Abstract: In order to fully consider the differences of data information extraction among different diagnosis models and bring the advantages of different diagnosis models,the paper proposes a strategy for constructing a two-layer cascade transformer fault diagnosis model based on ensemble learning. Firstly,the feature extraction of dissolved gas data is performed by using the non-code ratios. Then four classifiers at the first layer are used to classify the data in parallel. Finally,the diagnosis results are obtained by using the second layer classification model to extract the features from the combined diagnosis results of the first layer model. The case validation shows that compared with the four classifiers at the first layer including support vector machine,classification regression tree,K-nearest neighbor,and Naive Bayes,the cascade model achieves the optimal or suboptimal level of classification accuracy for each classifier for each fault type,and improves the comprehensive recognition accuracy by 6%,24.8%,8.96% and 4.99% respectively.