Abstract: In order to solve the influence of long time series of fault deterioration gradual process on gear box condition monitoring model and improve its decision-making accuracy,a combined modeling method based on SCADA data is proposed in this paper. Firstly,the input observation vectors which are closely related temperature of the gear box are selected by using principal component analysis（PCA）,and the temperature model of gear box under normal and abnormal conditions are established by long short-term memory（LSTM）neural network respectively. Secondly,the residual distribution eigenvectors are extracted by combining the model output with SCADA data,and the random forest residual distribution model is implemented to monitor the operation status of the gear box. Finally,the model is carried out and simulated in a large wind farm. The strong practicability and high accuracy of the LSTM neural network combined with random forest algorithm in monitoring the wind turbine gear box are demonstrated through the results,which provides a new method and idea for subsequent health evaluation of gear box.