Skip to Main Content
A new PCA adaptive algorithm is introduced, utilizing a rough fuzzy cluster-based granulation scheme for fault detection and diagnosis purposes. This granulated cluster-based algorithm can be used for segmentation of multivariate time series and initialization of other partitioning clustering methods that need to have good initialization parameters. The proposed algorithm is suitable for mining data sets, which are large both in dimension and size, in case generation. It utilizes Principal Component Analysis (PCA) specification and an innovative granular computing method for detection of changes in the hidden structure of multivariate time series data in a bottom up cluster merging manner. Rough set theory is used for feature extraction and solving superfluous attributes issue. Upper and lower approximations of rough set is calculated based on fuzzy membership functions. These approximations are updated after each granulation stage. Features of a pattern can hence be described in terms of three fuzzy membership values in the linguistic property sets as normal (N), abnormal (A) and ambiguity (a). The algorithm has been tested on an artificial case study and Tennessee Eastman (TE) benchmark process plant. The resulting performances show the successful and promising capabilities of the proposed algorithm.