Classification of anomalies in electroluminescence images of solar PV modules using CNN-based deep learning

The escalation of implementing photovoltaic (PV) power generation has paved the road to innovative remarkable applications. The technology of utilizing electroluminescence imaging (EL) has aided the early identification of faults and rapid classification of solar cells in PV panels. Recently, deep learning neural networks (DNNs) has been extensively utilized in the field of PV fault detection and classification. Despite of the good achievements in the field of DNN-based approaches, however, there is still a potential for further developments. This includes better data preparation, proper dataset categorization and designing of efficient light-weight DNNs. In this work, an efficient approach is proposed to be used for defect detection and malfunctions' classification in PV cells, based on utilizing EL-based imaging analysis. Here, three approaches were developed using multi-scale convolutional neural network (CNN) models, the former is based on deploying the pretrained SqueezeNet and the GoogleNet, in a transfer learning fashion, whereas the latter is a light-weight CNN approach (denoted as LwNet). The experiments were elaborated on the ELPV dataset after being properly modified and categized. Two scenarios were adopted: 4-class- and 8-class-classification procedures. Experimental validation of the developed CNNs have demonstrated very promising performances, especially when adopting the 8-class approach. An average accuracy of about 94.6%, 93.95%, and 96.2% was obtained using GoogleNet, SqueezeNet and LwNet, respectively. A privilege has been granted to LwNet over SqueezeNet and GoogleNet, in terms of classification performance and time saving efficiency.