Uso de meios computacionais na detecção de falhas incipientes em transformadores de potência como método alternativo à análise de gases dissolvidos.

Abstract

Power transformers are essential equipment for the operation of the electrical system, both economically and technically. In view of this, any effort devoted to identifying failures becomes relevant. This work aims to identify the computational methods with better performance in the classification of incipient failures in transformers based on the concentration of gases dissolved in oil. For this purpose, a database with 201 records of chromatographic analyzes (or dissolved gas analysis) available on the IEEE DATAPORT was used. The data were treated, segregated into two sets (one containing the original normalized dataset and the other containing the ratios between gases using a methodology defined by Rogers) and submitted to classification using traditional and computational methods (Knearest neighbor, Naïve bayes and Multilayer perceptron). For the original dataset, the Multilayer perceptron obtained a precision of 0.80, the K-nearest neighbor 0.60 and the Naïve bayes 0.36. In turn, for the ratios dataset, the Multilayer perceptron obtained a lower precision (0.70), but both the K-nearest neighbor and the Naïve bayes improved their performance reaching values of 0.70 and 0.63, respectively. It is possible to conclude that using the method of ratios in the pre-processing of data from the analysis of dissolved gases, improved the performance of two classifiers (K-nearest neighbor and Naïve bayes) and worsened the performance of one of them (Multilayer perceptron), in addition, of the implemented classifiers, the Multilayer perceptron and the K-nearest neighbor, obtained higher success rates, respectively 0.80 and 0.70, than the traditional method described by Rogers (0.69), which is only better than Naïve Bayes (0.67).