Advances in Classifier Evaluation: Novel Insights for an Electric Data-Driven Motor Diagnosis

Abstract

Fault diagnosis of inductions motors has received much attention recently. Most of the works use data obtained either from the time domain or by applying advanced techniques in the frequency domain. Some researchers have employed a considerable effort in designing sophisticated algorithms to achieve the best performance of the diagnosis system. However, some contributions in the field have not taken advantage of the benefits that a good evaluation stage can bring to the developing of classifiers for fault diagnosis. In this paper, novel insights for the classifier evaluation are presented to promote better assessment practices in the field of electric machine diagnosis based on supervised classification. A case of study consisting of a motor with a broken rotor bar is described to analyze the performance of two classifiers by using scores focused on the fault detection. Also, different error estimation methods are considered to obtain unbiased predictive performances. Two statistical tests are also discussed to confirm the significance of the results under a single data set.