Adaptive closed-loop production control for discrete manufacturing systems

Scott Andrew Moses, Purdue University

Abstract

Discrete manufacturing systems rarely reach steady state: they require constant control while operating in a transient phase. To date, very little research has considered the systematic aspects of closed-loop control of these systems. At the sequencing level for example, when a composite dispatching rule is used then a control loop can be defined that modifies the values of decision variables in the rule to control the tradeoff between multiple performance objectives. In this research an adaptive closed-loop production controller was developed for a discrete manufacturing system. The controller estimates system state and reasons with qualitative knowledge to infer the values of decision variables such that a satisfactory performance trajectory is achieved. The controller does not require an analytically tractable model of the manufacturing system, and it uses reinforcement learning to acquire control knowledge even though the reinforcements received are very noisy. Structurally it is analogous to a nonlinear proportional-derivative feedback controller since its antecedents are the performance and the performance trajectory. Estimating the precise trajectory from the observed performance is difficult, so a fuzzy algorithm for transient state detection was formulated that estimates the qualitative rate of change in system congestion. In closing the control loop unexpected results were encountered that led to revealing insights into the nature of planning, sequencing, and control. From these results three significant generalizations about planning and sequencing were developed: (1) Planning is a feasibility problem, (2) Discrete-event simulation is inappropriate for planning, and (3) Planning (not sequencing) determines the mean system performance.

Degree

Ph.D.

Advisors

Solberg, Purdue University.

Subject Area

Industrial engineering|Systems design|Operations research

COinS