Abstract: Early overflow detection and warning is an important means to prevent blowout accidents. Most of the current overflow identification methods use a single monitoring means, and the identification of drilling data is not accurate enough. The key of overflow identification is the extraction of abnormal data, which is not linear relationship with whether overflow occurs. In order to improve the efficiency of early onsite overflow warning, the Attention-GRU algorithm is used to build an overflow identification model to correct the outlet flow and drilling data. The input is arranged in time series. The output is the overflow discrimination result, and the overflow probability is calculated to achieve early overflow identification and early warning. Through comparative analysis, the optimal time interval suitable for the model is selected. And compared with three time series prediction models: RNN, LSTM and GRU, the accuracy of the monitoring results of the early warning model is verified, and it has high field application value.