Data-driven Performance Monitoring And Diagnosis Of Multivariate Control Systems

Nowadays, industrial process control system become more and more complex and highly sophisticated. Their stability and the reliability of instruments are also enhanced by the progress of advanced technology. But on the other hand, they become frailer than the simple but robust ones. For one thing, long term smooth running have to be guaranteed for return of heavy investment. For the other, large-scale plants have more chance to suffer faults or disturbances and hence fail. Usually, for large-scale plant, faults can propagate through the plant and affect a large number of variables which may lead serious disaster.For performance monitoring of multivariate systems, multivariate statistical process control (MSPC) has been well developed and lots of algorithms such as PC A, PLS and ICA have been sufficiently studied. In the past four decades, large literature for linear and nonlinear system emerged in the discipline of fault detection and diagnosis (FDD). But for plant-wide level, especially analysis of fault propagation and root cause diagnosis, it is still open for researchers. Starting with fault detection using PCA, in this thesis we propose a comprehensive framework of performance monitoring and diagnosis. Several key issues are studied and the main contributions are as follows:1. Among the methods of MSPC, PCA attracts much attention for its conceptually simple and dealing with recorded data directly. But it is only appropriate for identical and independent (i.i.d.) Gaussian variables which is a very restrict assumption in practical application. Furthermore, it is proved that it may not detect incipient changes in the variable covariance structure nor changes in the geometry of the underlying variable decomposition. Therefore it can not be used as a early warning tool of process. An improved PCA algorithm is proposed by integrating statistical local approach into PCA to settle this problem. A new monitoring statistic for the residual space is developed which can successfully deal with the case of zero eigenvalues. The new statistic can improve the robustness of numerical computation and decrease false alarm rate. The statistical properties of the residuals and also the monitoring statistics are well analyzed. It is proved that the improved PCA algorithm is more sensitive for different kinds of fault than the proper PCA. A new performance assessment method is proposed based on the analysis of improved residual directly. This simple but effective method can assess the performance trend in every eigen-directions.2. Method for identifying the propagation path of plant-wide faults is proposed. The problem is solved by two methods. For variables which can be analyzed through time series analysis, an ordinal point of view is introduced. A certian length of data points are coded as a symbol. By sliding the data vector, a new symbolic sequence corresponding to the original time series is derived. The statistical and relative analysis is based on the new symbolic sequence. Causality is a well established conception in physics and economics by Wiener and Granger. Its nonlinear extension is proposed by Schreiber as the transfer entropy. Based on the thought of causality analysis, a new measure of causality named directional symbolic mutual information is proposed. It has the property of robustness, conceptual simplicity and fast computational speed. A new causality index is then stated by subtracting the directional symbolic mutual information and its inversed counterpart. Coupling strength and direction of process variables are measured by the causality index.3. As a key factor in model-based control technique, model fidelity has significant influence on control performance. Model-Plant-Mismatch (MPM) detection is an important step in the procedure of control performance monitoring and system maintenance. For the majority of existing APC systems, two layers containing the MPC in the upper layer and the basic layer, e.g. PID control, are employed. All the methods are based on the use of statistical local approach and insensitive to the change of disturbance dynamics. The key step of local approach is finding a primary residual that sensitive to the fault concerned. First, residuals are derived using sub-space identification for MIMO system in the MPC layer. For the basic control layer, two methods are proposed. One need dithering signal in the set-point and the correlation between dithering signals and model residuals is used to construct primary residual. The other is non-invasive and the primary residual is obtained by using instrumental variable identification method.4. The existing MPM detection methods based on Kalman filter and correlation analysis are no longer efficient for nonlinear systems. To tackle this issue, mutual information as a general correlation measure is introduced. Mutual information, which is a well known concept in information theory can reflect the dependence of two stochastic variables, no matter nonlinear or linear. The conception of information flow in systems is introduced to analyze the relationship between the information transfer of exciting signals and model error. This method can help users locate the sub-system that has model-plant-mismatch. The estimation of mutual information, as well as the surrogate method to determine threshold are introduced.