Ensemble Learning-Based Data Augmentation for Condition Monitoring of Induction Machines

Abstract

This study addresses the challenges of machine learning-based condition monitoring of induction machines under varying load conditions, which can result in low accuracy at unmeasured loading levels. A hybrid data augmentation framework is developed that combines multiple regression models and ensemble learning techniques to generate feature values at any unmeasured loading levels. The proposed method requires feature computation from only four measured loading levels under healthy, one, two and three broken rotor bars conditions as training data, enabling feature values augmentation for other loading levels. In this study, the augmentation method is applied to generate feature values at two intermediate levels (50% and 75%) and one extreme level (100%) and the corresponding results are presented. This hybrid data augmentation method not only produces accurate feature values for intermediate loading levels but also performs exceptionally well in extrapolating feature values at extreme loading levels. Incorporating this generated data during the training phase resolves generalisation issues and substantially improves the classification accuracy of machine learning models. In particular, the integration of ensemble learning techniques helped to increase accuracy from 38.75%, 42.75% and 60%–100% for the K-nearest neighbours, support vector machine and decision tree models, respectively, at the 100% loading level.