Open Collections

Yang, Beverly

University of British Columbia

The past decade has seen a significant increase in the use of machine learning (ML) in rock engineering. While ML has the potential to revolutionize rock engineering by increasing efficiency and lowering costs, the ML models first need to be reliable. Developing reliable ML models requires good quality (i.e., objective) data and enough of this good quality data such that it represents the problem. These data requirements pose a challenge to rock engineering when considering rock mass characterization and classification data as there are often limited amounts of this data, and this data is subjective. This leads to the main research question that this thesis answers: can we develop reliable ML models with rock mass characterization and classification data? This thesis examines the reliability of commonly used rock mass characterization and classification parameters (rock quality designation, rock mass rating (RMR), the joint condition rating in RMR, Q-system, and geological strength index) by providing a rediscovery of their assumptions and limitations, as well as investigating their subjectivity. The results of this analysis demonstrate that these parameters are unreliable due to their misuse and inherent subjectivity. This thesis also examines the reliability of developing ML models on dataset sizes commonly encountered with rock mass characterization and classification data (i.e., a few hundred data points). Using a synthetic dataset, this analysis demonstrates that ML models trained on these smaller dataset sizes are unreliable as they are susceptible to how the data was split into training and test data. A practical methodology is proposed to assist rock engineers with determining if they have enough data to develop reliable ML models. The main conclusion of this research is that it is not possible to develop reliable ML models for predictive purposes from current rock mass characterization and classification data due to their subjectivity and often-times limited quantities. However, developing reliable ML models is possible when working with objective characterization and classification data. A summary of best practices for the development of reliable ML models using rock mass characterization and classification data is proposed and demonstrated in a case study.

Text

2024-10-28

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89560

Mining Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/