9. Intelligent methods for processing and exploration of signals and systems (IMPRESS)

Abstract

Results

In the Handwriting Recognition Research project, adaptive methods for on-line recognition of isolated characters have been developed. These methods either adapt a single prototype-based classifier to new writing styles by modifying the prototype set or, alternatively, refine the set of decision rules of a committee consisting of different classifiers. The two forms of adaptation can be carried out simultaneously or sequentially. Adaptation is performed in a self-supervised fashion. Methods for increasing the system's tolerance to erroneous learning samples have been studied. The recognition system is now being implemented in a palm-sized computer in order to collect experience on the actual use of adaptive on-line recognition for text input.

The results have been or will soon be published as one doctoral and one master's thesis, three chapters in edited books, one journal article, and ten refereed conference articles.

Project information

Participants, dates and volume

• ABB Pulp and Paper (Visual Fault Analysis of a Running Paper Web)
• Nokia Research Center (Handwriting Recognition Research)

Project leader

Publications

[2] Jukka Iivarinen and Ari Visa. Unsupervised Image Segmentation with the Self-Organizing Map and Statistical Methods. In D. P. Casasent (Ed.), Intelligent Robots and Computer Vision XVII: Algorithms, Techniques, and Active Vision, Proc. SPIE 3522, pages 516-526, 1998.

[3] Jukka Iivarinen and Juhani Rauhamaa. Surface Inspection of Web Materials Using the Self-Organizing Map. In D. P. Casasent (Ed.), Intelligent Robots and Computer Vision XVII: Algorithms, Techniques, and Active Vision, Proc. SPIE 3522, pages 96-103, 1998.

[4] Jukka Iivarinen. Texture Segmentation and Shape Classification with Histogram Techniques and Self-Organizing Maps. Acta Polytechnica Scandinavica, Mathematics, Computing and Management in Engineering Series No. 95, Espoo 1998.

[5] Jukka Iivarinen. Unsupervised Segmentation of Surface Defects with Simple Texture Measures. In Workshop of Texture Analysis in Machine Vision, pages 53-58, Oulu, Finland, June 14-15 1999.

[6] Katriina Heikkinen and Petri Vuorimaa. Computation of Two Texture Features in Hardware. In Proceedings of the 10th International Conference on Image Analysis and  Processing, Venice, Italy, pages 125-129, September 27-29, 1999.

[7] Jukka Iivarinen. Unsupervised Segmentation of Surface Defects with Simple Texture Measures. To appear in M. Pietikäinen (Ed.), Texture Analysis in Machine Vision, Series on Machine Perception and Artificial Intelligence, World Scientific.

[8] Jukka Iivarinen, Katriina Heikkinen, Juhani Rauhamaa, Petri Vuorimaa, and Ari Visa. A Defect Detection Scheme for Web Surface Inspection. To appear in International Journal of Pattern Recognition and Artificial Intelligence.

[9] Jorma Laaksonen, Matti Aksela, Erkki Oja, and Jari Kangas. Adaptive local subspace classifier in on-line recognition of handwritten characters. In Proceedings of International Joint Conference on Neural Networks 1999.

[10] Jorma Laaksonen, Matti Aksela, Erkki Oja, and Jari Kangas. Dynamically Expanding Context as committee adaptation method in on-line recognition of handwritten latin characters. In Proceedings of International Conference on Document Analysis and Recognition, pages 796-799, 1999.

[11] Jorma Laaksonen, Jarmo Hurri, Erkki Oja, and Jari Kangas. Comparison of adaptive strategies for on-line character recognition. In Proceedings of International Conference on Artificial Neural Networks, pages 245-250, 1998.

[12] Jorma Laaksonen, Jarmo Hurri, Erkki Oja, and Jari Kangas. Experiments with a self-supervised adaptive classification strategy in on-line recognition of isolated handwritten latin characters. In Proceedings of Sixth International Workshop on Frontiers in Handwriting Recognition, pages 475-484, August 1998.

[13] Jorma Laaksonen, Vuokko Vuori, Erkki Oja, and Jari Kangas. Adaptation of prototype sets in on-line recognition of isolated handwritten latin characters. In Seong-Whan Lee, editor, Advances in Handwriting Recognition, pages 489-497. World Scientific Publishing, 1999.

[14] Jorma Laaksonen, Vuokko Vuori, Matti Aksela, Erkki Oja, and Jari Kangas. Experiments with Adaptation Methods in On-line Recognition of Isolated Latin Characters. To appear in a book edited by Nabeel Murshed, 2000.

[15] Vuokko Vuori. Adaptation in on-line recognition of handwriting. Master's thesis, Helsinki University of Technology, 1999.

[16] Vuokko Vuori, Jorma Laaksonen, Erkki Oja, and Jari Kangas. On-line adaptation in recognition of handwritten alphanumeric characters. In Proceedings of International Conference on Document Analysis and Recognition, pages 792-795, 1999.

9.2 Analysis of annual reports by advanced neural network methods

Abstract

Results

Project information

Participants, dates and volume

• Ramse Consulting Oy
• Teollisuuden Voima Oy

Project manager

Publications

The following list of publications consists of those that have been published or have been accepted for publication.

