• To demonstrate the application of batch / semi-batch optimisation and optimising control on comprehensive process simulations and on pilot and production plants.
• To optimise production operations over a few batch runs where only minimal process data is available for model building.
• To provide a set of tools to meet the above objectives and demonstrate their use on comprehensive simulations of batch and semi-batch processes.

Achievements

Research has addressed both model-based and model-free (empirical) modelling approaches for property prediction and batch optimisation. Model-based approaches have focussed on the Extended Kalman Filter and a new Polynomial Approximation based filter for state and parameter estimation. However, due to the highly non-linear nature of many batch systems, work on neural network representations formed one of the major modelling approaches. One concern, however, is that for model based control using neural network models to become a practical industrial tool, not only must they be shown to be able to provide accurate models, but more importantly they must be shown to display high levels of robustness. An additional major industrial problem associated with the building of neural network representations is the availability of training data. There are numerous industrial situations where it is necessary to build an empirical model using minimal process data. The initial studies built on the established research at Newcastle in the areas of neural network and neuro-fuzzy network modelling. One Research Associate and one PhD student have taken up industrial jobs.

1. Robust Empirical Models with Minimal Data

The main achievements have been in the development of a novel approach to building robust neural network models from minimal process data through stacked generalisation and bootstrap aggregated regression approaches. The techniques were demonstrated on a comprehensive batch polymerisation reactor model / simulation in conjunction with Professor Costas Kiparissides (CPERI, University of Thessaloniki, Greece), a Visiting Professor to CPACT (Newcastle). Applications have demonstrated the on-line prediction of molecular weights, the estimation of impurities and fouling, and the optimal control of a MMA batch polymerisation reactor

2. Batch Reactor Optimisation

   A recurrent neural network based non-linear dynamic modelling and optimal control strategy for a batch emulsion co-polymerisation reactor has been explored. To avoid the excessive effort and time associated with the development of a detailed mechanistic model, recurrent neural networks were used to build empirical representations of the complex polymerisation process. Based on the developed neural network model, sequential quadratic programming (SQP) was then used to calculate the optimal temperature profile leading to a polymer product with a desired number average molecular weight, co-polymer composition and high conversion. By changing the constraints in the objective function, investigation of the process mechanisms was carried out to demonstrate the effect of temperature control on the overall polymer conversion. The proposed scheme exploited the advantages of neural network model based optimal control policies which were evaluated on two batch polymerisation process simulators. Sensitivity analysis of the initial conditions to the model parameters and the optimisation criterion was also carried out to illustrate the robustness of the computed optimal control profiles.

   It is well known that the molecular properties of polymers are sensitive to reactor operating conditions. This has implications on the achievement of a specified quality of the desired product, which is a more complex issue in polymerisation than in more conventional short chain reactions. A study was undertaken which explored the issues in the optimal control of batch polymerisation reactors with unknown disturbances such as reactive impurities and reactor fouling. A novel technique which consists of on-line estimation of reactive impurity and reactor fouling and dynamic on-line re-optimisation strategy was developed. The initial optimal control policy was again computed using Sequential Quadratic Programming based on the nominal reactor operating conditions. Once an impurity problem is detected and the amount of reactive impurity estimated during the early stage of the batch, on-line re-optimisation is applied to update the optimal reactor temperature profile for the remaining period of the batch reactor operation. This approach was illustrated by application to the optimisation of a batch methyl methacrylate (MMA) polymerisation reactor.

More recent work that is still on-going, is investigating model-based on-line state and parameter estimation and batch reactor optimisation. In a particularly novel development, Bayesian statistics is being integrated into both the state and parameter estimation algorithms as well as the batch optimisation algorithm. This will allow information to be included in the estimation of the initial states, model parameters and optimisation constraints thereby allowing the calculation of more accurate estimates and enabling the automatic generation of confidence bounds.

3. Neuro-Fuzzy Approaches

Model robustness and openness can be improved through the combined use of process knowledge and process input-output data to decompose the process operation into a number of local operating regions. Within each region, a reduced order linear model is used to approximate the local behaviour of the process. The aim of this part of the project was to integrate neural network representations within such a framework. Fuzzy sets provide an appropriate means to define operating regions since the definition of local operating regions is often vague in nature and there usually exists overlapping among different regions. The global model output is obtained through the centre of gravity defuzzification, which is essentially the interpolation of local model outputs. This modelling approach is very powerful in that it decomposes a complex system into several less complex subsystems.

