本文研究之個案公司為業界指標性電梯廠牌，在市場型態轉變下依然保有優良的銷售表現並且開發更多迎合客戶需求的機種。作為歷史悠久的老字號電梯品牌，個案公司難以避免面臨人力高齡化帶來的人工效益減退，以及多樣化機種帶來的生產排程混亂。 過去十年，個案公司在物料來源與管理層面上持續進行改善，大幅降低庫存壓力與製品成本，但相對帶來的是短促製程衍生欠件出貨與耽誤客戶需求頻繁發生，額外產生難以預測的成本負擔。 上述生產排程問題，過去有學者先賢針對零工式生產型態的傳統產業進行各種啟發式生產排程研究，比如童維新學者之基因演算法對於電梯生產排程之研究。該研究獲得顯著地工時降低成果，但逾期比例提昇75％。真實世界裡，客戶不一定配合工廠最佳排程而接受延期，使得基因演算法縮小可能解空間。故面對客戶對交期的要求，資深的業務人員會依據過去類似機種規格的製作與出貨經驗回覆妥善的交期，並不會勇敢地承諾交期再進行排程，避免推擠其他寶貴客戶的需求。因此，吾人應嘗試將這些經驗交給電腦進行深度學習，使電腦能藉由客戶的規格需要以及目前產線狀況預測可能完工時間，進而使業務能保有對客戶的承諾，而工廠也降低插單造成的成本負擔， 本論文使用個案公司2015年至2020年度所有工廠生產電梯之實際出貨日與工作站進度日期經由人為標籤遲延出貨與正常出貨後，題列各項電梯特徵後得到深度學習精準度達52％，預測差異天數約24天，這表示複雜的零工式排程問題可以深度學習方式得到嶄新參考方向，更代表傳統產業過去依賴經驗安排生產進度的狀況將不會隨著人員異動而面臨困難。 ;The case company studied in this thesis is the industry′s index elevator brand, and it still maintains excellent sales performance and develops more models that cater to customers′ cost needs under the changing market pattern. As a long-established and long-established elevator brand, the case company also faces the decline in labor benefits brought by the aging of direct personnel, and the chaos of production scheduling caused by diversified models. In the past ten years, the case company successfully reduced inventory pressure and product costs under various system and management attempts, but the relative situation is that the shipment of arrears and delays in customer urgent needs frequently occur, but additional Generate unnecessary shipping burden. For the above production scheduling problem, some scholar sages conducted various heuristic production scheduling studies on the traditional industry of the zero-work production type, such as Tong Weixin′s genetic algorithm research on elevator production scheduling. The study used genetic algorithms to target the research company ’s standard process for workstations during a specific period of time and performed calculations to obtain significant reductions in man-hours. However, in reality, customers will not adjust the construction progress according to the factory′s schedule, so if they can learn the production and shipment experience of specific models in the past through deep learning, they will be able to provide a certain amount of advance notice to the production materials personnel and carry out Workstation progress adjustment. This paper uses the actual shipment date and workstation progress date of all the elevators produced by the case company from 2015 to 2020 in the factory. After the delayed shipment and the normal shipment through the artificial label, the depth of the learning accuracy is as high as 52% after listing the elevator characteristics. This means that the complicated zero-work scheduling problem can get a reference direction through deep learning, and it also means that the traditional industry′s past reliance on experience to schedule production schedules will not face difficulties with personnel changes.
