Abstract: A new operation stage separation method is proposed for the substage separation of multistage batch processes. Based on the changes in principle component number, loading matrixes, and principle component matrixes, which reveal evolvement of the underlying process behavior, a three-step substage separation method is realized. First, rough separation of operation substage is executed by the difference of principle component number. To reflect objectively the similarity of the loading matrixes and the similarity of the principle component matrixes, two improved similarity distances, based on weighted cosine of the angle between loading vectors and weighted Euclidean distances, are introduced respectively. According to the criterion of minimum similarity distances, time-slice matrixes are sorted using the rival penalized competitive learning algorithm to realize separation of operation substage more particularly for batch processes. The effectiveness of the proposed method is illustrated by applying it to the MPCA modeling and on-line monitoring of the injection mold process.