Horizontal attitude control method for vibrating screen surface using deep reinforcement learning

High performance of vibrating screening is often required under dynamic material-feeding conditions. In this study, a control system was proposed for horizontal attitude of vibrating screen surface using the double deep Q-Network (DDQN) deep reinforcement learning. Taking the vibrating screening of rice grains as the research object, a multi-degree-of-freedom dynamics model was established using the discrete element method (DEM). Materials transportation on the screen surface was obtained through the screen apertures after DEM simulations. The evaluation index was proposed for the uniform distribution of material on the screen surface. There were the correlations between the materials distribution state on the screen surface and the screening performance. A systematic investigation was conducted to explore the influence of the material feeding distribution and the horizontal attitude angle of the screen surface on the screening performance, providing for the optimal range of the uniformity coefficient of the material distribution on the screen surface. Optimal regions below the vibrating screen surface were determined to monitor the materials passing through the screen apertures. A BP neural network was established to real-time predict the uniform distribution of materials on the vibrating screen surface. Neural network demonstrated excellent convergence and generalization after training. Action and state space of vibrating screening were determined under the DDQN framework. A reward function was constructed with the dense and sparse terms. Then, the main hyper-parameters of the DDQN model were optimized to construct the horizontal attitude control model of vibrating screen surface. The DDQN model was trained after DEM simulation, and then applied into the multi-degree-of-freedom hybrid vibrating screening. A series of tests were conducted on a multi-DOF vibration screening under different materials feeding conditions. The horizontal attitude angle of the screen surface was optimized after validation. Once the materials were uniformly fed onto the vibrating screen surface, the ideal horizontal attitude angle of the screen surface was 0°. Therefore, there was no significant difference in loss rate, compared with the conventional fixed horizontal attitude angle. Once the material feeding coefficient increased from 0 to 0.375, the loss rate increased rapidly from 0.59% to 0.98% under the fixed horizontal attitude angle. In improved control system, the horizontal attitude angle of the vibrating screen surface was adaptively adjusted for the movement direction of the materials on the screen surface. The uniformity of material distribution was effectively improved, whereas the loss rate only increased to 0.68 %. The loss rate decreased by approximately 30.6%. The feasibility of vibrating screen attitude was also verified in the control system using deep reinforcement learning. The finding can provide an effective solution for the materials screening performance under dynamic feeding conditions. The promising prospects can also offer in the field of intelligent control for agricultural screening equipment.