Jun Chen ¹ Yong Wang ¹ Lingming Kong ¹ Yilong Chen ¹ Mianzhi Chen ¹ Qian Cai ^2* Gehao Sheng ²

• ¹ Guangzhou Power Supply Bureau, Guangdong Power Grid Co., LTD., Guangdong, China
• ² Shanghai Jiao Tong University, Shanghai, China

Power transformer fault diagnosis methods based on machine learning struggle with the im-balanced distribution of fault cases across different categories, which may lead to decreased diagnosis accuracy. To improve the accuracy and operation efficiency of the power transformer fault diagnosis model, this paper constructed a fault diagnosis model that integrates neighborhood component analysis (NCA) and k-nearest neighbor (KNN) learning, with the introduction of correction factors. Firstly, a correction factor was introduced into the objective function of the NCA algorithm to mitigate the in-fluence of sample imbalance on model training. The sample parameter correlation quantization matrix was obtained by combining oil chromatography fault data through the association rule, which would be used as the initial value for the NCA algorithm's training metric matrix. The metric matrix obtained from the training was then used to perform a mapping transformation on the input data of the KNN classifier to reduce the distance between samples of the same type, thereby improving the KNN classification performance. Finally, the Bayesian optimization algorithm was employed to tune the hyperparameters of the model to obtain the model parameter set that can achieve the highest accuracy of the test set. Through the analysis of the transformer fault case library, the model proposed in this paper saved nearly half the time compared with traditional machine learning diagnosis models. Furthermore, the diagnosis accuracy of the minority sample classes was improved by at least 15% compared with other models.