Aiming at the existing deficiencies of probabilistic neural network (PNN) and genetic algorithm (GA) in internal faults of transformers, a fault diagnosis method based on radial basis probabilistic neural network (RBPNN) improved by particle swarm optimization (PSO) is proposed. Firstly, this paper introduces RBPNN and selects back-propagation as the learning algorithm as well as the content ratio of dissolved gases in oil as the characteristic quantity of fault. Then, since the network structure and the initial value have a great impact on RBPNN, this model is optimized and tested with GA, PSO and improved PSO. The comparison results show that improved PSO has more advantages in determining topology, decreasing error accuracy, accelerating the convergence speed and improving prediction accuracy, which also verify the correctness and feasibility of the proposed method in fault diagnosis. This work is supported by National Natural Science Foundation of China (No. 51577050) and Science and Technology Project of Jiangsu Province Electric Power Company (No. J2014055).