Abstract: In order to improve the robustness to noise and computational efficiency, and considering the correlation between features, a feature level data fusion method based on optimal estimation of low rank factor is proposed. By optimizing the cosine similarity measure between the potential feature vector and the vector extracted from the linear combination estimation, the estimation of the potential signal feature space was obtained. The stochastic matrix theory was used for feature and data fusion to solve constrained data-driven optimization problems with different noise levels. Experimental results on two datasets show that the effect of the proposed method is remarkable.