2. Dynamic systems modeling (DYSMO)

2.1 Elevator dynamics modeling and simulation

Abstract

The main goal of the project was to develop a method for building simulation models of elevators for the development of new control schemes.

VTT Automation was responsible of the experimental design that was carried out in close cooperation with Kone Elevators and the Automation and Control Institute of Tampere University of Technology (TTKK). The actual measurements were performed by Kone Elevators.

Results

In the simulation models, the friction forces are described by feedforward neural networks and the viscoelastic properties of the rope depend on rope length. The rope tension could be predicted with sufficient accuracy for simulation purposes.

Simulators of the both elevator types were implemented by using Matlab and are described in detail in a report also describing the models [1]. An analysis of the starting vibrations of an existing elevator was also performed and reported.

The incorporation of the very strongly nonlinear friction force models into the dynamic simulator revealed a general problem related to neural network and other empirical models that are based on process data. Real process data are always strongly correlated due to automated feedback controllers and operator control. Thus the interesting region in the data space is only partly covered. The neural networks models that accurately approximate strong nonlinearities may give unrealistic predictions even slightly outside the region covered by the available observations used for the training and validation.

A principle was developed for avoiding the use of a neural network model outside its validity domain [3]. The principle is based on clustering the regressor training data and projecting the input vector to the validity domain defined by the cluster centers. The confidence level for the projection can be specified. In the friction force model, Fuzzy C-means clustering was utilized. Another example of a lime kiln process was also reported in the paper describing the general method [3].

The model development methods and experimental designs can be applied to new elevator types. The simulators can be used in elevator design, ride planning, control development, analysis of vibration problems, and training. The developed instrumentation and experimental designs for model identification have already been used for the analysis of new elevator types.

Project information

Participants

• Kone Elevators
• VTT Automation

Project dates

Project volume

Project manager

Publications

2. Lehtinen H., J.J. Hämäläinen, P. Torenius and T. Tyni. Simulation of elevator dynamics. 2nd Tampere International Conference on Machine Automation, ICMA´98, 15-18-September 1998, Tampere, Finland, 1998.

3. Hämäläinen J.J and I. Järvimäki. Input projection method for safe use of neural networks based on process data. Proceeding of IEEE World Congress on Computational Intelligence, IJCNN ’98 , 4-9 May 1998, Alaska, 1998.

2.2 Modeling, analysis and control of combustion

Abstract

The goal of the project was to (1) develop methods for indirect estimation of fuel quality by using only the standard measurements available in the plant and (2) utilize the methods for fast compensation of the changes in quality and enhance plant control.

The project was carried out in close cooperation with VTT Automation and a Finnish power company, Imatran Voima Oy (IVO). The collection of data and testing of the methods was performed by IVO. The data were analyzed and methods developed by VTT Automation. IVO was responsible for implementing the new method in the process control system of an existing power plant.

Results

The calculation of FQV was tested by extensive off-line simulations and also on-line in the plant. The analysis of several variables (11) by using PCA (Principal Component Analysis) and SOM (Self Organizing Map) confirmed that FQV reliably estimated fuel quality. A method was developed for the utilization of the index in plant control. The method was tested by simulations and implemented in a power plant. The initial on-line testing of the new control scheme revealed that some changes are still necessary in the process control system. These changes can only be made during service intervals and will be completed in June 1998.

Two new models were developed and compared for the estimation of fuel power [1]. In the more traditional Hammerstein model the oxygen cost of combustion was estimated based on the total air flow feed and flue gas oxygen concentration. The steam power could then be predicted by using an ARX structure. The developed model predicted the power with sufficient accuracy.

A dynamic nonlinear feedforward neural network model was also identified. The model inputs were the air-flow and flue gas oxygen concentration signals with only standard filtering but without explicit calculation of the oxygen cost. The predictions of the neural network model were slightly more accurate.

The conclusion was that in spite of slightly better predictions given by the neural network the Hammerstein model is more convenient in this application. The recalibration of the Hammerstein model is straightforward and can be automated if necessary. The retraining of the neural network demands experienced R&D staff and is not so straightforward.

