Applications 1994-1999 - Final Report
Main page
Preface
Program
summary
Research
projects
1994-1998
Research
projects
1998-1999
Product
development
projects
1994-1999
8. Intelligent control of dynamic systems (DYHA)
8.1 Intelligent control and diagnostics of the harvester head
Abstract
Timber harvesters are used under a very wide range of operational conditions. The control algorithms of the harvester head have to be robust. Accurate positioning to the cutting window and damage-free stem feeding ensure the highest possible output and quality of logs produced.The control problem of the harvester head has non-linear and multi-variable characteristics. Therefore, advanced control methods have to be investigated including adaptive and neuro-fuzzy control.
The harvester head includes a wide range of instrumentation. Their proper operation is required for high throughput. Because of hard operation conditions, wear and soiling gradually impair the performance. The control system has to evaluate continuously the performance and alert the user if any component needs maintenance or repair.
Results
The harvester head was modeled in detail. The developed simulation model includes a very realistic description of the dynamical behaviour of the head. The model provides a test bench for control design. New control algorithms and diagnostic schemes can be easily tested with the model. The model is modular, so it can be expanded and it will serve future needs as well.A new control concept was developed for stem feeding. The design goals have been robustness of the controller, the ability to tune it, and the minimization of damage to logs. The controller makes the acceleration and the braking smooth. The acceleration mode detects the reduction in feeding speed caused by the delimbing of branches and accelerates more carefully.
Commonly, characteristics of hydraulic components are measured with pressure and flow sensors. A diagnostic algorithm was developed for the harvester head that makes additional instrumentation unnecessary. Features of feeding performance are analyzed in order to detect failure modes or their coming. Features that are selected to indicate a fault are not sensitive to the environmental conditions or the control parameters. A prototype was developed for the simulation model. The system will give a warning if it detects that a component will break. It will also give an estimate of the time by which the component should be repaired or replaced. Such messages will be utilized in the design of the maintenance schedule.
The results obtained will increase the competency of the harvester, use the raw material more effectively, increase the quality of produced timber and keep the operating rate at a high level.
Project information
Participants
- Tampere University of Technology, Automation and Control Institute
- Plustech
Project dates
March 1, 1998 - December 31, 1999
Project volume
Total budget FIM 394.000, 14 man months
Project manager
Professor Heikki KoivoHelsinki University of Technology, Control Engineering Laboratory
P.O.Box 5400, FIN-02015 HUT
Tel: (09) 451 5200 / (050) 515 1270
Fax: (09) 451 5208
E-mail: heikki.koivo@hut.fi
8.2 Intelligent control of switched-mode power supplies
Abstract
A switching power supply is a DC-DC converter, which can transform DC voltages to a desired level with very little losses. Because the dynamics of the system consisting of the power supply and the load are very fast, commercial controllers have been realized by using analog components. Because these are not very flexible, there is a growing need to replace the analog solutions with computerized ones. For switching power supplies, the study of output voltage control by discrete and intelligent controllers is going on worldwide at this moment, but commercial solutions are not yet available. The goal of the project was to study the possibility of replacing the existing analog control solutions of the switching power supply with digital ones. Different “intelligent” control algorithms have been developed, analyzed, and compared to traditional ones by using simulation as a tool. A test bench consisting of a real switching power supply connected to a digital signal processor (DSP) board has been built to test the operation of the controllers in practice.This 2-year project belonged in an essential way to the field of control theory, in which new control methods are developed and applied in different applications. The industrial partner has benefited from the project by gaining new information on current trends in control theory in order to be able to develop a product, which gives a considerable market benefit internationally. The topic of the project provided a good example of how to replace analog solutions with microprocessor algorithms, which contain intelligent control strategies. The research extended new control ideas into the field of electronics and narrowed the gap between theory and practice. It also demonstrated a fast way of developing the prototype of a new product by using efficient simulation tools such as Matlab/Simulink. All this improved the cooperation between control engineers and electrical engineers.
Results
The scientific and engineering results of the project consist of the following items- Modelling, analysis, controller design, and simulation of a DC-DC converter.
- Building a test bench to verify the operation.
A 500 W switching power supply was chosen as a target to study the control of the output voltage. The desired control accuracy was ±0.3 V in normal operation, ±2V for maximum load disturbances, and ±0.5 V for maximum line disturbances.
