The Dempster combination rule relies on fundamental assumption that evidences are mutually independent. However, this idealized prerequisite is frequently challenging to fulfill in real-world scenarios. How to measure belief mutual information to investigate dependence among evidence sources is a critical and significant issue. Inspired by fuzzy mutual information based on D–S evidence theory, a method of approximate belief mutual information (BMI) is developed to partially represent the dependence among the sources of evidence. In the BMI model, an interval correlation method considering the basic probability assignment (BPA) and its negation are proposed. Several examples are used to illustrate the availability of the BMI, which shows the developed BMI has better fault tolerance. Finally, the proposed method obtains higher recognition accuracy in pattern recognition tasks. The proposed method can maintain stable and reliable recognition results even when the data is disturbed, verifying the effectiveness of the proposed method.