Adaptive Soft Sensors for Non-Gaussian Chemical Process Plant Data Based on Locally Weighted Partial Least Square

Abstract

This thesis proposes an improved algorithm attributed to its abilities to deal with non-Gaussian distributed and nonlinear data and missing measurements. It was formulated through a modification on locally weighted partial least square by incorporating an ensemble method, Kernel function and independent component analysis and expectation maximisation algorithms. The algorithm was then tested using process data generated from six simulated plants. Simulation results indicate superiority of this algorithm compared to the existing algorithms.