On-line automatic controller tuning of a multivariable grinding mill circuit using Bayesian optimisation

Abstract

Process controllers are abundant in the industry and require attentive tuning to achieve optimal performance. While tuning controllers by the most primitive method of trial and error is possible, it often leads to sub-optimal performance if not conducted by a skilled expert. It is much more appealing to develop an on-line, sample efficient, automated tuner which can optimise the performance of a given controller to the task at hand. The automatic tuning procedure can be conducted during commissioning, when poor controller performance is observed or when process conditions have changed. The problem statement is formulated as the minimisation of an objective function constructed to achieve the desired controller performance. In this context the automatic tuning problem of multi-input multi-output (MIMO) controllers is considered within the framework of Bayesian optimisation and applied in simulation to an ore milling circuit with three manipulated and three controlled variables. Regulatory and set point tracking controllers are tuned automatically and are shown to achieve better performance than a reference controller.