The invention discloses a 
transformer fault diagnosis method, and the method comprises the steps: step1, a vibration 
signal of a 
transformer is acquired, and 
noise reduction preprocessing is performed on the vibration 
signal; step2, EMD 
decomposition is performed, obtained IMF components are selected according to set critical values, so that the IMF components with low amplitude are removed can be prevented; step3, 
wavelet analysis is performed, and feature vectors after de-noising are obtained; step4, a three 
layers BP 
neural system is constructed, feature vectors extracted from the 
wavelet analysis serves as input vectors of BP neural network training to 
train, and a fault 
signal is outputted. According to the invention, methods such as the 
wavelet analysis and the like are adopted to remove interference signal in the fault signal, 
transformer internal 
fault occurrence and development multi-process and multi-fault 
multiple modes are diagnosed through the neural network, the transformer fault can be quickly and accurately diagnosed, the fault diagnosis efficiency is improved, the service life of the transformer is prolonged, the unnecessary costs is reduced, and the cost is saved.