1 Introduction

Bioremediation is a process that uses microorganisms which are naturally occuring in the soil to decompose contaminants such as toxic or hazardous substances. Bioremediation can be done due to the special property of these microorganisms to transform the organic compounds that comprise hazardous wastes as part of their cell metabolism. The goal of this biological degradation can be to transform substances into non toxic ones or to effect a more rapid removal in combination with conventional remediation techniques, as it is described in [Smith and Aiken, 1991].

However, no matter what the explicit object of the application of biotechnology in the remedial process is, bioremediation heavily depends on environmental parameters (e.g. temperature and pH). Although this dependence especially on temperature is well known and it could be influenced externally, process control systems as they are commonly used for chemical production processes are not applied to remedial processes. One reason for this might be that optimal set-points and/or control sequences -- e.g. a temperature which leads to a maximum activity of the microorganisms -- are unknown in most cases.

Similar situations can be found in most biochemical production processes, which are the field of research the department of Control Systems Theory & Robotics at the Darmstadt University of Technology has been working on for several years. In the following section a short overview of our actual research and applications in the field of control and optimization of biotechnological processes is given. Section 3 describes the objective of a research project in which these approaches shall be applied for the control and optimization of a bioremediation process.

2 Control and optimization of biotechnological processes

The control and optimization of biotechnological processes is a complex task of industrial relevance, due to the growing importance attached to biotechnology. Therefore the number of modern intelligent approaches of computer and control engineering applied in the fields of development and optimization of bioprocesses increases. In biotechnological productions microorganisms are cultivated which have the special property to excrete or accumulate a desired product -- e.g. pharmaceuticals or food supplements -- as part of their cell metabolism.

The lack of a complete mathematical description that arises from the incomplete knowledge on the dominant biological pathways as well as the low availability of sensor information about the current physiological state are characteristic problems. Therefore an automatic control and optimization of biotechnological processes often appears to be very difficult. In industrial practice the development of biotechnological production processes is characterized by a big number of empirical test series that are expensive and time consuming. Organisms and substrate composition have to be selected or modified by microbiologists. Furthermore a large number of experimental runs is needed to find appropriate environmental process parameters (e.g. temperature, pH or feed-rate for fed-batch processes). During batch or fed-batch fermentations -- which are the most common operation modes for bioprocesses -- often significant alternations in the cell metabolism due to changes in extracellular conditions can be observed. Therefore it is obvious that the environmental parameters have to be changed during a fermentation in order to achieve optimal product yield. However in industrial practice usually constant set-points are applied to the whole fermentation which are chosen because they provided the best results during test series in laboratory scale -- see [Bailey and Ollis, 1986]. That is due to the fact that often appropriate methods for an analysis of experimental data are not available. Neither an analytical way exists in most cases to evaluate optimal temporal sequences for the environmental parameters by biological or physical reflections. For such processes computer based learning control approaches are an attractive way for automatic control and optimization.

The basic conception of learning control loops can briefly be explained by describing the system LERNAS [Tolle and Ersü, 1992], shown in fig. 1 .

Figure: Scheme of the learning control loop LERNAS [Tolle and Ersü, 1992].

Figure 2: Associative memory mapping for process prediction.

Basically there are two mechanisms working at the same time: First an associative memory -- this could be a neural network -- which is connected in parallel to the process learns the input/output mapping of a predictive process model (see fig. 2 ). In order to enable a pseudo dynamic mapping several history values taken from a short term memory are used beside the actual process values as inputs of the associative memory. Second by applying different control inputs to the predictive model and assessing the predicted outputs with respect to a predefined optimization criterion advantageous control strategies are generated. Finally these control strategies are stored into another associative memory. Both modelling and optimization can be operated off-line based on stored process data as well as on-line.

From this learning technique certain advantages arise compared to adaptive approaches, as they were proposed e.g. in [Bastin and Dochain, 1990] for bioprocess control. Adaptive approaches use simplified models in which the complex nonlinear dependence on environmental process parameters (cf. section 1) are not explicitly considered. Instead these dependence is interpreted by time-varying parameters. That means changes -- e.g. of temperature in the reactor -- lead to a new adaption of the model's parameters. Therefore based on such models there is no chance to find optimal set-point sequences due to the fact that the influence of variations of the environmental parameters cannot be predicted.

