Fpga realization of rbf neural network based transformer impulse fault classification scheme

This paper proposes simple hardware architecture for realizing a Radial basis function (RBF) network for transformer impulse fault classification. Fault conditions are applied on the lumped parameter model derived for the DUT and the model is simulated using PSPICE orcad software. The winding currents thus computed are analyzed using db5 wavelet and the statistical features namely mean and Variance are extracted from the third level approximation. The RBF network has a number of advantages compared with other type of ANN including simpler network structure and faster learning speed. The key point of RBFNN is to decide a proper number of hidden nodes. Here the possibilistic FCM algorithm is used to cluster the derived statistical features into 21 different clusters representing the defined fault types and the RBF network is constructed with 21 hidden nodes representing the clusters. The hardware implementation is carried out using Xilinx system generator for DSP on Spartan 6 FPGA. The overall classification accuracy of this scheme is 97%.