H.J.L. van Can

An efficient model development strategy for bioprocesses based on neural networks in macroscopic balances

In the serial gray box modeling strategy, generally available knowledge, represented in the macroscopic balance, is combined naturally with neural networks, which are powerful and convenient tools to model the inaccurately known terms in the macroscopic balance. This article shows, for a typical biochemical conversion, that in the serial gray box modeling strategy the identification data only have to cover the input-output space of the inaccurately known term in the macroscopic balances and that the accurately known terms can be used to achieve reliable extrapolation. The strategy is demonstrated successfully on the modeling of the enzymatic (repeated) batch conversion of penicillin G, for which real-time results are presented. Compared with a more data-driven black box strategy, the serial gray box strategy leads to models with reliable extrapolation properties, so that with the same number of identification experiments the model can be applied to a much wider range of different conditions. Compared to a more knowledge-driven white box strategy, the serial gray box model structure is only based on readily available or easily obtainable knowledge, so that the development time of serial gray box models still may be short in a situation where there is no detailed knowledge of the system available. (c) 1997 John Wiley & Sons, Inc. Biotechnol Bioeng 53: 549-566, 1997.

Comparison of intelligent control schemes for real-time pressure control

Abstract Direct fuzzy control of the Mamdani type, fuzzy supervised PI control and predictive control based on fuzzy and neural models were applied to a nonlinear pressure process. The main goal of this study is to compare these different control concepts in terms of the development time, the type and amount of prior information needed for the controller design, the tuning requirements and the closed-loop performance.

Semi-mechanistic modeling of chemical processes with neural networks

Abstract One of the major drawbacks of nonlinear black-box models is the scaling problem. A black-box model derived for a certain process scale or pilot plant cannot be used for other process scales. To avoid this problem, the combination of white-box and black-box modeling techniques is worth investigating. In this paper an approach based on a combination of white-box and black-box techniques based on neural networks is described. All the known parts of the process are based on first principles, and the remaining, unknown parts are modeled by black-box models consisting of a neural network. The black-box model is incorporated in the white-box model. Both modeling techniques are compared in an example. This comparison shows that for that particular example the semi-mechanistic modeling technique outperforms the straightforward nonlinear black-box model. The application of neural networks to the black-box modeling of nonlinear processes always depends heavily on the availability of enough informative data. If the process operation does not allow for the measuring of many states and outputs under varying operational conditions, then the application of neural networks is not admissable. The application of semi-mechanistic models is then preferable.

Fuzzy modeling of enzymatic penicillin-G conversion

Abstract Mathematical modeling is essential for the design, analysis and optimization of biotechnological processes, as well as for the development of their control systems. Because of the complexity and uncertainty associated with these processes, rigorous mathematical modeling often becomes a major bottleneck. This article presents a modeling approach that is based on a combination of first-principle modeling of known relationships, with fuzzy modeling of the unknown parts of a process. A Penicillin-G conversion process is used as an application example to demonstrate the methodology. A linguistic fuzzy model, which represents the kinetic term of the conversion, is developed from experimental data by means of fuzzy clustering. The model is then incorporated in macroscopic balance equations describing the overall conversion process. It is shown that the approach leads to an accurate prediction model, and, at the same time, allows for a qualitative interpretation of the unknown relationships learnt from data. The extracted knowledge base was presented to experts, who confirmed the overall correctness of the rules, and also the relevance of the membership functions for the particular process under study.

