Application of expert knowledge-guided adaptive multi-objective deep reinforcement learning to grinding-flotation process optimization

Abstract In industrial grinding–flotation circuits, non-stationary feed conditions, coupled system delays, and stringent safety constraints pose significant challenges for conventional control methods to simultaneously optimize concentrate grade, recovery, and energy consumption. To address these issues, this paper proposes a supervisory optimization framework based on expert-knowledge-integrated adaptive multi-objective deep reinforcement learning (DRL). The framework features a three-layer decoupled architecture: a policy layer (Actor–Critic) for generating adaptive control commands, a guardian layer for enforcing safety screening and action projection, and the underlying DCS/PLC for execution. To mitigate deployment risks, a ‘conservative offline → constrained online’ training paradigm is employed, significantly enhancing policy robustness against extrapolation errors in non-stationary environments. By incorporating state-driven dynamic objective weighting and expert-guided safe execution, the framework enables flexible trade-offs among competing production objectives. Experimental results demonstrate that, under a ‘no grade deterioration’ constraint, the proposed method increases recovery by 0.96 percentage points and reduces specific energy consumption by 7.0% compared to the best competing approach. Furthermore, it exhibits superior performance in disturbance rejection speed, Pareto-front advancement, and safety boundary preservation.