Status : Verified
| Personal Name | Madayag, Jan Vincent M. |
|---|---|
| Resource Title | Multi-kernel canonical variate analysis with optimal kernel designs for nonlinear system identification |
| Date Issued | August 2022 |
| Abstract | With the ever-increasing demand for safe and efficient industrial plant operations, system identification has become a prevalent issue, especially since the applicability of linear models is limited to only a small range of operating conditions; hence, this study focuses on nonlinear models. Kernel Canonical Variate Analysis (KCVA) is a popular nonlinear system identification algorithm where nonlinearities are learned by approximating the covariance of the data with a kernel function, typically the radial basis function (RBF) kernel. However, there are currently limited efforts that study the influence of multiple rather than single kernel designs in KCVA. In this study, three multiple kernel CVA (KCVA-MKL) models were developed to investigate its training and predictive performance via a rigorous optimization process using population-based search methods i.e., Particle Swarm Optimization (PSO) and Genetic Algorithm (GA); surrogate-assisted search method like Bayesian Optimization (BO); and Random Search (RS) method for a simulation dataset (Newell-Lee evaporator system) and a real-world plant dataset (Industrial dryer by Cambridge Control Ltd.). Overall, the KCVA-MKL models displayed enhancement in terms of their predictive performance against the base model MKCVA. Additionally, kernel designs optimized by PSO and BO were shown to exhibit better performance in terms of predictive accuracy than RS and GA. |
| Degree Course | MS Chemical Engineering |
| Language | English |
| Keyword | multiple-kernel learning; hyper-parameter optimization; nonlinear system identification; Newell-Lee evaporator system; industrial dryer |
| Material Type | Thesis/Dissertation |
Preliminary Pages
277.86 Kb
Category : I - Has patentable or registrable invention of creation.
Access Permission : Limited Access