Global Feature Perception Network for Open-Switch Fault Diagnosis in Grid-Connected Photovoltaic Inverters Under Imbalanced Data Conditions

Intelligent monitoring and fault diagnosis of photovoltaic (PV) grid-connected inverters are important. However, the accuracy of traditional fault diagnosis methods is challenged due to multiple data imbalances. To address this unique challenge, this study proposes the global feature perception network (GFPN) framework. First, to effectively capture the operational characteristics of PV inverters under complex conditions, a dilated feature boost block is designed to improve multilevel feature representation. Second, to ensure accurate judgments across various operating scenarios, a dilated feature mix transformer is designed to capture both causal and noncausal relationships between different features. Additionally, to enhance diagnostic performance for minority fault samples, the enhanced dual focal loss (EDFL) is proposed, improving the fault diagnosis effectiveness of the model under imbalanced conditions. Experimental results show that even under the most challenging data imbalance scenarios, GFPN achieves fault diagnosis accuracies of 96.09% and 96.25% on PV grid-connected inverters and the hardware-in-the-loop platform, respectively.