The filtering of the signals and the determination of the related delays was very demanding. The delays could only be estimated by fitting low order ARX models between the input and output signals and the determination of the delay by trial and error.

During the last year of the project, methods were tested for the monitoring and visualization of the process [2]. The goal was to evaluate the information content of the several temperature measurements of the bed and also the reliability of FQV as an indicator of fuel quality. The results showed that the temperature distribution was not significant and the first principal component of the temperatures could be used.

The analysis of eleven (11) variables including the estimated FQV by PCA and SOM component surfaces proved to give similar results. The expected relationships between the variables could be observed and the results confirmed FQV as a reliable fuel quality estimator. New ideas for operator support monitoring and visualization emerged, but were not realized in this project.

Two computer programs were developed: (1) off-line simulation of FQV with monitors to the selected most important variables and (2) off-line prediction of steam power by Hammerstein and NN models.

IVO is responsible for further testing of FQV and the control algorithms. The company has already decided to start a product development project if the results of the final tests are positive.

Project information

Participants

• Imatran Voima Oy
• VTT Automation

Project dates

Project volume

Project manager

Publications

Bärman R. and J.J. Hämäläinen. Modeling, Analysis and Comtrol of Combustion (Turve- ja biopolttoainevoimalaitosten mittaustiedon jalostaminen ja palamisen hallinnan kehittäminen). Report for Imatran Voima, 1998, 100 p. (confidential, in Finnish).

2.3 Multiparameter polynomial models for monitoring and management of fermentation and food processes

Publications

Visala A. & Halme A. (1996): Recognition of Functional States on the Basis of Laguerre Presentation. CESA'96 IMACS Multiconference Symposium on Control, Optimization and Supervision, Lille, France, July 9-12 1996, Vol 1, pages 263-268.

Visala A., Pitkänen H. and Halme A. (1997): Describing the dynamics of a chromatographic separation process with Wiener-type NN-model. A poster in 10th International Conference in Aqueous Two-Phase Systems (SCI), 10-15 August 1997, Reading, UK.

Visala A., Pitkänen H. and Halme A. (1997): Wiener-type SOM- and MLP-classifiers for recognition of the dynamic modes. In Artificial Neural Networks ICANN'97, 7th International Conference, Lausanne, Switzerland, October 1997, Proceedings, edited by W. Gerstner, M. Hasler, J.-D. Nicoud, Lecture Notes in Computer Science 1327, Springer.

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 Halme A. (1998): Describing the dynamics of a chromatographic separation process with Wiener-type NN-model. Practical Manuscript, approved for publication in Journal of Chromatography B.

Visala A., Pitkänen H. and Halme A. (199X): Describing the dynamics of a chromatographic separation process with Wiener-type NN-model. Manuscript, invited to resubmit after revision for publication in Journal of process control.

2.4 Neural network based system for forecasting district heat load

Abstract

Results

The main input arguments are the values of ambient temperature and district heat generation. The dynamics of the district heat network are brought into the model by using various filtered ambient temperatures as input variables. The problematic nature of human behavior was overcome by implementing time and weekday information into a sparse multimodel network structure. The multimodel structure com­bines a nominal model which functions as a forecaster for Tuesday to Thursday and three special mod­els for Monday, Friday and Saturday-Sunday time periods. The last three models take into account the rhythms caused by industry and regional habits e.g. heating up office buildings after a weekend and common sauna days.

The basic structure of the neural network model was built and verified during the first project year. The model is a standard nonlinear autoregressive model with a moving average (NARX). The implementa­tion is a standard, fully connected, multilayer perceptron network and the parameters were optimized using the Levenberg-Marquardt optimization algorithm.

During the second and third years it became obvious that the main goal from the industrial point of view was a model which can function reliably on a different heat load network without any training until the first winter has passed. This problem was solved by introducing a median filter construction and a database. The median filters are used to remove the one to two hour abnormal peaks from the heat load information. The database stores the previous corrections and functions as a memory for the forecaster enabling the model to function within a few weeks. The database collects the errors between the model and the current situation and then the error correction algorithm adds a correction into the model’s output. This construction enables adaptation of the model to the current situation whether it is a new district heat load region or a modification of an old network.

