Application of Neural Networks and Other Learning Technologies in Process Engineering

The following sections are included:

• Foreword

• Acknowledgements

• Contents

Liquid-liquid extraction is not understood well enough to allow acceptably accurate design calculations to be made. Modelling and simulation can be difficult and time-consuming and is usually heavily dependent on empirical correlations of restricted range of applicability. The use of artificial neural networks to achieve more precise simulation has therefore been examined. Application to multicomponent equilibrium and diffusion coefficient data, extraction column hydrodynamic data (drop sizes and hold-up), mass transfer stage efficiency and performance prediction of an industrial extraction column has been carried out with widely varying degrees of success. The lack of data for building a neural network is the largest problem faced.

Methods developed for radial basis function network (RBFN) identification are applied to a complex multiple-input, multiple-output (MIMO) simulation of a solution copolymerization reactor. For RBFN identification, k-means clustering and stepwise regression analysis are used. The practicality of applying these methods to large industrial identification problems is discussed, considering the restrictions of industrially practical input sequence design. The RBFN model has three inputs and two outputs, and the dimensionality of the identification problem poses some difficulties for nonlinear empirical model identification; specifically, the large amount of data required is a problem for plant testing and may cause computational difficulties for identification algorithms as well.

This paper demonstrates the applicability of unsupervised neural networks in the form of self-organizing maps (SOM) for process visualization and modeling. The structures of SOMs and learning algorithms are summarized. Unsupervised methods are applied to identify the different physiological states which exist during a yeast fermentation. The neural network model was able to predict accurately different physiological states.

In this paper we concern ourselves with a Bayesian approach to training radial basis function (RBF) networks, and in particular, nonlinear techniques for proper location of the basis functions in the input space. Thin plate spline networks are trained on simulated data, using both 'unsupervised' and nonlinear techniques for the basis function placements. It is observed that optimisation improves the performance of the networks by allowing the centres to spread well beyond the data space. This challenges the suitability of traditional unsupervised methods, which usually require the basis functions to be located near the input data.

In this work a recurrent neural network is applied as a non-linear identification tool of a fed-batch penicillin process. The recurrent neural network is trained by a multiple-stream extended Kalman filter, which allows the process to be identified in real-time. It is shown that the fed batch process can be estimated very accurately by a recurrent network and that the extended Kalman filter is a very efficient and rapid training algorithm for non-linear process identification. The recurrent neural network could be trained to give an one-step ahead prediction for a sample-time of six minutes. This ensures that the neural network can be used as an identification model in a model predictive control loop for calculation of the optimal feeding strategy in real-time.

This paper analyzes the combination of prior knowledge in the form of first principle models (parametric models) and neural networks. These models are called hybrid models. Neural networks and hybrid models were used to identify a fed-batch fermentation. Different neural networks were integrated into the hybrid model structure. The performance of these hybrid models is compared with "traditional" neural networks.

It is well understood that the optimal control policies can be significantly different with and without due consideration to the plant-model mismatches. In our previous work, the detailed dynamic model was assumed to be the exact representation of the plant while the difference in predictions of the plant behaviour using a simple model and the detailed model was assumed to be the dynamic plant-model mismatches. Theses dynamic mismatches were modelled using neural network techniques and were added to a simple model to produce a hybrid model. Previously, we developed a general optimisation framework based on the hybrid model for dynamic plants.

In this work, a hybrid model for an actual pilot plant batch distillation column is developed. However, taking advantage of some of the inherent properties of batch distillation process a simpler version (new algorithm) of the general optimisation framework is developed to find optimal reflux ratio policies which minimises the batch time for a given separation task. Finally, discrete reflux ratio used in most pilot plant batch distillation columns, including those used in industrial R&D Departments, does not allow a direct implementation of the optimum reflux ratio (treated as a continuous variable) obtained using a model based technique. Here a relationship between the continuous and the discrete reflux ratio is developed. This allows easy communication between the model and the plant and comparison on a common basis.

Hierarchical structures have been introduced in the literature to deal with the dimensionality problem, which is the main drawback to the application of neural networks and fuzzy models to the modeling and control of large-scale systems. In the present work, hierarchical neural fuzzy models are reviewed focusing on an industrial application. The models considered here consist of a set of Radial Basis Function (RBF) networks formulated as simplified fuzzy systems and connected in a cascade fashion. These models are applied to the modeling of a Multi-Input/Multi-Output (MIMO) complex biotechnological process for ethyl alcohol (ethanol) production and show to adequately describe the dynamics of this process, even for long-range horizon predictions.

Many difficult techniques involving non-linear models have been proposed and applied for the control of non-linear processes in the past. However, most of these techniques are complex and difficult to obtain and implement. They are also restricted to limited ranges of operations. In recent years, neural networks have emerged as an attractive method that is easily implemented in various model-based control techniques. One such technique is the internal model control method, which incorporates approximations of both system model and its inverse in the control algorithm. In this article, the application of this neural network based IMC strategy on a highly non-linear system such as the continuous fermentation process will be shown. The control strategy regulates the biomass concentration by manipulating the dilution rate within the reactor. Acceptable performance was achieved for the set point regulation under various internal and external disturbances but with some offsets in the output. An adaptive scheme using the modified sliding window approach was further applied, which eliminated these offsets completely from the system under the same disturbances. The comparison between the conventional and adaptive IMC technique will be further highlighted in this article.

