Feature Extraction, Supervised and Unsupervised Machine Learning Classification of PV Cell Electroluminescence Images

Lifetime performance and degradation analysis of laboratory and field deployed PV modules is paramount to the continued success of solar energy. Image characterization techniques capture spatially resolved macroscopic manifestations of microscopic mechanistic behavior. Automated data processing and analytics allow for a large-scale systematic study of PV module health. In this study, degradation features seen in periodic EL images taken during test-to-failure damp-heat, thermal cycling, ultra-violet irradiance, and dynamic mechanical loading accelerated exposures are extracted and classified using supervised and unsupervised methods. Image corrections, including planar indexing to align module images, are applied. On extracted cell images, degradation states such as busbar corrosion, cracking, wafer edge darkening, and between-busbar dark spots can be studied in comparison to new cells using supervised and unsupervised machine learning. The systematic feature groupings provide a scalable method without bias to quantitatively monitor the degradation of laboratory and commercial systems alike. The evolution of these degradation features through varied exposure conditions provides insight into mechanisms causing degradation in field deployed modules. The supervised algorithms used in this application are Convolutional Neural Networks (CNN) and Support Vector Machines (SVM). With the increase in data and diversity of features, unsupervised learning can be employed to find relations between inherent image properties. Feature extraction techniques help identify intrinsic geometric patterns formed inthe images due to degradation. Principal component analysis is then applied to the extracted set of features to filter the most relevant components from the set, which are then passed to an agglomerative hierarchical clustering algorithm. Google's Tensorflow library was utilized to enhance the computational efficiency of the CNN model by providing GPUbased parallel matrix operations. Using supervised methods on 5 features an accuracy greater than 98% was achieved. For unsupervised clustering, the classification was done into two clusters of degraded and non-degraded cells with 66% coherence.