The Auto Associative Kernel Regression (AAKR) algorithm does not consider the contribution of each element in the state vector to the Euclidean distance when calculating the similarity, and the model parameter is often calibrated based on subjective experience. As a result, the accuracy of the model is relatively low. A non parametric modeling method for establishing the normal behavior model of gearbox is proposed based on the SFO algorithm and the modified AAKR algorithm. Firstly, the memory matrix is constructed by full parameter equal interval partition method; Secondly, the distance weight coefficient is introduced into the AAKR model, and the width coefficient and distance weight coefficient in the AAKR model are optimized by SFO algorithm; Finally, the health index is constructed based on sliding window and residual data to realize the condition monitoring of wind turbine gearbox. Taking the measured data of a 2MW wind turbine as an example, the results show that compared with the traditional AAKR, weighted AAKR and robust state estimation model, the average accuracy of the proposed algorithm is improved by 1.55%, 0.6% and 0.76% respectively. In fault early warning, the constructed health index can more sensitively and accurately reflect the early fault and development trend of gearbox.