(57d) Detecting Change in Complex Process Systems with Phase Space Methods

Fault detection and identification in complex processes are particularly challenging, since the nonlinear behaviour of these processes may be difficult to model with sufficient accuracy to detect anomalous conditions timeously. For example, feedback control loops containing sticking valves may generate self-sustained and limit cycle oscillations that can be propagated plantwide through physical coupling and recycles. The source of these oscillations may be difficult to locate with conventional techniques. Likewise, in corrosion systems, a change in the nature of the process may result in changes in current and potential signals that may be difficult to detect with linear methods. As another example, in fluidized beds and bubble columns, pressure signals may be a rich source of information on the state of the system, but the detection of changes in these signals may be difficult, owing to the nonlinear and possibly chaotic behaviour reflected by these measurements. Different approaches to detecting change in such systems have been proposed over the last decade. Of these, methods based on a time-delayed phase space reconstruction of the system are particularly promising. The statistics derived from the embedded data typically include various topologically invariant measures of dimensionality that can serve as sensitive measures of change in the underlying dynamics of the system. In this paper, different phase space methods are consequently considered to detect change in simulated case studies and a real world case study from an industrial plant. The methods include a novel approach based on the use of empirical orthogonal functions, as well as bootstrapped estimates of the correlation dimension of the system.