Data Expansion Method of Transformer Oil Chromatogram Based on Generative Adversarial Net Model
Yaxin Li, Huijuan Hou, Gehao Sheng · 2020
The lack and imbalance of oil chromatographic data lead to problems including over-fitting, unsatisfactory effect and poor representativeness, making it difficult to accurately evaluate the state of power transformers. In this paper, a method of transformer fault case expansion based on policy gradient and generative adversarial networks is proposed to expand the number and diversity of transformer oil chromatographic cases, and the method is verified by case study. The quality of synthesized data is tested by applying a transformer fault classification model based on BP neural network. The dataset is divided into 9 classes of 700 cases. Compared with the classification model obtained by training only the real data as train set, the classification accuracy of the model obtained by mixing the synthesized data with the real data as the train set is improved. And the improvement is higher than other traditional expansion algorithms. These show the effectiveness of the proposed method.