PARTIAL DISCHARGE PATTERN IDENTIFICATION IN HIGH VOLTAGE EQUIPMENT USING CONVOLUTIONAL AND MULTILAYER PERCEPTRON BACK PROPAGATION NEURAL NETWORK

• Main Author: Puspitasari, Nuraida
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39662
• Institution: Institut Teknologi Bandung

The use of Neural Network as a tool for Partial Discharge (PD) Identification has been used since 1990s, where Backpropagation Neural Network (BPNN) with input in the form of data values was the most widely used method. However, for the past 5 years, Convolutional Neural Network (CNN) has been used to identify Partial Discharge with input in the form of images. The purpose of this study is to develop a PD diagnostic technique that could simplify the identification of PD sources by using a signal waveform image for identification. In our research, we used two-dimensional images of time domain single PD waveform obtained from experiment using 3 sensors (High Frequency Current Transformer (HFCT), Surface Current Sensor (SCS), and Transient Earth Voltage (TEV) sensor) for input to CNN. At the first layer namely Convolutional Layer, the images were extracted using Gaussian Filter on the convolutional stage, then searched the maximum value for each group of pixels to make computing easier on the pooling stage. The output images from Convolutional Layer were identified on the Fully Connected Layer. In order to evaluate more deeply regarding the CNNs performance, the trained CNNs are then tested using new PD data obtained from experiment laboratory that were not involved in the CNNs training process. Results reveal that the CNN prediction result probability achieves 92.17% in average. Moreover, the trained Multilayer Perceptron Backpropagation Neural Network (MLP BPNN) with nine waveform parameters (peak value, rise time, fall time, phase, pulse width, pulse area, even width, frequency of maximum amplitude, frequency from method of moment) are also tested using new PD data obtained from experiment. As a result, the trained MLP BPNN was able to predict PD defect with the average probability achieve 90.94%. These results indicate that both waveform images and parameters can be used as input data for the CNN and MLP BPNN to provide sufficient accuracy for identifying the PD source. The results suggest a possibility that developed CNN and MLP BPNN can be used to identify Partial Discharge in High Voltage (HV) equipment by comparing field data.