CGIR: A knowledge-guided integrated framework for fault diagnosis of power transformers under imbalanced DGA data

To improve fault diagnosis performance for power transformers under the challenge of imbalanced dissolved gas analysis (DGA) data, this study proposes a novel diagnostic framework named CGIR. This framework integrates CTGAN-CRS-based data generation, Guided Random Forest (GRRF)-based feature construction, an Improved Convolutional Neural Network (ICNN) for fault classification, and Red-tailed Hawk (RTH) algorithm-driven hyperparameter optimization. Specifically, a hybrid approach combining Conditional Tabular Generative Adversarial Network (CTGAN) with a domain knowledge-driven Cascade Reject Sampling (CRS) mechanism is employed to generate high-quality, balanced synthetic data. Subsequently, a comprehensive gas ratio method combined with random forest screening constructs a diverse and discriminative feature space. For classification, the proposed ICNN enhances feature learning and model optimization by integrating the Convolutional Block Attention Module (CBAM) and Kolmogorov–Arnold Network (KAN), with hyperparameters tuned by the RTH algorithm. Experimental evaluations demonstrate that CGIR substantially improves diagnostic accuracy, robustness to data imbalance, and classification stability across multiple fault types, offering a reliable and effective solution for practical transformer fault diagnosis scenarios.