Data-driven Optimal Operational Feedback Control Methods For Mixed Sorting Process

The mixed sorting process is an important and efficient benefication method,including grinding process,thickening process and flotation process,thickening process and flotation process are main points of this paper.Thickening process is a physical process to realize the solid-liquid separation of slurry according to gravity settlement principle.Flotation process is a physical process to separate useful metal and gangue using hydrophilicity or sparseness of metal itself or that coming from chemicals impacts.Thickening process is up-stream link of flotation process,industrial control goals are to keep the underflow slurry concentration of thickening process stable and within the constraints of the production process,and also guarantee the concentration grade and the tail grade of flotation process are within the target ranges after the underflow slurry is fed,besides,the relative process variables are also kept within the physical allowable ranges.Thickening process and flotation process are both have complex dynamic structures and dynamic characteristics.For these two complex processes,the main research work of this paper are as follows:1.The big difference in physical characteristics between variables in thickening process leads to two-time-scale problem,along with the uncertainty and time-varying of parameters in thickening system making the traditional model-based control methods lose their application value.Therefore,based on the two-layer dynamic model and industrial control goals,firstly,to solve two-time-scale problem existed in thickening system,this paper proposes singular perturbation method to separate fast variables and slow variables according to the differences between their rates of change and then decomposes the original system with stiffness into two independent subsystems called singularly perturbed system,so stiffness is eliminated.Next,since non-linear,uncertainty and other complex features make it quite difficult to establish accurate model of thickening process,establish the performance index and apply Q-learning to singularly perturbed system to learn the optimal control policy with only measurable data,without knowing the dynamics of thickening process.The above work achieves the combination of singular perturbation and reinforcement learning called industrial singular perturbation reinforcement learning,making it more efficient and easier to realize optimal operational tracking control of the thickening system with two-time-scale problem and guarantee the satisfactory tracking performance to keep desired and stable slurry concentration given to the flotation process and fluctuation of variables within the physical allowable ranges.2.As a typical complex industrial process,flotation process has complex characteristics like non-linear,multi-input and multi-output,strong coupling and time-varying parameters.At first,for its non-linear,time-varying parameters and random disturbance,taking advantage that flotation process operates in its operating points,this paper introduces unmodeled dynamics,establish the input and output dynamic model of flotation process and formulates its system as the sum of linear part and unmodeled dynamics,the latter includs high-order term after linearization and fluctuation caused by random disturbance and changes of parameters near their operating points.According to unmodeled dynamics at last time instant being measured accurately,compensate the current unmodeled dynamics by the last one and consequently reduce the current unmodeled dynamics in system to avoid the bad effects to system stability and control due to over big unmodeled dynamics value.Secondly,considering the negative effects brought by compensated unmodeled dynamics,this paper takes it as noise and proposes H infinity control based on reinforcement learning for compensated flotation system,and use off-policy reinforcement learning to learn the optimal control policy to achieve data-driven optimal operational control,enhance robustness of system and realize stable operating,keeping the concentration grade and tail grade within their ranges and obtaining satisfactory tracking performance within the allowable error range.Above work achieves data-driven optimal operational feedback control for flotation process based on the combination of unmodeled dynamics compensation and reinforcement learning H infinity control.3.Finally,to testify the effectiveness of data-driven optimal operational feedback control for thickening system with two-time-scale problem based on singular perturbation and reinforcement learning,and data-driven optimal operational feedback control based on unmodeled dynamics compensation and reinforcement learning H infinity control for flotation process proposed in this paper,MATLAB simulation platform is used to conduct simulation experiments based on thickening process and flotation process,and simulation results are given,including two groups of comparative experiments.