A multi strategy improved sparrow search algorithm (ISSA) based on sample equalization and fused feature optimization was proposed to optimize the eXtreme gradient boosting (XGBoost) transformer fault diagnosis model for the problem of low diagnostic accuracy caused by imbalanced fault samples and difficult effective feature extraction in oil immersed transformers. Initially, the conditional tabular GAN (CTGAN) was employed to balance the dissolved gas analysis (DGA) dataset. Subsequently, a 23-dimensional candidate set of transformer fault features was established using the correlation ratio method. The optimal feature set for input was then determined by combining the Spearman correlation coefficient with the feature importance derived from the XGBoost model. Furthermore, the SSA was enhanced by incorporating improved Henon_Circle chaotic mapping, the fish eagle algorithm, the differential evolution algorithm, the Cauchy distribution, and an elite opposition-based learning strategy. The superiority of the ISSA was validated through performance tests using four typical test functions. Ultimately, the ISSA-XGBoost diagnostic model was constructed, and its diagnostic performance was verified through multiple comparative experiments. The results show that the proposed method achieves a diagnostic accuracy of 97.8%, which can effectively improve the sample classification capability and is more suitable for transformer fault diagnosis with imbalanced samples.