At present, more than 7.6 million celestial spectra has been obtained by the LAMOST spectrum sky surveys, and the processing of low SNR spectra has been regarded as a difficulty in this domain. In this paper, a new method was proposed to deal with repeated observation spectra. The process of the method was: for every group of repetitive observation spectra, selecting that whose difference of red shift was within a certain range, and then an optimal stack algorithm based on SNR weighting was run to increase the SNR. The results of the processing of all repetitive observation spectra in LAMOST DR4 showed that this method was very effective for increasing the SNR for low SNR repetitive observation spectra. This method made the SNR of the 7 571 sets of stellar spectra reach the standard of parameter measurement; The SNR of the 3 357 quasar and Galaxy spectra improved at an average improvement rate of 56.38%; And 43 021 binary candidate candidates were obtained.