Fault diagnosis based on multi-source information fusion can improve equipment diagnosis rate. However, such methods need to obtain a variety of fault feature data obtained from different sensors. If the data of all features are used for diagnosis, then computation burden will be too large to realize real-time diagnosis. Based on evidence theory and rough sets, a belief interval-based attribute reduction theorem and the corresponding fault feature(attribute) reduction method are presented so as to use the reduced important features to make fast diagnosis decision. In detail, the random fuzzy variable and K-means are used to discretize the data of features. The core attribute set can be achieved by compressed binary matrix. Thus, the belief interval size of attribute is taken as the criterion of attribute selection in the iterative reduction process in which the important attributes are added to the core. Finally, the reduction result is obtained. In the experiment of features fusion diagnosis of motor rotor, the effectiveness of the proposed method is illustrated by comparing with the classical rough set reduction methods.