Back propagation neural network is applied in transformer fault diagnosis. And a comparative study of neural network efficiency is also presented according to five diagnosis criteria commonly used for dissolved gas analysis in transformer insulation oil. These criteria are modified Rogers

IEC

Doernenburg

Duval and CSUS. The study shows that the efficiency of diagnosis is different when the criteria or methods under consideration are different

with values in the range of 88.3%～96.7%. The four gas ratios （FGR） method and six key gases （SKG） method based on the criteria are designed

reaching the highest efficiency. The verification shows that the BPNN diagnosis method is fit for the potential fault diagnosis of transformer.