In the present paper, an automatic and efficient method for searching for dwarf nova candidates is presented. The methods PCA (principal component analysis) and SVM (support vector machine) are applied in the newly released SDSS-DR9 spectra. The final dimensions of the feature space are determined by the identification accuracy of training samples with different dimensions constrained by SVM. The massive spectra are dimension reduced by PCA at first and classified by the best SVM classifier. The final less number of candidates can be identified manually. A total number of 276 dwarf nova candidates are selected by the method and 6 of them are new discoveries which prove that our approach to finding special celestial bodies in massive spectra data is feasible. The new discoveries of this paper are added in the current dwarf nova template library which can contribute to constructing a more accurate feature space. The method proposed in this paper can also be used for special objects searching in other sky survey telescopes like Guoshoujing (Large Sky Area Multi-Object Fiber Spectroscopic Telescope -LAMOST) telescope.