Predictive models for equipment fault detection: Application in semiconductor industry

Abstract

Predictive Maintenance (PdM) in manufacturing using Machine Learning (ML) techniques has been received much attention of many researchers and practitioners in recent years especially for capital-intensive industry such as semiconductor manufacturing. Building a decision tool that detects problems in equipment or processes in semiconductor industry as promptly as feasible to maintain high process efficiencies is a crucial step in process control for cost reduction. However, imbalanced characteristic in dataset is a challenge to implement a predictive model. In this study, an experiment of comparing several different data balancing techniques for predictive maintenance for fault detection are studied. The purpose is to build a high-quality predictive model and investigate the influence of different approaches on data pre-processing phase and data balancing phase to the efficiency of classification. The procedure is examined on three different cases corresponding to three different data preprocessing approaches with five different machine learning algorithms. For this SECOM dataset, it is found that data with different balancing techniques outperforms when training model for this data and this dataset just performs remarkable results if the data is balanced. More importantly, the best model is found with the optimal performance in terms of evaluating by Recall score, as well as highest TPR and lowest FPR. Among that, Logistic Regression trained with SMOTE as data balancing method and Mutual Information for data preprocessing approach provides the best performance as a classification model in this dataset. This study's contribution is to provide a clear understanding and overview of several predictive approaches developed by ML algorithms using this imbalanced dataset, with a focus on semiconductor data applications. The research aims to assess the performance of some commonly used techniques and provides an overview of machine learning as well as predictive maintenance.