Abstract: The fuzzy c-means algorithm has several limitations: too sensitive to choose initial class center of divisions， too sensitive to noises and outliers， easy to be effected by data distributions. The improved algorithm uses the data field according to the theory of fields in physics, uses potential values of fault points in the data field to identify noise point and determine the initial class center, uses multicenters clustering algorithm that big cluster is cut into several small clusters, and then takes the separation measures as evaluation function to merge small clusters. Experiments show that the improved fuzzy c-means algorithm could make up the defects of fuzzy c-means algorithm, and be well suited to the non-uniform subject distributions．

Key words: clustering, fuzzy c-means algorithm, data field, initial cluster centers