Multivariate process modelling and monitoring for chemical processes

One of the most important challenges faced by control engineers is design and implementation of decision support systems that can assist operators to make supervisory control decisions. Operator failing to exercise the appropriate supervisory control decisions often has an adverse effect on product quality, process safety, occupational health and environmental impact. Thus, there exists considerable incentive in developing decision support systems that can provide automated operator assistance for complex plants.; Before design and implementation of decision support systems for process monitoring, one has to obtain a relatively accurate representation of the system under normal operating condition. This step is known as process modelling or system identification. It is widely regarded as a key step towards successful design of process monitoring systems. There are various approaches available in this area such as prediction error method (PEM), subspace identification method (SIM) and multivariate statistical regression method (MSRM). In this thesis, partial least squares (PLS) method was applied onto a bleached chemi-thermomechanical pulp (BCTMP) plant to obtain the process model with various considerations, e.g. nonlinear effects. Further, traditional canonical variate analysis (CVA) was extended to deal with ill-conditioned data via reduced Krylov space and Cayley-Hamilton theorem.; After the process model has been obtained through the above mentioned approaches, one can design process monitoring systems using variety of methods. However, the model quality, i.e. model plant mismatch (MPM), and unknown disturbances can affect process monitoring results significantly. By extending the Chow-Willsky ([19]) scheme, the effect of process uncertainties can be completely removed for sensor fault detection and diagnosis (FDD). For actuator FDD, the principle components of process uncertainties can also be eliminated so that FDD results are affected in the minimum manner.; A FDD method dealing with multiplicative faults is proposed by using data reconciliation (DR) and gross error detection (GED). The proposed method was then applied to a chemical tank inventory system and successfully identified the location and magnitude of a multiplicative sensor calibration error.; FDD problem under multirate situation is also investigated in this thesis. Using the well-known lifting technique, the multirate discrete time model can be obtained by extending subspace identification method. Once the multirate model is obtained, the structured residual vector approach is used for fault detection and diagnosis. An experimental case study proves the effectiveness of the proposed method.; The fundamental issue of fault detectability is also analyzed in order to provide the bases for FDD research. The general fault detectability problem can be categorized into two subproblems: fault detectability and strong fault detectability. The conditions for both the subproblems are provided and proved.