In this paper, a novel dimensionality reduction technique, named sparse representation preserving embedding (SRPE), is proposed by utilizing the sparse reconstruction weights and noise-removed data recovered from robust sparse representation. And a new dynamic process monitoring scheme is designed based on SRPE. Different from traditional manifold learning methods, which construct an adjacency graph from K-nearest neighbors or epsilon-ball method, the SRPE algorithm constructs the adjacency graph by solving a robust sparse representation problem through convex optimization. The delicate dynamic relationships between samples are well captured in the sparse reconstructive weights and the error-free data are recovered at the same time. By preserving the sparse weights through linear projection in the clean data space, SRPE is very efficient in detecting dynamic faults and very robust to outliers. Finally, through the case studies of a dynamic numerical example and the Tennessee Eastman (TE) benchmark problem, the superiority of SRPE is verified. (C) 2016 Elsevier Ltd. All rights reserved.