Adaptive and Intelligent Systems Applications 1994-1999 - Final Report | |
Main page Preface Program summary Research projects 1994-1998 Research projects 1998-1999 Product development projects 1994-1999 |
13 Signal and image processing13.1 Adaptive image analysis system for paper web inspectionGoalsThe goal of the project was implementation of adaptive algorithms and neural network techniques to detect low contrast paper flaws. This type of defect is embedded in paper formation and very difficult to extract from background noise.These subtle defects, e.g. blow wrinkles and ruptures, are very harmful to paper maker causing breaks at winders, supercalanders or, in the worst case, in the printing house. The early identification of them decreases the down time of paper machines and prevents returns of paper shipments from printing houses.
ResultsThe project has enhanced the market potential of web inspection systems by providing the latest image analysis technology in a commercial product.The test pilot system (Figure 1) was used to verify results with actual samples from paper mills. (Figure is missing, sorry.) Figure 1. Test pilot system with rotating cylinder, light source and signal processing unit.
Project informationParticipantsABB Industry Oy, Automation Division
Project schedule1.10.1997- 31.12.1999
Project volumeFIM 10,2 million
Project managerSeppo RantapuskaABB Industry Oy Automation Division Tel. +358 10 2222535 Fax +358 10 2222842 E-mail: seppo.rantapuska@abb.fi
Goals
Results
Figure 1. Installing the WIM scale in Luleå, Sweden.
MethodsThe whole system is based on a high frequency with excellent s/n ratio and neural computing.
Project informationParticipants
Project managerVesa KoivistoYhteistyönkatu 7 FIN-53300 Lappeenranta, Finland Tel. +358 5 623 5100 Fax +358 5 623 5110 E-mail: vesa.koivisto@iki.fi
More informationOy Omni Weight Control LtdYhteistyönkatu 7 FIN-53300 Lappeenranta, Finland Tel. + 358 5 623 5100 Fax +358 5 623 5100 E-mail: owc@wwnet.fi
GoalsVisy Oy had two different targets. The first was to collect and teach to the system most of the European licence plates and the second was to achieve a reliable OCR engine for the European market. One of our main goals and the most important message to the customers was that there is now a system reliable enough for the recognition of the licence plates of various countries.
Results and impactsThe project goals were achieved. The system recognises plates of various countries.We also discovered the power of nature - rain, storms, sunlight, fog, mud, snow and dust. We knew that we would face these problems when operating under the clear sky at the Russian border. Using a slightly bit different technique these problem were resolved. Now we have the keys to reliable operational system for customers all over the Europe.
MethodsIn this project the OCR system learnt EU and Russian plates (there were altogether nine different Russian plates). The system was based on the neural network technique. This means that we collected the material needed and used it to train the OCR program.This also means that we must have a lot of training material. We collected the material from across Europe, which was very expensive and demanded a lot of work. We also got test material at Vaalimaa (a splendid place to collect Russian plates). Because we had sufficient test material, we were able to repair the systematic bugs of the OCR engine. The advantage of the Visy system is that we can add as many countries as needed very easily. We also tested the use of different types of lights and cameras for achieving better results.
Project informationParticipantsThe system was tested with RGB colour images captured from a PAL video camera. Imaging was carried out at the Finnish border station and at the borders between different countries in Europe (including Germany, France, Switzerland, Spain, Greece and Italy).All algorithms were developed and implemented by Visy Oy. The advice of local authorities of the Finnish Customs was very useful when developing the features of the user interface and functions of the program.
Project datesThe project started on 1.11.1997 and ended on 30.09.1999.
Project volumeTotal budget was FIM 1 million.
Project managerKalevi VoutilainenVisy Oy Kauppatori 1 A FIN-78250 Varkaus, Finland Tel. +358 50 524 5450 Fax +358 17 366 9440 E-mail: kalevi.voutilainen@visy.fi
More informationInformation about Visy Oy can be found from the Internet page http://www.visy.fi .
Goals and resultsThe rapidly growing popularity of cellular phones in Finland has led to a very rapid construction and enlargement of the GSM network. Network capacity and services have been added by means of new network elements and new technologies. A dynamic and fast growing network is a really challenging task for network management. A quick reaction in maintenance work is usually crucial when an operator wants to secure the quality of service in its network. Information overflow causes problems in a large network and it is a challenge to obtain the most meaningful information as soon as possible.In Sonera’s NeuroFMS project we focused on developing an intelligent fault management system to operate in a multivendor cellular network. The result of the project was an entirely new fault management environment. This environment includes an intelligent diagnosis part that utilises e.g. traffic measurement information and experiential knowledge.
