Nonlinear process fault detection and identification using kernel PCA and kernel density estimation

Kernel principal component analysis (KPCA) is an effective and efficient technique for monitoring nonlinear processes. However, associating it with upper control limits (UCLs) based on the Gaussian distribution can deteriorate its performance. In this paper, the kernel density estimation (KDE) technique was used to estimate UCLs for KPCA-based nonlinear process monitoring. The monitoring performance of the resulting KPCA–KDE approach was then compared with KPCA, whose UCLs were based on the Gaussian distribution. Tests on the Tennessee Eastman process show that KPCA–KDE is more robust and provide better overall performance than KPCA with Gaussian assumption-based UCLs in both sensitivity and detection time. An efficient KPCA-KDE-based fault identification approach using complex step differentiation is also proposed.