Exploratory analysis of multivariate drill core time series measurements

Abstract

Demand for mineral resources is increasing, necessitating exploitation of lower grade and more heterogeneous orebodies. The high variability inherent in such orebodies leads to an increase in the cost, complexity and environmental footprint associated with mining and mineral processing. Enhanced knowledge of orebody characteristics is thus vital for mining companies to optimize profitability. We present a pilot study to investigate prediction of geometallurgical variables from drill sensor data. A comparison is made of the performance of multilayer perceptron (MLP) and multiple linear regression models (MLR) for predicting a geometallurgical variable. This comparison is based on simulated data that are physically realistic, having been derived from models fitted to the one available drill core. The comparison is made in terms of the mean and standard deviation (over repeated samples from the population) of the mean absolute error, root mean square error, and coefficient of determination. The best performing model depends on the form of the response variable and the sample size. The standard deviation of performance measures tends to be higher for the MLP, and MLR appears to offer a more consistent performance for the test cases considered.