Predicting finished product properties in mining industry from pre-extraction data

Abstract

An essential requirement of product quality control in the mining industry is to be able to reliably predict key quality properties of finished product from the data available before the extraction of the ore. From a production viewpoint, the unit of data collection is generally the input and output data set for each shift of crusher production but could be any period where mine pre-crusher data can be reliably matched with product data. Linear regression models can be used to predict crush grades from blast grades, even where the crush material is blended from multiple sources or pits for each of which differing regression models might apply. The best model for any application will be a balance between required predictability, available data and the tolerance of the business for complex models. The regression modelling approach has several advantages over the classic method of run of mine crusher trials. The models can use any predictor variable such as grade, geotype and in situ density provided the pre-extraction data can be reliably matched with post-crusher data and is significant as a predictor. The models have been used extensively in the generation of the daily crusher plan with the aim of maintaining finished product grade. This approach has also been used associated with exploration drilling and long term planning. It is acknowledged that there are inherent problems in fitting lump and fines grade to a linear model. However, these problems are minor when such information is used for interpolation within the window spanned by the shift blend records used to produce the model. This paper discusses some of the issues limiting linear regression models in this application, and suggests methods enabling consistent models to be formulated.