We purpose a method to refine the protein secondary structure prediction tools such as PSIPRED according the pattern analysis of 69534 protein data stored in PDB database. Besides, we use a bi-gram approach and classification to increase the accuracy of refinement. The method we purposed analyses the relation between twenty kinds of amino acid and protein secondary structures, and use the stasis results to refine the predicted secondary structures which are hard cases for PSIPRED. The experiment result reveals that our method is effective for all-beta entries.