A basic DC-DC converter topology (Buck) has been modelled in both voltage and current control modes. The operation of standard PID controller and its modification has been studied and simulated by using the software Matlab/Simulink. The results were verified by using Saber, which is a large simulation tool especially designed to simulate electrical systems. The results showed that Matlab/Simulink is an effective and accurate tool, and there is no question about its applicability in the simulation of devices containing high-frequency switching and dynamics. The discretized PID controller was designed and its operation was investigated by simulation. The design specifications were met as shown by the simulation results.
Troublesome operation modes like the cases with a constant power load and modified constant power load were studied both theoretically and from the control viewpoint. The theoretical problems related to these issues (e.g. regarding stability) were solved and suitable control solutions were provided. Because in the constant power load case the control problem becomes nonlinear, a fuzzy control algorithm was designed and tested by using the software FuzzyTech.
The nonideal characteristics of a real switching power supply was modelled by including equivalent series resistors (ESRs) in the model. Also the effect of line and load disturbances were studied and taken into account in the controller design to meet the design objectives. For comparison, an internal model control (IMC) algorithm was designed and tested to achieve the desired output voltage especially in the case of load disturbances.
A test bench for testing the control algorithms was constructed by removing the analog control part from a real power supply and connecting a DSP board to control the output voltage. At the construction phase and in the introduction of the DSP board unexpected problems were met, and the real tests have just begun at the time of writing.
Project information
Participants
- Helsinki University of Technology, Control Engineering Laboratory
- Efore Oyj
Project dates
A 2-year project (01.03.1998 - 31.12.1999), which was divided into two research periods in 1998 (01.03.1998 - 28.02.1999) and 1999 (01.03.1999 - 31.12.1999)
Project volume
Total budget: FIM 1.052.000Total number of man months: 42
Head of the project
Professor Heikki KoivoTel: (09) 451 5200
E-mail: Heikki.Koivo@hut.fi
Researchers
Idris GadouraE-mail: igadoura@cc.hut.fi
Tel: (09) 451 5221
Ammar Al-Khani (starting 01.02.99)
E-mail: ammar.al.khani@hut.fi
Tel: (09) 451 5223
Kai Zenger
E-mail: Kai.Zenger@hut.fi
Tel: (09) 451 5204
Kai Heikkinen (01.03.98 - 31.01.99)
Laboratory
Control Engineering LaboratoryHelsinki University of Technology
Konemiehentie 2
FIN-02150 Espoo, Finland
Contact information: Efore
Petri VallittuEfore Oyj
P.O. Box 61 (Piispanportti 12)
FIN-02211 Espoo, Finland
Phone: +358 9 478 46327
Fax: +358 9 478 46500
E-mail: petri.vallittu@efore.fi
Publications
Grigore, V., Hätönen, J., Kyyrä, J. and T. Suntio. Dynamics of a Buck Converter with a Constant Power Load. Proceedings of the PESC’98 Conference, Fukuoka, Japan, Vol. 1, pp. 72-78.Gadoura, I., Heikkinen, K., Zenger, K., Vallittu, P. and T. Suntio. New Methods for Controller Design, Analysis, and Validation in the Design of DC/DC Converters. Proceedings of the 1998 IEEE Nordic Workshop on Power and Industrial Electronics (NORPIE/98). Helsinki University of Technology, Espoo, Finland, August 26-27, 1998, pp. 164-169.
Gadoura, I., Grigore, V., Hätönen, J., Kyyrä, J., Vallittu, P., and T. Suntio. Stabilizing a Telecom Power Supply Feeding a Constant Power Load. Proceedings of the 20th International Telecommunication Energy Conference, San Francisco, California, USA, October 4-8, 1998, pp. 243-248.
Gadoura, I., Zenger, K., Suntio, T., and P Vallittu. New Methodology for Design, Analysis, and Validation of DC/DC Converters Based on Advanced Controllers. Proceedings of the 21st International Telecommunications Energy Conference (INTELEC’99), June 6-9, 1999, Copenhagen, Denmark, p. 23-1.
Gadoura, I., Suntio, T., Zenger, K., and P. Vallittu. Soft Computing-Based Controller Design for a Telecom Rectifier. Proceedings of the 1999 IEEE Midnight-Sun Workshop on Soft Computing Methods in Industrial Applications (SMCia/99), June 16-18, 1999, Kuusamo, Finland, pp. 80-85.