Figure 1. NeuroFMS architecture. Sonera is utilising Gensym's G2 Operations Expert (OpEx) software as the basis for the system. The NeuroFMS system includes four main functional parts (Figure 1): a diagnosis part (IDS), a performance measurement analyser (PMA), a Java based user interface application and an interface layer which adapts data for the system. The topology of a network has been modelled for the diagnosis system so that correlation and alarm filtering can be carried out based on the model. Analysis and diagnosis are implemented in IDS by graphical procedures that can be easily maintained and modified. The basic functions of IDS are described in Figure 2.
Figure 2. Some functions of IDS. The diagnostic system receives alarms from the network elements as well as from the PMA module. The alarms from the PMA module can be correlated with alarms that are generated by network elements. NeuroFMS system can also actively carry out maintenance tasks on network elements, such as testing and resetting BTS units, or forward alarms automatically to trouble ticketing systems without manual intervention.
Figure 3. PMA system. The PMA system carries out pre-processing of performance measurements collected from BSS. The PMA system looks for symptoms of decreased service grade and pinpoints problems for the operator. The use of fuzzy rules makes the system more adaptable. The NeuroFMS system supports centralised as well as distributed approaches to organise network management processes. Flexibility of the system is achieved by means of Java and Corba interfaces. We have gathered into NeuroFMS system the most useful features which operators need in daily network supervising tasks. Mobile networks are very dynamic, because plenty of new services and standards will be introduced with each passing year. Hence, we have had to take into account the adaptability of the NeuroFMS system, otherwise our application would be useless after a while. One interesting new application area could be monitoring the fulfilment of service level agreements between service providers and customers. At the moment, the NeuroFMS system needs some improvements to fulfil the requirements set for a SLA monitoring system.
Project informationParticipantsSonera subcontracted some research and implementation tasks to Control Software Oy and VTT Electronics.
Project datesThe NeuroFMS project was started in January 1998 and it was finished in December 1999.
Project volumeThe total work amount in the project was about 12 working years and its budget was about FIM 7 million.
ContactsThe contact person at Sonera is Jari Vormisto (Tel. 020401 and e-mail: jari.vormisto@sonera.com).
GoalsThis research project was part of Nokia Telecommunication´s GSM product development. It aimed to further improve the operations and maintenance (O&M) services and tools that Nokia GSM networks provide for operators.The GSM mobile telephone network is a complex system that includes mobile telephone exchanges, base station controllers, a great number of base stations and transmission facilities connecting all of these. The operation and maintenance of the network is centralized in a network management system (NMS). The NMS collects fault information (alarms) from the network elements, for instance. Quite often a single fault in the network causes an alarm burst, and it is difficult to distinguish the actual fault from secondary alarms. For example, disturbances in a transmission line may cause repeated alarm bursts from transmission equipment and all base stations and base station controllers that are using the services of that transmission line. Alarm processing was improved in the project by employing knowledge-based techniques. Improvements are designed for base stations and network management system. The goals of the project were:
Results and impactsAlarm processing in the base station was improved which resulted in fewer alarms sent to the network management system. In addition, alarm processing in the network management system was improved by utilizing a service model of the network.
MethodsBoth theoretical studies and prototyping were employed.
Project informationParticipants
Project dates1.1.95-31.12.97
Project volumeTotal budget FIM 2,6 million
Contact informationJuha HartikainenNokia Telecommunications P.O. Box 759, FIN-33101 Tampere, Finland Tel. +358 3 2577619 Fax +358 3 2577104 E-mail: juha.i.hartikainen@nokia.com GoalsThe aim of this project was to create a colour measurement and recognition system for outdoor conditions based on an RGB colour PAL video camera, digital image processing and analysis and an artificial neural network.The system can be used for colour measurement and recognition in different applications. Installation is possible without tailoring the program code by simply adjusting parameters and teaching the application specific colours. The pilot application was recognition of the colours of shipping and transportation containers. The system can measure a restricted set of colours reliably under outdoor lighting conditions varying with season, weather and time of day.
Results and impactsThe project goals were achieved. Tests showed the system to be accurate and reliable enough for the pilot application, the recognition of container colours.The system can be used for different applications and is now one part of Visy Oy machine vision “toolkit”. At the time of writing, the applicability of the colour measurement subsystem to a new application is being evaluated by Visy. If the system is found to be suitable, then an object oriented C++ implementation will be carried out.
