Nonlinear dynamic process monitoring using kernel methods

The application of kernel methods in process monitoring is well established. How- ever, there is need to extend existing techniques using novel implementation strate- gies in order to improve process monitoring performance. For example, process monitoring using kernel principal component analysis (KPCA) have been reported. Nevertheless, the e ect of combining kernel density estimation (KDE)-based control limits with KPCA for nonlinear process monitoring has not been adequately investi- gated and documented. Therefore, process monitoring using KPCA and KDE-based control limits is carried out in this work. A new KPCA-KDE fault identi cation technique is also proposed. Furthermore, most process systems are complex and data collected from them have more than one characteristic. Therefore, three techniques are developed in this work to capture more than one process behaviour. These include the linear latent variable-CVA (LLV-CVA), kernel CVA using QR decomposition (KCVA-QRD) and kernel latent variable-CVA (KLV-CVA). LLV-CVA captures both linear and dynamic relations in the process variables. On the other hand, KCVA-QRD and KLV-CVA account for both nonlinearity and pro- cess dynamics. The CVA with kernel density estimation (CVA-KDE) technique reported does not address the nonlinear problem directly while the regular kernel CVA approach require regularisation of the constructed kernel data to avoid com- putational instability. However, this compromises process monitoring performance. The results of the work showed that KPCA-KDE is more robust and detected faults higher and earlier than the KPCA technique based on Gaussian assumption of pro- cess data. The nonlinear dynamic methods proposed also performed better than the afore-mentioned existing techniques without employing the ridge-type regulari- sation.