Recurrent neuro-fuzzy networks have been developed which allow the construction of "global" non-linear multi-step-ahead models from the fuzzy conjunction of a number of "local" dynamic models. In common with feed forward neural network models, feed forward neuro-fuzzy network models can be constructed to provide both one-step ahead predictions as well as multi-step ahead predictions. The neuro-fuzzy network representations were used to develop two types of controllers – the first is based upon conventional model based predictive control approach whilst the second approach makes us of a heterogeneous control framework.

4. Batch Reactor Optimisation using Genetic Algorithms

A genetic algorithm was developed for the multi-objective dynamic optimisation of batch reactors. The optimisation problem was formulated as the minimisation of an objective functional subject to the dynamics of the reactions and the control was limited within an upper and lower boundary. The reaction time is divided into a finite number of stages and the control on each stage was approximated by a constant. Both the control magnitude and the length of the stages were subject to optimisation. The control profile was represented by a sequence of real numbers and a set of genetic operators was proposed for reproduction. The performance of the algorithm was demonstrated using simulation examples. The algorithm was shown to be effective in locating the switching times of the profiles, the resulting profile has a limited number of stages and this may be advantageous in batch reactor optimisation applications. The performance of the algorithm was demonstrated by application to an MMA batch polymerisation. The optimisation problem was formulated as the minimisation of several objective functions.

Subject to the reaction dynamics with the control excursions being constrained. The reaction time was divided into a finite number of stages and the control on each stage is approximated by a constant. Both the control magnitude and the length of the stages are optimised with the control profile being represented by a sequence of real numbers and a set of genetic operators.

5. Fault Detection and Isolation

A feasibility study into the use of model-based methods for fault diagnosis of stirred tank and batch reactor systems using Fault Detection and Isolation (FDI) methods has been carried out in collaboration with the Department of Electronic Engineering at the University of Hull. The problem of fault detection and isolation in linear time-invariant dynamic processes has received great attention during the last two decades and a wide variety of model-based approaches have been proposed. These different methods, however, can be brought down to a few basic concepts such as the parity space approach, the state estimation approach, the fault detection filter approach and the parameter identification approach. In every case for the detectability and distinguishability of faults, mathematical models of the process under investigation are required, either in state space or input-output form. State space descriptions provide general and mathematically rigorous tools for system modelling and residual generation that may be used in fault detection of industrial systems. Residuals are then processed to detect an actual fault condition, rejecting any false alarms caused by noise or spurious signals. These methods do not require mechanistic knowledge of the process under observation since the input-output relationships are obtained through an identification scheme, which uses Auto Regressive eXogenous (ARX) models.

The study addressed the application of fault detection and location procedures using system identification and residual analysis to data from a benchmark Continuous Stirred Tank Reactor (CSTR) system capable of simulating a raft of process and equipment malfunctions. The study demonstrated that simple linear models obtained from ‘black box’ identification procedures can be effectively used to solve fault detection and identification problems in the CSTR system. ARX models were identified for various process operating regions and fault detection of plant components were then obtained through observer residual analysis. This has produced some very encouraging results, which will be extended through an application to the EPSRC.

6. Associated Research

The contributions made by these studies, alongside those in Project 6, has helped win an EU Research Training Network "Knowledge Driven Batch Production" with the Technical University of Denmark; University of Porto, Portugal; Louvain a la Neuve, Belgium; ETH Lausanne, Switzerland; University of Stuttgart, Germany; CPERI Thessaloniki, Greece; and companies MDC Technology, Novo, Du Pont.

Manufacturing Excellence and CPACT

At recent meetings of Projects 5 and 6, views were expressed that there is a strong interdependence between Projects 5 and 6 and parts of Project 2. Through their conjunction, it is possible to tackle "plant-wide manufacturing problems". The "3P" issues of Product development, Process development and Product manufacturing (vertical integration) combined with academic research and in association with analytical and control systems providers provide access to new technologies across the "3P" industrial problems (horizontal integration).

In this respect the CPACT contributions come from:

• The importance of using multivariate statistical methods for data exploration, feature extraction and to obtain an enhanced understanding of the process (Project 6).
• The development of advanced linear and non-linear multivariate statistical methodologies (Project 6).
• The application of plant-wide process control and optimisation, as opposed to single unit optimisation (Project 5).
• The need to make real use of more detailed information on the chemical and physical properties of the material being processed (Project 2).
• The need to be able to develop process models and calibration models from small data sets (Project 2).
• The importance of being able to monitor the whole manufacturing route from product inception and development, through pilot studies to final product manufacturing and customer feedback (Project 6).