The forecaster with error correction can forecast 48 hours ahead with 10% accuracy. The forecaster parameters were optimized using the data available from years 1995-96 and in the validation data from year 1996-97 was used. The 10% accuracy was truly a good performance since there were some heat production numbers missing from the validation data. Encouraged by these results, final validations were decided and comparisons between the time series model, the model without the error correc­tion and the model with the error correction which are taking place. The validations are based on the earlier optimized forecasters and data from winter 1997-98. The industrial partner decides after these test in which form they will implement the forecaster into their new process plant information sys­tem.

The development of a research project into a successful product requires both complete documentation of the methodology and engineering solutions required for fault tolerant industrial application. These requirements were emphasized during the final project year to guarantee the means for direct develop­ment into a standard part of the district heat product plant’s optimization tools. This documentation will be completed during the first half of April.

Project information

Participants

Project dates

The project evolved during the work and also the milestone dates and goals evolved with the project. The project dates presented here are a modification of the original dates due the changed project goals.

Project year 1995-96

• project started 3/95
• Imatran Voima (IVO) selected the test plant and received the first data set 5/95
• validation of the first data set was unsuccessful due to data corruption 5/95
• IVO received the second data set 6/95
• the initial model was developed 8/95
• the final version of the neural network based forecasting system was finished 11/95
• comparison between traditional and neural network forecasters 12/95
• IVO decided to continue its support of this project 12/95

Project year 1996-97

• preparation of report on the previous year and planning details for the new project year 3/96
• coding of neural network algorithms in C 5/96
• prioritizing of training data for automated retraining 6-8/96
• coding the data prioritizing and validation algorithms 9-10/96
• defining the major features of the interface for forecaster training 8/96
• testing the portability of the neural network based forecasting system 11/96
• preparation of report on the project year 1996-97 12/96

Project year 1997-98

• testing new possible inputs 3/97
  These test were postponed due to the lack of necessary data. There was also a personnel change during the spring when Olli Lehtoranta (M.Sc.) was transferred to CCC Software Systems and Jari Seppälä (M.Sc.) took his place in this project.
• user interface development 6/97
  The user interface was deferred and the efforts were put into developing the error correction algorithm and database connection.
• testing on a real district heat plant 6/97
  The tests were not conducted because the error correction was not implemented in C, which is a requirement for real plant tests.
• implementing model improvements 10/97
  The model improvements were successfully implemented as scheduled.
• reporting 2/98
  The final reports are due to be completed in the middle of April. The delays are caused by additions for product development purposes.

Project volume

Project manager

Publications

[2] Lehtoranta, O. Seppälä, J. Koivisto, H. Koivo, H. 1997. Neural Network Based District Heat Load ForecastingPoster. Automaatiopäivät 97, Helsinki.

[3] Koivisto, H. 1995. A practical approach to model based neural network control. Dissertation. Tampere. Tampere University of Technology, Department of Electrical Engineering.

[4] Seppälä, J. Koivisto, H. and Koivo, H. 1998. Adaptive Neural Network Based District Heat Load forecaster. To be submitted.

2.5 Neural network model for elevator dynamics

Abstract

The work carried out was divided into four main components, a) designing a portable measurement strategy which holds all the necessary tests for the identification of an elevator, b) modeling the test eleva­tor, c) real-time verification of the model and d) designing a model based control strategy.

The results show that the empirical model implemented within a neural network framework was able to represent the real process in fine detail. Since both the physical and black-box model needed a certain amount of experimentation due to unknown physical dependencies, a neural network was identi­fied with a little extra effort compared to the first principles model.

A portable measurement plan was designed including all necessary measurements needed in the identification of a mechanistic and neural network model. Portability was tested when Kone Oy person­nel, encouraged by good results, decided to change the test elevator into a different type of construc­tion which has more non-linear aspects. During the design of new tests only minor changes needed to be made to the previously designed measurement plan.

