For non-convex, sparse and fuzzy nonlinear separable data, the clustering effect of K-means clustering is not good. Therefore, by introducing granule computing theory and using neighborhood granulation technology, a neighborhood granule K-means clustering method is proposed. The sample uses neighborhood granulation technology to construct neighborhood granules on a single feature, and to form neighborhood granule vectors on multiple features. By defining the size, measurement and operation rules of neighborhood granules and neighborhood granule vectors, two kinds of neighborhood granule distance measurements are proposed, and the axiomatic proof of the proposed neighborhood granule distance measurement is carried out. Finally, several UCI data sets are used to carry out experiments, the K-means clustering algorithm is combined with two neighborhood granule distance measurements respectively. It is compared with the classical clustering algorithm in two aspects of neighborhood parameters and distance measurement. The results show that the proposed neighborhood granular K-means clustering method is feasible and effective.