A feature space class balancing strategy-based fault classification method in solar photovoltaic modules

Photovoltaic (PV) power generation has become a primary method of energy production due to its clean and sustainable nature. Therefore, efficient fault detection and classification in PV systems are crucial for reducing energy losses and improving economic efficiency. This study addresses the challenge of class imbalance in PV fault data and proposes a fault classification method based on a feature space class balancing strategy. Firstly, a concurrent dual sampling method was developed to generate sample sets that are biased towards the majority and minority classes, respectively. Secondly, a PatchUp-based block-level feature mixing method was designed to fuse feature blocks of minority classes with those of the majority classes. Next, a multi-scale residual network with convolutional branches at different scales was constructed to capture image features effectively. The proposed method optimizes class balance by generating imbalanced feature-mixed samples, enhancing the network's attention to the minority classes. It achieved an average accuracy of 96.0% for 2class classification and 87.6% for 12-class classification on the Infrared Solar Modules dataset. Additionally, to address the issue of the limited PV modules sample set, we constructed a new dataset. On this dataset, our method achieved an average accuracy of 93.6% for 5-class classification. Furthermore, extensive comparative experiments demonstrated that the proposed method significantly improved the classification accuracy of minority classes without compromising the performance of the majority classes.