A new computational approach for inter-turn short circuit recognition in induction machines using currents analysis and multi-class Support Vector Machine

In this paper, the authors propose a new computational approach, which can detect, approximate (severity of the failure) and localize (which phase) the inter-turn short circuit in stator coils. This approach is based on a multi-class problem that is defined by several categories bounded with lower and upper limits of the severity in percentage, and the phases of the induction machine. To perform this recognition, the new method proposed relies on a modified three-phase currents analysis using the best-fit 3D-ellipse method, charged to extract different characteristics of the stator failure, and a multi-class Support Vector Machine (SVM) to classify these features. These features compose a vector used as input for the multi-class SVM generated by a training set and tested with a dataset by adding white Gaussian noise on the current signals. Finally, the multi-class SVM classifies the extracted characteristics, and the classification result represents the detection, the estimation of the severity and the localization of the failure. The results obtained show that this new computational approach is promising.