Deep learning feature-based setpoint generation and optimal control for flotation processes

Abstract

Computer vision-based control is a nonintrusive, cost-effective, and reliable technique for flotation process control. It is known that deep learning features can depict the complex behavior of the froth surface more comprehensively and accurately than handcrafted features. However, few studies have tried to use additional information to improve flotation performance through optimal control. To this end, we have attempted to develop a novel deep learning feature-based two-layer optimal control scheme. The first layer is proposed for setpoint generation of high-dimensional features using improved fuzzy association rule reasoning. Then, an offline conservative double Q-learning control layer that can learn from historical industrial records by mitigating bootstrapping error in action value functions is developed. The proposed method can adapt the setpoint to the change in process feeds. Meanwhile, in contrast to traditional approximate dynamic programming methods that need to interact with real/simulated process systems, this controller can work without any further interactions, which makes it possible to transfer the success of reinforcement learning algorithms to complex industrial process control where opportunities to explore are missing. Experiments demonstrate that the proposed method is effective and promising for practical flotation process control.