Refereed international conference papers

[2]  Back, B., Vanharanta, H., Toivonen, V., Visa, A., Toward Computer Aided Analysis of Text in Annual Reports, in Proc. of 2nd European Conference on Accounting Information Systems (CD-ROM), Bordeaux, France, May 3-4, p 8, 1999.

[3]  Visa, A., Toivanen, J., Back, B., Vanharanta, H., Towards Text Understanding - Comparison of Text Documents by Sentence Map, in Proceedings of EUFIT 99 (CD-ROM), Aachen, Germany, September 13-16, p.6, 1999.

[4]  Ari Visa, Jarmo Toivonen, Barbro Back, Hannu Vanharanta, Knowledge Discovery From Text Documents Based On Context Maps, in Proceedings of the 33rd Annual Hawaian International Conference on System Sciences (HICSS, CD-ROM), Maui, Hawaii, January 4-7, p. 9, 2000.

[5]  Ari Visa, Jarmo Toivonen, Barbro Back, Hannu Vanharanta, Toward text understanding: classification of text documents by word map, To be published in Proc of SPIE's International Symposium on Optical Engineering in Aerospace Sensing, Vol xxxx, Data Mining and Knowledge Discovery: Theory, Tools, and Technology II, Orlando, Florida, April 24-28, pp. xxx-xxx, 2000.

9.3 Data analysis and representation by neural networks (DAEMON)

Summary

(a) Document matching,  where the aim was to find similar documents from a database that consists of reports about data communications  network  fault diagnostics. Document matching was based on textual similarity between a new textual description and ca. 100 000 old written reports of fault symptoms and their corresponding corrections.

(b) Air quality prediction, where the aim was to estimate and monitor the development of urban air quality based on weather service information.

Main results

In substask (b) a prototype software was developed and tested with data from the cities of Kuopio, Imatra, Stockholm and Singapore. An experimental web-based monitoring system was also installed in Kuopio.

In addition, the project enabled technology transfer in the Jyväskylä and Kuopio areas that resulted the following outcomes of the research:

• A newly founded company, Visipoint Oy, that applies NDA for environmental technology and bioinformatics. See web pages at  http://www.visipoint.fi/ for more information.
• 2-3 doctoral theses that were partially supported by the project.
• Two new European (EC-funded) projects (Appetise and Euredit) were started to continue the research.

Figure 1. Gene pattern recognition in Kuopio with self-organizing maps.

Project information

Participants, dates and volume

Work was carried out in cooperation with Sonera Oyj, Greenwin Oy, Visipoint Oy,  Kuopio Centre of Expertise, Savon liitto and the Ministry of the Environment.

Project leader

Publications

Already published

Lensu, A. and Koikkalainen, P. (1998), 'Analysis of Gallup Questionnaires through Self-Organizing Maps'. In proc. STeP'98: 8th Finnish Conference on Artificial Intelligence. Picaset, pp. 171-180.

Häkkinen, E. and Koikkalainen P., (1998) 'Linked Data Representations: Analyzing and Visualizing Data and Knowledge'. In proc. STeP'98: 8th Finnish Conference on Artificial Intelligence. Picaset, Helsinki, 1998. pp. 11-20.

Rekkilä M., E. and Koikkalainen P. (1998), Configuration of a two-layer MLP-network, In proc. STeP'98: 8th Finnish conference on Artificial Intelligence, pp. 161-170.

Mohammed T., and Koikkalainen P. (1998), A Methodology for the use of Self Organizing Maps for Process Monitoring, In proc. STeP'98: 8th Finnish conference on Artificial Intelligence, pp. 181-188.

Lensu, A. and Koikkalainen, P.(1999). 'Laskennallisesti älykkäiden menetelmien käyttö laadullisen tutkimuksen apuna' (In Finnish). In proc. Tiedosta tutkittua - Raportti Tiedon tutkimusohjelman I tutkijaseminaarista. May 6, 1999, Tampere, Finland. Tampereen Yliopistopaino, Tampere, 1999. pp. 35-39. Koikkalainen, P. (1999): Tree Structured Self-Organizing Maps, Kohonen Maps, E.Oja and S. Kaski Eds., Elsevier Science, 121-130.

Hautamäki, J., Honkanen, T. and Koikkalainen, P. (1999), Visualization of the  Traffic Type in Customer Networks with Neural Networks, In proc. IEEE Malaysia international conference on communication, IEEE Press.

Lensu, A. and Koikkalainen. P. (1999). 'Similar Document Detection using Self-Organizing Maps'. In proc. KES'99: Third International Conference on Knowledge-Based Intelligent Information Engineering Systems, IEEE Press, pp. 174-177.

Törhönen, P., Kolehmainen, M., Wong, G., Castren, E. (1999), Analysis of Gene Expression Data using Self-Organizing Maps, FBES Letters 451, pp. 142-146.

