The metabolomics dataset is disturbed by various stimuli inevitably. The main task for metabolomics data preprocessing is to reduce the impacts of the disturbing factors. In present work, the formation of data variance and their distribution in feature space are analyzed. Furthermore, a new method to filtrate unknown disturbing factors is proposed and the significance of interesting factors is improved. The efficiency of the new filtering algorithm is estimated by real metabolomics dataset. Comparing with orthogonal signal correction (OSC) method, the experiment shows that the new method is superior in reducing unknown disturbing factors and retaining useful information and intrinsic individual differences in organisms. In addition, it can also prevent the overfitting of model and make the subsequent statistical analysis more reliable.