A Comparative Study for the Detection of Stator Inter-Turn Faults in Induction Motors Using Shallow Neural Networks and Non-Neural Based Techniques

Abstract

Traditional Induction Motor (IM) Fault Diagnosis (FD) relies on extracting features from original signals, which directly impacts FD performance. However, high-quality features require expert knowledge and human interaction. Therefore, this paper presents a comparative analysis of the detection of stator inter-turn faults in IMs using state-of-art classifiers. Experimental data was acquired and the class was divided, into up to three levels of severity for the stator fault. Thereafter, using a non-parametric approach, features from the three-phase currents were engineered relative to the characteristic fault frequencies for stator faults. Various families of classifiers were trained on the developed feature-set, and for each family of classifiers, Bayesian Optimization was applied to train its own variants and derive the best hyperparameters. The best results were obtained for Tree, Ensembled and Shallow Neural-based techniques using a separate test set.