In recent years, the wind power has developed rapidly, but the proportion of wind turbine operation and maintenance costs large, so the wind turbine installation and application-line monitoring system is essential. This paper mainly studies the wind turbine on-line monitoring methods, a large number of bearing vibration signal from 1.5 MW direct-drive models in wind farms were extracted. Select a group normal and fault data, use wavelet packet domain bispectrum analysis, and compared with the traditional bispectrum method. Experimental results show that wavelet packet domain bispectrum was significantly better than the traditional bispectrum, and it can well suppress non-Gaussian noise, and can effectively avoid the disadvantages of the wavelet analysis. It provides a good basis for the online monitoring reflection of further information and early warning of abnormal conditions.