MethodsThe heart of the system comprises two main parts: a colour measurement algorithm and a teachable neural network classifier.The colour measurement is based on calibration of the colour information contained in digital images captured from a colour PAL composite-video signal. Calibration is performed by using a reference white. A physical white reference board is positioned so as to be visible in a constant location in the images, typically in one specific corner. The calibration algorithm produces data values for reference white from the image data for the area of the calibration board by image processing and analysis. A robust method, not sensitive to noise, is used to extract colour features inside a specific region of interest. The extracted colour features are then transformed into calibrated features by using the data values of reference white. The calibrated colour features form a colour sample. The features are fed into a multilayer perceptron (MLP) based neural network classifier which outputs the colour (class) in which the input colour sample belongs. The network also outputs a classification confidence value. The implementation of the neural network comprises two main parts, the classifier network and a “teacher”. The number of layers in a network and the number of nodes in each layer can be set by parameters. The teacher creates a binary neural network configuration file describing the neural net. An application reads in the neural network configuration file when initialising recognition functionality. The very same neural network implementation is also applied to optical character recognition (OCR) by Visy.
Project informationParticipantsThe system was tested with RGB colour images captured from a PAL video camera signal. Imaging was carried out at the Länsisatama harbour site of Oy Finnsteve Ab.The algorithms were developed and implemented by Visy Oy and ATE Oy. ATE worked as a supplier for Visy. ATE developed the colour calibration and measuring algorithms and created the preliminary C-program language implementations. ATE also tested the colour measuring system with a simple decision tree classifier capable of separating a restricted set of known container colours. Visy enhanced the preliminary implementation of colour measuring system by ATE to conform to Visy software development conventions. Visy merged the colour measuring system software and an earlier artificial neural network software implemented by Visy to create a learning system for colour recognition. Adjustment of parameters and system testing was carried out by Visy.
Project datesThe project started on 1.2.1997 and ended on 31.12.1997.
Project volumeTotal budget was FIM 98.400.
Project managerKari Saarinen, Development ManagerVisy Oy Hermiankatu 6-8 H FIN-33720 Tampere, Finland Tel. +358 50 5245465 Fax +358 3 3165064 E-mail: kari.saarinen@visy.fi
More informationInformation about Visy Oy can be found on the Internet page www.sypal.fi .GoalsThe aim of this project was to develop an indoor navigation device which could be attached to typical commercially available powered wheelchairs. The goal was to develop an inexpensive attachment prototype that would enable semi-autonomous point-to-point movement of wheelchairs inside buildings, such as old persons' service centers, libraries, shops etc. The use of radio beacons and other similar artificial landmarks was ruled out, because the goal was freedom of mobility in any building, rather than just within the confines of special precincts. During the route-following process, the device should also have obstacle and falling avoidance functions for safety reasons.
MethodsAn important approach was the utilisation of a collision avoidance system as a 'radar' to detect navigationally important 'natural' landmarks such as door openings, columns, walls, and corners. The building and its landmarks should be digitally mapped to a reasonable degree of precision. The navigator uses the digital map to plan its route and to record the location of the landmarks along the route. During motion the navigator compares the detected environmental features with the landmarks marked on the map, trying to direct the wheelchair to follow the route correctly.
ResultsThe first prototype of the navigation system has been built. The main advantage of the system is that it requires fewer sensors than current navigation systems. In demonstrations the prototype has performed well enough, despite some unfinished software modules and imperfect maps. The concept also seems also applicable to other autonomous indoor navigation tasks, e.g. in the developing service robotics sector.
ImpactsUnfortunately the market for ‘moving aids for the elderly and disabled’ is not large enough to justify the efforts needed to turn this navigation prototype into a product. This situation may change as the growing proportion of elderly people creates an expanding market. Products that assist elderly and disabled persons to live more independently are receiving growing interest from the welfare sector authorities.
Project informationThis project was funded initially by Tekes, but further funding would have required a company, which would have developed the concept into a product. While looking for the company, VTT developed the concept onwards using its internal funding.
