SYNOPSIS Underground mining gives rise to geotechnical hazards. A formal geotechnical risk assessment can help to forecast and mitigate these hazards. Frequentist probability methods can be used when the hazard does not have many variables and a lot of data is available. However, often there is not enough data for probability distributions, such as in the case of new projects. The risk assessment is often subjective and qualitative, based on expert judgement. The purpose of this research is to present the use of Bayesian networks (BNs) as an alternative to existing risk assessment methods in underground mines by combining expert knowledge with data as it becomes available. Roof fall frequency forecasting using parameter learning is demonstrated with 1141 sets of roof fall data across 12 coal mines in the USA. The prediction is nearly identical for individual mines, but when multiple mines are evaluated it is difficult to find a single best fit distribution for annual roof fall frequency. The BN approach with TNormal distribution was twice as likely to fit the observed data compared to the Poisson distribution assumed in the past. A hybrid approach using BN combining multiple probability distribution curves from historical data to predict annual roof fall is proposed. The BN models can account for variability for multiple parameters without increasing the complexity of the calculation. BNs can work with varying amounts of data, which makes them a good tool for real-time risk assessment in mines. Keywords: Bayesian network; expert opinion models; geotechnical risk; incident forecasting; parameter learning; roof fall risk.
Interpretable deep learning for roof fall hazard detection in underground mines