In batch reactors, the optimal reactor temperature profiles which maximise the conversion to the desired product is obtained by solving dynamic optimisation problems off-line. The control Vector Parameterisation (CVP) technique is used to pose the dynamic optimisation problems as Non-linear Programming Problems which are solved using the Successive Quadratic Programming (SQP) based optimisation technique. Two different types of controllers are used here to track the optimal batch reactor temperature profiles (set points). They are Generic Model Control (GMC) and dual mode (DM) control with Proportional-Integral-Derivative (PID). The Neural Network technique is used as the on-line estimator to estimate the amount of heat released by the chemical reaction within the GMC strategy. The GMC controller coupled with the Neural Network based heat-release estimator is found to be more effective, robust and stable compared to PID controllers in tracking the optimal reactor temperature profiles of various reaction schemes. Two different exothermic reaction schemes are used in order to illustrate the idea.

Inferential estimation and optimal control of a batch polymerisation reactor using bootstrap aggregated neural networks are presented in this contribution. In responsive agile manufacturing, the frequent change in product designs makes it less feasible to develop mechanistic model based estimation and control strategies. Techniques for developing robust empirical models from a limited data set have to be capitalised. The bootstrap aggregated neural network approach to nonlinear empirical modelling is very effective in building empirical models from a limited data set. It can also provide model prediction confidence bounds, thus, provide process operators with additional indications on how confident a particular prediction is. Robust neural network based techniques for inferential estimation of polymer quality, estimation of the amount of reactive impurities and reactor fouling during an early stage of a batch, and optimal control of batch polymerisation process are studied in this contribution. The effectiveness of these techniques is demonstrated by simulation studies.

Conventional methods for batch chemical process optimisation and control depend on having both perfect process models and measurements available. Here, to avoid this, we apply a novel methodology centred on reinforcement learning (RL) whereby, unlike most forms of machine learning, an autonomous agent is not instructed on how to act by example but instead learns directly by trying control actions and seeking for those giving maximum reward. A central notion is the performance or value function that, in a given current state, signifies the contribution a specific action will make towards maximising the final performance or reward over an entire batch. For batch-to-batch, incremental learning and control, the initially unknown value function is here represented using wire fitting and a neural network. This is a simple yet powerful means of simultaneously learning and fitting the value function. The performance achieved in each completed batch can be propagated from the end point back through the intermediate states. With echoes of dynamic programming, this allows calculation of Bellman=s errors which can be minimised in neural network fitting. The higher level optimisation and control problem in batch processing thus fits neatly into this framework and some results of a case study illustrate the potential of the approach.

In process plant operation and control, modern computer control and automatic data logging systems create large volumes of data, which contain valuable information about normal and abnormal operations, significant disturbances and changes in operational and control strategies. The data unquestionably provide a useful source of information for supervisors and engineers to monitor the performance of the plant and identify opportunities for improvement and causes of poor performance. This contribution describes the use of data mining and knowledge discovery techniques for automatic analysis and interpretation of process operational data both in real time and over the operating history. Techniques studied include data pre-processing using wavelets and principal component analysis, multivariate statistical analysis, and unsupervised machine learning approaches as well as inductive learning for conceptual clustering. Examples and industrial case studies are used to illustrate these methods.

In this paper the problem of design and elaboration of artificial neural networks as direct process controllers is developed. The neural controller is a feedforward multi-layer network, and the controller design methodology is based on the modelling of the process inverse dynamics. The advantage of this method is that it is not necessary to perform initial closed-loop experiments with a classical controller to generate the learning data base. By this way, multivariable controllers can be easily developed, taking into account the dynamics and the interactions of the different control loops. The efficiency of such a control methodology is exemplified through its application to different chemical processes :

- a semi-batch pilot plant chemical reactor

- a liquid-liquid extraction column

- a low pressure chemical vapour deposition reactor

Artificial Neural Networks (ANN) have been used as black-box models for many systems during the past years. Specifically, neural networks have been used advantageously in the Chemical Processing Industries (CPI) in a number of ways. Successful applications reported range from enhanced productivity by kinetic modeling, to improved product quality, and the development of models for market forecasting. Typically, a main objective in ANN modeling is to accurately predict steady-state or dynamic process behavior to monitor and improve process performance. Furthermore, they also can help in process fault diagnosis. The black-box character of neural net models can be enriched by available mathematical knowledge. This approach has been extended to consider nonlinear time-variant processes. The potential of neural network technology faces rewarding challenges in two key areas: evolutionary modeling and process optimization including qualitative analysis and reasoning. Recent work indicates that evolutionary optimization of non-linear time-dependent processes can be satisfactorily achieved by combining neural network models with genetic algorithms. Industrial validation studies indicate that present solutions point to the right direction, but additional effort is required to consolidate and generalize the results obtained.