Improving Cable Fault Classification Using Optimization Algorithm

Chan, Jun Qiang (2020) Improving Cable Fault Classification Using Optimization Algorithm. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Chan Jun Qiang.pdf Restricted to Registered users only Download (2MB)

Abstract

Partial Discharges (PD) is one of the faults that will degrade the insulation quality in power cables. PD mostly occur at the weak point along the power lines, which is the cable joint. This insulation failure will initiate random pulses at an unpredicted time, and it happens when the conductors experience a partial short-circuit. If left untreated, the insulation defect will eventually cause an insulation breakdown. The works of identifying types of insulation defects can be challenging, even for an expert with years of experience. Lately, the Machine Learning technique had been used widely for insulation defect recognition. However, this technique required human effort for the feature extractions before using a neural network for training. And the selected features may not properly represent the insulation defect itself. Therefore, the technique of Deep Learning had proposed in this paper, utilizing the image classification model to execute the PD defect classification. This technique is more favourable than Machine Learning is because the feature extraction process will be done in an automated way. The PD defect used in this paper were measured from the commonly encountered cable joint defects in the power industry. The Phase-Resolve Partial Discharges (PRPD), Pulse Height Distribution (PHD) and Pulse Count Distribution (PCD) pattern will be used as the representation of PD sources. Different types of PD will provide a different PD pattern, so the Convolutional Neural Network (CNN) was chosen as the model to recognize the unique pattern of each defect efficiently. The Transfer Learning technique was exploited, which aims are to reduce the overall time in constructing and training a new model from scratch, most importantly is to repurpose a pre-trained image classification model. The pre-trained model used was excellent in executing image classification, which had trained more than a million images to classify 1000 of objects. VGG16 and VGG19 will be used to determine the performance of PD defects recognition from noise-free and noise-contaminated PRPD, PHD and PCD patterns. The Bayesian Optimization Algorithm (BO) will be used to fine-tuning the pre-trained model for better PD recognition performance. The fine-tuning hyperparameters included learning rate, number of dense layers, number of dense nodes and activation function. The dataset having 100 images for each defect and there is a total of 5 different defect types, and each type of PD source pattern constituted of fix scale and auto scale. The strategies of selecting suitable hidden layers to be fine-tuning in Convolutional CNN is set, to allow the model to learn in recognition of PD pattern. Others pre-trained models such as ResNet50, MobileNetV2, InceptionV3 and InceptionResNetV2 will be used to classify the PD dataset as well, to provide more performance comparison. Using the BO algorithm, VGG16 and VGG19 model showed overall classification performance over 80% in both fix scale and auto scale of PHD, PCD and PRPD datasets.