Practical Applications of Machine Learning to Underground Rock Engineering

Rock mechanics engineers have increasing access to large quantities of data from underground excavations as sensor technologies are developed, data storage becomes cheaper, and computational speed and power improve. Machine learning has emerged as a viable approach to process data for engineering decision making. This research investigates practical applications of machine learning algorithms (MLAs) to underground rock engineering problems using real datasets from a variety of rock mass deformation contexts. It was found that preserving the format of the original input data as much as possible reduces the introduction of bias during digitalization and results in more interpretable MLAs.

A Convolutional Neural Network (CNN) is developed using a dataset from Cigar Lake Mine, Saskatchewan, Canada, to predict the tunnel liner yield class. Several hyperparameters are optimized: the amount of training data, the convolution filter size, and the error weighting scheme. Two CNN architectures are proposed to characterize the rock mass deformation: (i) a Global Balanced model that has a prediction accuracy >65% for all yield classes, and (ii) a Targeted Class 2/3 model that emphasizes the worst case yield and has a recall of >99% for Class 2. The interpretability of the CNN is investigated through three Input Variable Selection (IVS) methods. The three methods are Channel Activation Strength, Input Omission, and Partial Correlation. The latter two are novel methods proposed for CNNs using a spatial and temporal geomechanical dataset. Collectively, the IVS analyses indicate that all the available digitized inputs are needed to produce good CNN performances.

A Long-Short Term Memory (LSTM) network is developed using a dataset for Garson Mine, near Sudbury, Ontario, Canada, to predict the stress state in a FLAC3D model. This is a novel method proposed to semi-automate calibration of finite-difference models of high-stress environments. A workflow for optimizing the hyperparameters of the LSTM network is proposed. The performance of the LSTM network predicting the three principal stresses is improved as compared to predicting the six-component stress tensor, with corrected Akaike Information Criterion (AICc) values of -59.62 and -45.50, respectively.

General recommendations are made with respect to machine learning algorithm development for practical rock engineering problems, in terms of how to format and pre-process inputs, select architectures, tune hyperparameters, and determine engineering verification metrics. Recommendations are made to demonstrate how algorithms can be rendered interpretable with the application of tools that already exist in the field of machine learning.