The non-negative matrix factorization model is a common data dimensionality reduction method. In the existing non-negative matrix factorization algorithm for clustering research， each category is generally represented by only one or more designated center points. However， this type of representation often fails to accurately describe the characteristics and structure of its category， which affects the clustering performance. In order to solve this problem， we proposed the peaks non-negative matrix factorization （PNMF） algorithm. The algorithm first finds multiple density peak points for the dataset， constructs a bipartite graph of the density peak points and sample points， then uses the bipartite graph to complete the clustering. In addition， the algorithm introduces a manifold regularization term to make full use of the manifold structure information between the data， and gives the iterative update rules of the algorithm. Sufficient experiments on real-world datasets demonstrate that the proposed method can effectively utilize the structural information of data and improve the clustering performance.