Machine learning-based software sensors for machine state monitoring - The role of SMOTE-based data augmentation

Abstract

A method for flexible vibration sensor-based retrofitting of CNC machines is proposed. As different states leave different fingerprints in the power spectrum plane, the states of the machine can be distinguished based on the features extracted from the spectrum map. Due to some states, like tool replacement, are less frequent than others, like production state, monitoring the machine states is considered an imbalanced classification problem. The key idea is to use Borderline-Synthetic Minority Oversampling Technique (Borderline-SMOTE) to augment the data set. The concept is validated in an industrial case study. Soft sensors based on four machine learning algorithms with and without SMOTE to predict the states of the machine were implemented. The results show that the SMOTE-based data augmentation improved the performance of the models by 50%.