Design of online soft sensors based on combined adaptive PCA and DMLP neural networks

Abstract

An accurate on-line measurement of important quality variables is essential for successful monitoring and controlling of chemical process. However, these variables are usually difficult to measure on-line due to the limitations such as the time delay, high cost and reliability considerations. To overcome this problem, two online soft sensors are proposed based upon a combined adaptive principal component analysis (PCA) and a dynamic multi-layered perceptron (DMLP) artificial neural network (ANN). For this purpose, a recursive PCA and a PCA based on a sliding window are presented to adaptively extract the inherent features inside the measurements with high dimensions. The extracted low-dimension features are then used recursively as the main inputs to the DMLP networks. The developed online soft sensors are finally tested on a highly nonlinear distillation column benchmark problem to illustrate their comparative performances. The simulation results demonstrate the superiority of the soft sensor based on the recursive PCA and the DMLP network.