Batch processes with multiple phases are commonly found in process industries. Process dynamics and correlations among variables also tend to change with the transitions across such phases. Traditional approaches where the model is constructed from data representing the whole batch process would not be sufficient to capture the varying process dynamics and correlation structure. Different ways of phase segmentation and modeling strategies have been previously reported that account for the multi-phase characteristics of such processes. For a given process dynamic or a particular phase in a multi-phase process, a Principal Component Analysis (PCA) model can project the maximum deviation with small number of principal components, and represent a certain percentage of the deviation with fixed number of principal components. As the process dynamics change, the percentage represented by the fixed number of PCs also changes if a single PCA model is applied. In this paper, a new phase identification method is proposed based on the change of the first cumulative contribution between different PCA models. Every phase is modeled separately based on the phase identification. The method is applied to fault detection in the fed-batch penicillin cultivation process. The results show that the method can better capture the process dynamics in different phases and detect process upset in an early stage. (C) 2010 Elsevier Ltd. All rights reserved.