During the project, non-linear output-error neural network (NOE) models for elevator dynamics and motor torque were developed. This type of recurrent non-linear dynamical model represents the deterministic part of the process and is especially suitable for simulation purposes like control design, for instance.

The elevator model is a hybrid linear-non-linear neural network. Non-linear part of the model describes the dynamics from selected inputs to the car acceleration which is essential to ride com­fort. The linear part implements the physical dependency of a = v’ = x’’, i.e. the integral of accelera­tion is velocity and the integral of velocity is distance. This ’gray-box’ modeling scheme incorporates physical process knowledge into the neural network structure, resulting in a less complex identification task and a more robust and reliable model. The identified network model is also suitable for on-line control. The motor torque model is a standard recurrent fully connected neural network model which is capable of forecasting the output torque of direct current motor and as such is an essen­tial part of the new torque based control scheme design.

The hybrid linear-nonlinear network was combined into one sparse network. This construction and parameter optimization scheme was carried out by using advanced software tool developed in an earlier project. Use of a recurrent neural network (NOE) was essential for obtaining reliable results. This requires both a proper recurrent gradient computation algorithm and an advanced parameter update algorithm. The Levenberg-Marquardt algorithm was used in this study.

The both models were combined into a modular simulation environment. The environment is imple­mented as a Matlab Simulink model and consists of three blocks. The first block contains the elevator control system, the second contains the model for direct current motor and the third contains the ele­vator dynamics model. The modular environment enables any of the blocks to be changed which results in a simulation system that combines the ability to easily use conventional models or sophisti­cated models in place of an elevator dynamics model and motor torque model. The module for the control system enables the design of new control schemes just by changing the block or testing and improving currently used schemes by offline simulation.

The results show that the empirical model implemented within a neural network framework was able to represent the real processes in fine detail. Since both the physical and black-box model required a certain amount of experimentation due to unknown physical dependencies, a neural network was identi­fied with a little extra effort compared to the first principles model.

Project information

Participants

Project dates

The design of the portable measurement strategy was finished slightly behind schedule due to the instrumentation problems affecting the measurement plan. The neural network model was finished on time although improvement of the model has continued. The real-time verification and model based control strategy design have been re-scheduled due to good results in modeling which led to chang­ing the test elevator. During the project year 1996 was decided to leave out the real time tests and concentrate on finishing the modular test environment.

Project year 1995-96

• actual project started 3/95
  test elevator was chosen during a feasibility study
• preliminary data set was created 5/95
• preliminary models were ready 6/96
• final model was created 11/95
  improvement of model continues until the end of project year
• Kone decided to continue project 12/95
• new test elevator was chosen (different type) 1/96
• measurement plan for new elevator was created 2/96

Project year 1996

• selection of the new test elevator 1/96
• preliminary tests with hybrid model 2/96
• design of test drives 3/96
• the instrumentation of the test elevator and test runs 5/96
• model identification and validation 9/96
• the design of new control schemes 10/96
• reporting 12/96

Project volume

Project manager

Publications

2. Seppälä, J. Koivisto, H and Koivo, H. Modelling Elevator Dynamics using Neural Networks. International Joint Conference on Neural Networks '98, Anchorage, Alaska, May 4-9, 1998.

3. Seppälä, J. Koivisto, H and Koivo, H. Neural Network Model for Elevator Dynamics. Automaatio-97, Helsinki, September 23-24, 1997.

4. Koivisto, H. 1995. A practical approach to model based neural network control.
Doctoral dissertation. Tampere. Tampere University of Technology, Department of Electrical Engineering.

5. Tarkiainen, M. 1995. Modelling of elevator and effect of speed control on car
vibrations (in Finnish). Master of Science Thesis. Tampere. Tampere University of Technology, Department of Electrical Engineering. Confidential.

jukka.iivarinen@hut.fi
http://www.cis.hut.fi/neuronet/Tekes/2.shtml
Tuesday, 28-Nov-2000 15:29:03 EET