The invention discloses a power transformer fault diagnosis method which comprises the following steps: S1, constructing a deep belief network model, and carrying out feature extraction and classification to obtain a trained deep belief network model; S2, processing the to-be-diagnosed data to obtain a corresponding output result; S3, calculating a basic probability value of the output result, andobtaining basic probability values of other evidences at the same time; S4, fusing all evidences by using the D-S evidence theory is fused, and calculating the likelihood, the trust degree and the conflict factor K of each focal element; S5, obtaining a confidence interval (Bel, pl) and a diagnosis conclusion according to the data obtained in the step S4; and S6, reconstructing a deep belief network model according to the fault part in the obtained diagnosis conclusion, and repeating the steps S1 to S5 to further analyze the fault cause until a final conclusion is obtained. According to the diagnosis method, a large amount of multi-source heterogeneous data can be well processed, and the diagnosis method is good in multi-evidence fusion and feature extraction, low in uncertainty and highin reliability.