Aiming at the low accuracy of fault diagnosis in oil-immersed transformers caused by redundant features in the sample data, this paper proposes a new transformer fault identification model based on a multi-strategy improved dung beetle optimizer (MSIDBO) and kernel extreme learning machine (KELM) is proposed. First, random forest (RF) and kernel principal component analysis (KPCA) are used to extract the features from the raw transformer data, effectively reducing dimensionality. Next, the dung beetle optimizer (DBO) is comprehensively improved through multiple strategies, including improved Circle chaotic mapping, variable spiral search mechanism, nonlinear control factor, fusion with the sine-cosine algorithm, and a hybrid multi-population differential evolution mutation strategy. These enhancements significantly improve the algorithm’s global search ability and convergence accuracy. Finally, the MSIDBO algorithm is applied to optimize the kernel and regularization parameters of KELM, constructing the KPCA-MSIDBO-KELM transformer fault diagnosis model. Experiments show that the constructed transformer fault diagnosis model achieves an accuracy of 94.07%, which is 2.54%, 3.39%, 5.93%, 7.63% and 13.56% higher than DBO-KELM, WOA-KELM, HHO-KELM, GWO-KELM, and PSO-KELM. Overall, the proposed method significantly improves the accuracy of transformer fault diagnosis.