Aiming at the weakness that the K-means algorithm cannot effectively suppress the noise attributes and realize irregular shape clustering on high-dimensional data, an improved K-means clustering algorithm based on feature selection and removal on target point is proposed. In the improved K-means algorithm, the Minkowski metric is adopted as the evaluation of distance for the classification of the target point. The weighting adjustment parameter a is added and the weighting coefficient α is reset for feature selection and removal, which can reduce the effect of non-clustering index noise features. The UCI real datasets and artificial datasets are used for clustering analysis in the algorithm validation experiment. And the effectiveness of suppressing the noise features is validated. Compared with the WK-means and iMWK-means algorithms in the validation experiment, the applicability of feature selection in clustering learning process is analyzed. At the same time, the optimal distance coefficient beta and the weighting coefficient α are found.