The invention discloses an unbalanced data set-oriented extreme learning machine based transformer fault diagnosis method. The method specifically comprises the following steps: step 1, dividing a collected unbalanced sample set S={(x1, t1), (x2, t2)...(xn, tn)} with class labels of an oil-immersed transformer into training samples and test samples by a ratio of 6:1, wherein xi represents a sampleproperty, i may be equal to 1, 2, 3, 4, 5, 6, and specifically comprises six attributes of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, ti represents a class label, i may be equalto 1, 2, 3, 4, 5, 6, 1, 2, 3, 4, 5, 6 respectively corresponding to normal state, middle temperature overheat, high temperature overheat, partial discharge, spark discharge, arc discharge, and the tiis clustered by a PAM algorithm; step 2, for minority classes, taking the cluster center of the PAM algorithm as a central point; and step 3, during the classification output stage of the extreme learning machine, firstly establishing a DAG-ELM model, secondly dividing a new data set generated in step 2 into training sets and test sets by the ratio of 6:1, wherein 6 parts are used for training modeling, and 1 part is used for verifying the classification effect. According to the unbalanced data set-oriented extreme learning machine based transformer fault diagnosis method, the influence of the unbalanced data set on the transformer fault diagnosis result is solved.