Abstract: Objectives: Under the traditional satellite observation data processing operation mode,outlier detection methods in gravity gradiometry have the problems of insufficient accuracy,low efficiency,and so on. This paper constructs an intelligent outlier detection method,which combines variational autoencoder (VAE) and gated recurrent unit (GRU),based on significant characteristics of variational autoencoders in multi-modal data integration analysis. Methods:Firstly,on the basis of the origin and propertier of satellite gravity gradionmetry outliers,the gravity gradiometry dataset with outlier are simulated. Secondly,the network model capture effective features of dataset by variational autoencoder,make predictions on dataset by combining with gated recurrent unit,and automatically find optimal convergence of the loss function by designing adaptive moment estimation as optimizer. Finally,the tested training model will be applied to actual satellite gravity observation data from China’s civilian gravity satellites. Result: The results show that the accuracy of the model in outlier detection reach more than 98%,and have good detection effect on both discrete and regional gross errors. Conclusions:The trained network quickly and accurately constructs the gravity gradiometry simulation dataset sample features,achieving fast and efficient gross error detection capabilities,and can be effectively applied in data preprocessing of China’s autonomous satellite gravity missions.