To ensure the accuracy of fault diagnosis for oil-immersed transformers and to enhance the convergence speed and generalization of the diagnostic process, a new model based on the SMOTE-Tomek-ZOA-DNN method is proposed. This model first employs the SMOTE algorithm to expand the dataset, which increases sample diversity and mitigates class imbalance. Then, Tomek Links are used to remove boundary redundancies, thereby improving data quality by ensuring clearer class separation. To further enhance data accuracy, k-fold cross-validation is applied. Finally, the feature data is fed into the ZOA-DNN model for parameter optimization, creating an integrated fault diagnosis model. As a result of these enhancements, the optimized model’s accuracy improves from 77.55% to 93.33%, significantly increasing the precision and reliability of fault prediction and providing a robust tool for transformer maintenance.