State Key Laboratory of Electrical Insulation and Power Equipment, Department of Electrical Engineering, Xi'an Jiaotong University, Xi'an 710049, China.
Department of Computer Science, Xi'an Jiaotong University, Xi'an 710049, China.
ISA Trans. 2023 Mar;134:268-277. doi: 10.1016/j.isatra.2022.08.009. Epub 2022 Aug 19.
Data-driven diagnosis methods have been systematically investigated for the diagnosis of gas-insulated switchgear (GIS) partial discharge (PD). However, because of the scarcity of samples on-site, an operational gap exists between the diagnostic methods and their actual application. To settle this issue, a novel metric-based meta-learning (MBML) method is proposed. First, a hybrid self-attention convolutional neural network is constructed for feature extraction and trained through supervised learning. Then, the episodic MBML is used to train other parts, and the metric classifier is employed for diagnosis. The proposed MBML exhibits an accuracy of 93.17% under 4-way 5-shot conditions, which is a significant improvement over traditional methods. When the number of support sets is small, the benefits of MBML are more prominent, providing a viable solution for the on-site diagnosis of PD in GISs.
针对气体绝缘开关设备（GIS）局部放电（PD）的诊断，已经系统地研究了数据驱动的诊断方法。然而，由于现场样本稀缺，诊断方法与其实际应用之间存在差距。为了解决这个问题，提出了一种新的基于度量的元学习（MBML）方法。首先，构建了一个混合自注意力卷积神经网络进行特征提取，并通过监督学习进行训练。然后，使用基于事例的 MBML 来训练其他部分，并使用度量分类器进行诊断。所提出的 MBML 在 4-way 5-shot 条件下的准确率为 93.17%，明显优于传统方法。当支持集数量较小时，MBML 的优势更加明显，为 GIS 局部放电的现场诊断提供了可行的解决方案。
