The mixed vibration signal of transformer oil tank surface is composed of winding and core vibration signals.Separation of mixed signals can improve the accuracy of using vibration signals to diagnose internal mechanical state.RBF（radial basis function） neural network has the advantages of approaching nonlinear function with arbitrary precision. In this paper

a separation model is established by using radial basis neural network. It takes the frequency domain signature of mixed vibration signal as input and frequency domain signature of source signal as output.The collected vibration signals are divided into a training set and a verification set. The training set is used to train the network

and the verification set is used for verification.Then set the target value of training error. When the network iteration reaches the specified accuracy

the network training is completed. So it can be used to realize the separation of vibration signals of core and winding.The calculation results of the waveform similarity coefficient show that the similarity coefficient of the separated signal and the vibration source is higher than 0.7. Separation effect is ideal. Compared with the traditional blind source separation method

the RBF neural network based separation method overcomes the two limitations of the former:One is the uncertainty of the sequence of source signals

that is

the sequence of separated core and winding vibration signals cannot be determined. The other is the uncertainty of the signal amplitude

that is

the signal is similar to the source signal and the amplitude difference is large.The vibration signals of the windings and cores can be obtained more accurately

and the diagnosis of the mechanical state of the windings and cores can be realized.