In this paper, a novel Bayesian Network (BN) learning method is proposed, in which Genetic Algorithm (GA)and structure-parameter restrictions are combined to optimize the BN's structure and parameters simultaneously. We firstlytransferred the domain knowledge into structure and parameter restrictions, which can be considered ‘hard’ restrictions in the sense that they are assumed to be true forthe BN representing the domain of knowledge. In order to use these restrictions in conjunctionwith Genetic Algorithm for learning Bayesian networks, gene restrictions table is designed to kick out the unsatisfied candidate genes, so that more accurate results and less convergence times can be achieved. Experiments show that the proposedalgorithm can contribute to the global optimum of the system, and can improve the value of the evaluation function more than 15% while keeping the same detection rate.