Although decentralized modeling has been widely employed in monitoring large-scale processes, the dynamic property in process data is rarely investigated. Meanwhile, fault diagnosis in a way similar to pattern recognition is still challenging. To handle these issues, a dynamic decentralized fault detection and diagnosis framework based on dynamic feature selection and informative fault pattern (IFP) dissimilarity is presented. The proposed method accounts explicitly for the dynamic property in process data, while handling the challenging fault diagnosis task at the same time. First, a dynamic feature selection method is proposed to interpret the dynamic relations through characterizing the auto- and cross-correlation for each variable individually. As a consequence, multiblocks are derived for decentralized modeling and monitoring purposes. Second, a novel classification-based fault diagnosis approach on the basis of the dissimilarity analysis and filtered monitoring statistics (termed as IFP) is formulated. This sort of method can be feasible even in a condition that the training samples for each reference fault class are insufficient and also overlapped with each other. Finally, the salient performance in terms of fault detection and recognition capabilities that can be achieved is validated by two simulated examples. The comparisons clearly demonstrate the superiority and feasibility of the proposed monitoring scheme.