Detection of busbar defects in photovoltai̇c electrolumi̇nescence i̇mages by deep learning methods

The energy needs of countries are increasing day by day. Therefore, interest in renewable energy sources has become inevitable. Among these energy sources, solar energy is renewable and sustainable, unlike fossil fuels. In this context, the importance of photovoltaic panel production, which produces electricity with sunlight, is gradually increasing. Like any other technology, photovoltaic panels can experience various defects or problems that can affect their performance. It is worth noting that regular maintenance, periodic inspections and quick repairs can help eliminate or prevent these defects in photovoltaic panels, ensuring their longevity. It is much more critical to detect these problems during the production phase, and problems that may occur during the operating phase can be prevented. There are some commonly used methods to detect defects in FV cells. One of the most used methods today is detecting defects with Electroluminescence imaging . It is worth noting that detecting defects in FV cells often requires special equipment and expertise. FV panel defects are very diverse and only qualified professionals can perform these tests and inspections to accurately detect defects. In this study, considering such difficulties, studies have been carried out to automatically detect busbar slippage defect, which is one of the photovoltaic panel defects, with deep learning methods. The study uses a dataset containing 60-cell panel images taken from a real production line. These images are labeled as bar shift defective and perfect and consist of 500 cell EL images in total. Feature extraction was made using two different pre-trained deep learning architectures, and the effect on success parameters was observed by combining the obtained features with different combinations. The obtained qualities were classified with Naive Bayes and k-NN classification methods and success parameters were obtained. The results show that the deep features obtained from Relu34 and Relu25+Conv10 layers give the most successful results, providing an accuracy rate of 0.9920. In terms of classification methods, distance-based methods have been found to classify bus drift defects in photovoltaic EL images more successfully. Additionally, it has been revealed that combined deep features are more successful in bar slippage defects identified by machine learning methods. In the classification made using a single architecture, it has been understood that the deeper layers the attributes are taken from, the more successful a detection can be made. All studies show that bus drift defect can be successfully detected from photovoltaic panel EL images with pre-trained deep learning architectures.