In [Gehlen et al., 1992] an extension of LERNAS with respect to the specific properties of biotechnological processes was introduced. The system BioX shown in fig. 3 includes a number of special solutions for an integrated knowledge based and learning control of bioprocesses:

Figure: Scheme of the system for integrated knowledge based and learning control of bioprocesses BioX.

For the control of fermentations the major task is not to establish given set-points for environmental parameters at the reactor but to choose these set-points. Therefore the optimization module is used to generate optimal inputs for underlying conventional control loops.
As already mentioned in section 1, fermentation processes are characterized by a temporal sequence of process phases, in which process behavior can be very different. By a classification of the current physiological state [Konstatinov and Yoshida, 1989] combined with the use of phase specific models and control strategies, easier generation and better quality of the predictive model can be achieved [Gehlen, 1993]. For this reason a rule based phase classifier -- see also [Halme, 1989] -- was realized.
Finally a rule based fault detection and a plausibility check for the generated control action was supplied to the process control system.

BioX was successfully applied to the process control of an -amylase production with Bacillus subtilis (cf. fig. 4 ). By optimized dynamic variation of the environmental process parameters the product yield was increased by more that 100% [Gehlen, 1993].

Figure: Comparison of results with optimized and normal constant environmental parameters: process input pH (left) and product yield -amylase (right).

During the last few years this fundamental approach was systematicly investigated and improved. An extended concept was first presented in [Bettenhausen and Tolle, 1993]. A homogeneous object oriented implementation concept was chosen to overcome the explicite separation between the knowledge based and the learning layer. Since that time a number of new techniques were developed to provide a better transparency and to make the system more user-friendly. They are summarized in [Bettenhausen et al., 1995a]. These new techniques include aspects of self-organizing generation of structured dynamic nonlinear process models based upon the ideas of genetic programming -- see [Bettenhausen and Marenbach, 1995] and [Bettenhausen et al., 1995b] -- as well as the transparent generation of fuzzy rules in a particular NeuroFuzzy approach. The latter is used for the classification of physiological states during batch and fed-batch fermentations -- see [Bettenhausen et al., 1993] -- and for the long time strategy generation to optimize the achievable product yield by dynamic variation of the organism's environmental conditions -- see [Bettenhausen et al., 1995c].

3 Objective of a cooperation with ENSR International

ENSR International is currently doing in-situ bioremediation at a site in Frankfurt, Germany. In the near future ENSR International will start to apply a new remedial technique which was developed in cooperation by ENSR Consulting and Engineering and AT&T. This new approach consists of a combination of biotransformation, stream injection, as well as more conventional pump-and-treat techniques and vapor extraction. The goal of this system is to effect a more rapid removal of dissolved and non-dissolved solvents than it can be achieved using standard techniques.

As pointed out in section 1 environmental parameters -- especially the soils temperature -- can heavily influence the remedial process. E.g. a higher temperature could lead to an increasing solubility of substances which shall be removed. At the same time the activity of microorganisms in the soil depends on the temperature and choosing a temperature which is to high even can kill them. Therefore -- like it was described for biotechnological production processes -- the optimization of or adaptation to these environmental parameters is an interesting task which is important for the optimization of the overall remedial process. By improvement of the environmental conditions it should be possible to decrease the amount of energy consumed by stream injection as well as the amount of time needed to achieve the remedial goals. Both of these factors will finally lead to a decrement of the anticipated cost of remediation.

However the success of the approach described in section 2 which can be described in a few words as a powerful technique for computer based analysis of measured process data depends on quality and quantity of the data which are presented to the system. Therefore it is necessary for this kind of a system to continuously receive data that enables it to assess the current state of the remedial process or a value indicating the achieved success of remediation. This does not mean that the frequency of complete chemical analysis of the soil has to increase dramatically, but a type of process data has to be provided which allows at least an approximation of the process performance. That means it must be possible for the system to decide for example whether the current temperature has a good or a bad influence on the remedial process.

