Bayesian network is an uncertainty inference network based on probability. Its structure learning is one of the main research techniques in the field of data mining and knowledge discovering, while constructing Bayesian network structures from data is NP hard. According to the information theory and conditional independence test, a new algorithm is presented for the construction of optimal Bayesian network structure, and numerical experiments show that the structure with highest degree of data matching can be much faster determined by the new algorithm, thus the study of Bayesian network structures becomes more efficient.