Cross-array fault diagnosis of photovoltaic arrays with different configurations based on endpoint-dense gram feature encoding and mixup-enhanced domain adversarial network

Recently, photovoltaic (PV) arrays fault diagnosis technology has advanced rapidly. However, existing PV array fault diagnosis models typically rely on large datasets collected under specific array configurations. Operating on the premise that training and testing data follow the same distribution, these algorithms fail to address feature distribution discrepancies caused by varying array configurations, resulting in poor transferability and limited generalization when applied to unseen arrays with different structural topologies or PV modules. To address this, considering inter-array relationships, this study proposes a novel fault diagnosis method for cross-array scenarios, i.e., mixup-enhanced domain adversarial network (MDAN). To our knowledge, this study represents an early investigation into unsupervised model transfer across heterogeneous PV arrays to mitigate the resultant domain shifts. The method features three key innovations. First, a two-dimensional Gram feature matrix (2D-GFM) encoding method based on endpoint-dense resampling is designed to extract fault-related similarities from I-V curves. Second, a dual-objective adversarial framework is established, utilizing a Gradient Reversal Layer (GRL) to align feature distributions between the source (labeled) and target (unlabeled) domains. Third, a feature-wise mixup layer is integrated to enhance the decision boundary's robustness against inter-domain variations. Experimental results demonstrate that the proposed method effectively handles scenarios with extremely scarce, unlabeled target- domain samples and enables robust cross-domain transfer across arrays of diverse configurations (e.g., 3 × 4, 2 × 8, 5 × 5, etc.), outperforming existing methods in accuracy and reliability.