Constrained identification for neural network models

Presents an identification scheme and practical experiences for incorporating existing process knowledge into an identification task to obtain more reliable nonlinear process models with a reduced amount of required process measurement data. The scheme is implemented within a neural network modelling environment, although it is applicable to most nonlinear model types. The available process knowledge is incorporated as constraints for localized model behaviour resulting in a constrained identification task. The efficiency of the proposed approach is demonstrated using experiments made with two laboratory pilot processes. The case studies clearly show the usability of the proposed approach