Nonlinear process monitoring based on kernel dissimilarity analysis

To overcome the disadvantage of linear dissimilarity analysis (DISSIM) when monitoring nonlinear processes, a kernel dissimilarity analysis algorithm, termed KDISSIM here, is presented, which is the nonlinear version of DISSIM algorithm. A kernel dissimilarity index is introduced to quantitatively evaluate the differences between nonlinear data distribution structures, which can reflect the changes of nonlinear process correlations and operating conditions. In KDISSIM algorithm, the input space is first nonlinearly mapped into a high-dimensional feature space, where the initial nonlinear correlations are changed into linear ones. Then the process operating condition can be effectively tracked by investigating the linear data distributions in the feature space. The idea and effectiveness of the proposed algorithm are illustrated with respect to the simulated data collected from one typical nonlinear numerical process and the well-known Tennessee Eastman benchmark chemical process. Both the results show that the proposed method works well to capture the underlying nonlinear process correlations thus providing a feasible and promising solution for nonlinear process monitoring.