Accurate separation of fault leakage current from residual current was a typical new data prediction problem, the methods of fault leakage current separation were scarce and the accuracy was low. In this paper, we proposed a construction strategy of small scale cross auto encoder deep network, and applied the model to separate fault leakage current from the residual current. First, two independent auto encoder networks were learned on the residual current dataset and the fault leakage current dataset respectively. Then, the feature encoding module of residual current and the feature decoding module of fault leakage current were cascaded to form a cross auto encoder network. Finally, separation mapping model of residual current to fault leakage current was obtained by using the paired residual current and fault leakage current for finetuning training of the crossauto encoder network. Experiment results showed that the average separation accuracy was 7733% when the error threshold was set to 5. When the error threshold was 15, the accuracy was up to 8867%. Obviously, the method can realize the separation of fault leakage current and provide the technical support for the design of intelligent current separation residual current protection device.