In this study, we propose a transformer fault diagnosis model based on adaptive integrated oversampling (ADASYN) and kernel-based extreme learning machine of improved whale algorithm (IWOA-KELM). This model aims to examine the effect of transformer unbalanced data on the recognition accuracy of the transformer fault diagnosis model. The imbalance between transformer fault data sets results in a bias problem. So, to resolve this, we first use the ADASYN algorithm to optimize the equalization process of transformer fault data. Secondly, we use a multi-strategy combination to improve the search speed, convergence ability, and escape ability of local extremums of the whale optimization algorithm (WOA). Finally, the WOA is used to optimize the KELM regularization coefficient and kernel function parameters. It also constructs a fault diagnosis model of the optimized IWOA-KELM. When the model is applied to the field of transformer fault diagnosis, its diagnostic accuracy is found to improve by 14.17%, 12.5%, and 8.34%, respectively, for particle swarm algorithm KELM (PSO-KELM), gray wolf algorithm KELM (GWO-KELM), and WOA-KELM. Our findings proved that the proposed fault diagnosis model has higher breaking accuracy and generalization ability.