Design and real time testing of a neural model predictive controller for a nonlinear system

This paper presents the design of a one step ahead model predictive controller based on a neural network for a single input single output nonlinear system. The design procedure is mainly concerned with the identification of the nonlinear system using a neural network. In the absence of a consistent theory for nonlinear systems an heuristic approach is proposed, based on three main subjects: selection of a suitable training signal, calculation of the network parameters and selection of the most appropriate network configuration. The neural network model was trained using a carefully selected training signal that contained all the relevant process and controller dynamics. The parameters of many network configurations, varying in dimension and composition of the input vector and number of hidden nodes were calculated with a fast noniterative method. The predictive power of all these configurations was compared for the one step ahead prediction and for the multistep ahead prediction over the whole horizon of the test set. This way it was possible to make a rational choice for the most appropriate network configuration. The proposed controller based on the obtained network configuration was tested with real time experiments using a pressure vessel. The presented heuristic approach proved to be successful. The controller was capable of tracking various setpoint changes, even under circumstances that were not present during the identification of the model. The performance of the controller was compared with the performance of a conventional PI controller.

Fuzzy Modeling of Enzymatic Penicillin-G Conversion

Abstract A fuzzy model of the kinetics of enzymatic Penicillin-G conversion has been developed from data. The model consists of If-Then rules that describe qualitatively how the enzyme kinetics depend on the concentrations of the components involved in the conversion. The rules and the membership functions were extracted from experimental data, using an identification method based on fuzzy clustering. Although no prior knowledge about the process was involved in the identification phase, the obtained model provides a correct semi-qualitative description of the process. The rules and membership functions were presented to experts from the field of biotechnology who confirmed the overall correctness of the rules and also the relevance of the membership functions for the particular process under study. The model also provides accurate numerical predictions of the process behavior.

Two-step approach in the training of regulated activation weight neural networks (RAWN)

Abstract Feedforward neural networks with a single hidden layer of neurons and a linear output layer are a convenient way to model a nonlinear input-output mapping. If the activation weights, i.e. the weights between input and hidden-layer neurons, are known, an estimation problem remains that is linear in the parameters. This can easily be solved by standard least-squares methods. The problem thus reduces to finding appropriate activation weights. This paper describes a method to obtain the activation weights, based on local linear approximations, which also can be solved with standard least-squares techniques. The local linear models can be obtained by fuzzy clustering methods. The method is demonstrated on a simple example. With the proposed method the weights are obtained very fast, and the results are good. The method is also flexible with respect to the incorporation of a priori process knowledge.

Neural Models in Predictive Control

Abstract Model Based Predictive Control (MBPC) has not yet been widely used biotechnology, although in other areas successful application have been reported. This mainly because accurate nonlinear models are usually not available for biotechnological processes. A neural network is an adequate tool for modelling nonlinear systems and can be applied straightforwardly in the MBPC structure. The design procedure described in this paper presents a way to deal with the lack of theory for nonlinear systems in case one needs to identify a neural network that will be used in a MBPC structure. The approach is demonstrated successfully on the control of the pressure in a fermenter.

Semi-Physical Modeling of Chemical Processes with Neural Networks

Abstract One of the drawbacks of nonlinear black-box models is the scaling problem. A black-box model derived for a certain process scale can not be used for other process scales. To avoid this problem, the combination of white box and black box modeling techniques may be worthwhile to investigate. In this paper an approach based on the combination of white-box and black-box modeling techniques will be described. All known parts of the process will be based on first principles (physical and chemical laws) and the remaining, unknown parts will be modeled by black-box models describing unknown parts, Comparison with a pure nonlinear black-box modeling strategy shows that for the particular example the semi-physical modeling technique outperforms the straightforward nonlinear black box model.

Nonlinear Predictive Control with Neural Models

Nonlinear black-box modeling techniques are opening new horizons for modeling and control of nonlinear processes. These kind of models can be used in Model Based Predictive Control (MBPC). These techniques include Wiener Models, Fuzzy Modeling, Recurrent and Feedforward Neural networks and combinations of these. In MBPC, a process model is used to predict process response to alternative controller outputs. There are practically no restrictions with respect to the model structure, so that MBPC can very well deal with process nonlinearities. Model-based predictive control has become an important research area of automatic control theory and, moreover, it has been accepted also in industry [3]. A number of successful applications to industrial processes based on linear techniques has been reported, see [3] for a survey. The ability to handle input and output constraints straightforwardly is one of the reasons for this success.