Método de mapa auto-organizável probabilístico para classificação de descargas parciais

Abstract

In this work, a methodology called Probabilistic Self-Organizing Map (SOPM) is proposed for the automatic classification of multiple sources of partial discharges (PDs), with an emphasis on dubious patterns, in hydrogenerator stator windings, combining Self Organizing Maps (SOM) with statistical analysis. The database is composed of PRPD (Phase-Resolved Partial Discharges) maps from online PD measurements of hydrogenerators operating in a real environment at the Tucuruí and Coaracy Nunes Hydroelectric Power Plants. The main contributions of this work are: (i) the development of a classifier with a probabilistic output, capable of quantifying uncertainty and classifying patterns with characteristics of multiple classes simultaneously; (ii) the application of a semi-supervised methodology, which uses unsupervised training for the topological organization of the data and a reduced set of labeled samples for optimization, making the method robust for scenarios with few labeled data; (iii) the validation of the method on a real database containing doubt classes, which represent the overlap of different types of faults. The classification task is performed in two stages: first, a SOM organizes the input data into a low-dimensional map; then, the SOPM calculates the probabilities of a sample belonging to each class based on the average distance between the neurons and the samples in the feature space. Overall accuracy rates exceeding 90% were obtained, with emphasis on the classification of the doubt classes, which reached up to 100% accuracy on the test set. Finally, the methodology was validated with the database of real measurements, demonstrating its effectiveness for diagnosis under operational conditions