Zenger, K., Heikkinen, K., Gadoura, I., Suntio, T., and P. Vallittu. System Modelling and Control in the Design of DC-DC Converters. In the Proceedings of the 14th IFAC World Congress, Beijing, P.R.China, 1999, (CD-ROM format). In preprints Vol.O.
Gadoura, I., Suntio, T., Zenger, K., and P. Vallittu. Internal Model Control for DC/DC Converters. In the proceedings of the 8th European Conference on Power Electronics and Applications, September 7-9, 1999, Lausanne, Switzerland.
Zenger, K., Heikkinen, K., Gadoura, I., Suntio, T. and P. Vallittu. Hakkuriteholähteen älykäs säätö. Automation days 14-16.9, 1999, Helsinki, pp. 308-313. (In Finnish.)
Thesis
Gadoura, I. Design of Intelligent Controllers for Switching-Mode Power Supplies. Licentiate thesis, Helsinki University of Technology, Control Engineering Laboratory, October, 1999.
Figure 1. The test equipment.
Figure 2. The switching power supply (left), DSP board with interface logic (top), and the power supply for the DSP board (bottom).
8.3 Intelligent visualisation of dynamic process data
Abstract
The most complicated problem emerging from an industrial process is the vastness of the measurement information. For example, in a paper process over 250 measurement channels generate a measurement instance every minute, and more and more often every second. This, together with multivariate time-varying process, leads to complicated task of trying to see the forest from a very thick wall of trees. However, this is the task operators must face in everyday life. The human ability to comprehend dynamic multidimensional data is quite weak, so a visualisation system which preprocesses the measurement data for operators should be an essential part of automation system design. Currently there are few automation systems or visualisation systems that can or do deal with this commonly acknowledged problem.This project was carried out in close cooperation with Automation and Control Institute of Tampere University of Technology and UPM Kymmene Jämsänkoski Paper Mills. The collection of data was performed together with UPM Kymmene Jämsänkoski and Automation and Control Institute. The collected measurement data from this project was also used by University of Oulu and Åbo Academi as part of other Tekes research projects and by a General Electric research project.
Results
The target was to create a plant-tested prototype for both distributed and intelligent analysis together with operator friendly visualisation of dynamic process data. The complete prototype will consist of distributed mathematical analysis engine, a fuzzy logic engine for operator knowledge embedding, and a visualisation engine. The mathematical analysis engine collects simple analysis tools for online analysis, e.g. trend analysis and statistical process control analysis, and utilises temporal shapes which are used to detect different process states and features. The operator knowledge was collected via interviews and discussion sessions where the project ideas were analysed and commented on by operators. The decision-making, or the 'intelligence' in the feature combining phase, is implemented using fuzzy logic. The fuzzy logic engine is essential to be able to identify the unusual behaviour from the normal phenomena induced by the process feedback loops.The main scientific contribution of this project was to study how the match value (degree of certainty) between analytic models identified from process data and simple primitives, e.g. lines and parabola, can be used as fuzzy values for fuzzy inference engine. This degree of certainty value can be used as an input for fuzzy rules. The fuzzy rules created based on process knowledge can then be used not only to make predictions but also by "reverse engineering" the rules to find sources of problems in the process.
The main industrial contribution was the discussions with process personnel and automation system manufacturer. The need for new adaptive alarm methods has been an important topic in these discussions. The ideas and methods of this project will be seen in future versions of the automation system.
The process proved to be a difficult target for this type of method. The vast number of controlled signals are too stable and contain no fruitful phenomena to find with this method. The uncontrolled signals are too noisy and unstable for model identification. Although the paper machine is a process where a huge number of measurements are available, a problem arises from the nonstandard interfaces between the process and the mathematical tools used for methodology development.
The prototype has only a simple user interface and visualisation. The final visualisation environment and graphical user interface will be formed as part of the production phase of the visualisation environment. The ideas for the final outlook together with the pros and cons of different approaches will be addressed in the Licentiate thesis by Jari Seppälä.
Project information
Participants
This project is a joint effort of Tampere University of Technology Automation and Control Institute and UPM-Kymmene Jämsänkoski Paper Mill. Neles Automation has been also an important contributor in project meetings as the automation system provider.
