Abstract: The operation state of a wind turbine will be affected by its generator，a key component of a wind turbine.The hybrid kernel-support vector machine（HK-SVM）was used to train the temperature model of a generator under normal working condition to make prediction.Firstly，appropriate sample data were screened，and the parameters with the highest correlationship with the generator bearing temperature at driving end were selected from them by using correlation calculation.Then，temperature range for a generator at normal working condition was established.When the generator ran abnormally，its dynamic characteristics deviated from that under normal working state，resulting in residual difference between the actual temperature and the predicted temperature obtained by HK-SVM. The residual distribution characteristics in the sliding window were calculated dynamically，and a reasonable alarm logic and a threshold were set to realize the early warning for the abnormal state of generators. The results show that this method is of great guiding significance for fault early warning and remote diagnosis of wind turbine.

Key words: wind turbine unit, support vector machine, hybrid kernel function, wind turbine generator, temperature warning