Issues in process monitoring and control: Identification, model predictive control, optimization, and fault detection

The objective of this thesis is to explore a number of issues in process monitoring and control. We address problems related linear system identification, stability and performance of two-stage model predictive controllers and monitoring the accuracy of sensor measurements with a view of detecting faults.; Process model is at the heart of most advanced control algorithms. Accurate identification of a process model is time consuming and often difficult. We propose two new methods for system identification. Both employ polynomials. The first method utilizes polynomials to approximate the impulse response in time domain. In the second method, the imaginary part of the frequency response is modeled by polynomials. Both methods have advantages of robustness and high accuracy. Simulation and experimental results confirm the superiority of these methods.; Model Predictive Control (MPC) is widely used for multi-input-multi-output systems in process industry. Typically, the setpoints of MPC are from infrequent Real Time plant wide Optimization (RTO). However, industry also uses a two-layered cascade structure for implementing MPC. The top stage could be a Linear Program (LP) or Quadratic Program (QP) with an economic objective. The properties or the design method of such cascade systems are not established. In this work, we analyze the stability and dynamics of LP-MPC and QP-MPC. Simulation studies are used to validate the properties described.; Control systems depend on the accurate measurement of process variables. Thus it is important to develop tools that enable on-line monitoring of the health of sensors in order to make sure that the control system gets the right information. Here we propose a new method for sensor fault detection, based on the concept of noise finger printing and characterization. The measurement noise is characterized first under normal, i.e., well calibrated sensor operation. Power spectrum density (PSD) of the noise is monitored over a range of frequency bands. Principal component analysis is used as a multivariate monitoring tool. T2 statistics is adopted to detect the fault. Experiments using a temperature sensor and a strain gage sensor prove the feasibility and effectiveness of this approach.