Abstract: Condition monitoring of wind turbine can greatly raise the operation of unit and reduce the maintenance cost. Nonlinear state estimation technique (NSET) was used to construct the behavior model of gearbox bearing temperature to complete bearing temperature prediction; grey correlation analysis method was used to verify the rationality of variable selection aiming at the lack of theoretical basis for observation vector choose; similarity analysis method was used to structure simple process memory matrix to shorten the modeling time for the data redundancy of process memory matrix. Prediction residual distribution of model will change when gear box works abnormally and gives warning when the mean or standard deviation of residual error exceeds the threshold. Verification results show that: model based on sample optimization can predict the bearing temperature accurately and has better timeliness, so it can monitor the operation of the gear box.

Key words: wind turbine, gear box bearing temperature, sample optimization, prediction