The stability of a rolling mill plays a decisive role in the quality of rolled products, but signals monitoring the status of the rolling mill have very strong coupling. In this case, to separate the complex signals into independent status signals of each mill is essential to realize the status monitoring and fault diagnosis for rolling mill. In this paper, a technical scheme for separating fault signals of continuous rolling mills based on sparse features is proposed, and simulation and field verification are carried out. First, the weak impulses of the mixed signals are extracted through the method of sparse decomposition based on spectrum segmentation; second, the selected atoms are sorted out according to certain rules to get the sparse matrix of mixed signals; then, the selected atoms are clustered according to the similarity of sparse atoms to determine the number of blind sources; finally, the mixing matrix is estimated according to the sparse matrix and the number of blind sources to realize the blind source separation of mixed signals.