The invention discloses a transformer fault classification method based on an improved density peak clustering algorithm. The method comprises the steps that (1) transformer oil chromatograph monitoring system data is collected and forms a dataset omega after being preprocessed; (2) a non-parameter density estimation method is adopted to acquire density distribution characteristics of the dataset omega; (3) the density distribution characteristics rho and distance distribution characteristics delta of the dataset omega are utilized to construct a decision graph G, and a sample density center is recognized; (4) a similarity matrix gamma of the dataset omega is calculated, and a density skeleton of all fault type data subsets is constructed based on the similarity matrix gamma; and (5) remaining points are classified according to the distance distribution characteristics delta of all the fault type data subsets to complete fault type recognition. Through the method, the problem that a traditional fault diagnosis method is poor in adaptability due to a complicated transformer fault mechanism is solved, and the method can be applied to fault diagnosis and classification of a power transformer in a targeted mode.