Deep Learning Berbasis CNN Untuk Pengenalan Pola Partial Discharge Isolasi Silicone Rubber

Abstract

Partial discharge (PD) activity measurements have been carried out by selecting noise signals (de-noising) using Support Vector Machine (SVM)and then recognized using Convolutional Neural Network (CNN). CNN testing was carried out using various models such as activation methods: Sigmoid, Softmax, Relu, Tanh, Swish. Number of layers used is 1, 2, 3, 4 with filter sizes of 32, 64, 128, 256 and kernel sizes 3x3, 2x2, 1x1, 1x2, 1x3 in the MaxPooling and AveragePooling pooling methods. The results obtained, On sigmoid method the MaxPooling and AveragePooling with 1 layers having a low accuracy around 14.40% but the other layers configurations gets a high accuracy around 98.99% both has been done with or without de-noising. In Softmax activation method, MaxPooling pooling method has an accuracy around 84.94% and has de-noising 90.66%. The AveragePooling pooling method has an accuracy 65.25% and around 75.29% with de-noised. The result shows that SVM de-noising increases the accuracy around 11.12% in the Softmax activation method. In the Tanh, Relu, and Swish activation methods, a low level of accuracy is obtained with an average of 14.40%, and SVM de-noising doesn’t increase the accuracy, so CNN-based deep learning with SVM de-noising is more suitable using the Sigmoid and Softmax.