Augmenting few-shot fault diagnosis in photovoltaic arrays using Siamese-enhanced generative adversarial network

To address limited fault diagnosis samples in photovoltaic power plants, this research introduces a novel methodology based on the Siamese-Enhanced Generative Adversarial Network (SiaGAN), integrating Siamese network and Multiple Kernel Maximum Mean Discrepancy (MK-MMD) algorithms. SiaGAN effectively alleviates overfitting in conventional Generative Adversarial Networks (GAN) on small datasets by refining generator parameters using synergistic similarity distance and data distribution metrics. Time-series analysis of input data ensures comprehensive feature extraction. The proposed architecture creates high-fidelity samples for photovoltaic array fault diagnosis in small-sample scenarios. Additionally, a sophisticated loss function balances the similarity of generated samples to actual ones and the diversity among synthetic samples. When synthetic samples are highly similar to genuine ones, the loss function adaptively prioritizes generating more diverse samples, broadening the synthetic data spectrum. Moreover, to circumvent limitations of traditional GAN metrics Fréchet Inception Distance and Kernel Inception Distance, which require abundant source data and pre-trained networks, a novel evaluation metric, Similarity-Yet-Different (SYD), is introduced. Unlike conventional metrics, SYD leverages the Siamese network and MK-MMD to simultaneously evaluate sample quality and diversity without external data sources, suitable for few-shot learning. Empirical findings indicate SiaGAN yields superior quality samples, with the fault diagnosis network achieving 95.00% test accuracy.