Improved PCA methods for process disturbance and failure identification

Principal component analysis (PCA) is a powerful technique for constructing reduced-order models based on process measurements, obtained by the rotati on of the measurement space. These models can be subsequently utilized for chemical-process monitoring, particularly for disturbance and failure diagn osis. Since the standard PCA procedure does not account for the time-depend ent relationships among the process variables, this leads to poorer disturb ance isolation capability in dynamic applications. A simple idea, in which the last s PCA scores are recursively summed and used to construct descript ive statistics for process monitoring, is presented. Analytically, it is sh own that the disturbance resolution afforded is enhanced as a result. Resol ution is improved further through the use of an algorithm that enhances the correlations between the input and output variables through optimal time s hifting. An overall strategy for on-line monitoring developed includes dist urbance identification through mapping. The approach is demonstrated by two industrially relevant case studies.