An improved genetic algorithm (IGA) is proposed in this paper for structure learning of Bayesian Network (BN). Compared with the traditional GA, two new operators named optimized mutation and illegal figure modification are proposed in the improved GA, which aim to solve the BN structure learning problem. The two new operators can simultaneously maintain the diversity and correctness of BN structure learning as well as the algorithm convergence speed of searching the global optimal network structure. In simulation, compared with the traditional algorithms such as K2 algorithm, GS/GES algorithms, normal GA, PSO, etc., the proposed GA shows better performance in global searching and convergence speed.