Abstract: Aimed at the gearbox fault problem of offshore wind turbines, a gearbox fault early warning method based on maximum information coefficient- long and short term memory neural network (MIC- LSTM) is proposed. The SCADA system data was preprocessed and the MIC- LSTM network model was constructed to obtain the residual of the gearbox oil temperature under normal working conditions and the predicted output value of the wind turbine. The residual was exponentially weighted moving average (EWMA) processed to determine the threshold range for temperature warning and to achieve gearbox fault warning. Experiments were conducted using data from a 3 MW offshore wind turbine at Guishan Island in Zhuhai City. The results show that compared with other models, this model has an RMSE of 0.34 and an R^2 of 90.7%. The warning effect is significantly better than other models, and it can effectively improve the accuracy of fault warning for offshore wind turbines.