Using SMOTE techniques for industrial defects diagnosis to feed Machine Learning models

Abstract

Industrial process automation has become a priority for most modern enterprises over the past years. Product inspection for defect detection is a commonly automated process since is a repetitive pattern seeking activity. Regarding this, Machine Learning based solutions has been applied for industrial defect detection. However, most of the available solutions approach defect diagnosis using Computer Vision based solutions, which implies the installation of infrastructure for capturing production line images. This is not always a viable option for small young business because of budget limitations. In this work we approach industrial defect diagnosis using tabular history defect diagnosis records. We propose using Synthetic Minority Oversampling Technique for augmenting insufficient original data to feed Machine Learning algorithms for defect detection. We present the results of training with original and SMOTE oversampled data, reaching better results with SMOTE, approximately 69% of accuracy.