Predictive System of Semiconductor Failures based on Machine Learning Approach

Abstract

Maintenance in manufacturing has been developed and researched in the last few decades at a very rapid rate. It’s a major step in process control to build a decision tool that detects defects in equipment or processes as quickly as possible to maintain high process efficiencies. However, the high complexity of machines, and the increase in data available in almost all areas, makes research on improving the accuracy of fault detection via data-mining more and more challenging issue in this field. In our paper we present a new predictive model of semiconductor failures, based on machine learning approach, for predictive maintenance in industry 4.0. The framework of our model includes: Dataset and data acquisition, data preprocessing in three phases (over-sampling, data cleaning, and attribute reduction with principal component analysis (PCA) technique and CfsSubsetEval technique), data modeling, evaluation model and implementation model. We used SECOM dataset to develop four different models based on four algorithms (Naive Bayesian, C4.5 Decision tree, Multilayer perceptron (MLP), Support vector machine), according to the five metrics (True Positive rate, False Positive rate, Precision, F-Mesure and Accuracy). We implemented our new predictive model with 91, 95% of accuracy, as a new efficient predictive model of semiconductor failures.