Enhancing DGA Transformer Fault Diagnosis Using Data Reconstruction and GD-AHBA-SVM Model

The application of machine learning method for oil-immersed transformer fault diagnosis and with monitoring dissolved gas analysis (DGA) is an effective engineering method. This paper proposes a novel framework for improving accuracy of power transformer incipient fault diagnosis. Firstly, this paper reconstructs the DGA primitive data with Adaptive Synthetic Sampling (ADASYN) synthesizing minority class samples and Kernel Principal Components Analysis (KPCA) extracting features. Furthermore, we propose a novel GD-AHBA optimization method for enhancing SVM performance and build the GD-AHBA-SVM model. On the one hand, we improve the control parameters of population motion and enhance the ability to escape from local optimum in the later periods. On the other hand, Good Point Set theory and Differential Evolution (DE) are incorporated to optimize the spatial distribution of the population, which improves the convergence accuracy and speed of Honey Badger Algorithm (HBA), and reduces the computational overhead of invalid populations. Various diagnostic methods are evaluated by experimental comparison and results show that the framework proposed in this paper significantly improves the accuracy of transformer DGA fault diagnosis.