Improved Statistical Method Based Exponentially Weighted GLRT Chart and Its Application to Fault Detection

This paper deals with fault detection (FD) of chemical processes. Our previous study [1] has proved the effectiveness of multiscale principal component analysis (MSPCA)-based Moving Window (MW)-Generalized Likelihood Ratio Test (GLRT) to detect faults by maximizing the detection probability for a particular false alarm rate with different values of windows. However, the conventional PCA method is not suitable in nonlinear processes. In fact, this lack affects the monitoring system. To address this problem, we propose, first, to use multistage kernel PCA (MSKPCA) technique to extract the deterministic features and compute the principal components (PCs) in the original space. Second, integrate exponentially weighted moving average (EWMA), that has shown better abilities to reduce the false alarm rates and enhance the (FD) performances. Therefore, this work focuses on extending MSKPCA, and developing a MSKPCA-based EWMA-GLRT technique in order to improve the (FD) performance. The performances of the MSKPCA -based EWMA- GLRT are illustrated using Tennessee Eastman benchmark process.