ABSTRACT
It is often time prohibitive for geotechnical professionals to examine large datasets to investigate complex rock mass phenomena in detail. Machine learning algorithms (MLAs) are gaining momentum for data processing in rock engineering, however research into their practical applications are just emerging. These multivariate rock mass datasets are ideal for developing MLAs to forecast rock mass behaviour. An Input Variable Selection (IVS) approach is presented for a Convolutional Neural Network (CNN) that predicts tunnel liner yield due to squeezing ground conditions at the Cigar Lake Mine. A IVS method called Input Omission (IO) is modified and applied to the CNN to enhance its performance. The IO method ranks the CNN inputs in terms of usefulness for forecasting the output. The IO findings for this CNN indicate that none of the available inputs may be omitted, and that the geotechnical zones and radial tunnel displacement inputs contain the strongest signals for forecasting the severity of tunnel liner yield at Cigar Lake Mine.
