Dissolved gas analysis (DGA) is an effective online fault diagnosis technique for large oil-immersed transformers. However

due to the limited number of DGA data

most deep learning models will be overfitted and the classification accuracy cannot be guaranteed. Therefore

this paper has introduced the idea of deep neural networks into the multi-grained cascade forest (gcForest)

which is a tree-based deep learning model

and proposed an improved gcForest that can be accelerated by GPU. Firstly

in order to extract features more effectively and reduce memory consumption

the multi-grained scanning of gcForest is replaced by convolutional neural networks. Secondly

the cascade forest (CasForest) is replaced by cascade eXtreme gradient boosting (CasXGBoost) to improve the classification ability. Finally

235 DGA samples are used to train and evaluate the proposed model. The average fault diagnosis accuracy of the improved gcForest is 88.08 %

while the average recall

precision

and Fl-score are 0.89

0.90

0.89

respectively. Moreover

the proposed method still has high fault diagnosis accuracy for datasets of different sizes.