A comparison between PCA and PLS-based process system monitoring

J.L. GODOY; J.R. VEGA; ALEJANDRO H. GONZÁLEZ; JACINTO L. MARCHETTI

Oro verde

Congreso; XIV Reunión de Trabajo en Procesamiento de la Información y Control (RPIC 2011); 2011

Monitoring of a stochastic process that includes collinearities between input and output variables can be performed through a model based on Principal Component Analysis (PCA), which makes no difference between outputs and inputs. In contrast to PCA, a Partial Least Square Regression (PLSR) model is closer to the intrinsic system structure because it allows eliminating some undesired input variables from the original data sets (e.g., those interfering the regression model) [1]. In this work, a different PLSR model is computed by simultaneously deflating the input and output data matrices, by means of the classical PLS-NIPALS algorithm [2]. This procedure gives better results for multivariate prediction and for process monitoring than others alternative PLS algorithms, such as the one presented in Gang et al. (2010) [3]. Besides, the simultaneous deflation on input and output data matrices allows the detection of predictor variables that can play an interfering effect on the system.