Kernel Canonical Variate Dissimilarity Analysis for Fault Detection

Almost all of the existing fault detection statistics only consider the input data, but neglect the information contained in the out put data. In this paper, a novel method based on KCVA, called kernel canonical variate dissimilarity analysis (KCVDA) is presented for nonlinear dynamic process monitoring The new statistic proposed by the KCVDA method can take into account both input and output data. In addition. the KCVDA method absorbs the merits of kernel method in dealing with nonlinearity and fully utilizes the advantage of CVA in tackling the dynamic issue. The proposed dissimilarity -based statistic is implemented under the Tennessee Eastman process, and is compared with the T-s(2), T-r(2) statistics of KCVA and the T-2,Q statistics of KPCA to demonstrate its effectiveness in fault detection. The outcome illustrates the accuracy of the proposed dissimilarity-based statistic.