DEEP LEARNING BASED DETECTION OF CRACKS IN ELECTROLUMINESCENCE IMAGES OF FIELD PV MODULES

In this paper, we have proposed a deep learning network for classification of electroluminescence (EL) image of solar cells into good or cracked cells for low resolution images captured in the field. Such crack detection in EL images is becoming important to reduce the time required for manual inspection and to minimize human error. Most implemented deep learning methods use high-resolution images captured in controlled environment inside a lab for training the network model. This model, when used for field data will not perform well due to mismatched conditions. In our work we use, EL images collected from PV power plants located in different parts of India during ‘All India Survey 2018’, which include 5 climatic zones and modules from 20 different module manufacturers to train the deep neural network. There EL images were captured using customized CMOS-based camera having comparatively lower resolution. Thus the deep neural network model is trained for on field variations and the accuracy of crack detection is 94.77% even for the lower resolution. The preprocessing of raw EL image to extract individual cells from the PV module image, training the model on the cell image dataset using supervised learning and testing of the trained model on testing data are the 3 steps involved in the analysis.