The performance of power transformer directly affects the operation reliability of power system. Accurate and efficient transformer fault diagnosis can help to discover the unsafe factors of power grid in time. In this paper, a distributed neural network decision method of transformer fault type diagnosis based on massive data driven is presented. The ratio method is used to obtain the ratio type eigenvector of DGA. According to the Pearson correlation coefficient and the Euclidean distance calculation, the representative data of each class are selected as the training samples. The BPNN algorithm is parallelized on Spark to adapt to mass data processing. The samples are divided into pieces by interpolated random sampling, and sub classifiers with different performance are constructed through BPNN learning different training pieces. Finally, the final diagnosis type is obtained by majority voting of sub classification results. The example shows that the method has a good effect on diagnosis of transformer fault type. It is more accurate than the IEC three ratio codes and the standalone BPNN diagnosis result. It also proves the validity and applicability of the method for transformer fault type diagnosis.