At present, the method of learning Bayesian network structure with missing data is mainly based on the search and scoring method combined with EM algorithm. The algorithm has low efficiency and easily gets into local optimal structure. In this paper, a new method of learning Bayesian network structure with missing data is presented. First, unobserved data are randomly initialized. As a result, a complete data set is got. Based on the complete data set, the maximum likelihood tree is built as an initialization Bayesian network structure. Second, unobserved data are reassigned by combining Bayesian network with Gibbs sampling. Third, on the basis of the new complete data set, the Bayesian network structure is regulated based on the basic dependency relationship between variables and dependency analysis method. Finally, the second and third steps are iterated until the structure goes stable. This method can avoide the exponential complexity of standard Gibbs sampling and the main problems in the existing algorithm. It provides an effective and applicable method for uncertain knowledge representation, inference, and reasoning with missing data.