Statistical process monitoring with measured data corrupted by noise and gross error

Principal component analysis (PCA) is an efficient method to extract relationships between correlated variables and thus has been widely applied to the multivariate statistical process monitoring. However, the validity of the PCA is highly depending on the quality of measured process data which is usually contaminated by noises and gross errors in practice. In this paper, an improved PCA is presented to minimize the influences of corrupted data. The original measured data is first processed using conventional PCA to partially eliminate noises by abandoning principal components with small eigenvalues. Then the retained principal components are decomposed and rectified online by boundary-corrected wavelets, combined with techniques of shift-invariant transform and median filtering. The monitoring results of a simulated binary distillation column show that the proposed method has superior performance and more robust to noises and gross errors than conventional methods