To solve the drawbacks of constraint-based method for learning Bayesian networks(BN) and the unreliability of
the conditional independence(CI) tests as the conditioning sets become too large, this paper proposes a structural learning algorithm based on maximal prime decomposition(MPD). Firstly, MPD technique is used to transform the moral graph of BN into its sub-graphs. Then, only zero order and first order CI tests are used to identify V-structures in part of sub-graphs and takes scoring function searches to optimize local structure, so that the number of conditional independence tests can be decreased. Redundancy tests can be avoided and the time performance can be greatly enhanced. Finally, theoretical and experimental results show that the new algorithm is effective and reasonable.