Adaptive Control for the Surface Attitude of a Multi-Degree-Of-Freedom Vibration Screen Using Deep Reinforcement Learning

Abstract

The screen surface attitude is an essential factor that determines the vibration screening performance. Due to the complexity and uncertainty of the material screening process, it is still a challenging task to achieve optimal control of screen surface attitude. In this paper, reinforcement learning (RL) was used to solve this problem. Firstly, the dynamic process of grain vibration screening was simulated using the discrete element method (DEM). Two indicators, uniformity coefficient K and dispersion coefficient P, were proposed to describe the distribution of the material on the screen surface. Secondly, reasonable value ranges of coefficients K and P were determined according to the DEM simulation results, and a novel reward function was proposed using coefficients K and P. Then, based on the double deep Q-network (DDQN) architecture, a control system for screen surface attitude, including inclination angle α and horizontal angle β, was developed. After training, the designed DDQN model showed good convergence. An intelligent control system was developed, and validation tests were carried out on a multi-degree-of-freedom (multi-DOF) vibration screening test rig under different material feeding conditions. The results indicated that the control system was effective and robust. Compared with the conventional reciprocating vibration screening with a fixed screen surface attitude, the screening loss rate was effectively reduced.

Keywords: Vibration screening, Screen surface attitude, Control method, Double Deep Q-Network