The traditional fault diagnosis methods of power transformer can not detect the power fault accurately and ensure the normal operation of power transformer. Therefore, the fault diagnosis method of power transformer based on wavelet packet transform and support vector machine is studied. For the collected power transformer power signal, the improved minimum noise separation transform denoising (MNF) is used to denoise, and the noise matrix is estimated by the weighted neighborhood mean method. After the estimation, the improved MNF transform is used to effectively realize image dimensionality reduction and denoising, extract the signal characteristics, and divide the signal into low-frequency part and high-frequency part by wavelet packet transform, In order to obtain the wavelet packet energy feature vector, the obtained wavelet packet energy feature vector is input into the support vector machine classifier, and the output results of the support vector machine classifier are used to realize the state recognition and fault diagnosis of power transformer. The experimental results show that this method can effectively diagnose the faults in power transformer, such as iron core short circuit, coil interlayer short circuit, bushing to ground breakdown, coil insulation resistance drop and bushing to bushing discharge, and the fault diagnosis accuracy is higher than 98.5%.