A Bayesian network (BN) is a probability graph model which can be used to express the causal relationship between variables. Due to the solution space growing exponentially with the increasing number of nodes, Bayesian network structure learning (BNSL) has become an NP-hard problem. The genetic algorithm (GA) can efficiently search more possible structural combinations in the search space, and it has achieved many satisfying results in recent years in the BNSL problem, but there are still some problems, such as premature convergence and low accuracy. To solve these problems, we propose a diversity and mutual information mixed guidance GA for Bayesian network structure learning (DM-GABN). In the cycle removal phase, this paper uses flip edges instead of deleting to preserve more diverse genotypes. In the selection phase, we remove older individuals and maintain a reasonable population age structure to improve population diversity. We introduce the concept of biological genotype frequency into the crossover strategy, and protect the structure with low genotype frequency. In addition, we use mutual information to guide the search process. Experimental results on ten widely used benchmark networks and compared with ten BNSL algorithms including the most advanced methods show the proposed algorithm DM-GABN outperforms other algorithms regarding structural accuracy and convergence speed.