Enhanced Image Classification for Defect Detection on Solar Photovoltaic Modules

Wiliani, Ninuk (2023) Enhanced Image Classification for Defect Detection on Solar Photovoltaic Modules. Doctoral thesis, Asia e University.

	Text Thesis Ninuk Wiliani.pdf Restricted to Repository staff only Download (7MB)
	Text Thesis Ninuk Wiliani-1-24.pdf Download (469kB)

Abstract

Solar photovoltaic modules are a technology that utilizes solar energy. Solar photovoltaic modules have many advantages, such as clean electric energy without pollution, very simple to channelling energy, and the most important is that it does not produce greenhouse gas emissions and can be built in remote areas because it doesn't require energy transmission. In actuality, solar photovoltaic module systems are minimal maintenance and do not require any moving parts, but they still have more chances to get various defects by the environment or human beings. Once PV modules are electrically linked, the performance of the entire system might be impacted by any problem between them. Error-prone areas may be difficult to locate or recognise in a big solar photovoltaic module. A solar photovoltaic modules system can hide it until the whole system collapse or breakdown. On the surface of the photovoltaic modules, solar cell defects are identified based on cell shapes and textures. However, high similarity of characteristics among the shapes and textures has been a major challenge in defect classification process. The objective of this research was to develop and analyse feature extraction used for classification techniques for defect detection of solar photovoltaic modules surfaces. Methodologically, the entire study used a quantitative experiment technique. This research uses the Gaussian Naïve Bayes Algorithm using a ratio of training data and testing data of 70:30 resulting in an accuracy value of 46%. The second algorithm uses K Nearest Neighbour using a ratio of training data and testing data of 95:05 resulting in an accuracy value of 62%. Both methods combine Statistical Feature Extraction and GLCM. Statistical tools provide quantitative information about the intensity distribution of pixels in an image, capturing important statistical properties such as mean, standard deviation, skewness and kurtosis. GLCM, on the other hand, analyses the spatial relationship between pixel pairs and extracts texture features such as contrast, correlation, energy and homogeneity. The accuracy value shows that the KNN algorithm is better when compared to the Naïve Bayes algorithm. Using the same data, these results are compared again using Convolutional Neural Network. The architecture used uses Le Net which is then modified into 3 2D layers and 1 Maxpooling screen. Experiments also compare the size of the image as input, using relu activation and adam optimization. The experiment results in the highest accuracy value at a ratio of 70:30 for training data and test data, which is 68%.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Enhancing, Naïve Bayes, K-Nearest neighbour, Convolutional Neural Network, feature extraction
Divisions:	School of Graduate Studies
Depositing User:	Siti Nor Fairuz Rosaidee
Date Deposited:	06 Jan 2025 03:29
Last Modified:	06 Jan 2025 03:29
URI:	http://ur.aeu.edu.my/id/eprint/1271

Actions (login required)

View Item