A SWIN TRANSFORMER IN THE CLASSIFICATION OF IR IMAGES OF PHOTOVOLTAIC MODULE

Abstract

The faults occurring in the photo voltaic module as to be detected in order to increase its efficiency. The infrared images, electroluminescent images, and photo luminescent images of the photo voltaic modules have been used to detect and classify the faults. The infrared data set is used for classification and it is highly imbalanced. To make it balanced, generative adversarial networks are used to generate images for each fault category. If the fault classification is done using a convolution neural network, the feature maps are generated by convolution operation with filters on images. The convolution neural network model is to be trained for 1000 epochs and the time required for training is greater than 24 hours. The computational cost of a convolution neural network (CNN) is reduced in transformers through the attention mechanism. The Swin (shifted window) transformers which are used for classification as to be trained for 40 epochs and the maximum training time is less than 6 hours. The computation time, categorical classification accuracy, and top-5-accuracy obtained using Swin classifier are compared with the existing methods. It is found that the computational cost is very much reduced by the Swin transformer and top-5-accuracy of 99.04% is obtained by it while classifying 11 faults of IR images.