*Corresponding author. Computers and Chemical Engineering 23 (1999) 301 — 314 Estimation of impurity and fouling in batch polymerisation reactors through the application of neural networks J. Zhang, A.J. Morris, E.B. Martin *, C. Kiparissides Centre for Process Analysis, Chemometrics and Control, Merz Court, University of Newcastle, Newcastle upon Tyne NE1 7RU, UK Department of Chemical Engineering, Aristotle University of Thessaloniki, Thessaloniki 54006, Greece Received 17 November 1997; accepted 31 August 1998 Abstract The estimation of the amount of reactive impurities and the level of fouling in a batch polymerisation reactor is of strategic importance to the polymerisation industry. It is essential that the level of impurities and reactor fouling are known (estimated) in order to be able to develop robust and reliable monitoring and control strategies. This paper describes two approaches based upon stacked neural network representations. In the first approach, an inverse neural network model of the polymer process is constructed and the initial reaction conditions are predicted. The amount of impurities and reactor wall fouling are then calculated by comparing the predicted values with the nominal initial conditions. In the second approach, a neural network is used to model the dynamic behaviour of the polymer process. The predicted trajectories are then compared with the on-line measurements of conversion and coolant temperatures. The techniques are compared on a first-principles-based simulation of a pilot scale batch methyl methacrylate (MMA) polymerisation reactor.  1999 Elsevier Science Ltd. All rights reserved. Keywords: Fault detection; Reactive impurities; Reactor fouling; Stacked neural networks; Polymerisation 1. Introduction The advanced monitoring and control of polymeris- ation processes, in particular the properties of the poly- mer, is of major strategic importance to the polymer manufacturing industries. Polymer production facilities face increasing pressures to reduce the costs of produc- tion whilst maintaining or increasing the yield and also ensuring that the quality and consistency of the final product is maintained through safe operation of the process. These goals are difficult to achieve without the provision of efficient and reliable polymerisation charac- terisation techniques. Currently, the main factors limiting the development of comprehensive policies for control- ling the properties of the polymer include the limited availability and the cost of on-line instrumentation, a lack of detailed understanding of the dynamics of the process and, finally, the highly sensitive and non-linear behaviour of the process. Appropriate process control technology and optimisation techniques provide leverage for making cost reductions and improvements in product consistency by enabling processes to be operated closer to economic, plant and safety constraints. The development of the control of polymerisation re- actors has been reviewed by a number of authors includ- ing Ray (1986), Elicabe and Meira (1988) and, more recently, Kiparissides (1996). Much of the research con- ducted in the area of polymerisation has focused upon the estimation of infrequently measured polymer quality variables through non-linear state estimation techniques (Kozub and MacGregor, 1992; Zhang et al., 1995); the development of optimal trajectories for the manipulated variables (Thomas and Kiparissides, 1984); and the evaluation of specific feed back controllers (Congalidis et al., 1989; Peterson et al., 1992). The optimal control of batch polymerisation reactors, in particular for the pro- duction of a product, in the minimum possible time and with desirable physical and mechanical properties has been studied by a number of researchers (e.g. Secchi et al., 1990; Jang and Yang, 1989). The primary focus in these studies has been the calculation of optimal profiles for the reactor temperature and/or the addition of initiator 0098-1354/99/$ — see front matter  1999 Elsevier Science Ltd. All rights reserved. PII: S 0 0 9 8 - 1 3 5 4 ( 9 8 ) 0 0 2 7 5 - 0