Submitted manuscripts

9.4 Data-fusion and neural networks in complex models

Abstract

The main application problem in the project was to develop a computational model for consumer behavior in the choice of store, based on the store sales, profiles and locations, road networks, and other geographical data. An essential requirement was that questionnarie or polling based data is not required, to make the system easily adaptable to new locations. This also invalidated most of the existing studies in this domain. The practical goal was to support the industrial partners in developing an application for aiding in the design of the communal infrastructure (e.g. to minimize traffic loads) or to aid in determining optimal store locations. A further goal was to replace the GIS based data by remote sensing information (satellite images), so that the system would also be applicable in developing countries in eastern Europe and Asia, which lack highly-developed GIS systems.

Another task was related to estimation of velocity profiles from observed Doppler spectra. We have studied a completely novel approach, which has considerable potential by facilitating the use of narrow band radar signals. The task has required the development of an accurate model for the formation of the Doppler spectrum in the studied application (in collaboration with the Dept. of Mathematics, HUT), and a novel inverse method to recover the velocity profile from the spectrum.

A small task in the project was to support the Laboratory of Metallurgy, HUT, in using neural network methods in various steel production problems. There we have applied the advanced neural network modelling tools developed in earlier projects.

Results

In the satellite image task we have anticipated IKONOS satellite data by aerial images with matching 1m resolution. We have developed a method for segmenting and classifying the images to recover the population and road information required in the modeling task. The results are, however, more generic, with potential application in, e.g., quick analysis of destructions in cathastrophy areas, such as earthquakes.

In the Doppler radar task the major result has been a novel statistical inverse method to recover the  velocity profile from the Doppler spectrum. The profile is parameterized as a neural network (MLP)  curve, and Maximum Likelihood and Bayesian MCMC methods are used to estimate the network parameters, given the Doppler spectrum and the forward model which transforms the velocity profile to the spectrum. The results are very promising, but further work is needed to assess the effect of uncertainties in the forward model.

Project information

Participants

• Laboratory of Computational Engineering, HUT (Dept. of Mathematics, HUT,  Lab. of Metallurgy, HUT)
• Vaisala Ltd
• Soil and Water Ltd
• Kouvola Region Federation of Municipalities
• Finnish Metal Industry Consortium (Rautaruukki Ltd, Outokumpu Polarit Ltd, Fundia Wire Ltd, Imatra Steel Ltd)

Project dates

Project volume

Project manager

Publications

Figure 1. The figure shows an example analysis of grocery store demand distribution in the Kouvola region. The black dots indicate the stores in the region, and the colored house icons show the distribution of population. Based on these data and total sales of the stores, the developed system estimates the utility parameters for different store types. The figure shows an example output of  the model. The red box with white asterisk shows the largest hypermarket in the region, and the colors of the house icons indicate the predicted share of money spent in that hypermarket of all the funds used in grocery stores. The width of the map area is about 50 km.

9.5 Data mining and analysis using the self-organizing map

Abstract

The research was focused on the Self-Organizing Map (SOM), which is an unsupervised neural network method that can be efficiently used to form a visual display of high-dimensional process data. Different presentations of the SOM make it possible to detect non-linear and local dependencies between process measurements. The SOM can also be used for clustering data without knowing the class memberships of the input data. In process data analysis this means extraction and identification of different operational states of the process.

Use of the SOM makes it possible to efficiently investigate the behaviour of an industrial process and, for instance, to optimise and control the system. By considering the material flows and output products, the state of the process can be monitored from different points of view.

Results

The scientific results of the project include 14 international publications: (1) 5 journal articles, (2) 4 book chapters and (3) 5 conference papers listed below in section 4. Note that the results reported in the publications have not always been obtained only during the IMPRESS project.

Project information

Participants

• UPM-Kymmene / Wisaforest Oy
• Jaakko Pöyry Consulting Oy
• Metsäteho Oy
• Rautaruukki Raahe Steel, Outokumpu Polarit Oy, Fundia Wire Oy Ab, Imatra Steel Oy Ab

Project dates

Project volume

Project manager

Publications

2. Esa Alhoniemi. Prosessin mittauksiin perustuva sulfaattisellun keiton analyysi. Licenciate's thesis, Helsinki University of Technology, August 1998. (In Finnish.)

3. Esa Alhoniemi, Johan Himberg, and Juha Vesanto. Probabilistic Measures for Responses of Self-Organizing Map Units. In H. Bothe, E. Oja, E. Massad, and C. Haefke, editors, Proceedings of the International ICSC Congress on Comutational Intelligence Methods and Applications (CIMA '99), pages 286-290. ICSC Academic Press, 1999.

4. Esa Alhoniemi, Jaakko Hollmen, Olli Simula, and Juha Vesanto. Process Monitoring and Modeling Using the Self-Organizing Map. Integrated Computer-Aided Engineering, 6(1):3-14, 1999.

5. Esa Alhoniemi. Analysis of Pulping Data Using the Self-Organing Map. Submitted to Tappi Journal.

6. Johan Himberg, Jussi Ahola, Esa Alhoniemi, Juha Vesanto, and Olli Simula. Feature analysis, clustering and classification: soft computing approaches, chapter: The Self-Organizing Map as a Tool in Knowledge Engineering. World Scientific. To appear.

7. Pekka Hippeläinen. Cluster Analysis of Forest Parcel Data. Master thesis, Helsinki University of Technology, October 1999.

