Fault Detection in Manufacturing Companies with Ensemble Machine Learning Method

Abstract

Purpose – Multivariate control charts cannot be indicative of which variable is the cause of the out-ofcontrol signal. To keep the process under control, the cause of the out-of-control signal must be determined correctly. The study, it is aimed to predict the variable that causes the out-of-control with the highest accuracy when there is 2 sigma and 3 sigma shift from the mean. Design/methodology/approach – The method used in the study is machine learning-based detection analysis. The data set was taken from a company that produces furniture connecting part. Sample values were collected from the enterprise. Then the under-control samples were detected from these. According to these samples' mean and standard deviation values, data was produced in such a way that 2 sigma and 3 sigma shifts occur from the mean for training the machine learning algorithms. To predict the out-of-control samples three individual machine learning algorithms and three ensemble methods (Bagging, Boosting and Stacking) were used. In addition, 3 stacking models were developed using combinations of the individual algorithms. Findings – When the results are examined, higher accuracy has been reached by using a model developed with the stacking method than individual algorithms. The highest accuracy rates have been achieved as 69.00% for 2 sigma and 85.75% for 3 sigma shift with the stacking 3 models developed based on the stacking method. Discussion – The issue of detecting out-of-control signals in the quality processes in manufacturing companies with the least error was examined and the results were discussed.