Gaussian process modelling of an industrial flotation bank

Abstract

A control-oriented Gaussian process regression (GPR) model of froth flotation is developed and compared to a previously developed parametric model. The model aims to predict the behaviour of froth flotation, taking into consideration which state variables are available from measurements: air recovery, top of froth bubble size, and pulp level. The framework encodes prior knowledge of a published flotation model. Each state is modelled using a separate GP, with a custom covariance function whose form is given by the flotation model. These kernels capture the interaction between the relevant state variables and manipulated variables. The model aims to balance the complexity required to explain such a complex process with the uncertainty of its instrumentation. To evaluate the ability of the GPR model to capture the process dynamics, the GP model is assessed using an industrial data set, demonstrating its capacity to improve the performance of state prediction. The purpose of the GPR model is to enable supervisory and advanced model-based control.

HIGHLIGHTS • A Gaussian process regression (GPR) model is developed using industrial online froth flotation data. • The kernels for the GPR model are based on modelling insights. • The predictive capacity of the GPR model is better than that of a dynamic semi-mechanistic model. • The GPR model shows potential for use in model predictive process control.