Abstract:
As an alternative to fossil fuels, renewable energy can help reduce pollution and save the planet's natural
resources. Solar energy is considered one of the most promising renewable energy sources due to its low
cost and high efficiency. It is no surprise that the use of photovoltaic (PV) systems for energy harvesting
has skyrocketed in recent years, given the widespread recognition of the technology's potential. However,
constant monitoring of the system's health is required to guarantee its best performance. The methods
currently used for monitoring are laborious, costly, time consuming and error prone. Two main approaches
are used in the proposed work to overcome the shortcomings of the aforementioned methods. First, a
robust deep-learning method for distinguishing between microcracks and deep cracks in the PV cell is
proposed. The orientation of microcracks plays a significant role in output power efficiency therefore
classified accordingly. The cracks' severity is then used to perform the computation of power. It has been
observed that the output power efficiency is proportional to crack size in the case of deep cracks; therefore,
using digital image processing size of the deep crack is calculated. Using a publicly available online
dataset of electroluminescence images of solar cells, four different deep-learning models are trained,
evaluated, and compared to determine which is the most efficient in detecting cracks. These models are
U-net, LinkNet, FPN, and attention U-net. Different metrics, including intersection over union (IoU) and
precision and recall, are used to compare the models' effectiveness. These models are subjected to an
ensemble learning technique to improve the segmentation IoU score and robustness further. The attention
U-Net model with a mIoU score of 48.5% is observed to outperform all other trained models. In terms of
accuracy, precision, and recall, it also achieved superior numerical values and surpassed other models. UNet is second best in terms of mIoU, while LinkNet performance is inferior among all models. Regarding
optimization, LinkNet outperforms all other models with a loss value of 0.002, whereas U-Net has the
highest loss value of 0.12. The ensemble learning utilized in this method significantly improved the
system's performance, achieving a mIoU of 54.19 %.
