Monitoring of chemical processes using improved multiscale KPCA

Statistical process monitoring charts are critical in ensuring safety for many chemical processes. Principal Component Analysis (PCA) is often used, due to its computational simplicity. However, many chemical processes may be inherently nonlinear, and this degrades the performance of the linear PCA method. Kernel Principal Component Analysis (KPCA) is an extension of the conventional PCA chart, which can help deal with nonlinearity in a given process. Additionally, PCA assumes that process data are Gaussian and uncorrelated, and only contain a moderate level of noise. These assumptions do not usually hold in practice. Multiscale wavelet-based data representation produces wavelet coefficients that possess characteristics that are able to handle violations in these assumptions. A multiscale kernel principal component analysis (MSKPCA) method has already been developed to tackle all of these issues, but it usually provides a high false alarm rate. In this paper, an improved MKSPCA chart is developed in order to deal with the false alarm rate issue, by smoothening the detection statistic using a mean filter. The advantages brought forward by the improved method are demonstrated through a simulated example in which the developed fault detection method is used to monitor a continuous stirred tank reactor (CSTR). The results clearly show that the improved MSKPCA method provides lower missed detection and false alarm rates as well as ARL1 values compared to those provided by the conventional methods.