Uma abordagem baseada em perceptron multicamadas para detecção de faltas no estator de geradores eólicos do tipo PMSG

Abstract

Wind generators have recurring operating interruptions due to internal failures. Internal failures are difficult to detect and may silently lead to machine damages since they occur between turns, being named turn-to-turn, or between turn and machine housing, being named turn-to-ground. Thus, these plants must be constantly monitored so that these faults are detected in their initial stage. This early detection makes it possible to reduce maintenance costs while decreasing wind turbine downtime. This work proposes a strategy for noninvasive detecting stator failures in its initial stage through a classifier module that analyzes the stator current patterns. This classifier is based on a Multilayer Perceptron (MLP), that is a class of feedforward artificial neural network (ANN), which was trained using a dataset generated by a mathematical model of the PMSG-based wind turbine. The results show that the MLP classifier is able to detect the proposed problem with 97.62% global accuracy. In addition, detection was performed at a initial stage of 1% to 4% of faulty turns with 100% accuracy, contributing to continuous and noninvasive detection of internal wind turbine stator faults.