Rough set theory is an efficient mathematical theory for data reduction and knowledge discovery of various fields. However, classical rough set theory is not applicable for knowledge induction of incomplete information systems. In this paper, a concept of interval granule is presented. Based on this concept, the hierachical structure of knowledge granularity and approximation of rough sets in incomplete information systems are studied, and related properties are given. An example show that the interval granule have better results than existing models for knowledge induction and approximation.