機器學習之物料搬運時間預測系統

Abstract

本研究探討生產排程給定下的自動物料搬運系統的運輸時間預測。研究對象是TFT-LCD面板製造的cell廠區。為了對時間進行預測，本研究從歷史資料中分析並研擬四種具有不同特徵的資料供機器學習。機器學習過程，藉由多層感知機和深度神經網路的基本神經網路模型與集成式的CatBoost模型為基礎，開發出可以預測當派貨系統發出指令到指令完成所需的時間的模型。通過誤差函數計算輸出結果與正確答案間的誤差，並使用選定的優化器與算法更新神經網路的權重和CatBoost的殘差。反覆的更新權重以及修正殘差，直到CatBoost模型和神經網路的計算結果與正確答案家的誤差值收斂，得到預測模型的最佳權重。數值測試結果顯示CatBoost模型和深度神經網路模型為基礎的運輸時間預測模型雖然具有預測時間的效力，但最終無法改善歷史資料顯示的預測準確率。此外，運輸時間模型雖未能達到預期的準確率，但依然可以提供使用者一項做決策的工具。

This research focuses on the prediction of the transportation time of an automated material handling system for a given production schedule in a cell plant for TFT-LCD panel manufacturing. In order to predict the time, four types of data with different features are analyzed and developed from historical data for model training. A basic neural network model based on multi-layer perceptrons and deep neural networks, and an ensemble model called CatBoost are used to develop a model that predicts the time required between the dispatch system issuing an order and its completion. During the training process, the loss function is used to calculate the error between the output and the answer, and the selected optimizer and algorithm are used to update the weights of the neural network and the residuals of CatBoost. The weights are repeatedly updated and the residuals are corrected until the error between the CatBoost model and the neural network is converged with the answer, and the optimal weights for the prediction model are obtained. The results show that the transit time prediction model based on the CatBoost model and the deep neural network model is effective in predicting time, but ultimately fails to improve the prediction accuracy shown by the historical data. In addition, although the transit time model does not achieve the expected accuracy, it can still provide a tool for users to make decisions.