An intelligent controller for manufacturing cells
Abstract
A flexible scheduling system is essential if multiple performance objectives are to be met and uncertainty handled during production. In this study, a learning-based, real-time scheduling system for controlling a manufacturing cell is presented. The issues regarding the decision time are discussed first and analyzed by simulation. To develop the controller, an off-line module utilizes simulation to generate training samples to initialize the knowledge bases within the control system. During the operation, the difference between the performance level and the output from knowledge bases will be used to update and maintain the knowledge bases through reinforcement learning. To accomplish both off-line training and reinforcement learning, a CMAC (Cerebellar Model Articulation Controller) network is adopted to develop each knowledge base. A heuristic is proposed to select the parameters during the batch training so that a network with better generalization ability can be obtained. The experimental results from simulation show that the controller performs well under multiple criterion environments. Also, the control system can adapt itself to new production conditions. The simulation results show significant improvement in system performance when reinforcement learning is incorporated in the feedback loop. In addition, the proposed system demonstrates its adaptability to changing objectives in a dynamic production environment.
Degree
Ph.D.
Advisors
Yih, Purdue University.
Subject Area
Industrial engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.