Linear-quadratic level control for flotation through reinforcement learning

In the mining industry, flotation is a commonly used process to separate valuable minerals from waste rock in a concentrator. The rougher flotation is the first stage of the process and in Boliden AB’s concentrator at Aitik, it consists of two lines of four flotation cells each. In this paper we consider one line and the buffer tank upstream of it. Modeling this process step, and maintaining an updated model over time, is a challenge. The process itself changes over time as equipment degrades and parts are replaced. Additionally, the operating conditions in the flotation process change as the ore quality varies. We address these challenges by using reinforcement learning (RL) to design a state feedback controller for level control, without the need of an explicit process model. Using simulations, we compare the performance of the resulting controller to that of the cascade coupled PI-control structure that operates the real plant today. The RL-based controller improves the performance and shows good potential. However, convergence to an admissible control law requires careful hyper-parameter tuning. Industrial deployment thus requires further work to ensure the required reliability.