Building knowledge from plant operating data for process improvement applications

Large amounts of data collected and stored in process control computers are rich in information but poor in knowledge. Careful and systematic selection and analysis of data can provide more insight (knowledge) into the equipment/process. This knowledge in the form of mathematical models (empirical or semi-empirical) provide the basis for the process improvement applications such as system identification for control, process monitoring, fault detection, soft sensor development, etc. In this paper, four case studies have been presented to illustrate the potential for building knowledge from plant operating data using multivariate statistical analysis and neural networks. In the first case study, a MIMO parsimonious orthonormal basis filter based prediction model has been developed for a pilot scale distillation column. The second example illustrates the detection of control valve stiction using nonlinear principal component analysis (NLPCA) using data collected from an operating plant. In the third example, data from a refinery crude preheat train is analyzed for monitoring the thermal efficiency of the heat exchangers and a fouling prediction model was developed. The last case study illustrates the development of a soft sensor in a pilot scale distillation column. In conclusion, the potential of historical operating data in providing information to build knowledge which in turn can be used for the process operational excellence has been demonstrated.