Oil immersed transformer will inevitably have various latent faults during operation, it is very important to diagnose the state of transformer in time and correctly. The traditional three ratio method based on DGA data has insufficient coding, which limits the effect of fault diagnosis. In this paper, an improved manta ray foraging optimization (MRFO) is proposed to optimize the back propagation (BP) fault diagnosis model. Firstly, the algorithm uses the multi-stage algorithm integrating logic mapping and opposition based learning (OBL) to provide the initial position for MRFO and strengthen the global optimization ability of the algorithm. At the same time, three foraging strategies of manta ray algorithm are optimized by orthogonal experiment to adjust the individual exploration and development of manta ray, so as to strengthen the optimization ability of the algorithm on specific problems. Finally, the optimal solutions obtained by the improved manta ray algorithm are given to the weight and bias of BP network, and the transformer fault diagnosis system is established. Experiments are carried out using IEC TC 10 fault data, and the results are compared with the other algorithms, which show that the accuracy of the algorithm proposed in this paper is 16%, 8% and 24% higher than those using BPNN, unmodified MRFO-BP and three ratio method respectively.