Electroluminescent Image Processing and Cell Degradation Type Classification via Computer Vision and Statistical Learning Methodologies

A data set of 90 60-cell module images from 5 commercial PV module brands over 6 exposure steps of damp-heat testing were analyzed. An automated data analysis pipeline was developed using the open source coding language Python to parse the module images into individual cell images. As the original raw images are not directly suitable for modeling, this algorithm implements techniques which include filtering, thresholding, convex Hull, regression fitting, and perspective transformation to pre-process the original image. After cell extraction, 5400 individual cell images as a function of brand and exposure time were obtained. From the data set, 3 initial degradation categories were observed: good, cracked, and heavily busbar-corroded. For supervised machine learning classification, these images were manually sorted into these 3 categories yielding 3550 images. To increase the data set size, the cell images were augmented by flipping the images about the x-axis and y-axis as well as rotated 180 degrees. This increased the total sample size to 14,200 images with good, cracked, and heavily corroded counts of 12,004, 492, and 1704, respectively. A training and testing framework was generated using stratified sampling with a training to testing data ratio of 80:20. The statistical learning algorithms Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Network (ANN), were independently trained on the training set and then given the remaining data images to predict their classification. The results showed model prediction accuracies of 98.77%, 96.60%, and 98.13% for the SVM, RF, and ANN models, respectively.