A monitoring method is proposed for batch processes, starting with limited reference batches and then updating model data structure with accumulation of new normal batches. On the basis of analysis of the unfolded matrix ((time x batch) x variable), a generalized moving window method is introduced for exploring local covariance structure, phase division, and the development of monitoring models. Then, an adaptive updating algorithm, based on phase-specific dissimilarity analysis, is developed for model updating to accommodate additional information from the accumulation of new batch data and to explore time-varying behaviors from batch to batch. The proposed method is illustrated with two processes, an industrial scale experimental injection molding and a simulated fed-batch penicillin fermentation. Both results show that the proposed method is effective.