8. Olli Simula, Juha Vesanto, Petri Vasara, and Riina-Riitta Helminen. Industrial Applications of Neural Networks (L.C. Jain and V.R.Vemuri, eds), chapter 4: The Self-Organizing Map in Industry Analysis, pages 87-112. CRC Press, 1999.

9. Olli Simula, Juha Vesanto, Esa Alhoniemi, and Jaakko Hollmen. Neuro-Fuzzy Techniques for Intelligent Information Systems (N. Kasabov and R. Kozma, eds), chapter: Analysis and Modeling of Complex Systems Using the Self-Organizing Map. Springer, 1999.

10. Olli Simula and Esa Alhoniemi. SOM Based Analysis of Pulping Process Data. In Proceedings of International Work-Conference on Artificial and Natural Neural Networks (IWANN '99), volume II, pages 567-577. Springer, 1999.

11. Olli Simula, Jussi Ahola, Esa Alhoniemi, Johan Himberg, and Juha Vesanto. Kohonen Maps (E. Oja and S. Kaski, eds.), chapter: Self-Organizing Map in Analysis of Large-Scale Industrial Systems. Elsevier, 1999.

12. Juha Vesanto, Johan Himberg, Markus Siponen, and Olli Simula. Enhancing SOM Based Data Visualization. In T. Yamakawa and G. Matsumoto, editors, Proceedings of the 5^th International Conference on Soft Computing and Information/Intelligent Systems, pages 64-67. World Scientific, 1998.

13. Juha Vesanto, Esa Alhoniemi, Johan Himberg, Kimmo Kiviluoto, and Jukka Parviainen. Self-Organizing Map for Data Mining in Matlab: the SOM Toolbox. Simulation News Europe, (25):54, March 1999.

14. Juha Vesanto. SOM-Based Data Visualization Methods. Intelligent Data Analysis, 3(2):111-126, 1999.

15. Juha Vesanto and Esa Alhoniemi. Clustering of the Self-Organizing Map. Submitted to IEEE Transactions on Neural Networks.

16. Juha Vesanto, Johan Himberg, Esa Alhoniemi, and Juha Parhankangas. Self-organizing map in Matlab: the SOM Toolbox. In Proceedings of the Matlab DSP Conference 1999, pages 35-40, Espoo, Finland, November 1999.

Figure 1. A SOM is presented using color coding that visualizes cluster structure of the map trained with steel slab temperature measurements. The histograms on units indicate variable values of each model vector.

Figure 2. A steel slab is dyed using the coloring in order to analyze the temperature behaviour during casting.

9.6 Intelligent signal processing

Abstract

The main goal in this research project was to utilize adaptive and intelligent methods especially for signal processing related applications. The following applications were considered: speech recognition, speech coding, image interpolation, bar-code recognition, channel equalization, enhancement of digital x-ray images and passive detection of moving targets. The obtained results have provided valuable information about the performance of the new methods, and understanding of the main problems of each industrial application has increased substantially.

Results

In speech recognition research, a hybrid speech recognition system for continuous spoken digit recognition has been implemented. In this context techniques for noise compensation and speaker adaptation have also been investigated.  In speech coding research, a waveform interpolation coder has been implemented, and several modifications have been proposed for the basic system. In image interpolation research, the performance of neural networks has been investigated for this problem. In bar-code recognition research, a system that recognizes bar codes from digital images has been developed and implemented. In channel equalization research, neural networks and other nonlinear machine learning methods have been considered for adaptive equalization of binary data bursts. In enhancement of digital x-ray images research, several advanced computational methods have been developed and implemented in order to improve the quality of the x-ray images and simplify the image acquisition system. In passive detection of moving targets research, systems for passive helicopter detection and for passive underwater acoustic detection were developed. The results of these studies have been published in several conference proceedings and journals. In addition, a number of theses have appeared as a result of the studies. However, some of the results are yet to be published.

Project information

Participants, dates and volume

The research was performed in the Signal Processing Laboratory at Tampere University of Technology, Finland during 1.3.1998 – 31.12.1999. The budget of this two-year project was roughly FIM 3.1 million, and the amount of work performed was about 140 person months. The work was carried out in cooperation with the following industrial companies:

• Nokia Research Center (speech coding and recognition, image interpolation, bar-code recognition)
• Nokia Mobile Phones (channel equalization)
• Oy Imix Ab (enhancement of digital x-ray images)
• Patria Finavitec Systems (passive detection of moving targets)

Project leader

Prof. Jukka Saarinen
Signal Processing Laboratory
Tampere University of Technology
P.O. Box 553, 33101 Tampere, Finland
Tel: +358-3-365 3397
Fax: +358-3-365 3095
E-mail: jukkas@cs.tut.fi

Publications

To date, two international journal papers and fourteen international conference papers have been published as a result of this project. However, some of the results are yet to be published. Information about the future (as well as the past) publications can be found at the following web address: http://www.cs.tut.fi/~mikkol/algo.html

The following is a list of those articles that have so far been published.

