Transformer Fault Diagnosis Method Based on Pseudo‐Image Processing and Improved Random Forest

The power system operation is becoming challenging due to growing power demand. The detection and classification of transformer faults, as the core equipment in the power system, are crucial for the stable operation of the grid. However, there is no uniform standard for transformer fault feature selection based on dissolved gas analysis (DGA). This paper proposes a transformer fault diagnosis method based on parameter migration feature extraction and improved random forest (IRF) feature selection. Firstly, the gramian angular field (GAF) is introduced to transform the one‐dimensional gas sequence into a three‐channel map, and the sample data are balanced using image processing methods. Next, the parameters of the pre‐trained VGG16 feature extraction layer are utilized to establish a model that can extract GAF image features automatically. Then, to obtain optimal features, the IRF algorithm is improved by comprehensively considering the Pearson correlation coefficient. The results indicate that the proposed method is more effective in extracting fault features than the conventional approach. After filtering out the optimal features with IRF, the diagnostic rate of the LR, SVM, MLP, and SGD transformer fault model is improved by 4.27%, 12.2%, 6.7%, 10.97%, and F1_score is enhanced by 4.53%, 12.55%, 6.43%, and 10.92%, respectively. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.