應用類神經網路與田口方法建構鋁壓鑄生產預測模型之研究

A Study on the Construction of Die-Casting Production Prediction Model Using Machine Learning with Taguchi Methods

黃翊涵(Yi-Han Huang)

指導教授 ： 周永燦

中原大學/電機資訊學院/工業與系統工程研究所/碩士(2021年)

https://doi.org/10.6840/cycu202100077

摘要

器學習、大數據、工業4.0等技術，已成為近幾年各國科技發展的趨勢，而在臺灣多數企業仍使用人工生產的模式，使得企業競爭力較同業不足，如何將生產模式轉型智慧化，已然成為近幾年重要研究課題之一。然而轉型是無法一蹴而成的，如何利用資訊了解實際生產狀況及預測未來即將發生的事，對於智能工廠顯得非常重要。壓鑄具有高生產力、高精度尺寸、高機械強度等多種特性，是鑄造工藝中最為廣泛使用之一。因此本研究希望透過數據分析建立預測模型，協助壓鑄廠判別鑄件是否存在缺陷，提高國內壓鑄廠生產競爭力。
本研究與國內汽車工業的壓鑄廠合作，利用實際生產之數據作為分析依據，透過文獻與專家訪談，探討關鍵製程參數因子並歸納缺陷造成原因。本論文將壓鑄製造相關參數作為自變數，鑄件是否存在瑕疵作為應變數，建構類神經網路(Artificial Neural Networks, ANNs)、支持向量機(Support Vector Machine, SVM)及隨機森林(Random Forest)等三種預測模型，並利用田口方法(Taguchi Methods)找出各模型之最佳參數配置，最後透過混淆矩陣(Confusion Matrix)的F-指標(F-Measure)及ROC曲線下面積(Area Under the Curve of ROC, AUC ROC)評比三種模型之優劣，找出最適配之網路預測模型。
研究結果顯示，三種模型在此分類問題中，皆有傑出之表現，而隨機森林模型之準確率高達99%，因此本研究以隨機森林作為鋁壓鑄件品質特性預測的模型。

關鍵字

鋁壓鑄 ； 機器學習 ； 類神經網路 ； 支持向量機 ； 隨機森林 ； 田口方法

並列摘要

Industry 4.0, Big Data, and other technologies have become the trend of science and technological development in various countries. However, most Taiwanese enterprises still use the traditional mode of manual production, which makes enterprises' competitiveness less than that of the same industry. Transforming the production model into an intelligent production model has become an important topic in recent years. However, the transformation cannot be achieved overnight. So, the production data should practice digitalization, transform data into information through analysis to assist production, and uses Artificial Intelligence (AI) related applications, gradually transformed into a smart factory. Die-casting is a metal casting process. Compared with other casting processes, it has many characteristics such as high productivity, excellent dimensional accuracy, and high mechanical strength, making die-casting one of the most widely used casting processes. Therefore, this research hopes to establish a predictive model through data analysis to assist the die-casting factory in determining whether there are defects in the castings and improve the die-casting factory's product competitiveness.
In this paper, we cooperate with a die-casting factory in the domestic automobile industry, using the actual production data as the basis for analysis, discuss key process parameter factors, and summarize the causes of defects. The independent variables are relevant parameters of die-casting manufacturing. And, find the presence or absence of defects in the casting process as the dependent variable. Afterward, we construct three kinds of forecasting models: Artificial Neural Networks (ANNs), Support Vector Machine(SVM), and Random Forest, and use Taguchi Methods to find out the best hyperparameters for each model. Finally, we compare the three models through the F1-Measurex and AUC to determine the best network prediction model for die-casting.
The result of the research shows that, three models have good results. The accuracy of the random forest model is as high as 99%, so this study uses random forest as the die-casting production prediction model.

並列關鍵字

Die-Casting ； Machine Learning ； Artificial Neural Network ； Support Vector Machine ； Random Forest ； Taguchi Methods

參考文獻

參考文獻

英文文獻

Adke, M. N., Karanjkar, S. V. (2014). Optimization of die-casting process parameters to identify optimized level for cycle time using Taguchi method. Int. J. Innov. Eng. Technol, 4(4), 365-375.

Al-Badarneh, I., Habib, M., Aljarah, I., Faris, H. (2020). Neuro-evolutionary models for imbalanced classification problems. Journal of King Saud University - Computer and Information Sciences.

Bach, M., Werner, A., Żywiec, J., Pluskiewicz, W. (2017). The study of under- and over-sampling methods’ utility in analysis of highly imbalanced data on osteoporosis. Information Sciences, 384, 174-190.