Sensor network and soft sensor design for stable nonlinear dynamic systems

Abstract

In chemical processes, online measurements of all the process variables and parameters required for process control, monitoring and optimization are seldom available. The use of soft sensors or observers is, therefore, highly significant as they can estimate unmeasured state variables from available process measurements. However, for reliable estimation by a soft sensor, the process measurements have to be placed at locations that allow reconstruction of process variables by the soft sensors. This dissertation presents a new technique for computing an optimal measurement structure for state and parameter estimation of stable nonlinear systems. The methodology can compute locations for individual sensors as well as networks of sensors where a trade-off between process information, sensor cost, and information redundancy is taken into account. The novel features of the approach are (1) that the nonlinear behavior that a process can exhibit over its operating region can be taken into account, (2) that the technique is applicable for systems described by lumped or by distributed parameter models, (3) that the technique reduces to already established methods, if the system is linear and only some of the objectives are examined, (4) that the results obtained from the procedure can be easily interpreted, and (5) that the resulting optimization problem can be decomposed, resulting in a significant reduction of the computational effort required for its solution. The other issue addressed in this dissertation is designing soft sensors for a given measurement structure. In case of high-dimensional systems, the application of conventional soft sensor or observer designs may not always be practical due to the high computational requirements or the resulting observers being too sensitive to measurement noise. To address these issues, this dissertation presents reduced-order observer design techniques for state estimation of high-dimensional chemical processes. The motivation behind these approaches is that subspaces, which are close to being unobservable, cannot be correctly reconstructed in a realistic setting due to measurement noise and inaccuracies in the model. The presented approaches make use of this observation and reconstruct the parts of the system where accurate state estimation is possible.