以梯度提升樹之演算法預測自動搬運設備之搬運時間–以面板產業為例

Abstract

本研究為探討連接多個製程機台間之自動搬運設備其搬運時間預測，由於搬運工作量分布的不平衡，導致部分機台上的在製品無法及時搬離上游機台輸出端或無法及時將物料搬運至下游機台進行加工，因回堵及缺料等原因造成加工機台的停機，間接影響部分廠區生產排程規劃。因此本研究透過梯度提升樹架構之演算法預測自動搬運設備之總搬運時間，以利掌握整體產線狀況，進而規劃出更為緊湊、更高產能之生產排程。在特徵構造的階段根據分析原始資料、製程特性與指令派工邏輯，創造出更多有利於模型訓練之特徵，如指令產生當下在製品堆積在機台輸出端數量、搬運設備之間暫存區狀態、搬運路徑過往搬運平均時間、搬運路徑長等特徵。本研究為預測數值之迴歸問題，其資料型態包含連續型的特徵與類別型的特徵，與類神經網路架構模型，如長短期記憶網路(LSTM)、深度神經網路(DNN)比較後，得出梯度提升樹(GBDT)架構之演算法更有利於同時處理這兩類型特徵值，因此採取GBDT作為預測模型。預測模型分為全範圍設備預測模型與單一設備預測模型，結合兩者每當指令完成一站搬運後再度進行預測以修正誤差，並將平均誤差由原本使用原始特徵的88%透過特徵構造、調整模型與分段預測後下降至17%。

The study is aim to predict the transportation time of AMHSs between numbers of process machines in panel industry. Some of the AMHSs have heavier workload which usually can’t finish the tasks in time. It will cause the upstream machine blocked or downstream machine shut down due to lack of material. This research creates kinds of new features based on original features to help us improve the prediction accuracy, such as moving distance, buffer status, queue length. In this study, data include not only numerical type but also categorical type. Gradient boosting decision tree (GBDT) has the greatest result in two types of data when compare with long short-term memory (LSTM) and deep neuron network (DNN). This is the main reason why GBDT be chosen as prediction model. With generate new features and multi-phase prediction, which combined with single machine model and the whole area model will renew the prediction time when a task arrives at the next AMHS. Finally, MAPE can be reduced from 88% to 17% by feature construction, fine-tuning model and multi-phase prediction.