In this thesis, Fuzzy Neural Networks pattern classifier structure and learning method were proposed. The purpose of this study was focused to redesign the structure of the neural networks and to develop a design methodology to improve the learning method of the neural network based on K-means Clustering.
The key features of the proposed Fuzzy Neural Networks pattern classifier were listed as follows.
1) To acquire the partial discharge data, the original data was obtained using the Epoxy Mica Coupling(EMC) sensor equipped with the Phase Resolved Partial Discharge Analysis(PRPDA) method and processed through the Motor Insulation Monitoring System(MIMS) program. The processed data reduced the input variable of the high dimensional input to low dimension through the Independent Component Analysis preprocessing algorithm. And the processing speed became faster. At this time, the data was processed with characteristics of maximum value, average value, and median.
2) The cost function used the cross entropy error function instead of the sum of error squares and applied the softmax function to the number of nodes in the output layer to normalize the output value to a probability value between 0 and 1.
3) In order to adjust the connection weights of the hidden layer, K-means Clustering and Gaussian function were used. The center values generated by the K-means Clustering was applied to the Gaussian function and thus the membership values were generated. The connection weights of the output layer were controlled by the nonlinear least squares method using the Newton’s method.
4) L2 regularization is considered to prevent overfitting and improve generalization capability. By adding the L2 penalty term to the cross entropy error function, the performance of the proposed Fuzzy neural networks pattern classifier is superior to that of the previous classifier reported in the literature.