Abstract
M.Ing. (Electrical and Electronic Engineering)
Dissolved Gas-in-oil analysis (DGA) is used to monitor the condition of bushings on large power transformers. There are different techniques used in determining the conditions from the data collected, but in this work the Artificial Intelligence techniques are investigated. This work investigates which gases in DGA are related to each other and which ones are important for making decisions. When the related and crucial gases are determined, the other gases are discarded thereby reducing the number of attributes in DGA. Hence a further investigation is done to see how these new datasets influence the performance of the classifiers used to classify the DGA of full attributes. The classifiers used in these experiments were Backpropagation Neural Networks (BPNN) and Support Vector Machines (SVM) whereas the Principal Component Analysis (PCA), Rough Set (RS), Incremental Granular Ranking (GR++) and Decision Trees (DT) were used to reduce the attributes of the dataset. The parameters used when training the BPNN and SVM classifiers are kept fixed to create a controlled test environment when investigating the effects of reducing the number of gases. This work further introduced a new classifier that can handle high dimension dataset and noisy dataset, Rough Neural Network (RNN). This classifier was tested when trained using the full dataset and how it is affected by reducing the number of gases used to train it. The results in these experiments showed that ethane and total combustible gases attributes are core attributes chosen by the four algorithms as gases needed for decision making. The average results of the classification performance showed that the reduction of attributes helps improve the performance of classifiers. Hence the science of transformer condition monitoring can be derived from studying the relations and patterns created by the different gases attributes in DGA. This statement is supported by the classification improvements where the RNN classifier had 99.7% classification accuracy when trained using the three attributes determined by the PCA.