Improved Kernel Canonical Variate Analysis for Process Monitoring

This paper proposes a kernel canonical variate analysis (KCVA) approach for process fault detection. The technique employs the kernel principle to map the original process observations to a high dimensional feature space on which canonical variate analysis is performed. The aim is to obtain an effective monitoring technique that accounts for non-linearity and process dynamics simultaneously. The kernel principle accounts for non-linearity while the CVA accounts for serial correlations widely encountered in dynamic processes. The kernel CVA algorithm proposed in this work is based on QR decomposition in order to avoid singularity problems associated with kernel matrices which require a regularisation step. The technique is evaluated using the Tennessee Eastman Challenge process. Tests show the effectiveness of the proposed kernel CVA approach.