Refereed international journal papers

[1]  A. Kantsila, M. Lehtokangas and J. Saarinen, "Burst Adaptive Equalization of Binary Data," Journal of Intel­ligent Systems, vol. 9, no. 2, 1999.

[2]  P. Salmela, M. Lehtokangas and J. Saarinen, "Neural Network based Digit Recognition System for Voice Di­alling in Noisy Environments," International Journal of Information Sciences (to appear).

Refereed international conference papers

[3]  M. Vuolahti and V. Ruoppila, "Performance of Predic­tive Vector Quantizers in Noisy Channels," Proceed­ings of IEEE Nordic Signal Processing Symposium, NORSIG’98, Jun. 8-11, 1998, Vigso, Denmark.

[4]  A. Kantsila, M. Lehtokangas and J. Saarinen, "Im­proved Neural Network Equalization by the use of Maximum Covariance Weight Initialization," Proceed­ings of European Signal Processing Conference, EU­SIPCO'98, Sep. 8-11, 1998, Rhodes, Greece.

[5]  P. Salmela, M. Lehtokangas and J. Saarinen, "Noisy Digit Recognition with MLP and HMMs using Modified Cost Function in MCE Training," Proceedings of Inter­national Conference on Neural Information Process­ing, ICONIP'98, Oct. 21-23, 1998, Kitakyushu, Japan.

[6]  M. Harju, P. Salmela, M. Lehtokangas and J. Saarinen, "Speaker Adaptation using Neural Networks in Hidden Markov Model based Speech Recognition System," Proceedings of International Conference on Modelling, Identification and Control, MIC’99, Feb. 15-18, 1999, Innsbruck, Austria.

[7]  A. Kantsila, M. Lehtokangas and J. Saarinen, "Adap­tive Equalization of Binary Data Bursts with Cascade-Correlation Trained Multilayer Perceptron Networks," Proceedings of International Conference on Modelling, Identification and Control, MIC’99, Feb. 15-18, 1999, Innsbruck, Austria.

[8]  M. Lammila, J. Yrjänäinen, M. Lehtokangas and J. Saarinen, "Edge Recognition by MLP Neural Net­work," Proceedings of International Conference on Modelling, Identification and Control, MIC’99, Feb. 15-18, 1999, Innsbruck, Austria.

[9]  V. Onnia, V. Varjonen, M. Lehtokangas and J. Saarin­en, "Removal of Grid Lines from Digitized X-ray Imag­es," Proceedings of International Conference on Modelling, Identification and Control, MIC’99, Feb. 15-18, 1999, Innsbruck, Austria.

[10]  M. Tammi, S. Kuusisto and J. Saarinen, "Modelling of Quasi-Periodic Signals using Waveform Interpolation," Proceedings of International Conference on Modelling, Identification and Control, MIC’99, Feb. 15-18, 1999, Innsbruck, Austria.

[11]  H. Haverinen, P. Salmela, J. Häkkinen, M. Lehtokan­gas and J. Saarinen, "Enhancement of Noisy MFCC Vectors using Context Dependent Multilayer Percep­tron Networks," Proceedings of Workshop on Robust Methods for Speech Recognition in Adverse Condi­tions, ROBUST'99, May 25-25, 1999, Tampere, Fin­land.

[12]  M. Tammi, V. Ruoppila, S. Kuusisto and J. Saarinen, "Coding Distortion caused by a Phase Difference Be­tween the LP Filter and Its Residual," Proceedings of 1999 IEEE Speech Coding Workshop, June 20-23, 1999, Porvoo, Finland.

[13]  H. Haverinen, P. Salmela, J. Häkkinen, M. Lehtokan­gas and J. Saarinen, "MLP Network for Enhancement of Noisy MFCC Vectors," Proceedings of European Conference on Speech Communication and Technolo­gy, EUROSPEECH'99, Sep. 5-9, 1999, Budapest, Hun­gary.

[14]  M. Harju, P. Salmela, O. Viikki, M. Lehtokangas and J. Saarinen, "Robust Speaker Adaptation of Continuous Density HMMs using Multilayer Perceptron Network," Proceedings of European Conference on Speech Communication and Technology, EUROSPEECH'99, Sep. 5-9, 1999, Budapest, Hungary.

[15]  P. Salmela, M. Lehtokangas, K. Laurila and J. Saarin­en, "On String Level MCE Training in MLP/HMM Speech Recognition System," Proceedings of 1999 IEEE Systems, Man, and Cybernetics Conference, IEEE-SMC'99, Oct. 12-15, 1999, Tokyo, Japan.

[16]  S. Viitanen, J. Syrjärinne, M. Lehtokangas and J. Saarinen, "Neural Networks in Identification and Clas­sification of Acoustic Helicopter Signals," Proceedings of Artificial Neural Networks in Engineering, AN­NIE’99, Nov. 7-10, 1999, St. Louis, U.S.A.

9.7 Nonlinear disturbance analysis

Abstract

This project concentrated on applying linear and nonlinear methods in two areas in process disturbance analysis.

1. Detection of abnormal process behavior
2. Analysis of the sources of process variation

Detection of abnormal behavior is handled by multivariate statistics. Principal components analysis (PCA) and partial least squares (PLS) were the linear methods and the self organizing map (SOM) was the nonlinear method used.

