A Wrapper-based Feature Selection Approach for Accurate Fault Detection of Rotating Diode Rectifiers in Brushless Synchronous Generators

Abstract

Fast and accurate fault detection is a vital issue for all engineering systems, especially for those which generate power for sensitive applications. To achieve rapid and accurate fault detection, an appropriate signal must be measured and the best features of that signal should be extracted. In this paper, an approach is proposed to detect the healthy and faulty states of a rotating diode rectifier in a brushless synchronous generator by using three-phase terminal voltage. There is not any direct access to the rectifier, and therefore, fault detection must be performed using signals from the machine itself. In order to extract the correct voltage signal, fast Fourier analysis is performed on the signal, and to select the best frequency to reduce redundancy and increase classification accuracy, a wrapper-based feature selection approach is utilised. In this approach, the best frequency is selected from among the frequencies generated according to the accuracy of the classifier, a Support Vector Machine (SVM). The other frequencies are then added one by one according to their accuracy in each step, and the best subset, that is, the one with the best accuracy for the classifier, is selected. The proposed approach is then evaluated using experimental data.