Abstract: Aiming at the problem of low accuracy of existing power transformer fault identification algorithms, in order to realize the accurate identification of power transformer faults, by analyzing the characteristics of transformer oil chromatographic data, a power transformer fault identification algorithm based on multi-depth neural network was proposed. The algorithm consists of two parts: The first part is based on Spark and constructed multiple deep neural networks (DNN) recognizers, and each recognizer is added with a Dropout layer to reduce the over-fitting phenomenon and to enhance the generalization ability of network; After that, using Spark computing framework to assign multiple deep neural network recognition tasks to the slave nodes in the Spark cluster to improve computing efficiency. The second part is the fusion of recognition results and decision-making algorithms. The Reduce module of the Spark framework was used to gather the recognition results of each recognizer. The trust evaluation mechanism and the vector similarity evaluation mechanism were introduced to construct the multi-recognizer fusion decision model, and then get the final comprehensive decision results. The results of actual calculation examples show that the multi-deep neural network integrated fault recognition method can improve the diagnosis accuracy by more than 5% on the basis of the traditional DNN algorithm, and it also has different improvement degrees of recognition accuracy compared with other comparison algorithms.