Analysis of the sources of process variation was handled by multivariate time series methods; multivariate (or vector)  autoregressive modeling (MAR, VAR) was the linear method and multi layer perceptron network model (MLP) was applied nonlinear method.

The linear methods were used for reference.

Results

The nonlinear methods were developed and applied to some test cases where they were compared to linear methods. The test case for detecting abnormal process behavior was a non-trivial real papermill quality change. Both PLS and SOM were found to be useful detection methods. The preferred method in practice depends on the case under study. Both methods will probably be implemented in KCL-WEDGE process analysis software represented by KCL Development Oy.

The test case for nonlinear variation analysis was a simulated process model of paper machine short circulation. The partial process was selected because it contains known nonlinearity. The results showed that the linear method (MAR) can still be applied to some extent. The nonlinear method (MLP) is much more expensive in computational terms. The nonlinear methods can be applied but they need further development.

Project information

Participants, dates and volume

The research was performed in the Laboratory of Measurement and Information Technology at Tampere University of Technology, Finland during 1.3.1998 – 31.12.1999. The budget of the project was FIM 500.000. M.Sc. Pekka Kumpulainen was the main researcher. The industrial partner was KCL Development Oy.

Project leader

M.Sc. Heimo Ihalainen
Tampere University of Technology
Measurement and Information Technology
P.O. Box 692, FIN-33101 Tampere, Finland
Tel: +358-3-365 2472
Fax: +358-3-365 2171
E-mail: Heimo.Ihalainen@mit.tut.fi

Publications

A conference paper on the detection of abnormal process behavior was presented and published in the proceedings of XV IMEKO World Congress. The final results, also containing the analysis of process variation, will be published in Measurement Journal.

[1]Kumpulainen P., Ihalainen H. and Ritala R., "Multivariate Disturbance Analysis", XV IMEKO World Congress, Osaka, June 1999.

9.8 Optical characterization of microstructures

Abstract

The goal in the project was to apply neural networks as statistical models in optical scatterometry. This is a technique in which the scattered intensity distribution of a coherent laser beam from a regular microstructure is measured, and the parameters of the structure are determined from the measurements. Since the problem is typically highly nonlinear and no analytical solution exists, statistical nonlinear regression has to be used instead. Neural networks provide a natural solution in this case.

Optical scatterometry is noncontact and utilizes only the intensity, not the phase, of the scattered light distribution. These aspects make this method a versatile and inexpensive tool in comparison to, ellipsometry, atomic force microscopy, and profilometry, for example, which either require expensive equipment or are inherently invasive. This project also aimed to build an experimental setup for optical scatterometry.

Results

We have proposed and successfully demonstrated numerically the utilization of a multi-layer perceptron neural network model to predict surface parameters of micro-structured plates with a nanometer-level accuracy. We have also introduced a hierarchical two-level system of neural networks in order to reduce the required amount of prior information of the samples. This is necessary to keep the computational toll low while calculating the training data for the neural network models. The numerical results have been presented in several international publications.

An experimental set-up has been built that enables the measurement of both the reflected and transmitted scattered light intensity distribution from one-dimensional microstructures. At the moment the accuracy of the method is approximately 5 promille of the intensity. The setup allows the user to change the angle of incidence of the probing laser beam on the sample and the state of polarization of the beam. Furthermore, lasers with other wavelengths can be easily used if required. The set-up has been tested with samples from the collaborating companies and the feasibility of the method has been verified.

Project information

Participants, dates and volume

The research was carried out in the Materials Physics Laboratory at Helsinki University of Technology from March 1, 1998 to December 31, 1999. The total budget of the project was FIM 455.000, and the total amount of work done was 26 person months. In this research we have cooperated with the following industrial companies:

• Terapixel Oy (contact person Arto Salin)
• Planar Systems, Inc. (contact person Martti Sonninen)

Project leader

Dr. Jyrki Saarinen
Heptagon Oy
Tekniikantie 12
FIN-02150 Espoo, Finland
Tel: +358 9 2517 2041
Fax: +358 9 2517 2042
E-mail: Jyrki.Saarinen@heptagon.fi

Project manager

Ilkka Kallioniemi
Materials Physics Laboratory
P.O. Box 2200
FIN-02015 HUT, Finland
Tel: +358 9 451 3157
Fax: +358 9 451 3164
E-mail: Ilkka.Kallioniemi@hut.fi

International publications

So far two articles have been published in refereed international journals and three papers have been published in international conferences concerning the results of this project. Furthermore, one Licentiate thesis and one Master’s thesis have been completed by persons involved in this project. Research results have been also been presented at the Annual Meetings of the Finnish Physical Society and the Finnish Optical Society.

[1] I. Kallioniemi, J. Saarinen, and E. Oja, ”Optical scatterometry of subwavelength diffraction gratings: neural network approach”, Applied Optics, Vol. 37 (1998).

[2] I. Kallioniemi, J. Saarinen, and E. Oja, ”Characterization of diffraction gratings in a rigorous domain with optical scatterometry: hieararchical neural network model”, Applied Optics, Vol. 38 (1999).

