Using Convolutional Neural Networks to Detect In-Field PV Module Glass Cracks

Double-glass photovoltaic modules are being increasingly deployed, and there is a growing concern about glass cracking in these modules. To confirm this phenomenon, to quantify the rates of glass cracking, and to alleviate the laborious task of finding these cracked modules in the field, this article considers the use of convolutional neural networks for glass crack detection. Seven models are tested: a six-layer model, a four-layer model, VGG16, VGG19, ResNet18, ResNet34, and ResNet50. Using a nonstandardized image acquisition method in two photovoltaic (PV) fields with two module types, seven labeled datasets have been created, ranging in size from 3540 to 12 600 images. The six-layer model can achieve crack versus no-crack classification accuracies of up to 97.7%, while the inference time per image using a CPU is 1.5 s. The more complex ResNet50 model achieves classification test accuracies up to 99.7%, but it has an inference time of up to 10.5 s per image. Thus, this article demonstrates that using a relatively simple convolutional neural network is a viable approach to detecting PV glass cracks with low computational cost. In addition, the YOLOv11 instance segmentation model is used to segment the glass cracks, and it achieves up to 95.7% in bounding box precision, mask precision, recall, and mean average precision@50. This demonstrates that segmentation can also be implemented on an automated inspection system without increasing computational costs. Finally, the water index on two modules was measured using a near-infrared absorption spectrometer, and the results suggest that there is higher water ingress along the cracks when compared with the rest of the modules.