Project dates
1.3.1998-31.12.1999
Project volume
FIM 752.000, 21 man months
Project manager
Prof. Heikki N. KoivoTel: +358 9 4513320
E-mail: heikki.koivo@hut.fi
Publications
[1] Seppälä, J. Intelligent Visualisation of Dynamic Process Data. Licentiate thesis. Tampere. Tampere University of Technology, Department of Automation Engineering, 2000.[2] Nissinen, Ari S. 1999. Neural and Evolutionary Computing in Modelling of Complex Systems. Doctoral Dissertation. Helsinki. Helsinki University of Technology, Department of Automation and Systems Technology, Control Engineering Laboratory, Tech. Rep. 118/1999.
[3] Seppälä, J. Nissinen, A. Koivo, H. and Laitila M. 2000. Intelligent visualisation of dynamic process data. Control Systems 2000, Victoria BC, Canada, May 1-4, 2000. Accepted for publication.
Figure 1. Wiener MLP model with feedback.
Figure 2. Wiener NN with separate system and state MLPs.
The Wiener NN models can be presented in state-space form on
the basis of the state space of Laguerre dynamics (F and G matrices in Figure
2). The Kalman filter framework makes it also possible to obtain smoothed
estimates of the state trajectories, especially useful when there are missing
or sparse measurements. With Extended Kalman filter it is possible to calculate
the covariance of the state estimates, useful in fault diagnostics. The models
are identified from data gathered during normal production fermentations. No
controlled experiments were performed. The paucity of the training data poses
strong limits on the complexity of the models that can be identified
empirically. The Levenberg-Marquardt algorithm, Extended Kalman filter and
Iterated Extended Kalman filter have been used in parameter estimation. Wiener
NN models were identified for two different enzyme fermentations. In parameter
estimation Wiener-NN models showed faster convergence compared to traditional
state-space neural network models. A factory floor implementation of online
state estimation has been planned and will be implemented for amylase
fermentation. Factory implementation will also act as final model validation.
Dynamic input-output models of Wiener NN type have been
identified for columns of an industrial sequential ion-exclusive
chromatographic separation unit. A separate delay model is needed for
describing the movement of the front. The delay model adapts on variations of
the process flow rate. The form transformation of the front is described with a
dispersion model, which is a smoother type of
Wiener-MLP model. Model identification was based on the analysis data
from a factory experiment taken during normal production. The paucity of data
made the modeling experiments in latter phase difficult. The models were tested
both as predictors for use in process monitoring and as a simulator to simulate
the whole separation process, consisting of two separation columns, several
cycles ahead for testing control parameters. The model works quite well when
the scope of training data is sufficiently large. However, when simulation
advances outside the scope of training data problems occur as supposed. The
collection of a rich enough data set from a production process can be very
difficult. A time invariant model version having flow volume as the
"time" variable was also tested successfully. The tested linear
Laguerre approximation model, more suitable for simulation outside the scope of
training data, was not sufficient to model the inherently nonlinear process.
The robust identification principles of linear systems on
the basis of n-width measure can be applied in dynamic nonlinear systems by
using a Wiener NN structure. The space spanned by the continuous Laguerre
functions is an optimal n-dimensional subspace in the n-width sense for a set
of transfer functions having poles inside a certain disk (Wahlberg and Mäkilä
1995). When a damped or slightly oscillating nonlinear system is linearized in
all possible operating points, the corresponding poles define a certain closed
set. If the disk referred above is parameterized so that it covers this set of
poles, the corresponding Laguerre function is an (heuristically) optimal
n-dimensional subspace in the n-width sense for this nonlinear system in the
Wiener context. Kautz functions form an (heuristically) optimal basis for a
nonlinear dynamic system having dominating resonant mode.
8.4 Monitoring and validation of control for
biotechnological and food processes
Abstract
Generic Wiener NN -based methods have been developed and
demonstrated for management, monitoring and control validation of batch type
and cyclical continuous processes, particularly fermentation processes and
chromatographic separation processes. The classical input-output Wiener
representation consists of linear dynamics (Laguerre filters) and static
nonlinear polynomial mapping. In Wiener NN, the static nonlinear mapping is
realized with NN, usually with MLP. By adding feedback in the Wiener NN, a
model capable of modeling autonomous systems, such as batch processes, can be
realized. Other types of modified Wiener NN models have also been tested.