It was suggested the CPACT company members all had these issues in one form or another – i.e. they were generic and CPACT was unique in being able to address these with a range of tools capable of being used to provide generic solutions.

The Japanese manufacturing engineer will first look for a statistical tool to identify the cause of the problem. The European Process Engineer will attempt to solve the problem by modeling from first principles. CPACT is unique in the UK, and possible Europe, in that it embraces both of these approaches

Projects 5 and 6 recommend that CPACT embraces the best of the Eastern and Western approaches to problem solving. This can be achieved through the establishment of an industrially led generic programme focussing on contributions to "Manufacturing Excellence".

Deliverables

Batch Reactor Optimisation Review and Use of Genetic Algorithms
Document Ref: 99/P5/1
Issued: 10 May 1999
incorporating three Project Reports. The Executive Summaries are undernoted.

• "Multiple Optimisation of Batch Polymerisation of Methyl MethAcylate (MMA) using Genetic Algorithms"
  "A genetic algorithm is developed for the multiobjective optimisation of batch reactor operation. The optimisation problem is formulated as the minimisation of several objective functions subject to the dynamics of the reactions and the control is limited within upper and lower boundaries. The reaction time is divided into a finite number of stages and the control on each stage is approximated by a constant. Both the control magnitude and the length of the stages are subject to optimisation. The control profile is represented by a sequence of real numbers and a set of genetic operators are proposed for reproduction. The performance of the algorithm is demonstrated by multiobjective optimisation of batch polymerisation of Methyl MethAcrylate (MMA)."
• "Batch Reactor Optimisation using Genetic Algorithms"
  "A genetic algorithm is developed for the dynamic optimisation of batch reactors. The performance of the algorithm is demonstrated using simulation examples found in literature. The algorithm is very effective to locate the switching times of the control profiles, the resulting control profile only has a few number of stages and may prove to be useful in batch reactor optimisation applications."
• "Batch Reactor Optimisation - a Review"
  "Dynamic optimisation methods for batch reactors are reviewed. The report addresses actual dynamic optimisation methods rather than operations optimisation of specific types of reactors. The problem is defined as the optimisation of general dynamic system described by a set of differential equations or a kinetic model. Only the most promising methods in the view of the authors are included in review. The methods reviewed may not necessarily be drawn from batch reactor optimisation research and some approaches proposed in other areas are also described. Batch reactors are characterised by unsteady state operation, imperfect measurements together with various changes in operating conditions, model uncertainties and imperfect measurements. After a review of optimisation methods for nominal models, optimisation methods for models with uncertainties are described. Possible directions of research are discussed briefly."

Batch Reactor Optimisation
Document Ref: 99/P5/2
Issued: 26 May 1999

Summary
Dynamic optimisation methods for batch reactors are reviewed. The report addresses actual dynamic optimisation methods rather than operations optimisation of specific types of reactors. The problem is defined as the optimisation of general dynamic system described by a set of differential equations or a kinetic model. Only the most promising methods in the view of the authors are included in review. The methods reviewed may not necessarily be drawn from batch reactor optimisation research and some approaches proposed in other areas are also described. Batch reactors are characterised by unsteady state operation, imperfect measurements together with various changes in operating conditions, model uncertainties and imperfect measurements. After a review of optimisation methods for nominal models, optimisation methods for models with uncertainties are described. Possible directions of research are discussed briefly.

Document Ref: 01/P5/P6/Execsum
Issued: 26 February 2001

• Dynamic On-Line Re-optimisation of a Batch MMA Polymerisation Reactor using Hybrid Neural Network Models
• Estimation of Impurity and Fouling in Batch Polymerisation Reactors Using Neural Networks
• Neural Network based Optimal Control of a Fed-batch Reactor
• Optimal Control of a Batch Emulsion Co-polymerisation Reactor based on Recurrent Neural Network Models
• Prediction of Polymer Quality in Batch Polymerisation Reactors Using Robust Neural Networks
• Long Range Predictive Control of Nonlinear Processes based on Recurrent Neuro-Fuzzy Network Models
• Fault Diagnosis in a Continuous Stirred Tank Reactor using Identification Techniques