An integrated neural-network and expert-system approach to the supervision of reactor operating states in polyethylene terephthalate production

Abstract

With the emphasis on quality indices in polyethylene terephthalate (PET) production, it is highly desirable to assess, and subsequently to maintain, individual reactors at satisfactory operating states. To fulfil these aims, it is proposed to use an approach that integrates artificial neural networks (ANNs) with an expert system (ES); the purpose of the former is to estimate the quality indices from reactor process variables, whilst that of the latter is to assess current operation, and thence to advise on, for instance, the application of optimisation procedures to any of the reactor controllers. In addition, an expert system is used to filter plant measurements before they are input to the ANNs; the aim is to suppress gross errors that can cause the ANNs to output incorrect conclusions. The ANN training algorithm is based on a Quasi-Newton method with a self-scaling variable metric (SSVM), because simulation results show that the algorithm has high performance especially in terms of its speed of convergence. The work was implemented on a large-scale PET plant, with the software installed as an Application Module of a Honeywell TDC-3000 distributed control system.