model. The fourth chapter includes three machine models in the rock burst prediction task: The Support Vector Machine (SVM), the Generalized Regression Neural Network (GRNN), and the Decision Tree (DT). SVM is mainly used to explore the feasibility of using machine learning in rock burst prediction since the SVM is a fundamental classification model. The GRNN and DT are employed to predict rock bursts while considering some different characteristics of training data in the rock burst prediction task. Our simulations showed that the GRNN performs well with a small dataset while the DT works well with an incomplete dataset. Finally, at the end of chapter four, we compare two categories of fundamental classification models, the generative model and the discriminative model, and we draw a conclusion that the discriminative model is more suitable for rock burst prediction task. The fifth chapter presents a special situation when we can not trust the training labels because of various reasons. In the chapter six we talk about the essential role of the backfill in rockburst control. As backfilling process reduces the rock surface exposure and reduces mining induced stress concentrations. In this chapter a Gaussian process model was built to predict the required strength of cemented rockfill for a backfill. Essentially, The research in this thesis systematically introduces machine learning approach into rock burst prediction. The prediction results at a diamond mine can be matched with the observations of the rock burst cases from the field, which verifies the success of proposed methodology of this research.

• Subjects / Keywords

  • Rock burst
  • Machine learning

• Graduation date

  Fall 2019

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-7dwn-5c22

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.