The use of neural nets to detect systematic errors in process systems

Abstract

The monitoring of plants and the verification of process models depend crucially on reliable sets of steady state component and total flow rate data. These measurement data are generally subject to random noise (and possibly systematic errors) and typically violate the process constraints of the system. It is consequently necessary to adjust the data, and also to account for systematic or gross errors in the data prior to this reconciliation procedure, or as part of it, in order to avoid severe impairment of the adjustment process. This can be accomplished by using a back propagation neural net to form an internal representation of the relationship between the distributions of the measurement residuals and the residuals of the process constraints. The major advantage of using neural nets instead of conventional statistical methods, is that neural nets can also be used to detect systematic errors in process systems subject to non-linear process constraints (a common situation in the chemical and mineral industry that is not accommodated satisfactorily by traditional statistical procedures).