應用深度學習於AMHS最短傳送時間之分析

Deep Learning Analysis for AMHS Shortest Delivery Time

許哲綱(Che-Kang Hsu)

指導教授 ： 顏家鈺

國立臺灣大學/工學院/機械工程學系/博士(2023年)

https://doi.org/10.6342/NTU202300376

摘要

自動化物料搬運系統 (AMHS) 在超大晶圓廠被廣泛運用，可降低人力搬運的風險也增加傳送的速度，晶圓製造中通常利用高架懸掛式運輸系統 (OHT, Over-head hoist transport )傳送晶圓盒(FOUP, Front Opening Unified Pod)搭配自動化倉儲(Stocker)連結成完整的傳送網絡。自動化物料搬運系統(AMHS)的控制中需要在最短時間內完成傳送避免生產機台空轉，最先進的系統包括數千公尺的軌道、數百個交叉路口以及數千個上下貨點和緩衝區。新的晶圓廠更大、更複雜。儘管有這種複雜性，傳輸的路由仍然相當簡單。大多數系統都基於最短路徑，將長度或權重分配給每條軌道，然後起始點和目的地之間的最短路徑進行計算。而無考慮當前流量或軌道負載，從起始點到目的地的每次傳輸都採用接近的路徑。在軌道擁塞或特殊軌道設計的情況下，這種路徑計算傳送平衡的方式限制了整個系統的傳送量。

在過去的文獻中主要研究傳送的最短路徑以降低運送時間並增加產能，在實際晶圓廠管橋的傳送行為中觀察到必須考慮動態傳送路徑中的車數及事件，利用平均分流的方式分散OHT避免塞車。本論文以深度類神經網路 (DNN) 產生一個有效的深度學習模型，於網路架構中加入貼近實際生產環境的函數。利用深度類神經網路建立高效率傳送系統的預測模型，當傳送條件改變的時候，即時計算適合的命令特徵，選擇最快路徑，使 AMHS 發揮最大的傳送效能，進而提升產能。

關鍵字

自動化物料搬運系統 ； AMHS ； 高架懸掛式運輸系統 ； 自動化倉儲 ； 深度類神經網路

並列摘要

Automated material handling system (AMHS) is widely used in ultra-large fabs, which can reduce the risk of manual handling and increase the speed of transmission. In wafer manufacturing, overhead hoist transport (OHT) is usually used to transfer FOUP (Front Opening Unified Pod) combined with automated storage (Stocker) to form a complete delivery network. The control of the automated material handling system needs to complete the transmission in the shortest time to avoid idling of the production machine. The most advanced system includes thousands of meters of tracks, hundreds of intersections, and thousands of loadings and unloading points and buffer zones. The new fabs are larger and more complex. Despite this complexity, the routing of transmissions remains fairly straightforward. Most systems are based on the short-est path, assigning a length or a weight to each track, and then calculating the shortest path between origin and destination. Regardless of current traffic or track load, every transfer from origin to destination takes the same path. In the case of track congestion or special track designs, this lack of path calculation balance limits the throughput of the overall system.

In the past literature, the shortest path of transmission was mainly studied to re-duce the transportation time and increase the production capacity. It was observed in the transmission behavior of the actual fab that the number of cars and event in the dy-namic transmission path must be considered, and the average distribution method was used to disperse the OHT to avoid traffic jams. In this paper, a deep neural network (DNN) is used to generate an effective deep learning model, and functions close to the actual production environment are added to the network architecture. Utilize deep neu-ral network to establish a prediction model of high-efficiency transfer system. When the transmission conditions change, the appropriate command characteristics are cal-culated in real time, and the fastest path is selected, so that AMHS can maximize the transmission performance and increase production capacity.

並列關鍵字

Automated material handling system ； AMHS ； overhead hoist transport ； automated storage ； deep neural network

參考文獻

參考資料

[1] Jacky Tung; Tina Sheen; Merlin Kao; C.H. Chen, "Optimization of AMHS de-sign for a semiconductor foundry fab by using simulation modeling," in 2013 Winter Simulations Conference (WSC),2013.

[2] Kim, H., Lim, DE. , "Deep-Learning-Based Storage-Allocation Approach to Im-prove the AMHS Throughput Capacity in a Semiconductor Fabrication Facility," AsiaSim 2018: Communications in Computer and Information Science, vol 946, pp. 232 - 240,2018.

[3] P. Cowling and M. Johansson, "Using real time information for effective dynam-ic scheduling," Eur. J. Oper. Res., vol. 139, no. 2, pp. 230-244, 2002.

[4] Moyne, James, and Jimmy Iskandar. "Big data analytics for smart manufacturing: Case studies in semiconductor manufacturing," Processes 5.3,39, 2017.