طريقة الكشف عن أخطاء الكهروضوئية القائمة على التعلم العميق للصور الجوية

Conceived an international research group, the proposed model uses the convolutional neural network (CNN) architecture U-Net for image segmentation and the the CNN architecture InceptionV3-Net for fault classification.

An international research team has developed a novel PV fault detection method based on deep learning of aerial images.

The proposed methodology utilizes the convolutional neural network (CNN) architecture U-Net for image segmentation and then applies the CNN architecture InceptionV3-Net for fault classification.

“The presence of dust, snow, bird droppings, and other physical and electrical problems on the surfaces of the solar panels may lead to energy losses,” the academics said. “The need for efficient monitoring and cleansing protocols in solar energy systems cannot be emphasized enough. Based on that purpose, we have selected research subjects to enhance the image processing and classification tasks related to various types of damage to solar panels.”

For the model’s segmentation step, the group used a publicly available annotated database of 4,616 images. The aerial pictures were divided into six categories: ground cropland, grassland, saline-alkali, shrubwood, water surface, and rooftop. The database was split in a 60%-20%-20 % ratio for training, validation, and testing, respectively.

A different database containing 885 images was split in the same ratio for fault classification. The dataset comprises six categories of PV issues: clean, dusty, birds’ drop, electrical damage, physical damage, and snow-covered. Aside from the InceptionV3-Net model—which applies the InceptionV3 base with ImageNet weights—the researchers also tested other classification models for compression. Those were Dense-Net, MobileNetV3, VGG19, CNN, VGG16, Resnet50 and InceptionV3.

“Initially, aerial satellite images are processed using the U-net model architecture with a 256X256X3 input shape, undergoing three stages: decoding the input, combining encoding and decoding, and generating the output,” the group explained.

It also stressed that the InceptionV3-Net architecture uses the InceptionV3 base with ImageNet weights, enhanced by convolutional layers, Squeeze-and-Excitation (SE) blocks, residual connections, and global average pooling. The model includes two dense layers with LeakyReLU and batch normalization, ending with a Soft-Max output layer. It also utilizes data augmentation techniques such as rotation, shift, shear, zoom, and brightness adjustments.

“The model is trained using the Adam optimizer with a learning rate of 0.0001 and categorical cross-entropy loss,” they also said.

Their analysis showed that the proposed InceptionV3-Net achieved validation accuracy of 98.34% and an F1 score—which represents the balance between precision and recall—of 0.99%. That is compared with validation accuracy in the range of 20.9%- 89.87% and F1 of 0.21-0.92 in the competing models.

The testing results also showed that the proposed InceptionV3-Net achieved a validation accuracy of 94.35% and an F1 score of 0.94. This is compared with validation accuracy in the range of 21%—90.19% and F1 of 0.19-0.91 in the competing models.

“Future work could address several open scopes to further improvement in the InceptionV3-Net model’s capabilities,” concluded the researchers. “Applying the model to other renewable energy systems, such as wind turbines or hydroelectric plants, would test its versatility. Further optimization of the model for real-time fault detection could be outlined as future work to improve its practical utility.”

The novel method was presented in “SPF-Net: Solar panel fault detection using U-Net based deep learning image classification,” published in Energy Reports. The team included cientists from Bangladesh’s American International University-Bangladesh, Saudi Arabia’s King Saud University, and India’s GMR Institute of Technology have conducted the study.