Realtime Process Monitoring and Fault Propagation Path Estimation using Principal Component Analysis and Granger Causality

Abstract

Modern industrial process is a complex device industry consisting of a combination of numerous unit processes. Numerous process parameters such as flow rate, temperature, pressure, concentration and composition have strong linear or nonlinear correlation. Since improvement of computing power and process control systems in industrial processes, several board operator and field operator can manage huge amounts of data and whole process information from industrial plant. However, the number of processes and devices to be handled by a single operator will increase, and operators meets a limitation of cognitive ability due to flood of information, causing problems such as process malfunction or instrumental failure. To solve this problem, we propose a PCA modeling procedures that aims to improve monitoring performance by variable selection, removing noise, operation mode classification and mode change detection. Fault diagnosis and causal analysis is also introduced. We calculated the causal relationship matrix between the process variables and find out the root cause of the unexpected process changes. The proposed approach was applied and validated to LNG plant located in Incheon and plasma condition monitoring in plasma etcher. Chapter 2 discusses the application methodologies of signal processing to eliminate noises from OES signal and multivariate statistical techniques to improve monitoring sensitivity. Among the plasma sensors, optical emission spectroscopy (OES) has been widely utilized and its high dimensionality has required multivariate analysis (MVA) techniques such as principal component analysis (PCA). PCA, however, might devaluate physical meaning of target process during its statistical calculation. In addition, inherent noise from charge coupled devices (CCD) array in OES might deteriorate PCA model performance. Therefore, it is desirable to pre-select physically important variables and to filter out noisy signals before modeling OES based plasma data. For these purposes, this chapter introduces a peak wavelength selection algorithm for selecting physically meaningful wavelength in plasma and discrete wavelet transform (DWT) for filtering out noisy signals from a CCD array. The effectiveness of the PCA model introduced in this paper is verified by comparing fault detection capabilities of conventional PCA model under the various source power or pressure faulty situations in a capacitively coupled plasma etcher. The PCA model introduced in this chapter successively detect even extremely small variation such as 0.67% of source power change even though the conventional PCA model fails to detect all of the faulty situations under the tests. Chapter 3 discusses the application methodology of operation mode identification and multimode PCA to improve the performance of LNG mixed refrigeration (MR) process and prevent process shutdown. LNG MR process is usually used for liquefying natural gas. The compressors for refrigerant compression are operated with the high-speed rotating parts to create a high-pressure. However, any malfunction in the compressors can lead to significant process downtime, catastrophic damage to equipment and potential safety consequences. The existing methodology assumes that the process has a single mode of operation, which makes it difficult to distinguish between a malfunction of the process and a change in mode of operation. Therefore, k-nearest neighbor algorithm (k-NN) is employed to classify the operation modes, which is integrated into multi-mode principal component analysis (MPCA) for process monitoring and fault detection. When the fault detection performance is evaluated with real LNG MR operation data, the proposed methodology shows more accurate and early detection capability than conventional PCA. Chapter 4 discusses PCA based fault amplification algorithm to detect both the root cause of fault and the fault propagation path in the system. The developed algorithm project the samples on the residual subspace (RS) to determine the disturbance propagation path. Usually, the RS of the fault data is superimposed with the normal process variations which should be minimized to amplify the fault magnitude. The RS containing amplified fault is then converted into the co-variance matrix followed by singular value decomposition (SVD) analysis which in turn generates the fault direction matrix corresponding to the largest eigenvalue. The fault variables are then re-arranged according to their magnitude of contribution towards a fault which in turn represents the fault propagation path using an absolute descending order functions. Moreover, the multivariate granger causality (MVGC) algorithm is used to analyze the causal relationship among the variables obtained from the developed algorithm. Both the methodologies are tested on the LNG fractionation process train and distillation column operation where some fault case scenarios are assumed to estimate the fault directions. It is observed that the hierarchy of variables obtained from fault propagation path algorithm are in good agreement with the MVGC algorithm. Therefore, fault amplification methodology can be used in industrial systems for identifying the root cause of fault as well as the fault propagation path. The application results show that the proposed multivariate statistical method can improve productivity and safety by providing useful information for process monitoring and fault diagnosis in various processes with distributed control system.