考慮等候時長限制之深度學習動態允入方法

Abstract

本研究針對串聯式生產系統中具等候時長限制的動態環境，開發基於深度學習的允入控制機制。在半導體產業中，產品在工序間等候時長過長可能導致品質劣化或報廢，造成巨大損失，因此精確的允入控制決策對維持生產效率與降低總成本至關重要。研究首先利用線性規劃分解方法，將高維度優化問題分解為可處理的子問題，克服傳統動態規劃的局限性，並運用動態規劃技術對系統中的隨機事件建模，制定最優的派工和預防保養計劃。設計深度神經網絡模型，以捕捉動態製造系統中的複雜非線性關係，並預測產品於上下游工作站間的等候時長，作為允入決策的核心依據。此方法有效處理生產系統中的動態變化與不確定性，降低產品違反等候時長限制的風險並減少報廢數量，從而提升整體生產效率並最小化總成本。

This research develops a deep learning-based admission control mechanism for serial manufacturing systems with time constraints in dynamic environments. In the semiconductor industry, prolonged inter-process waiting times can lead to quality deterioration or scrapping, resulting in substantial losses. Therefore, precise admission control decisions are crucial for maintaining production efficiency and minimizing total costs. The research first employs linear programming decomposition to break down high-dimensional optimization problems into manageable sub-problems, overcoming the limitations of traditional dynamic programming, and utilizes dynamic programming techniques to model random events in the system for optimal dispatching and preventive maintenance planning. Building upon this foundation, a deep neural network model is designed to capture complex nonlinear relationships in dynamic manufacturing systems, predicting waiting times between upstream and downstream workstations as the core basis for admission decisions. This method effectively handles dynamic changes and uncertainties in the production system, reducing the risk of time constraint violations and minimizing scrap rates, thereby enhancing overall production efficiency and minimizing total costs.