Participants
Project dates1.1. ‑ 30.6.1997, 1.10.-31.12.1997VTT: 1.1.1998 - 31.12.1998
Project volumeFIM 495.000 + FIM 133.000VTT: FIM 500.000
Project managerPertti PeussaTel. +358 3 316 3606 E-mail: pertti.peussa@vtt.fi
More informationPresentation video "Smart Functions for Wheelchairs" 10 min. 1998.Pertti Peussa, Ari Virtanen, and Tiina Johansson, "Improving the Mobility of Severely Disabled". in Proc. The 2nd European Conference on Disability, Virtual Reality and Associated Technologies, Skövde, Sweden, Sep. 10-11, 1998, pp. 169-176. Pertti Peussa and Ari Virtanen, "A General Purpose Obstacle Avoidance System". in Proc. 2nd Tampere International Conference on Machine Automation (ICMA '98), Tampere, Finland, Sep. 15-18, 1998, pp. 211-220. Pertti Peussa and Ari Virtanen, "Make your wheelchair autonomous". Automation Technology Review 1998, VTT Automation, Helsinki, Dec. 1998, pp. 18-22. (Figure is missing, sorry.) Figure 1. The navigating wheelchair is being tested by the Finnish Association for the Disabled.
GoalsThe purpose of the project was to study possible neural solutions for various algorithms in digital communication systems, especially those that could be applicable in the third generation wireless communications systems. The particular areas of interest were developing algorithms for coding and decoding, study of receiving techniques with distorted signals and detection in the presence of interference. The outcome of the project was to be an increase in the understanding of this new technology. The ultimate goal was to find new circuit structures which could provide either simpler receivers or better performance in a difficult radio environment.
Results and impactsThe project developed a new neural decoding algorithm for convolutional codes. Two patent applications have been submitted on these. The benefits seem to be that roughly the same performance as with conventional algorithms can be achieved with less complexity. The implementation aspects will be studied in a continuation project together with Nokia Mobile Phones and VTT Microelectronics. The findings have been reported at several scientific conferences.Studies on neural detection methods in difficult environments led to conference papers, one journal article and finally to one doctoral thesis. These have increased Nokia’s visibility in scientific forums and will indirectly pave way for Nokia products onto the markets.
MethodsThe novel decoding algorithm for convolutional codes has been derived at the Nokia Research Center. It uses a recurrent neural network which has some similarities to the Hopfield network. The network is dependent on the code structure and has to be derived for each code separately. The performance of the codes has been verified using extensive computer simulations.The studies on detection in a hostile environment concentrated on using self-organizing networks (SOMs). The main finding was that neural methods may provide improved performance in detection problems where nonlinearities are involved. The performance comparisons were made using computer simulations.
Project informationParticipantsNokia Research Center, Nokia Mobile Phones and Laboratory of Computer and Information Science (Helsinki University of Technology).
Project dates1.3.1997 –30.6.99
Project volumeAbout FIM 2,3 million
Project managerJukka HenrikssonNokia Research Center P.O. Box 407, FIN-00045 Nokia group, Finland E-mail: jukka.henriksson@nokia.com
PublicationsKimmo Raivio, Jukka Henriksson and Olli Simula: Neural Detection of QAM signal with strongly nonlinear receiver. In Proceedings of the WSOM, pp. 20-25, Espoo, Finland, June 4 – 6, 1997.Kimmo Raivio, Ari Hämäläinen, Jukka Henriksson and Olli Simula: Performance of Two Neural Receiver Structures in the Presence of Co-Channel Interference. In Proceedings of the ICNN, volume 4, pp. 2080-2084, Houston, Texas, USA, June 9 – 12, 1997. Kimmo Raivio, Jukka Henriksson and Olli Simula: Neural receiver structures based on Self-Organizing Maps in nonlinear multipath channels. In Proceedings of the IWANNT, pp. 241-247, Melbourne, Australia, June 9 – 11, 1997. Kimmo Raivio, Jukka Henriksson and Olli Simula: Neural detection of QAM signal with strongly nonlinear receiver. Neurocomputing (21) 1-3 (1998), pp. 159-171. Ari Hämäläinen, Jukka Henriksson: A Recurrent Neural Decoder for Convolution Codes. To be published in Proceedings of ICC'99, S 33.5. Vancouver, June 6-10, 1999. A Hämäläinen and J Henriksson: Menetelmä ja järjestely konvoluutiokoodauksen toteuttamiseksi. Patenttihakemus FI-981168. 26.6.1998. (In Finnish.) Kimmo Raivio: Receiver Structures Based on Self-Organizing Maps. Ph.D. thesis, Helsinki University of Technology, January 1999. Acta Polytechnica Scandinavica, Mathematics, Computing and Management in Engineering Series No. 96.
Figure 1. An example of a neural decoder for a convolutional code.
jukka.iivarinen@hut.fi http://www.cis.hut.fi/neuronet/Tekes/13.shtml Wednesday, 29-Nov-2000 10:39:42 EET |