Abstract
•A Reduced Rank kernel PCA and GLRT chart (IRR-KGLRT) for interval-valued data is developed.•The proposed technique is applied for fault detection of uncertain nonlinear process.•The proposed technique is validated evaluated using a Tennessee Eastman process.•The results show the monitoring effectiveness of the proposed method.
In advanced industrial fields such as chemical processes, faults must absolutely be detected; we cannot afford to operate with failing operative parts. It is therefore, necessary to take into consideration the uncertainties in order to establish a guaranteed failure detection. For this propose, several approaches have been introduced in the literature based on the Principal Components Analysis (PCA) approach for interval-valued data. However, these developed approaches treated only the linear process models. To overcome this drawback, recently, the researchers proposed to extend monitoring process based on Kernel Principal Components Analysis (KPCA) approach to the nonlinear uncertain case, in which two models are developed: model based on the lower bound (LB) and upper bound (UB) and model based on midpoints and radii.
Nevertheless, the number of variables that adequately represent the normal operating condition of system can be very large. To deal with both high calculation costs and the memory storage, in this paper, we suggest an improved kernel method, in order to ameliorate failure detection process for nonlinear uncertain system. The suggested method, fuse the benefit of Interval Kernel Generalized Likelihood Ratio Test (IKGLRT) index with the Exponentially Weighted Moving Average (EWMA) filter and the Interval Reduced Kernel Principal Component (IRR-KPCA).
The failure detection performance of the suggested method is valued using a Tennessee Eastman Process (TEP).