Condition monitoring of transformer bushings using Rough Sets, Principal Component Analysis and Granular Computation as preprocessors

This paper introduces the adaption of Rough Neural Networks (RNN) in bushings dissolved gas analysis (DGA) condition monitoring. The paper extended by investigating the RNN, Backpropagation Neural Networks (BPNN) and Support Vector Machine (SVM) classifiers´ performance when Principal Component Analysis (PCA), Rough Sets (RS) and Incremental Granular Ranking (GR++) are used as preprocessors to reduce the attributes of the DGA training data. The performance of RNN classifier was benchmarked against the performance of BPNN since RNN was built using Backpropagation. The RNN classifier had higher classification accuracy than BPNN and SVM when trained using PCA and RS reduct dataset. RNN had a lower training time than BPNN and SVM when trained using RS and GR++ reduct dataset. PCA reducts dataset increased the classification accuracy of the BPNN, RNN and SVM classifiers, while RS reducts dataset only increased the classification accuracy of RNN classifiers. GR++ reduced the classification accuracy of BPNN, RNN and SVM but increased their training time.