Linear state space models and recurrent neural networks have been identified
for comparison.Results
In this project, nonlinear black-box models have been used
for empirical modeling of autonomous systems, large-scale industrial fed-batch
fermentations which grow filamentous fungi under aerobic conditions to produce
enzymes. The main models used and presented are variations of the Wiener NN
model with output feedback as in Figure 1, or with state feedback and
measurement correction as in Figure 2. Other black-box models, including linear
state-space and recurrent neural networks, have been tested for comparison. The
dynamics contains MLP and Laguerre system in the feedback loop. The MLP can be
interpreted as the measurement equation of the system. 
Project information
Participants
Project dates
March 1, 1998- December 31, 1999Project volume
FIM 925.000, 39 man months Project manager
D.Tech Arto Visala
E-mail: Arto.Visala@hut.fi
Publications
Visala A. (1998): Identification of Wiener-MLP with Feedback NOE-Model with Extended Kalman Filter. 1998 IEEE World Congress on Computational Intelligence (IJCNN '98) Anchorage, Alaska, May 4-9, 1998.Visala A. and Halme A. (1998): Dynamic Wiener-MLP with feedback NOE-model for Tricoderma Fermentation. 7th Int. Conf. on Computer Applications in Biotechnology – CAB7, May 31-June 4 1998, Osaka, Japan.
Visala A., Pitkänen H. and Paanajärvi P. (1999): Wiener-NN Models and Robust Identification. IJCNN'99, July 10-16, 1999, Washington DC, USA
Visala A., Paanajärvi J., Pitkänen H. and Halme A. (1999):Wiener-tyyppisten NN-mallien käyttö prosessin valvonnassa, Automation Days 1999, Helsinki, September 14-16 1999, Suomen Automaatioseura. (In Finnish.)
Paanajärvi J. (1999): MSc Thesis, HUT Automation Technology Lab.
Visala A., Pitkänen H. and Halme A.: Modeling of chromatographic separation process with Wiener-MLP representation. Accepted for publication in Journal of Chromatography B.
The goal of
this work was to develop a general method for the generation of fast neural
network (NN) models approximating thermodynamic properties. In order to be
useful, the method should be applicable to various substances, the accuracy
given by the user should be achieved, and the resulting NN model should be fast
in the simulation mode.
The method
was first applied to the description of temperatures, partial pressures,
densities, and mass fractions of different components of an air-H2O
mixture (Fig. 1). A total of 33 neural network models were identified for
different phases of H2O, air, and the mixtures. The preset accuracy
limit 1% of the output range was achieved at every point of the training and
validation sets. The NN models proved to be 5000 times faster than the
conventional iterative procedure that was used for the generation of the
training and validation sets. The execution time for one typical update was 1
ms when 17 separate property values were calculated for air and H2O,
and the mixture (199 MHz processor, Windows NT).
Figure 1.
Partial pressure (p*) of H2O steam in the air-H2O mixture
for H2O mass fractions 0.02 (left) and 0.1 (right). Separate
networks with 10-21 hidden layer neurons were developed for the gas-liquid
phase border, the temperatures in the gas and gas-liquid phases, the enthalpy
of liquid H2O, and the partial pressure of H2O in the gas
phase. The ranges of the inputs were 0.01-220 bar, 10-4000 kJ/kg and 0-1.0
(only part shown).
Feedforward
networks with one hidden layer of sigmoid neurons were used. The generalization
feature made it possible to train the network with a representative set of
input-output pairs and then also obtain sufficient accuracy with new inputs of
the validation set. The use of the same neural network architecture in all
cases made it possible to develop a general training procedure. The
Lewenberg-Marquardt optimization algorithm was used for the estimation of the
weights and cross-validation for choosing the number of hidden neurons. When
the preset accuracy limit was not achieved at all test points, the training was
repeated with a new training set. In the new set, the points of poor accuracy
were multiplied. The multiplication increased the contribution of the errors at
these points to the value of the cost function and thus also increased
accuracy.
The
developed neural network models were incorporated into a general-purpose
dynamic process simulation program APROS. Previous methods based on interpolation
require large tables and are not convenient for cases with several inputs.
Properties of the flowing medium consisting of air and H2O in
different phases are needed, e.g. in the condenser-circuit of a steam turbine,
the cold start-up of steam pipe systems, many processes in paper mills and in
air conditioning systems. The new method enables fast simulation of these and
many other previously intractable processes.
