This master's thesis presents the experiments with applications of computational intelligence techniques for the characterization of partial discharges in hydrogenerators. The classification of the partial discharge contributes to a prior analysis problems and allows predictive maintenance on machinery, reducing the possibility of failures in them. Data were collected online mode (operation machine) in the Tucuruí Hydroelectric Power Plant, observed the internal discharge standards, delamination and between bars. The IMA-DP software, developed in partnership with Eletronorte and Cepel, allowed these data were measured and recorded quickly, and organized through PRPD maps (Phase resolved Partial Discharges). Binarization techniques, ANOVA (Analisys of Variance), ICA (Independent Component Analysis) and PCA (Principal Component Analysis) were applied to the signals to adapt them to the use of computational intelligence techniques. The study was developed in IPython environment using scikit-learn library, which has efficient intelligence algorithms. The experiments were performed making use of techniques: KNN (K-Nearest Neighbors), Random Forest and MVS (Support Vector Machines). Such techniques showed good results with the experiments, highlighting those obtained for MVS that showed the best results, achieving an accuracy of 96.07%, due possess selection mechanisms of the main variables during the training process.
|
