深度強化學習處理 T/C 平衡生產排程於面板製造商組立製程

Abstract

隨著薄膜電晶體液晶顯示器（TFT-LCD）需求量的增加，對於生產排程的效率要求也隨之提升。本研究聚焦於TFT-LCD 組立製程，並將此製程歸納成為動態彈性零工式排程問題（DFJSS）。此排程問題主要解決薄膜電晶體（TFT）陣列基板與彩色濾光片（CF）基板之間的生產平衡，稱為T/C平衡。為了貼近真實生產環境，我們考慮了工作插單與生產時間之不確定性。本研究提出一個深度強化學習（DRL）框架以同時處理多個生產排程目標，包含總加權延遲時間、總完工時間、過度等候時間和T/C平衡。我們基於實證資料進行數值實驗，以評估本研究所提出的深度強化學習框架，並將其排程表現與傳統的最佳化模型和基因演算法進行比較。實驗結果顯示，相較於基準模型，本研究的方法在目標值上提升了約30%，並且每次處理工作插單只需約莫5秒鐘。這證實了在動態的生產情境下，此框架具有優良的效力與適用性，同時在面對生產環境的不確定性時，也能夠保持一定程度的穩健性。

The rising demand for Thin-Film Transistor Liquid Crystal Display (TFT-LCD) has amplified the need for efficient manufacturing process. In this study, we focus on TFT-LCD cell process, generalized as a dynamic flexible job shop scheduling problem (DFJSS). The scheduling problem tackles the production balance between Thin-Film Transistor (TFT) array substrate and Color Filter (CF) substrate, known as T/C balance. To align with real-world manufacturing environment, new job arrival and uncertain processing time are considered. A deep reinforcement learning (DRL) framework is proposed to address multiple objectives simultaneously, including total weighted tardiness, makespan, over-queued time and T/C balance. To validate the proposed DRL framework, numerical experiments based on empirical data are conducted to compare its performance with traditional optimization-based models and genetic algorithm. The result shows proposed framework achieves about 30% improvement in objective value compared to benchmark models, while handling each new job arrival around 5 seconds. This demonstrates the effectiveness and applicability in dynamic manufacturing scenario, while maintaining a certain degree of robustness against uncertainty.