[3] I. Kallioniemi, J. Saarinen, and E. Oja, ”Neural network as a statistical model for optical scatterometry”, 1998 OSA Technical Digest Series, Vol. 10 (1998).

[4] I. Kallioniemi and J. Saarinen,”Optical scatterometry with neural network model for nondestructive measurement of submicron features,” Proceedings of SPIE, Vol. 3743 (1999).

[5] J. Saarinen, I. Kallioniemi, A. Niinistö, and A. T. Friberg, “Surface roughness measurement with optical scatterometry,” Advanced Photonic Sensors and Applications, R. A. Lieberman, A. K. Asundi, and H. Asanuma eds., 29.11.-3.12.1999, Singapore (in Press for Proceedings of SPIE).

Ilkka Kallioniemi, ”Analysis and Characterisation of Diffractive Microstructures,” Licentiate thesis (Helsinki University of Technology, 1998).

Anu Huttunen, ”Rigorous Analysis and Characterisation of Continuous Profile Diffraction Gratings,” Master’s thesis (Helsinki University of Technology, 1999).

9.9 The development of spatio-temporal neural networks for solving time-dependent problems

Abstract

This project developed signal processing methods based on neural and adaptive computing for applications in telecommunication and medical measurement. The research was divided into two main areas where the goals were to find methods for self-management of health and the quality of life and to develop new adaptive filtering methods to improve the quality of speech signals in digital audio devices.

Health monitoring system

The focus of the research was on spatially and temporally adaptive learning applications and techniques. A static and deterministic application cannot cope with all the temporally changing situations in, for example, industrial environments or notably dynamic environments, such as the human physiological system. Spatial and temporal adaptability is a solution to the dynamic environmental modelling problems. The research problem is not easy, however, and requires techniques to give ratings for different types of relevant information in different situations.

In the project, the TERVA data was used for algorithm development. The data includes a huge amount of registered bio-signals, such as heart rate, temperature, blood pressure as well as diary record about the person’s overall subjective health status. One goal was to find explanations for phenomena visible in bio-signals from the diary record. A software application has been developed for an experimental learning and adaptive system to recognise, for example,  stress situations in a person’s daily life.

The large amount of data becomes a problem, especially when collecting new data. The collected data archive should therefore be maintained with intelligent tools. This smart archive concept has already been implemented on a small scale by utilising short history records when solving different problems.

Results

The purpose of this research was to monitor a person´s health in everyday life with a simple and descriptive device.

Pre-processing

The pre-processing state was time consuming because of the huge amount of data. The data was analysed with different statistical methods, including cross-correlations, scatter plots, spectral analysis, decision trees and SOM. Medical expertise was also used in choosing the significant variables for the application.

Because adaptive systems need certain amount of history, missing data may cause difficulties. Method development was our main goal, and missing values were therefore replaced with short history averages. The diurnal variability of the signal was also taken into account.

Methods

Dynamic control limits were used to detect drastic changes in the signals. Another view was obtained by dividing the signals adaptively into three categories: low, medium and high. The advantage of the latter view was that more data were obtained into the extreme classes. This classificatory information was utilised in the final model as (0,1) indicators.

The limits are altered every time a new observation enters the system and only a short history of the signal is preserved for these calculations. In order to maintain processing without supervision, several parameters are selected automatically, depending on the particular signal. These concepts are part of the more extensive realisation of the smart archive.

Association rules and Bayesian networks were used to find relationships between the variables and their alarm indicators. This information was used later with SOM, which was applied to alarm combination and visual monitoring. 

Achievements

Physical exercise was easily recognised, because it affects different people in a similar way. It would be desirable to find explanations for the diary records and emotional feelings as well. The development of a more discriminatory system continues.

A significant difference has been found between morning and evening measurements. The daytime measurements  are usually similar to either the morning or the evening observations, depending on whether they were made earlier or later. This difference was due to measurements made without supervision. Also, working days made up independent clusters on the map, whereas weekends were not so clearly recognisable.

New methods for monitoring human physical signals in everyday life have been developed, but in order to obtain more reliable results, longer measurement periods with different life situations (stress, illness, holidays, etc.) are needed. In the pre-processing state, variables obtained with spectral analysis cannot be utilised with this data for the same reason. More data from different populations is being gathered continuously. 

Digital filtering of speech signal

The objective of the research was to develop adaptive digital filtering techniques which remove cyclostationary interference from speech signals. The audio interference caused by the electromagnetic field of a TDMA standard based transmitter induces interference to both the microphone and the headphone parts of a mobile phone. The main difference between these two cases is that the digital microphone signal can be processed by digital filtering methods, but the analog signal in the headphone parts cannot. The developed solution estimates adaptively a compensating digital signal and transmits it to the headphone to compensate for the interfering signal.

Results

Inference in the microphone signal

A new separation method for periodic signals was developed. The developed method will be patented in the near future, and detailed information about the method cannot be published yet.

Compared to the conventional methods, one advantage of the new separation method is its increased adaptability, especially in speech segments. It also enables filtering at 8 kHz, which makes it possible to avoid changing the sampling rate, because interpolation and decimation were overly heavy operations for the implementation.

