Abstract: In order to deal with the uncertainties in feature extraction and decision-making, an information fusion algorithm of fault diagnosis is presented based on random set metrics of fuzzy features and evidence reasoning. Firstly, membership functions are used to describe the fault templates in model database and features extracted from sensor observations. Secondly, a random sets model of fuzzy information is introduced to give a likelihood function, which can be transformed into a Basic Probability Assignment (BPA) function. A BPA numerically shows the support degree of the hypotheses that the machine has certain faults under the fuzzy features. The proposed fuzzy feature is not extracted from single observation but from continuous observations. The fusion diagnosis results based on this proposed feature are more accurate than that based on traditional single observation feature. Finally, the diagnosis results of machine rotor show that the proposed method can enhance diagnostic accuracy and reliability.