Abstract
The mining industry is encountering a multitude of challenges from an economic, environmental, operational, management and social perspective. The way forward for the mining industry is the principles of 4IR and digitalisation, which include analytics and AI. The value of analytics lies in the ability to take data from various systems and functions and transform it for making decisions. Combining analytics and AI, particularly machine learning, is termed AI-Powered Analytics and is extremely powerful and beneficial. Review of literature indicates the use of AI-Powered Analytics in the mining industry. The current limitation is that most applications are isolated to specific mining operational and business functions. An engineering management approach is needed to bring together core mining engineering with overall business management. Based on this opportunity, this research aims to assess if AI-powered analytics on a mining architecture will enable data-driven decisions for optimisation of holistic business performance.
A TMSA is defined for open-pit mining, indicating the various operational and business systems and how data flows. Published values of mining variables are obtained and with normal sampling methods, a data set of 8,000 records for 42 variables across the TMSA is established. The following operational and business functions are used to indicate holistic business performance: economic including total fuel cost per tonnes handled and ore tonnes handled; mining including production drilled meters and volume blasted; and maintenance including tyre life. The data is pre-processed to remove outliers via the elliptic envelope algorithm and discretized using the quantile method for input into the FP-growth algorithm. The algorithm identifies relationships between variables which are represented as association rules. The rules indicate that with the given variables, target performance of the operational and business functions is achieved. There were no rules identified for the total fuel cost per tonnes handled and ore tonnes handled functions, meaning that in the actual mining process target performance is not achieved often. The production drilled meters function has 151 rules, volume blasted has 29 rules and tyre life has one rule.
Extracting the relationship of variables from the rules a mapping of input variables to the operational and business functions is defined for regression analysis. First, LASSO regression is applied to identify significant variables and then OLS regression for obtaining the output equations. The models are evaluated using the p-value, R2 and RMSE. The production drilled meters model is overfitted to the data with an R2 and RMSE of approximately 100% and 0% respectively. The input variables need to be reviewed for multicollinearity to avoid model complexity or alternate regression methods suitable for multicollinearity such as partial least squares or ridge regression need to be considered. The volume
9
blasted model indicated acceptable model verification parameter ranges. Additional input variables need to be considered for the tyre life model as the R2 is low at 35.15%.
The study resulted in output equations for operational and business functions. With these output equations and multi-objective optimisation methods, decision making for optimisation of holistic business performance can be guided.