The
developed procedure can probably be easily adapted for the description of
thermodynamic and other material properties of several substances and also
equilibrium states in chemical reactions. By using the neural network model,
large data tables of previous interpolation methods are not needed and the
approximations previously requiring several iterations for solving complicated
functions reduce to a single function call.
8.5 Neural networks for thermodynamic properties
Abstract
A new
method based on neural networks was developed for the modeling of thermodynamic
properties of substances. The neural network models enable new process
simulation applications. Wide operating ranges are covered by single networks
that give the desired properties for fast dynamic simulation. When applied to
the mixture of air and H2O, the preset accuracy of 1% was obtained
at every test point and the neural networks proved to be 5000 times faster than
a conventional iterative algorithm.Results
Thermodynamic
properties of substances are required in numerical simulation of industrial
processes. Simulation is used in process and control design, safety analysis,
and operator training. Properties that are often required are, e.g.
temperature, density, mass fractions, enthalpies, and partial pressures under
different process conditions. Fast calculation of these properties is also
necessary in measurements of several process variables needed for process
control in distributed control systems (DCS). Conventional algorithms for the
calculation of thermodynamic properties typically consist of several iterative
steps including calls to complex functions and interpolations in tables based
on both theory and experimental data. The iterative algorithms are too slow to
be used, e.g. in dynamic process simulators that should run faster than
real-time. Methods based on tables and linear interpolations are fast, but in
practice limited to cases with only two or three inputs.
Project information
Participants
Project dates
March 1, 1998 – December 31, 1999Project volume
FIM 1.000.000, 18 man monthsProject manager
Dr. Jari Hämäläinen
VTT Automation
P.O. Box 1301, FIN-02044, VTT
Phone: +358 9 456 6467
E-mail: jari.hamalainen@vtt.fi
Publications
Lilja R. and J. J. Hämäläinen (1999) Modeling of Thermodynamic Properties of Substances by Neural Networks. Proc. International Joint Conference on Neural Networks (IJCNN ’99), Washington, DC, July 10-16, 1999. The International Neural Network Society and The Neural Networks Council of IEEE, CD-ROM, ISBN 0-7803-5532-6, 1999.
The main problem is that the ore weighing has to be carried
out while the machine is moving. The movement indirectly corrupts measurements
with noise making data preprocessing one of the key issues in this project.
Moreover, the system-induced nonlinearities have to be compensated for in order
to obtain satisfactory results.
The results show that, especially with neural networks, a
very good weighing accuracy can be achieved. However, before neural
network-based algorithms can be used in the weighing system there are some
problems that have to be solved. Such problems are for example the huge amount
of training data needed for both neural networks and fuzzy systems. Also,
calibration is problematic. However, during the project, new innovative means
for solving the weighing problem were discovered and these are leading towards
a patent application.
8.6 Neuro-Fuzzy application for on-line weight
determination of a moving loader
Abstract
The objective of this project was to find a neural network
or fuzzy-based method for determining the payload in the bucket of a loader.
Information on the payload can be used in several ways. For example, weighing
provides information on the material flows coming out of the mine, which means
better controllability of the whole mine production process. In addition,
weighing provides means to monitor individual tasks and individual machines. If
individual transportation units are monitored, the payment of salaries can be
based on weighing. Information on the transported ore tons also helps the
preplanning of the maintenance actions of the machines. Results
The used ore weighing method was based on measurements of
hydraulic pressure in a boom lifting cylinder. Additional measurements such as
slope angle, boom position and the temperature of hydraulic oil along with the
pressure signal, are carried out to take into consideration the effects caused
by the movement of the machine during the pressure measurements. Before the
measured values are inputted into the neural network or fuzzy system an
efficient data preprocessing method is used to identify accurate signal levels.Project Information
Participants
Project dates
March 1, 1998 – December 31, 1999Project volume
Total budget FIM 230.000, 8 man monthsProject manager
Professor Heikki Koivo
Helsinki University of Technology, Control Engineering
Laboratory
P.O. Box 5400, FIN-02015 HUT
Tel: (09) 451 5200 / (050) 515 1270
Fax: (09) 451 5208
E-mail: heikki.koivo@hut.fi
jukka.iivarinen@hut.fi
http://www.cis.hut.fi/neuronet/Tekes/8.shtml
Wednesday, 29-Nov-2000 10:29:56 EET