SILVA, N. P; http://lattes.cnpq.br/4955391782870684; SILVA, Nathanael Pereira.
Résumé:
Insulators are equipment responsible for providing electrical insulation and mechanical support to
conductors, and are widely used in the Electric Power Systems. System reliability and continuity
depend on these equipment operating correctly. Due to the importance of insulators for the
electrical system, studies that develop techniques capable of performing the diagnosis/prognosis
of this equipment are necessary. This work presents a technique for classification or diagnosis
of insulators based on leakage current parameters. Initially, the database was built, through
electrical tests to obtain the leakage current with artificially polluted glass insulators at five
different levels: clean, very light, light, moderate and heavy. The leakage current parameters
that presented the greatest correlation with pollution were: peak value, RMS value and the
fundamental harmonic component. The classification models used to determine the level of
pollution were based on artificial intelligence techniques, namely: Support Vector Machines,
Kohonen Self-Organizing Maps and Naive Bayes. Performance indicators were used to assess
the reliability of each of the techniques as well as perform a comparative analysis between them.
The indicators used were: accuracy, precision, recall, f1-score and confusion matrix. It was
verified that the classification models presented excellent results, mainly the algorithm based on
unsupervised learning, Kohonen Maps. Thus, we obtained a classification model with satisfactory
performance that can be used in an insulator monitoring and diagnosis system capable of assisting
in decision-making processes related to asset management.