If the speech signal is much stronger than the interfering signal and if the signals have overlapping frequencies, separation may be insufficient. The algorithm indicates these segments and adapts the amplitude when signal separation is not sufficient. The following adaptation levels were examined:

1. Separation of a speech signal and a periodic interference signal using the new separation method.
2. The estimate is taken from a non-speech segment. Amplitude is adapted in speech segments.

Both of these adaptation levels can be used independently for periodic interference elimination. But both methods have their drawbacks. As mentioned earlier, separation is not perfect in all cases. The shortcoming of the amplitude adaptation method is the need for a non-speech segment at the beginning of filtering. The interference model may also change during filtering. The best results were obtained by combining the best properties of the two methods. The choice of the best adaptation level was done by adaptive thresholds. To switch automatically between adaptation levels we adopt a parameter, SNR, which gives the power ratio of a separated speech component to a interfering signal component. To control amplitude adaptation we adopt parameter RMS which gives the reliability of a new amplitude value. If the residual speech component is significant, previous amplitude information is being used.

The algorithms were simulated in a Matlab environment with authentic audio data. The filtering results were evaluated using hearing tests and error signal RMS values. The developed method helped to reduce the interfering signal to an acceptable level even when the amplitude was strongly variable.

Inference in the headphone signal

The interfering signal was measured with acoustic methods and then examined and analysed. A Matlab simulation model was made based on the system model of mobile phone headphone and microphone connections to a digital signal processing unit. Adaptive feedback was included to allow adaptation. The proposed solution estimates adaptively the digital signal that is opposite to the interference waveform and transmits it to the headphone to compensate for the interfering signal. The algorithm is designed to be adaptive, enabling learning of a compensating pulse model based on the interference signal in the headphone.

The effects of known transfer functions shape the pulse model in order to minimise the residual interference signal in the headphone. The transfer function of the headphone parts is unknown and it has to be estimated automatically. Two different methods for automatically determining transfer functions were examined, the adaptive filter method (LMS) and the spectral estimation method (Welch). The results obtained with the Welch method were more reliable. The data length needed was not the limiting factor, but in some cases the residual error remains bigger than the original effect of the estimated transfer function.

The low sampling rate of 8 kHz and the rising time of the interference pulse in the headphone make a theoretical limit for how well the interfering signal can be eliminated. The other error sources examined were synchronisation error, amplitude error, and transfer function estimation error. The most critical of these is the synchronisation error. If the compensating pulse and the interference pulse are not synchronised, the error grows quickly. It is difficult to evaluate the effect of transfer functions because we did not have measured data from the electrical parts of the headphone. To find out how the compensating pulse reacts upon reaching the headphone, more measurements must be made. A realistic goal would be to eliminate the main peak of interference. Because the level of the interfering signal is lower in the headphone parts than in the microphone parts, it is still possible to bring the level of the interfering signal under the hearing threshold.

Project information

Participants and volume

This research was a cooperative project with the Computer Engineering Laboratory of the University of Oulu and VTT Electronics. The budget of the project was FIM 1.5 million and the amount of the work performed was about 44 person months. The work was carried out in cooperation with Polar Electro Oy and Nokia Mobile Phones.

Project leader

Professor Juha Röning
Infotech Oulu and Department of Electrical Engineering
University of Oulu
P.O. Box 4500, FIN-90014 UNIVERSITY OF OULU
Phone: +358-8-5532794
Fax: +358-8-5532600
E-mail: juha.roning@oulu.fi

Publications

Tamminen S., Pirttikangas S., Röning J. (1999): The Self-Organizing Maps in Adaptive Health Monitoring. International Joint Conference on Neural Networks (IJCNN2000), 24-27 July, Como, Italy (submitted).

Tamminen S., Pirttikangas S., Nissilä S., Pentikäinen V., Väinämö K., Röning J. (1999): Multiple Alarm Management with Self-Organizing Maps. Proc. of Intelligent Data Analysis in Medicine and Pharmacology (IDAMAP99), November 6, Washington, DC, USA, pp. 135-144.

Tamminen S., Laurinen P., Röning J. (1999): Comparing Regression Trees with Neural Networks in Aerobic Fitness Approximation. Proc. of Advances in Intelligent Data Analysis (AIDA99), June 22-25, Rochester, NY, USA, pp. 414-419.

Väinämö K., Mäkikallio T., Tulppo M., Röning J. (1998): A Neuro-Fuzzy Approach to Aerobic Fitness Classification: A Multistructure Solution to The Context-Sensitive Feature Selection Problem. Proc. WCCI98, May 4-9, Anchorage, Alaska, USA, pp. 797-802.

Lipponen S., Mäkikallio T., Tulppo M., Röning J. (1998): Finding Structure in Fitness Data. Proc. 2^nd International Conference on The Practical Application of Knowledge Discovery and Data Mining, March 25-27, London, UK, pp. 101-109.

In the Digital  Filtering of Speech Signal no publications are available because patenting is in process.

jukka.iivarinen@hut.fi
http://www.cis.hut.fi/neuronet/Tekes/9.shtml
Thursday, 30-Nov-2000 10:32:30 EET