This paper proposes a definition for the fuzzy entropy of clustered result , i. e. , the fuzziness measure of a fuzzy set, and based on this definition, the definition for a measure of the distribution divergence of the sampled data vectors is also given in this paper for the first time. The divergence definition offers an ideal and objective criterion about whether there exist heavily interfered data vectors in the sliding data window for the Adaptive Windowing Filter. Thus, the noise discrimination ability has been much improved, and this increases the robustness of the fault detection algorithm to the strong noises and the sensitivity to subtle fault transients. In addition, the directional similarity between the two clustered centroid vectors (corresponding to the sliding data window) has been employed to detect faults and distinguish the intermittent faults from the failure. The practical testing data with strong noises have been used to verify its efficiency and objectivity in the fault detection and discrimination from the noised data vectors.