For the problem of irrational initial structure of belief rule base (BRB), the existing solving approaches still have deficiencies in many aspects such as non-repeatability, the completeness of data and the constraint with the associated level utility. In view of this, through theoretical analysis and experimental verification for parameter learning approaches of BRB, this paper summarizes that the irrational structure of BRB may lead to the problem of over-complete or incomplete structure. This paper takes the application of DBSCAN algorithm and error analysis to the existing parameter learning methods, and brings forth the structure learning approach for best decision structure. The experiments verify the new approach under over-complete and incomplete structures of BRB, and make a comparative analysis of the changes of error when the structure is varying. The results show the feasibility and effectiveness of the proposed approach.