Abstract: A single pulse search on the data generated by the large-scale radio astronomy survey project will generate hundreds of millions of pulsar candidates. This paper processed 320 000 ultra-wideband data files of the CRAFTS (The Commensal Radio Astronomy FAST Survey) survey, and obtained 50 000 pulsar candidates. In order to solve the problem that only 1.5% of the pulsar candidates in the results of CRAFTS single pulse search experiment really have pulsar signal, three features that distinguish pulsar signals from interference are proposed and applied to the single pulse sifting process. Experiments show that the improved single pulse sifting method will reduce the number of pulsar candidates that need to be manually checked by 20%, which improves the single pulse search efficiency of CRAFTS survey data to a certain extent.