Transformer DGA Fault Diagnosis Based on SMOTEand SSA-Optimized XGBoost

Abstract

In the daily operation and maintenance of power systems, the
condition monitoring of oil-immersed transformers has always been a
significant issue. Dissolved Gas Analysis (DGA) is one of the effective
methods for detecting faults in oil-immersed power transformers. However, in
practical applications, factors such as the uneven distribution of faulty samples,
inappropriate selection of feature dimensions, and difficulty in determining the
parameters of the classifier limit the accuracy of traditional diagnostic methods,
making the diagnostic results vulnerable to interference. This paper presents a
hybrid fault diagnosis scheme that integrates the Synthetic Minority
Over-sampling Technique (SMOTE) for oversampling of minority class
samples, feature importance selection based on LightGBM (Light Gradient
Boosting Machine), and the eXtreme Gradient Boosting classifier (XGBoost)
optimized by the Sparrow Search Algorithm. The experimental results on the
public dataset show that the proposed method achieved an accuracy of 97.42%,
a Macro-F1 score of 97.36%, and a Kappa coefficient of 0.9677. Compared to
the unoptimized XGBoost, SSA optimization improved accuracy by 0.42% and
Macro-F1 by 0.48%, This demonstrates that automatic hyperparameter tuning
is effective. The proposed method also outperforms SVM (80.26%), Random
Forest (93.56%), and PCA + SSA-LightGBM (88.41%). In addition, the
ablation experiment further validated the role played by each component.