Transformer Fault Diagnosis Based on Dissolved Gas Analysis in Oil and Ensemble Learning

Abstract

Dissolved Gas Analysis (DGA) in oil is one of the most commonly used methods for transformer fault diagnosis, and artificial intelligence is an effective means to realize transformer faults. However, due to the large difference in the number of fault samples of different types of transformers, a single artificial intelligence method combined with DGA is prone to bias the parameter update of the majority class samples while ignoring the correct classification of the minority class samples during the goal realization process. In order to improve the transformer fault diagnosis accuracy in the unbalanced sample scenario, this paper proposed an ensemble model based on DGA and AM2-CT (Adaboost.M2-Classiflcation Tree). Meanwhile, IEC three ratio method, Duval triangle method, Rogers ratio method, percentage method and nine ratio method were selected to preprocess the original fault data of the transformer, and the obtained noncode ratios were used as the input feature sets of the model respectively. The proposed ensemble model effectively alleviates the training problem of imbalanced data at the level of algorithm optimization. Compared with Support Vector Machine (SVM), Naive Bayes (NB), K-Nearest Neighbor (KNN), Decision Tree (DT), the accuracy of AM2-CT increased by 7.2%, 8.2%, 18.6%, 4.4% respectively.