The original rough set model is concerned primarily with the approximation of sets described by single binary relation on the universe. In the view of granular computing, classical rough set theory is researched by single granulation. The article extends the rough set model based on tolerance relation to incomplete rough set model based on multi-granulations, where the set approximations are defined by using multi tolerance relations on the universe. Its some basic mathematical properties are investigated as well. It is shown that some properties of rough set model based on tolerance relation are special instances of this new model, the approximation measure of a target concept described by using multi-granulations is always better than by using single granulation, which is suitable for describing more accurately the concept and solving problem according to user requirement in incomplete information systems. Keywords: Rough set; incomplete information systems; Multi-granulations; Approximation measure.