The estimation of platinum flotation grade from froth image features by using artificial neural networks

Abstract

The use of machine vision in the monitoring and control of metallurgical plants has become a very attractive option in the last decade, especially since computing power has increased drastically in the last few years. The use of cameras as a non-intrusive measurement mechanism not only holds the promise of uncomplicated sampling but could provide more consistent monitoring, as well as assistance in decision making and operator and metallurgist training. Although the very first applications of machine vision were in the platinum industry, no automated process control has been developed for platinum group metals (PGMs) as yet. One of the reasons is that to date froth features could not be related to key performance indicators, such as froth grade and recovery. A series of laboratory experiments was conducted on a laboratory-scale platinum froth flotation cell in an effort to determine the relationship between the platinum grade and a combined set of image features and process conditions. A fractional factorial design of experiments was conducted, investigating 6 process conditions, namely air flow rate (x1), pulp level (x2), collector dosage (x3), activator dosage (x4), frother dosage (x5) and depressant dosage (x6), each at levels. Videos were recorded and analysed to extract 20 texture features from each image. By using artificial neural networks (ANN), the nonlinear relationship between the image variables and process conditions and the froth flotation grades could be established. Positive results indicate that the addition of image features to process conditions could be used as sufficient input into advanced model based control systems for flotation plants.