We at the Darmstadt University of Technology believe that in near future a research project sponsored by a local institution with the objective to investigate the topics described above will be initiated. Within a cooperation with ENSR International, Alzenau, Germany, the approaches which were originally developed for the control and optimization of biotechnological productions could be applied to the remediation of a soil contaminated with hydrocarbons at a site in Frankfurt, Germany. Since the sponsoring only covers personal costs and computer equipment we are actually looking for a competent partner for establishing an improved availability of sensory information

References

Bailey and Ollis, 1986: Bailey, J. E. and Ollis, D. F. (1986). Biochemical engineering fundamentals. McGraw-Hill, New York, 2 edition. ISBN 0-07-003212-2.
Bastin and Dochain, 1990: Bastin, G. and Dochain, D. (1990). On-line Estimation and Adaptive Control of Bioreactors. Elsevier Science Publishers B.V., ISBN 0-444-88430-0, New York.
Bettenhausen et al., 1995a: Bettenhausen, K. D., Gehlen, S., Marenbach, P., and Tolle, H. (1995a). BioX -- new results and conceptions concerning the intelligent control of biotechnological processes. In Munack, A. and Schügerl, K., editors, 6th International Conference on Computer Applications in Biotechnology, pages 324--327, Garmisch-Partenkirchen, Germany. IFAC.
Bettenhausen and Marenbach, 1995: Bettenhausen, K. D. and Marenbach, P. (1995). Self-organizing modelling of biotechnological batch and fed-batch fermentations. In EUROSIM CONGRESS '95, Vienna, Austria. accepted.
Bettenhausen et al., 1993: Bettenhausen, K. D., Marenbach, P., and Flügel, A. (1993). Fuzzy-Logik zum strukturierten und transparenten Wissenserwerb. In Fuzzy-Systeme: Management unsicherer Informationen, pages 116--124, Braunschweig.
Bettenhausen et al., 1995b: Bettenhausen, K. D., Marenbach, P., Freyer, S., Rettenmaier, H., and Nieken, U. (1995b). Self-organizing structured modelling of a biotechnological fed-batch fermentation by means of genetic programming. In GALESIA '95 -- Int. Conf. on Genetic Algorithms in Engineering Systems: Innovations and Applications. IEE/IEEE. accepted.
Bettenhausen et al., 1995c: Bettenhausen, K. D., Möller, S., and Tolle, H. (1995c). Autonomous and transparent generation of control strategies. In Third European Control Conference ECC 95, Rom, Italy.
Bettenhausen and Tolle, 1993: Bettenhausen, K. D. and Tolle, H. (1993). BioX -- extended learning control of biotechnological processes. In IFAC World Congress, volume 7, pages 77--80, Sydney, Australia. IFAC.
Gehlen, 1993: Gehlen, S. (1993). Untersuchungen zur wissensbasierten und lernenden Prozeßführung in der Biotechnologie. PhD thesis, TH Darmstadt, FG Regelsystemtheorie & Robotik. Fortschritt-Berichte VDI, Reihe 20, Rechnerunterstützte Verfahren, Nr. 87, VDI-Verlag, ISBN 3-18-148720-1.
Gehlen et al., 1992: Gehlen, S., Tolle, H., Kreuzig, J., and Friedl, P. (1992). Integration of expert systems and neural networks for the control of fermentation processes. In IFAC Symposium on Modelling and Control of Biotechnological Processes, Keystone, Colorado, USA.
Halme, 1989: Halme, A. (1989). Expert system approach to recognize the state of fermentation and to diagnose faults in bioreactors. In Fish, N. M., Fox, R. I., and Thornhill, N. F., editors, Computer Applications in Fermentation Technology. Elsevier.
Konstatinov and Yoshida, 1989: Konstatinov, K. and Yoshida, T. (1989). Physiological state control of fermentation processes. Biotechnology and Bioengineering, 33:1145--1156.
Smith and Aiken, 1991: Smith, G. J. and Aiken, J. W. (1991). In-situ remediation of chlorinated solvents in soils and groundwater at an electronics manufacturing facility. Technical report, ENSR Consulting and Engineering, Westmont, IL 60559, USA.
Tolle and Ersü, 1992: Tolle, H. and Ersü, E. (1992). Neurocontrol. Number 172 in Lecture Notes in Control and Information Sciences. Springer-Verlag. ISBN 3-540-55057-7.

