Abstract: Transformer fault diagnosis often lacks a unified criterion for feature selection, with redundant features reducing the diagnostic performance. To address this, we propose a novel transformer fault diagnosis method based on image feature extraction and dual-layer feature selection. Using dissolved gas in transformer oil as the research object, we transform the concentration of dissolved gas into images using the Gramian Angular Field (GAF) method, followed by data equalization through image processing algorithms. Second, we transfer the feature parameters of the VGG16 (Visual Geometric Group 16) algorithm to the GAF images to construct a fault diagnosis model that automates feature extraction. After comprehensively considering the importance score and correlation coefficient, we improve the random forest algorithm to filter important features and establish a dual-layer feature preference model to enhance the fault diagnosis ability of power transformers. On this basis, the effectiveness of the proposed method is verified by four classifiers: linear regression, support vector machine, multilayer perceptron, and stochastic gradient descent. The experimental results show that the visualization-based method extracts fault features more effectively than traditional approaches. After the dual-layer optimization of the features, the accuracy of transformer fault diagnosis is improved by 4.27%, 11.2%, 6.1%, and 10.97%, respectively, and the F₁ value is improved by 4.53%, 12.55%, 6.08% and 11.1% respectively. The transformer operation state is identified more accurately.