Artificial Dataset Generation and Bagging Ensemble Learning for Enhanced Fault Detection in Induction Motors

Abstract

Induction motors are widely used in industrial settings because of their robustness, reliability, and efficiency in terms of energy consumption. However, like other types of machinery, induction motors are susceptible to faults that can result in costly interruptions and production downtime. In this work, we discuss and compare the performance of bagging ensemble learning techniques for enhanced fault detection using an artificial dataset that mimics real scenarios of induction motors. The generated artificial dataset contains data depicting conditions of normal motor functioning, as well as open-circuit and short-circuit faults under various load conditions. Our experiments show that for fault detection with three-class modeling, Random Forests outperform bagging ensemble configurations of SVM, ELM, and MLP. This is most likely because the model manages to find the intrinsic structure in the dataset through its decision trees. Nevertheless, as the number of fault conditions increases, the MLP bagging ensemble shows promise for greater performance. However, the reliability of models trained exclusively on synthetic data remains an open question, highlighting the need for validation with real-world industrial data. Integrating this methodology into predictive maintenance systems could significantly improve the reliability and operational efficiency of industrial equipment.