Aimed for the poblems of high feature dimensions of the samples in the process of pattern recognition and fuzzy rough set dimension reduction method easily affected by the outlier,a dimension reduction method based on fuzzy rough set combined with outlier detection was proposed. The method was eliminated of outlier by the calculation of instability coefficient values changing based on gravity among samples, which dispelling the influence of outliers disturbing the fuzzy rough set feature dimension reduction, and selected the feature set sensitive to pattern classification. The proposed method was tested by the simulation data and the gear failure data analysis. The experimental results show the effectiveness of the proposed feature dimension reduction method.