A hybrid convolutional-transformer neural network model for photovoltaic fault detection and localization

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Energy retention from losses is the primary goal of fault detection methodology for photovoltaic (PV) solar systems. A fault detection model should be designed effectively to minimize power and cost waste. We propose a novel fault detection and localization method that leverages deep learning techniques for PV systems. The model is a hybrid semantic segmentation method that combines Convolutional Neural Networks (CNN) and multi-level transformer neural networks. We examine our model on four different segmentation datasets, which have varied characteristics and different capturing conditions, to ensure our model generalization. Using aerial inspection, the first thermal dataset images give a high performance in locating the faulty cells in PV arrays with an accuracy and global precision of 99.89%, a mean average precision (mAP) of 88.73%, a mean intersection over union (mIoU) of 76.29%, and 84.48% of mean dice (mDice, or mean/unweighted F1 score). We use three electroluminescence datasets to investigate the precise location and class of different fault types and minor cracks at the cell level. The second dataset result achieved perfect detection and segmentation of anomaly areas in cells with 99% accuracy, 96.36% mIoU, 98.13% mDice, and 98.41% mAP. The first two datasets contain binary segmentation images with fault and no-fault classes; to accommodate multiple classes, we have utilized the third and fourth datasets. The third dataset, comprising five PV failure classes, is evaluated against other models, yielding a superior performance of 96.27% precision, 77.64% mAP, 57.3% mIoU, and 68.45% mDice. The final dataset has 29 segmentation classes for testing 25 classes, for which it achieves a precision of 95.05%, 66.7% mAP, 49.3% mIoU, and 59.3% mDice.

Links and sources

• PDF from original source

Comments