Clustered Multi-Task Sequence-To-Sequence Learning For Autonomous Vehicle Repositioning

Clustered multi-task learning, which aims to leverage the generalization performance over clustered tasks, has shown an outstanding performance in various machine learning applications. In this paper, a clustered multi-task sequence-to-sequence learning (CMSL) for autonomous vehicle systems (AVSs) in large-scale semiconductor fabrications (fab) is proposed, where AVSs are widely used for wafer transfers. Recently, as fabs become larger, the repositioning of idle vehicles to where they may be requested has become a significant challenge because inefficient vehicle balancing leads to transfer delays, resulting in production machine idleness. However, existing vehicle repositioning systems are mainly controlled by human operators, and it is difficult for such systems to guarantee efficiency. Further, we should handle the small data problem, which is insufficient for machine learning because of the irregular time-varying manufacturing environments. The main purpose of this study is to examine CMSL-based predictive control of idle vehicle repositioning to maximize machine utilization. We conducted an experimental evaluation to compare the prediction accuracy of CMSL with existing methods. Further, a case study in a real largescale semiconductor plant, demonstrated that the proposed predictive approach outperforms the existing approaches in terms of transfer efficiency and machine utilization.