Transformer Fault Sample Generation Method Based on Embedding Typical Transformer Fault Evolution Laws
Peng Zhang, Xiaoyu Fan, Zexu Du, Weishang Xia · 2025
To ensure the safe and stable operation of power transformers, mining the variation characteristics of state quantities in typical fault samples of power transformers based on machine learning approaches and constructing diagnostic analysis models is a significant research direction. Nevertheless, transformer diagnostic models encounter challenges such as a shortage of fault samples and subpar sample quality, which impedes the enhancement of model accuracy. The samples generated by existing methods based on data interpolation and general generative models are typically of inferior quality and fail to fulfill the actual requirements of model construction. To address the aforementioned issues, this paper proposes a nearest neighbor sample segment splicing generation method embedded with the evolution patterns of typical faults in power transformers. Firstly, various typical faults are simulated on the power transformer fault simulation platform, and the variations of key state quantities are measured to fit the evolution laws of each state quantity during the fault development process. Subsequently, the state quantities during the typical fault development process are segmented, with the evolution laws of the state quantities retained in each segment. The state quantity segment data are regarded as similar image patches and embedded in the GPNN (Generative Patch Nearest-Neighbor) generative model. Finally, a GPNN model for power transformer sample generation is constructed, and a large number of similar transformer fault samples are generated through the generative model and local block matching technology. Utilizing the method proposed in this paper, with 117 samples obtained from simulation experiments as training data, 536 samples were generated through the GPNN model. The matching degree of the 536 samples with the original samples was evaluated from multiple dimensions such as completeness, coverage, overlap, and identification quality, reaching$\mathbf{9 0. 0 3 \%}$.