Self-growth learning-based machine scheduler to minimize setup time and tardiness in OLED display semiconductor manufacturing

Abstract

This paper focuses on a scheduling problem of the evaporation and encapsulation (EVEN) line in organic light-emitting diode (OLED) display manufacturing towards minimizing total setup time and tardiness of allocated jobs. This scheduling problem is an unrelated dedicated parallel machine scheduling (UDPMS) problem with practical complex constraints including machine availability, family setup, preventive maintenance (PM), and job splitting. There are three major points to this problem. First, the processing of a job on a machine should be completed before its due date. Second, setups should be reduced to minimize labor costs. Finally, a job is able to be split into one or more sub-jobs for processing on machines independently. To obtain a schedule to minimize total setup time and tardiness of allocated jobs, we suggest two novel machine schedulers such as deep learning-based and self-growth learning-based schedulers, called DLS and SGLS, respectively. Although the structure of the proposed schedulers is identical, especially, SGLS incrementally learns to explore more effective allocation patterns by preserving the obtained knowledge from the trained DLS based on self-growth learning. The comprehensive experiments show that SGLS produces satisfactory schedules in terms of minimizing total setup time and tardiness. In particular, when the number of jobs and the ratio of job sizes are larger, SGLS achieves a schedule with both less setup time and tardiness compared to others.