Comparison of Transfer Learning and Conventional Machine Learning Techniques for Cable Fault Recognition

Chong, Choon Fatt (2019) Comparison of Transfer Learning and Conventional Machine Learning Techniques for Cable Fault Recognition. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Chong Choon Fatt.pdf Restricted to Registered users only Download (3MB)

Abstract

This project describes using transfer learning method on cable fault recognition using pretrained neural network. The pretrained neural network used is deep convolutional neural network which has been trained on ImageNet dataset with more than one millions of images with 1000 categories. These pretrained neural network has been participated in ImageNet Large Scale Visual Recognition Challenge which are AlexNet, GoogleNet, ResNet18, ResNet50 and ResNet101. The method of transfer learning on pretrained neural networks is implements by replacing the fully connected layer, Softmax Layer and Classification Output layer and are adjusting for cable fault recognition. At first, the hyperparameter values for those pretrained neural networks is randomly selected and is tested with noise contaminated PD images. Then, Grid Search hyperparameter optimization is implemented for fine-tuning the pretrained networks. The hyperparameter values found after implemented grid search is uses to train the pretrained networks. The trained network is then test on noise contaminated PD images to evaluate it performance. Shallow Neural Network – MLP-ANN is implement to compare the performance with modified pretrained neural network. The input for MLP-ANN is images data which same as the input uses for modified pretrained neural networks which is a 3DPRPD image with colormap and view from top. The result shows that MLP-ANN has better performance to classify noise-contaminated PD images than using transfer learning. Grid search shows the result is not able to optimize some of the type of pretrained neural network.