Aiming at the shortcoming of the low accuracy of transformer fault diagnosis, the PSOSOMLVQ（particle swarm optimization,self organizing maps,learning vector quantization） mixed neural network algorithm is presented in this paper. Firstly, the weight of SOM neural network is optimized by the method of PSO algorithm to obtain the more effective topology. Based on that, LVQ neural network is combined to cover the shortage of unsupervised learning SOM neural network. The mixed neural network algorithm combined with PSO, SOM and LVQ can improve the accuracy and reduce the error of transformer fault diagnosis. Through simulation, the three algorithms of SOM, PSOSOM and PSOSOMLVQ are compared. The comparison result show that the PSOSOMLVQ mixed neural network algorithm has the highest accuracy, and the fault diagnosis accuracy rate is 100%. Thus it can be seen, the PSOSOMLVQ mixed neural network algorithm can enhance the performance of transformer fault diagnosis effectively.