Nonlinear dynamic process monitoring using kernel CVA Raphael T. Samuel Supervisors - Dr Yi Cao, Dr Giorgos Kopanos Cranfield University UKACC PhD Presentation Showcase, London, 24 November 2015
UKACC PhD Presentation Showcase, London, 24 November 2015 Outline Background and motivation Methodology Application study Conclusion UKACC PhD Presentation Showcase, London, 24 November 2015
Background and motivation Effective monitoring to timely detect abnormal process conditions improve plant operation, safety, and product quality. Mathematical models used for this purpose are difficult to develop in large complex processes. Data-driven approaches are an alternative in such cases. Linear and static assumptions of traditional statistical algorithms and their extensions show limited performance in monitoring nonlinear dynamic processes. Kernel-based dynamic approaches have received limited attention in the literature despite their potential. UKACC PhD Presentation Showcase, London, 24 November 2015
Methodology Obtain past and future matrices Data Kernel Function Kernel Matrix Kernel PCs Calculate monitoring statistics and control limits Compute state variables & residuals Obtain past and future matrices Calculate transformation matrices UKACC PhD Presentation Showcase, London, 24 November 2015
UKACC PhD Presentation Showcase, London, 24 November 2015 Application study Monitoring results for Fault 3 (step change in D feed temp) Tennessee Eastman process (a) (b) (c) Control charts (a) DKPCA (b) CVA-KDE (c ) KCVA UKACC PhD Presentation Showcase, London, 24 November 2015
UKACC PhD Presentation Showcase, London, 24 November 2015 Conclusion A novel CVA algorithm using kernel PCs as collected data has been developed. The approached out-performed the DKPCA and CVA with KDE. Performing CVA on a reduced kernel component space prevents singularity problems and precludes the need for regularisation. Study on adapting the developed method to monitor processes with multiple operating modes is ongoing. UKACC PhD Presentation Showcase, London, 24 November 2015