Defect Classification Of Solar Photovoltaic Modules Using Image Processing

Abstract

ABSTRACT newlineIn recent studies, it has been noticed that to identify the defects in the solar photovoltaic newline(PV) modules, the manufacturers more rely on the automatic defect detection techniques newlineinstead of manual detection of defects in Electroluminescence (EL) images of PV newlinemodules. Sometimes the defects are so minute like a small crack such that the manual newlinedetection of PV modules cannot easily detect which leads to slow detection and thus newlineinvolves human error, as a result of which the overall accuracy is affected and hence newlinedegrading the performance and quality of modules. The manual detection method is newlineinvasive, time consuming and prone to human errors. Automatic defect classifications of newlineEL images of PV modules are much significant these days. Despite of the fact, it is a newlinedifficult task to perform automatic classification of defects due to the complexity and newlineinhomogeneity of the crystalline silicon solar cells. The Automatic defect detection is a newlinefast and reliable method to identify the defects from the large dataset. Deep learning is newlineefficient technique to identify these defects with greater accuracy. The present study is newlinecarried out for automatic defect classification of EL image dataset of PV modules. In this newlineresearch, AlexNet, a pretrained deep learning CNN i.e. Convolution Neural Network newlinemodel has been used to classify EL images into two classes i.e. defected and non-defected newlineand achieved the greater accuracy of 85.16 % with minimum training and validation loss newlinefor 0.0001 learning rate. Further, the comparison of the results has been done by different newlinepre-trained models to achieve the best accuracy. From the different models, we conclude newlinethat at different epochs ResNet-18 gives the maximum accuracy with minimum loss at newline0.01,0.001,0.0001 learning rates respectively, the training progress also shows the newlineminimum loss (both training and validation) with highest accuracy of 90.90 %. So, newlineResNet -18 is the best suited model for accuracy prediction and from AlexNet model newlineidentified the number of defected panels. Aut