Standard particle swarm optimization (PSO) algorithm just takes simple linear attenuation way to update the inertia weight, so it can not get the global optimum value. In order to solve this problem, in the paper, an improved particle swarm optimization (IPSO) algorithm is proposed, which is combined with the error back propagation neural network (BPNN), then a new transformer fault diagnosis method based on IPSO-BPNN is presented. The method gets the number of times for which the individual particle is continuously selected as the optimal point, which is taken as an adaptive variable and is used to adaptively adjust the inertia weight along with the particle's performance classification, so as to balance the local and global search capabilities. A large number of simulation shows that the algorithm is better than the BPNN and PSO-BPNN based transformer fault diagnosis system, and it can get a higher correct rate of transformer fault diagnosis.