The failure frequency and maintenance cost of wind turbine gearbox are relatively high. It is necessary to monitor its operation condition in realtime. Nonlinear state estimation technique (NSET) has the problems of high dependence on memory matrix, low accuracy caused by insufficient data utilization, bad realtime performance, etc. Therefore, a condition monitoring method based on soft fuzzy Cmeans clustering (SFCM) and ensemble NSET is proposed. SFCM is adopted to divide the historical data into different classes with overlapping boundaries to achieve the soft condition division. NSET models under different conditions are constructed as individual learner. The parametric regression method is used as the combiner. Lots of data can be used to train the parameters without affecting the realtime performance and the accuracy can be improved accordingly. The gearbox fault data of a 2 MW wind turbine are taken to evaluate the method. Compared with NSET, experimental results show that the proposed method has better accuracy and realtime performance. Through the means of predicted residual and the health index based on residual, it can reflect the early fault and its development trend of gearbox sensitively and accurately.