Fault diagnosis of photovoltaic array under service-cycle degradation evolution via semi-supervised conditional domain adversarial network with dynamic label smoothing

The inevitable performance degradation of photovoltaic modules during long-term service induces a dynamic drift in their output feature distributions over time. However, existing fault diagnosis methods predominantly rely on static datasets acquired from undegraded arrays or those with fixed degradation levels, largely overlooking the dynamic feature variations throughout the module lifecycle. Consequently, these models suffer from severe domain shift when confronted with unknown degradation states, leading to a significant decline in generalization capability and diagnostic accuracy. To address this challenge, a novel semi-supervised fault diagnosis framework is proposed, termed the dynamic label smoothing-based conditional domain adversarial network. Specifically, a conditional domain adversarial architecture is constructed, wherein a conditional mapping-based discriminator is employed to align data features from progressive degradation stages at the class level into a unified feature space, thereby effectively eliminating degradation-induced distribution discrepancies. Furthermore, to combat the few-shot overfitting problem caused by the scarcity of fault samples in degraded arrays, a dynamic label smoothing strategy is innovatively introduced. Coupled with consistency constraints, the proposed approach effectively enhances feature discriminability and the robustness of decision boundaries. Experiments are conducted under various unknown degradation scenarios, with the influence of dynamic dust accumulation further distinguished. The results demonstrate that the proposed method achieves a diagnostic accuracy of up to 98.55% using only an extremely limited number of target domain samples, exhibiting substantial superiority over comparative methods in both data efficiency and generalization robustness across life-cycle degradation stages.