Andritz: Thickener Optimization

Control of complex, dynamic mining machinery has always been a difficult and labor intensive task. Developing an automated controller usually requires an extensive development cycle and countless domain experts to codify rulesets and operating conditions. Thickeners in particular are tough to control due to the large amount of time before changes to the system are observed. A controller that can learn from data alone would be highly beneficial to this domain.

PSIORI is currently developing an algorithm that employs the latest advances in Deep Reinforcement Learning to solve the thickener control problem. By using reinforcement learning, the operation of the thickener can be learned from a simulation, allowing the controller to encounter all operating situations and react effectively.

A reinforcement learning approach allows a controller to be developed and deployed for countless thickeners without the need for a completely new redesign. Such a controller optimizes underflow while reducing or even eliminating emergency conditions and thickener downtime. Avoiding these situations saves the plant weeks of downtime and therefore avoids the costs associated with such thickener disruptions.