During the distribution system fault period, usually the explosive growth signals including fuzziness and randomness are too redundant to make right decision for the dispatcher. The volume of data with a few uncertainties overwhelms classic information systems in the distribution control center and exacerbates the existing knowledge acquisition process of expert systems. So intelligent methods must be developed to aid users in maintaining and using this abundance of information effectively. An important issue in distribution fault diagnosis system (DFDS) is to allow the discovered knowledge to be as close as possible to natural languages to satisfy user needs with tractability, and to offer DFDS robustness. At this junction, the paper describes a systematic approach for detecting superfluous data. The approach therefore could offer user both the opportunity to learn about the data and to validate the extracted knowledge. It is considered as a "white box" rather than a "black box" like in the case of neural network. The cloud theory is introduced and the mathematical description of cloud has effectively integrated the fuzziness and randomness of linguistic terms in a unified way. Based on it, a method of knowledge representation in DFDS is developed which bridges the gap between quantitative knowledge and qualitative knowledge. In relation to classical rough set, the cloud-rough method can deal with the uncertainty of the attribute and make a soft discretization for continuous ones (such as the current and the voltage). A novel approach, including discretization, attribute reduction, rule reliability computation and equipment reliability computation, is presented. The data redundancy is greatly reduced based on an integrated use of cloud theory and rough set theory. Illustrated with a power distribution DFDS shows the effectiveness and practicality of the proposed approach.