...Bettenhausen: Darmstadt University of Technology , Institute of Control Engineering, Department of Control Systems Theory & Robotics, Landgraf-Georg-Strasse 4, D-64283 Darmstadt, Germany. E-Mail:

Peter "Mali" Marenbach Thu Jul 20 09:17:44 MET DST 1995

Abstract:

1 Introduction

2 Control and optimization of biotechnological processes

The basic conception of learning control loops can briefly be explained by describing the system LERNAS [Tolle and Ersü, 1992], shown in fig. 1 .

Figure: Scheme of the learning control loop LERNAS [Tolle and Ersü, 1992].

Figure 2: Associative memory mapping for process prediction.

In [Gehlen et al., 1992] an extension of LERNAS with respect to the specific properties of biotechnological processes was introduced. The system BioX shown in fig. 3 includes a number of special solutions for an integrated knowledge based and learning control of bioprocesses:

Figure: Scheme of the system for integrated knowledge based and learning control of bioprocesses BioX.

For the control of fermentations the major task is not to establish given set-points for environmental parameters at the reactor but to choose these set-points. Therefore the optimization module is used to generate optimal inputs for underlying conventional control loops.

Finally a rule based fault detection and a plausibility check for the generated control action was supplied to the process control system.

BioX was successfully applied to the process control of an -amylase production with Bacillus subtilis (cf. fig. 4 ). By optimized dynamic variation of the environmental process parameters the product yield was increased by more that 100% [Gehlen, 1993].

Figure: Comparison of results with optimized and normal constant environmental parameters: process input pH (left) and product yield -amylase (right).

3 Objective of a cooperation with ENSR International

References

Bailey, J. E. and Ollis, D. F. (1986). Biochemical engineering fundamentals. McGraw-Hill, New York, 2 edition. ISBN 0-07-003212-2.

Bastin, G. and Dochain, D. (1990). On-line Estimation and Adaptive Control of Bioreactors. Elsevier Science Publishers B.V., ISBN 0-444-88430-0, New York.

Bettenhausen, K. D. and Marenbach, P. (1995). Self-organizing modelling of biotechnological batch and fed-batch fermentations. In EUROSIM CONGRESS '95, Vienna, Austria. accepted.

Bettenhausen, K. D., Marenbach, P., and Flügel, A. (1993). Fuzzy-Logik zum strukturierten und transparenten Wissenserwerb. In Fuzzy-Systeme: Management unsicherer Informationen, pages 116--124, Braunschweig.

Bettenhausen, K. D., Möller, S., and Tolle, H. (1995c). Autonomous and transparent generation of control strategies. In Third European Control Conference ECC 95, Rom, Italy.

Bettenhausen, K. D. and Tolle, H. (1993). BioX -- extended learning control of biotechnological processes. In IFAC World Congress, volume 7, pages 77--80, Sydney, Australia. IFAC.

Gehlen, S. (1993). Untersuchungen zur wissensbasierten und lernenden Prozeßführung in der Biotechnologie. PhD thesis, TH Darmstadt, FG Regelsystemtheorie & Robotik. Fortschritt-Berichte VDI, Reihe 20, Rechnerunterstützte Verfahren, Nr. 87, VDI-Verlag, ISBN 3-18-148720-1.

Gehlen, S., Tolle, H., Kreuzig, J., and Friedl, P. (1992). Integration of expert systems and neural networks for the control of fermentation processes. In IFAC Symposium on Modelling and Control of Biotechnological Processes, Keystone, Colorado, USA.

Halme, A. (1989). Expert system approach to recognize the state of fermentation and to diagnose faults in bioreactors. In Fish, N. M., Fox, R. I., and Thornhill, N. F., editors, Computer Applications in Fermentation Technology. Elsevier.

Konstatinov, K. and Yoshida, T. (1989). Physiological state control of fermentation processes. Biotechnology and Bioengineering, 33:1145--1156.

Smith, G. J. and Aiken, J. W. (1991). In-situ remediation of chlorinated solvents in soils and groundwater at an electronics manufacturing facility. Technical report, ENSR Consulting and Engineering, Westmont, IL 60559, USA.

Tolle, H. and Ersü, E. (1992). Neurocontrol. Number 172 in Lecture Notes in Control and Information Sciences. Springer-Verlag. ISBN 3-540-55057-7.

Darmstadt University of Technology , Institute of Control Engineering, Department of Control Systems Theory & Robotics, Landgraf-Georg-Strasse 4, D-64283 Darmstadt, Germany. E-Mail: