基于FEG – MLViT的草原与牧区光伏电站故障诊断方法

Fault Diagnosis System for Photovoltaic Power Stations in Grassland and Pastoral Areas Based on FEG – MLViT

摘要:

针对草原与牧区光伏电站特定故障频发,生物损害对光伏发电设备造成损坏且实际场景中故障样本不足及数据标签缺失等问题,提出一种融合伪标签的编码生成对抗视觉元学习(Fusion pseudo-label and encoder-generator adversarial network with meta-learning vision transformer,FEG – MLViT)故障诊断方法。首先,将采集的时序数据利用连续小波变换(CWT)从一维时序信号中学习数据结构的时频特征,增强非线性特征的表达能力;之后,设计编码生成对抗网络结合伪标签机制生成标签故障数据,同时,设计一种三维度综合评价指标对原始样本和生成样本进行综合评估,解决了传统评价指标单一的问题;最后,构建一种元视觉(MLViT)诊断网络,利用其全局特征提取与快速自适应能力实现高精度故障诊断。利用实际光伏发电数据对所提方法进行验证,并通过与最新的方法进行对比,验证了所提方法的可行性,最终诊断准确率在5类故障诊断中平均达到95.56% ,且所提方法在小样本(仅9条数据)下就达到收敛,表明所提方法可有效解决小样本的故障诊断问题。

Abstract:

Aiming to address the specific issues of frequent faults in photovoltaic (PV) power stations located in grassland and pastoral areas, particularly those caused by biological damage to PV equipment, as well as the challenges of insufficient fault samples and missing data labels in real-world scenarios, a novel fault diagnosis network was proposed: the fusion pseudo-label and encoder-generator adversarial network with meta-learning vision transformer (FEG – MLViT). Firstly, the collected time-series data was transformed by using continuous wavelet transform to extract time-frequency features from onedimensional signals, thereby enhancing the representation of nonlinear characteristics. Then an encodergenerator adversarial network integrated with a pseudo-labeling mechanism was designed to generate labeled fault data. In addition, a novel three-dimensional comprehensive evaluation metric was proposed to assess both real and generated samples, addressing the limitations of traditional single-dimensional evaluation indicators. Finally, a meta-learning-based vision transformer (MLViT) diagnostic network was constructed, leveraging its global feature extraction and rapid adaptation capabilities to achieve highprecision fault diagnosis. The proposed method was validated by using real PV operational data. Comparative experiments with state-of-the-art approaches demonstrated the feasibility and effectiveness of the proposed model, achieving an average diagnostic accuracy of 95.56% across five fault categories. Moreover, the model converged with as few as nine samples, indicating its strong performance in few-shot fault diagnosis scenarios.