To solve the problem of low classification accuracy due to unbalanced transformer fault samples and a local optimal fault model, a transformer fault diagnosis method based on improved synthetic minority oversampling technique and an optimized deep belief network (DBN) is proposed. First, the K-means algorithm of clustering fusion is used to select the unstable minority class samples by clustering and matching to improve the center point synthetic minority oversampling technique (CP-SMOTE) algorithm and expand the minority class samples to solve the problem of unbalanced distribution of transformer fault data. Secondly, the algorithm of predator optimization is improved by adding stochastic reverse learning and adaptive inertial weight technology, and the internal parameters of deep confidence network are optimized and adjusted by the improved algorithm, which improves the accuracy of the model. Finally, the transformer fault models in different data preprocessing conditions and data scales are simulated and compared. The results show that the transformer fault identification accuracy can be increased to 98% after data preprocessing and model optimization. This effectively solves the problem of low classification accuracy caused by fault data imbalance.