A knowledge-based approach for data reconciliation of process networks

Because the area of gross error detection and data reconciliation has received a significant amount of attention in the past two decades, the problem is fairly well-defined but the solutions have yet to be perfected. Also, despite the proliferation of new methods that offer particular strengths and the potential economic value of this data analysis step, it has not been as widely performed in chemical manufacturing plants because of the difficulties, both mathematical and instrumental, that plant operators encounter. Hence, the focus of this dissertation is the simplification of the gross error detection and data reconciliation problems at hand using additional process knowledge, developing heuristics for a more systematic and effective use of currently available gross error detection and data reconciliation techniques, and the creation of a computational framework to provide the mechanism for their use;New methods for detecting and identifying gross errors in dynamic and bilinear systems are presented. These new strategies improve the identification performance by optimally selecting statistical tests for determining balance closure by using special heuristics developed for the purpose. Furthermore, they have the advantage of being computationally simpler than most other available methods. Being less computational, these techniques promise to be good candidates for incorporation into data reconciliation expert systems or malfunction diagnosis expert systems. This could facilitate more use of process knowledge such as known or highly suspect faults and can also improve the partitioning search function especially when applied to large-scale process networks. The simulations involve changing the type, magnitude and location of biases and the significance levels, among others. Results show that the new strategies indeed successfully unproved the identification performance;A prototype expert system for gross error detection, identification and estimation, with a unique distributed knowledge structure incorporating process network hierarchies and functional relationships is developed and tested. The resulting framework provides advantages such as allowing the use of different forms of plant information, advising on proper selection of the appropriate gross error detection method to use, and allowing partial analysis of process networks, among others. The structure can also take advantage of some existing malfunction diagnostic structures;Results of test runs on two simulated systems, a 12-stage distillation column and a NASA experimental test bed of a crop growth chamber system successfully demonstrate these advantages of using a knowledge-based system for data reconciliation