Special stars are stars with anomalous metal abundance, the information of which is of great importance to the study of the origin of the universe, the evolution of the solar system and the evolution of life. Therefore, the search of special stars is an important goal in the large-scale survey project at home and abroad. Stellar spectra contain a wealth of information on the chemical composition, the physical property, and the movement state of stars, which is an important basis for conducting stellar studies. Stellar identification, classification, and the discovery of special stars are largely based on stellar spectral data. With the development of large-scale digital survey projects at home and abroad, such as LAMOST and SDSS, the data amount of stellar spectra has reached an unprecedented height. Such a large amount of data provide strong support for the discovery of special stars. Therefore, how to use these data to find the special, rare and even unknown types of stellar spectra rapidly and accurately is an important issue in astronomical research. Data mining is a technology that combines the pattern recognition, machine learning, statistical analysis and background knowledge of relevant experts to extract the potential unknown valuable information in the past. It has a natural advantage in dealing with big data. More and more data mining methods are applied to the survey data processing and analysis. At present, the data mining algorithms for special stars search mainly include stochastic forest, cluster analysis and outlier detection and so on. However, as the depth of the survey is expanded, the target of observation is getting darker and the signal-to-noise ratio of the observed spectrum accordingly lowers. There is a lot of useless information in the low signal-to-noise ratio spectrum, and the results obtained by directly analyzing and processing the relevant algorithms often have great deviations. Therefore, how to efficiently search out the special stellar spectra from a large number of low-SNR stellar data is an important issue nowadays. Due to the characteristics of the low-SNR stellar spectra themselves, a few studies are being done to search for the special stellar spectra. In order to solve this problem, a method based on principal component analysis (PCA) and the density peak approach is proposed to search special stellar spectra in low-S/N stellar data on the basis of careful study of the relevant methods. In this method, firstly, various types of high-SNR star spectra of O, B, A, F, G, K and M are selected, and then characteristic spectra are obtained by principal component analysis after wavelength unification and flux interpolation; secondly, the stellar spectra are reconstructed to obtain high-SNR spectra by using the first few characteristic spectra; finally, high-SNR spectra are clustered, and the outlier data is the special stellar spectrum. When clustering, this method uses a clustering method based on density peak for clustering and outlier mining with taking into account the characteristics of stellar spectral data itself. Experiments show that the proposed method can accurately search for a relatively smaller number of special stars in the low-SNR stellar data. At the same time, the proposed method can be applied to the spectral data analysis and mining of various galactic survey such as LAMOST and SDSS.